Editors

Zhang
Wang
Wang
Li
Geotechnical Safety and Risk IV contains the contributions presented at the
4th International Symposium on Geotechnical Safety and Risk (4th ISGSR,
Hong Kong, 4-6 December 2013), which was organised under the auspices of
Geotechnical Safety and Risk IV
the Geotechnical Safety Network (GEOSNet), TC304 on Engineering Practice
of Risk Assessment and Management and TC205 on Safety and Serviceability

Geotechnical Safety and Risk IV

in Geotechnical Design of the International Society of Soil Mechanics and
Geotechnical Engineering (ISSMGE). Geotechnical Safety and Risk IV
covers six themes:
(1) Geotechnical uncertainty and variability
(2) Geohazards such as landslides, earthquakes and climate changes
(3) Reliability and risk analysis
(4) Reliability-based design and limit-state design in geotechnical
engineering
(5) Risk assessment and management in geotechnical engineering and
infrastructural projects, and
(6) Practical applications.

Geotechnical Safety and Risk IV is an excellent source of reference on the

latest developments and practices of geotechnical risk and safety, and will
prove to be invaluable for civil engineers (particularly geotechnical engineers),
geologists, and natural hazards reduction agencies.

ISGSR IV is a continuation of a series of symposiums and workshops

on geotechnical risk and reliability, starting with LSD2000 (Melbourne,
Australia), IWS2002 (Tokyo and Kamakura, Japan), LSD2003 (Cambridge,
USA), Georisk2004 (Bangalore, India), Taipei2006 (Taipei, Taiwan), 1st
ISGSR (Shanghai, China, 2007), 2nd ISGSR (Gifu, Japan, 2009) and 3rd
ISGSR (Munich, Germany, 2011).

Editors: L.M. Zhang, Y. Wang, G. Wang and D.Q. Li

an informa business
 GEOTECHNICAL SAFETY AND RISK IV

ISGSR2013.indb i 10/18/2013 9:36:13 AM

This page intentionally left blank
 PROCEEDINGS OF THE FOURTH INTERNATIONAL SYMPOSIUM ON GEOTECHNICAL SAFETY
AND RISK – 4TH ISGSR, HONG KONG, 4–6 DECEMBER 2013

Geotechnical Safety and Risk IV

Editors
L.M. Zhang
The Hong Kong University of Science and Technology, HKSAR

Y. Wang
City University of Hong Kong, HKSAR

G. Wang
The Hong Kong University of Science and Technology, HKSAR

D.Q. Li
Wuhan University, China

ISGSR2013.indb iii 10/18/2013 9:36:15 AM

 CRC Press/Balkema is an imprint of the Taylor & Francis Group, an informa business

© 2014 Taylor & Francis Group, London, UK

Typeset by V Publishing Solutions Pvt Ltd., Chennai, India

Printed and bound in Great Britain by CPI Group (UK) Ltd, Croydon, CR0 4YY

All rights reserved. No part of this publication or the information contained herein may be reproduced,
stored in a retrieval system, or transmitted in any form or by any means, electronic, mechanical, by
photocopying, recording or otherwise, without written prior permission from the publisher.

Although all care is taken to ensure integrity and the quality of this publication and the information
herein, no responsibility is assumed by the publishers nor the author for any damage to the property or
persons as a result of operation or use of this publication and/or the information contained herein.

Published by: CRC Press/Balkema

P.O. Box 11320, 2301 EH Leiden, The Netherlands
e-mail: [email protected]
www.crcpress.com – www.taylorandfrancis.com

ISBN: 978-1-138-00163-3 (Hbk + CD-ROM)

ISBN: 978-1-315-79734-2 (eBook PDF)

ISGSR2013.indb iv 10/18/2013 9:36:15 AM

 Geotechnical Safety and Risk IV – Zhang et al. (eds)
© 2014 Taylor & Francis Group, London, ISBN 978-1-138-00163-3

Table of contents

Preface xi
Organisation xiii
Acknowledgements xv

1 Wilson Tang lecture

An homage to Wilson Tang: Reliability and risk in geotechnical practice—how
Wilson led the way 3
S. Lacasse, K. Høeg, Z.Q. Liu & F. Nadim

2 Keynote lectures
Advances in geotechnical risk and reliability for offshore applications 29
R.B. Gilbert, S. Lacasse & F. Nadim
Homogenization of geomaterials using the random finite element method 43
D.V. Griffiths, J. Paiboon, J. Huang & G.A. Fenton
Selecting optimal probability models for geotechnical reliability analysis 53
H.W. Huang, W.W. Su & J. Zhang
Robust design of geotechnical systems—a new design perspective 69
C.H. Juang, L. Wang & S. Atamturktur
Integrated geo risk management: Crossing boundaries 79
M.Th. van Staveren
Is landslide risk quantifiable and manageable? 101
H.N. Wong

3 Geotechnical uncertainty and variability

Quantifying epistemic uncertainty and aleatory variability of Newmark displacements
under scenario earthquakes 115
W. Du & G. Wang
Reliability analysis of 640 m long soil retaining wall for an embedded highway construction 121
Y. Honjo, Y. Otake, T. Kusano & T. Hara
CPT-based Bayesian identification of underground soil stratigraphy 127
K. Huang, Z. Cao & Y. Wang
Characterization of three-dimensional random crack network in soil using CT test 133
J.H. Li, L.M. Zhang & Y. Zhao
Linear regression and statistical analysis of shear strength parameters 139
X. Li, L.H. Chen & N. Zhang
Uncertainty quantification by polynomial chaos expansion for geotechnical
reliability-based analysis 145
S.H. Marques, A.T. Gomes & A.A. Henriques

ISGSR2013.indb v 10/18/2013 9:36:15 AM

 4 Influence of spatial variability on geotechnical design decisions
Mobilized shear strength of soils with constrained slip curves 153
Y. Hu & J. Ching
Reliability of heterogeneous slopes with cross-correlated shear strength parameters 161
T.M.H. Le
Finite element analysis of compression behavior of ground improvement with spatial variability 169
T. Namikawa
Stochastic finite element analysis for ground improvement by vertical drains
of spatially variable soils 175
M.W. Bari, M.A. Shams & M.A. Shahin
Regularity of the variance reduction function in Tianjin Port 185
S.-W. Yan, L.-P. Guo & Y.-H. Cao
Seepage through anisotropic non-stationary random fields 193
H. Zhu & L.M. Zhang

5 Geohazards
Integrating seismic hazard analyses with geotechnical site characterization for liquefaction
potential assessment in Kaohsiung area 201
D. Huang & J.P. Wang
Landslide considerations for low cost remedial works for the Karakorum Highway, Pakistan 207
A.D. Mackay
Landslide mitigation considerations for low cost highway remedial works, East Timor 215
A.D. Mackay
The landslide of Kirf: A chain of governance failures 221
S. Van Baars, M. Sosson, S. Jung & R. Becker
A risk-based active fault classification 225
J.P. Wang
Introducing non-stationary earthquake process concept: Including an analytical model and
a case study in Central Taiwan 229
Y. Xu & J.P. Wang
Modelling techniques of submarine landslide in centrifuge 235
C. Zhao, W.Z. Zhang, J.H. Zhang, K.Z. Wang & Z.S. Xiong

6 Reliability and risk analysis

The schedule optimization for subway station construction based on critical chain 241
G.X. Chen, J.L. Song, Z.R. Xiao, J. Wang & L.F. Yue
A cell-based model for predicting runout distances of detached materials
in rainfall-induced slope failures 247
H.X. Chen & L.M. Zhang
Dike failure mechanisms from the perspective of risk assessment 251
J. Danka & L.M. Zhang
Reliability assessment of earth pressure on rigid non-yielding retaining walls 257
S.M. Dasaka, T.N. Dave & V.K. Gade
Seismic damage occurrence probability in an existing estuary dam 263
T. Hara, M. Iwata, Y. Otake, Y. Honjo, T. Kato, A. Nishida & H. Yukimoto
Probabilistic analysis of dry soil mix columns 271
J. Huang, R. Kelly & S.W. Sloan

vi

ISGSR2013.indb vi 10/18/2013 9:36:15 AM

 Soft soil subgrade’s reliability and risk assessment under incomplete probability conditions 277
C. Liu & M. Zhao
Localised metamodelling techniques for geotechnical reliability-based analysis 283
S.H. Marques, A.T. Gomes & A.A. Henriques
Flood simulation considering probability of heavy rains and uncertainty of soil properties
of earth-fill dams 291
S. Nishimura, T. Shuku, H. Nagao & K. Fujisawa
Reliability analysis of 20-km river dike against liquefaction failure 299
Y. Otake, Y. Honjo & Y. Hiramatsu
Effect of slenderness ratio on the reliability-based serviceability limit state design
of augered cast-in-place piles 305
S.C. Reddy & A.W. Stuedlein
Uplifit model for the Red River dikes of Vietnam 311
P.Q. Tu, B.V. Truong & P.H.A.J.M. van Gelder
Effect of rainfall on the reliability of an infinite slope 317
J. Yuan, I. Papaioannou, C.M. Mok & D. Straub
Influence of particle transport on slope stability under rainfall infiltration 323
L. Zhang & L.L. Zhang

7 Reliability- and risk-based monitoring and site investigation

Effect of inadequate site investigation on the cost and time of a construction project 331
A.H. Albatal, H.H. Mohammad & M.E.A. Elrazik
Probabilistic characterization of Young’s modulus using Markov Chain Monte Carlo simulation 337
Z. Cao & Y. Wang
Site investigation approaches for the proposed Ukhuu Khudag to Gashuun Sukhait Railway,
South Gobi desert, Mongolia 345
A.D. Mackay & N.R. Wightman
Bayesian model updating of a tunnel in soft soil with settlement measurements 351
I. Papaioannou, W. Betz & D. Straub
A role of monitoring to reduce the uncertainty in the performance of pile foundations 357
S. Ryuo, Y. Asaka & K. Soga
Using head monitoring for reliability updating of levees 365
T. Schweckendiek

8 Performance evolution and risk assessment of high slopes in hydropower

engineering in China
Numerical simulation of blasting induced damage of high rock slope 373
Y.G. Hu, W.B. Lu, M. Chen & P. Yan
Slope reliability analysis using a non-intrusive stochastic finite element method 379
S.H. Jiang & D.Q. Li
A novel bond contact model for rock and its calibration 385
T. Jiang, M.J. Jiang, H. Chen, F. Liu & Z.M. Shi
Study on cracking risk of Jinping high and steep slope 391
P. Lin & X.L. Liu
Mechanism and numerical simulation on geological mechanical model test 397
X.Q. Luo, J.F. Bi & H. Shen
Time-dependent behaviour modeling of geomaterials using a discrete thermodynamic approach 403
W. Wang, J.F. Shao, Q.Z. Zhu & W.Y. Xu

vii

ISGSR2013.indb vii 10/18/2013 9:36:15 AM

 Study of time-dependent behavior of rocks with polycrystalline approach 409
T. Zeng & J.F. Shao

9 Reliability-based design and limit state design in geotechnical engineering

Serviceability limit state design of lime-cement columns—a reliability-based design approach 417
N. Bergman, R. Ignat & S. Larsson
Cost-effective framework for simplified geotechnical reliability-based design 423
J.Y. Ching & K.K. Phoon
Reliability index of pile foundations designed to SANS 1016-5 431
M. Dithinde & J.V. Retief
Codified reliability-based design of shallow foundations in Shanghai 437
J.P. Li, S.N. Liu, S.N. Hou & J. Zhang
Development and reliability of a pile driving formula for the MnDOT 443
S.G. Paikowsky, C.M. Marchionda, S. Amatya, M.C. Canniff & A.S. Budge
On the validation of reliability and partial safety factors for axially loaded piles in dense sand 455
K.A. Schmoor & M. Achmus
Reliability-based evaluation of vertical bearing capacity of piles using FORM and MCS 463
A. Teixeira, A.G. Correia & A.A. Henriques
Reliability-based design for the serviceability state design of an excavation with cross walls in clays 471
S.H. Wu, J. Ching & C.Y. Ou

10 Risk assessment and management in geotechnical engineering and infrastructural projects

Indicator Kriging for locating risk zones: An application to buildings at risk
in the Barcelonnette Basin 479
S.A. Arnaouti, S. Fotopoulou, K. Pitilakis, Th. Chatzigogos, A. Puissant & J.-P. Malet
Development of a probabilistic model for the prediction of building damage
due to tunneling induced settlements 485
C. Camós, C. Molins, O. Špačková & D. Straub
A three-level framework for multi-risk assessment 493
Z.Q. Liu & F. Nadim
Rockfall risk management based on survey data of real slopes 499
S. Moriguchi, Y. Otake, M. Iwata, Y. Honjo, A. Takagi, F. Kurauchi, T. Hara, K. Sawada,
A. Yashima & N. Asano
Implementing geo risk management in a client organization 505
M.Th. van Staveren, P.P.T. Litjens & J.J. Heerema
Implementing geo risk management in the construction industry 511
M.Th. van Staveren, P.P.T. Litjens & P.M.C.B.M. Cools
The buildup and assessment of environmental indices for storm induced disaster prone areas 519
M.-H. Wu, J.P. Wang, H.-R. Liao & Y.-R. Chen

11 Geotechnical applications and case studies

Ultrasound as a new approach for the prediction of collapsible soils 529
K. Abbeche & M.S.L. Mohamed Salah
Effect of riverbed scouring on bridge piles during earthquake 539
Z.H. Khan, M.R.I. Khan, M.F. Raiyan & K.M. Amanat
Risk minimization by an adapted dewatering scheme at the construction of the new ship
lock in Minden 545
H. Montenegro, B. Odenwald & R. Kauther

viii

ISGSR2013.indb viii 10/18/2013 9:36:15 AM

 Geotechnical instrumentation monitoring for the construction of the West Kowloon Terminus
of the Express Rail Link 551
A.K.O. So, P.W.L. Ko & V.K.W. Man
The impact of geometry bedding toward slope stability in coal mining 559
Supandi & H. Hidayat
Risk identification and mitigation for construction of a subway transfer station in Beijing 563
H. Wang, Z. Wu & D. Wang
Ground anchor considerations for crane footings adjacent deep rock excavations 569
N.R. Wightman & A.D. Mackay
Upper bound finite element method for seismic stability limit analysis of rockfill dams 575
X.G. Yang & S.C. Chi

Author index 581

ix

ISGSR2013.indb ix 10/18/2013 9:36:16 AM

This page intentionally left blank
 Geotechnical Safety and Risk IV – Zhang et al. (eds)
© 2014 Taylor & Francis Group, London, ISBN 978-1-138-00163-3

Preface

The 4th International Symposium on Geotechnical Safety and Risk (4th ISGSR) was organised by the
Hong Kong University of Science and Technology under the auspices of the Geotechnical Safety Net-
work (GEOSNet; Chair, Daniel Straub; Co-chair Limin Zhang), Technical Committee TC304 on Engi-
neering Practice of Risk Assessment and Management (Chair, K.K. Phoon) and Technical Committee
TC205 on Safety and Serviceability in Geotechnical Design (Chair, Brian Simpson) of the International
Society of Soil Mechanics and Geotechnical Engineering (ISSMGE). The Symposium was also sup-
ported by Hong Kong Geotechnical Society, the Geotechnical Division of the Hong Kong Institution of
Engineers, Chinese Institution of Soil Mechanics and Geotechnical Engineering, the Engineering Risk
and Insurance Branch of China Civil Engineering Society, and American Society of Civil Engineers—
Hong Kong Section.
The 4th ISGSR was a continuation of a series of symposiums and workshops on geotechnical risk and
reliability starting with LSD2000 in Melbourne, Australia, IWS2002 in Tokyo and Kamakura, Japan,
LSD2003 in Cambridge, USA, Georisk2004 in Bangalore, India, Taipei2006 in Taipei, 1st ISGSR in
Shanghai, China in 2007, 2nd ISGSR in Gifu, Japan in 2009 and 3rd ISGSR in Munich, Germany in
2011.
Safety, reliability, and risk assessment and management have attracted growing interests of the geotech-
nical community in recent years due to the frequent occurrences of natural and man-made disasters and
the needs for safe and cost-effective design, construction and operations of infrastructures. At the same
time there is an increasing expectation of the general public that requires the engineering community to
provide quantitative information concerning risks posed by geotechnical hazards. The 4th ISGSR pro-
vided an excellent opportunity to better understand the geotechnical safety and risk management issues
in engineering practices and research. The proceedings contain seven invited keynotes and 69 accepted
papers from 28 countries and regions. Each accepted paper in the conference proceedings was subject to
review by two peers. These papers cover six themes: (1) geotechnical uncertainty and variability, (2) geo-
hazards such as landslides, earthquakes and climate changes, (3) reliability and risk analysis, (4) reliability-
based design and limit-state design in geotechnical engineering, (5) risk assessment and management in
geotechnical engineering and infrastructural projects, and (6) practical applications.
One of the highlights of this symposium was the 3rd Wilson Tang Lecture. The lecture was inaugurated
during the 2nd ISGSR in Gifu to recognize and honor the significant contributions of the late Professor
Wilson Tang, who was one of the founding researchers in geotechnical reliability and risk. The first lec-
ture was given by Prof. T. H. Wu of the Ohio State University and the second lecture by Prof. Y. Honjo of
Gifu University. The 3rd lecture was given by Prof. Suzanne Lacasse of Norwegian Geotechnical Institute
during the 4th ISGSR.
The credit for the proceedings goes to the authors and reviewers. The publication of the proceedings
was financially supported by the National Basic Research Program of China (Grant No. 2011CB013500)
and the National Natural Science Foundation of China’s Oversea Collaborative Research Program (Grant
No. 51129902).

Limin Zhang
Chairman of the Regional Organising Committee
August 2013, The Hong Kong University of Science and Technology, HKSAR

xi

ISGSR2013.indb xi 10/18/2013 9:36:16 AM

This page intentionally left blank
 Geotechnical Safety and Risk IV – Zhang et al. (eds)
© 2014 Taylor & Francis Group, London, ISBN 978-1-138-00163-3

Organisation

REGIONAL ORGANISING COMMITTEE

Limin Zhang, Chairman of the Committee and Conference

Y.C. Chan and Dianqing Li, Co-Chairs of the Technical Subcommittee
Yu Wang and Gang Wang, Secretaries
J.P. Wang, Treasurer

Members
Ivan S.K. Au Jinhui Li
K.T. Chau Jack Pappin
Tony C.T. Cheung L.G. Tham
John Endicott Jun Yang
Albert Ho Kelvin K.V. Yuen

INTERNATIONAL ADVISORY COMMITTEES

GEOSNet Executive Board Members

Daniel Straub (Chair) Andrew Bond
Limin Zhang (Co-chair) Samuel Paikowsky
Yusuke Honjo (Past Chair) Iason Papaioannou (Secretary)
Richard Bathurst Dongmei Zhang

ISSMGE-TC304 Members
K.K. Phoon (Chair) B.K. Low
Gordon Fenton (Vice Chair) Dagang Lu
Jianye Ching (Secretary) Farrokh Nadim
Y. Ashkey Laszlo Nagy
Gregory Baecher Shin-ichi Nishimura
Hansgeorg Balthaus Ali Noorzad
Robert Berkelaar Lars Olsson
Laura Caldeira Lance Roberts
Paul Cools Adrian Rodriguez-Marek
Federica Cotecchia Nick Sartain
Guillermo Espin Bernd Schuppener
P. Ganne Timo Schweckendiek
Anthony TC Goh V.I. Sheinin
Vaughan Griffiths Kenichi Soga
Ken Ho Abdul-Hamid Soubra
Hongwei Huang M. Suzuki
Mark Jaksa Pavlos Tyrologou
Rafael Jiménez Martin van Staveren
Leena Korkiala-Tanttu Yu Wang
Kishor Kumar Jianfeng Xue
Giuseppe Lanzo Limin Zhang
Dianqing Li

xiii

ISGSR2013.indb xiii 10/18/2013 9:36:16 AM

 ISSMGE-TC205 Members
Brian Simpson (Chair) Tim Länsivaara
Colin Smith (Secretary) Kerstin Lesny
Sam Clarke (Secretary) Iacint Manoliu
Tony M. Allen Trevor Orr
Maria J. Avecillas Sam Paikowsky
Lech Bałachowski K.K. Phoon
Richard Bathurst Agustin Popa
Andrew Bond W.K. Pun
Luigi Callisto Rodrigo Salgado
Peter Day Bernd Schuppener
Franco Francisca Adriaan van Seters
Roger Frank Masahiro Shirato
Takashi Hara Antonio Soriano
Yusuke Honjo Jørgen S. Steenfelt
Hongwei Huang Marco Uzielli
Petr Koudelka Balazs Vasarhelyi
Kristian Krabbenhoft Aida S. Zhakulin
Anders Kullingsjö Limin Zhang

xiv

ISGSR2013.indb xiv 10/18/2013 9:36:16 AM

 Geotechnical Safety and Risk IV – Zhang et al. (eds)
© 2014 Taylor & Francis Group, London, ISBN 978-1-138-00163-3

Acknowledgements

MANUSCRIPT REVIEWERS

The editors are grateful to the following people who helped to review the manuscripts and hence assisted
in improving the overall technical standard and presentation of the papers in these proceedings:

Patrick Arnold Farrokh Nadim

M.W. Bari T. Namikawa
N. Bergman I.T. Ng
Zijun Cao Y. Otake
Y.C. Chan Iason Papaioannou
Dongsheng Chang K.K. Phoon
Hongxin Chen Nick Sartain
R.H. Chen Timo Schweckendiek
Jianye Ching Masahiro Shirato
Jozsef Danka H.W. Sun
Haijian Fan Alex Tang
Roozbeh Geraili Mikola Gang Wang
Duruo Huang J.P. Wang
Shuihua Jiang Yu Wang
Rafael Jimenez Jianfeng Xue
Suzanne Lacasse Ryan W.M. Yan
Andy Y.F. Leung Jun Yang
A.J. Li Zhaohui Yang
Chao Li K.V. Yuen
Dianqing Li Jie Zhang
Jinhui Li Limin Zhang
Victor Li Lulu Zhang
Xu Li Shuai Zhang
Jinchi Lu Zhenhua Zhang
Gang Ma Jidong Zhao
S.H. Marques W.H. Zhou
Madhusudhan Murthy Hong Zhu

xv

ISGSR2013.indb xv 10/18/2013 9:36:16 AM

This page intentionally left blank
 1 Wilson Tang lecture

ISGSR2013.indb 1 10/18/2013 9:36:16 AM

This page intentionally left blank
 Geotechnical Safety and Risk IV – Zhang et al. (eds)
© 2014 Taylor & Francis Group, London, ISBN 978-1-138-00163-3

An homage to Wilson Tang: Reliability and risk in geotechnical

practice—how Wilson led the way

S. Lacasse, K. Høeg, Z.Q. Liu & F. Nadim

Norwegian Geotechnical Institute, Oslo, Norway

ABSTRACT: The paper is in homage to Professor Wilson Tang for his inspiration to fellow engineers
in the area of geotechnical engineering. The role of statistics, probability and reliability in geotechni-
cal engineering is first outlined. Examples of solutions based on Wilson Tang’s pioneering work are
presented: uncertainties in soil parameters; Bayesian updating applications; reliability of tailings dam;
model uncertainty and calibration of safety factor. Two aspects of special interest to Wilson Tang are
also briefly discussed: improving the cost-effectiveness of site investigations and the reliability of offshore
structures.

1 INTRODUCTION The keynote speakers come from three conti-

nents, have very different backgrounds and dif-
This paper is in homage to Professor Wilson Tang ferent career profiles and are at different stages
(1943–2012) for his inspiration to fellow engineers in their engineering profession. Yet, each of these
to pursue his pioneering work in the application of recognized keynote lecturers has been influenced
reliability and risk in geotechnical engineering. The by Wilson’s work, as witnessed by the list in
paper illustrates how the work initiated by Wilson Table 1.
Tang led the way to further developments by his After introductory comments on the role of
colleagues, research partners, friends and practi- statistics, probability and reliability in geotechni-
tioners in the geotechnical profession. Case studies cal engineering, the paper emphasizes four topics
based on Wilson Tang’s learnings are provided for with solutions in large part developed thanks to
several geotechnical applications. the foundations laid by Wilson Tang:
Wilson Tang’s work covered a wide range of
– Uncertainties in soil parameters in practice.
expertise areas within statistics, probability and
– New applications for Bayesian updating.
reliability. These include: characterization of soil
– Reliability of containment facility.
properties and random field models, reliability meth-
– Model uncertainty and calibration of safety
ods, structural reliability-based design, Bayesian
factors.
updating and decision-making. Wilson applied reli-
ability concepts to, for example, site investigation
and geotechnical anomaly characterization, the
analysis of slopes and offshore structures, earth- Table 1. Keynote contributions at ISGSR 2013.
quake hazard, the analysis of foundation solutions, Author Title of keynote paper
model uncertainty and the calibration of safety
factors. Wilson’s work also covered the reliability of Gilbert et al. Advances in geotechnical risk and
landfill systems, accident hazard analysis and pre- (2013) reliability for offshore applications
diction, and road network reliability. Griffiths et al. Homogenization of geomaterials
Wilson Tang was a graduate student of the sec- (2013) using the random finite element
ond author, post-doctoral fellow at NGI, the exter- method
nal doctoral examiner for the third author, and an Huang et al. Selecting optimal probability models
inspiration and friend to all four authors. This is (2013) for geotechnical reliability analysis
only a random cross-section of three generations Juang et al. Robust design of geotechnical
of engineers at NGI. His radius of influence is so (2013) barriers—A new design perspective
much wider, as he touched the lives of many in so van Staveren Integrated geo risk management:
(2013) crossing boundaries
many ways.
Wong (2013) Is landslide risk quantifiable and
Examples of Wilson Tang’s lasting influ- manageable?
ence are the invited papers for the 2013 ISGSR.

ISGSR2013.indb 3 10/18/2013 9:36:16 AM

 Furthermore, two additional aspects of special erations, and as exemplified by Wilson Tang’s
interest to Wilson Tang are briefly discussed: the long list of publications, can be extended to the
use of probabilistic concepts to improve the cost- entire chain of geotechnical design steps, from site
effectiveness of site investigations and to estimate investigation and soil testing, selection of design
the reliability of offshore structures. parameters to design calculations, reliability of a
design method and selection of required safety
factor(s).
2 ROLE OF STATISTICS, PROBABILITY The examples presented in this paper illustrate
AND RELIABILITY IN GEOTECHNICAL the role of statistics, probability and reliability in
ENGINEERING geo-engineering.
In the books “Probability concepts in engineer-
Wilson Tang and his co-author A. H-S Ang firmly ing, planning and design” (Volume I and II 1975;
believed that the best and most effective way for 1984) and “Probability concepts in engineering—
engineers to learn the concepts of probability, sta- Emphasis on applications to civil and environmen-
tistics and risk was through applications of the tal engineering” (2nd ed. of Volume I–2007), Ang
principles to engineering problems. It was impor- and Tang published two of the first books that
tant for them to be able to show the usefulness of made the probability concepts easily accessible to
the method in physically meaningful terms. geotechnical engineers.
The motivation for probabilistic and statisti- From an engineering standpoint, the Ang and
cal decision theory is multi-fold: uncertainties are Tang books, together with Benjamin & Cornell
unavoidable, and they need to be considered and (1970) were instrumental in pointing the way for
reduced where possible; the need for a systematic most users, including the authors of this paper. As a
development of design criteria for engineering doctoral student at Stanford University, the young
designs; and quantitative risk assessment offers a Wilson Tang greatly benefitted from the lectures
logical framework for decision-making and docu- and discussion with Professor Jack Benjamin.
mentation of the steps towards the decisions. In Later books, especially Baecher & Christian
light of uncertainties, the role of probability and (2003) and Fenton & Griffiths (2008) are of special
statistics ranges from the description of the basic relevance for geotechnical engineers. Vick (2002)
information to the development of formulations as and Jordaan (2008), for example, published books
basis for design and decision-making (Ang & Tang on decisions under uncertainty and continue the
2007). Especially in geotechnical engineering, our legacy of Wilson Tang. Yet, the first Ang and
knowledge is imperfect. Tang’s books have the far-reaching influence of
As part of design and decision-making under being the pioneers for geotechnical engineers.
uncertainty (Høeg 1996), the properties of inher-
ently inhomogeneous and highly variable soil
materials must be considered. Natural deposits 3 UNCERTAINTIES IN SOIL
typically have irregular layers of clay, silt, sand, PARAMETERS IN PRACTICE
gravel or a combination thereof. The soil proper-
ties that affect strength and compressibility often The terms ‘aleatory’ uncertainties (those associated
have a wide range of variation. The information with natural randomness) and ‘epistemic’ uncer-
comes from the local geology, and limited soil or tainties (those associated with uncertainties in
rock sampling and limited coverage of the area of prediction and estimation) are known today. The
concern with in situ tests. terms ‘aleatory’ and ‘epistemic’ were first used by
The calculated bearing capacity (stability) can Hacking (1975) and Cornell, C.A (1982, Personal
vary widely according to the analysis parameters comm., Pau, France).
and the calculation method selected. The calcu- The importance of quantifying the variability in
lation will therefore involve some possibility of geotechnical design parameters is not adequately
overestimating the actual resistance provided by recognized in practice. Quantifying variability is
the soil, or leading to unnecessarily high costs due a positive contribution as its consistent model-
to overly conservative design. There will therefore ling and utilization lead, with limited additional
always be a finite probability that the forces on a effort, to more rational and economic designs. The
structure founded on or in soil or rock can cause modelling of soil variability belongs to one of two
damage, or the total collapse, of the structure. categories: (a) geostatistics, focusing on the inter-
Statistics, probability, reliability and the deci- polation of available data to estimate other values
sions made on the basis of these concepts offer at the same location; and (b) reliability-based engi-
remarkable tools that can quantify the trade-off neering, focusing on characterization for reliability/
between cost and tolerable probability of non- risk assessment.
performance (failure) and risk (sentence slightly A soil variability analysis can include three steps,
modified from Ang & Tang 2007). Such consid- each with increasing level of complexity:

ISGSR2013.indb 4 10/18/2013 9:36:16 AM

 data from artificial and natural deposits (Fig. 2),
1. statistical analysis of mean, variance (standard Popescu et al. (1998) observed that the distribution
deviation) and probability density function; of soil strength in shallow layers were prevalently
2. analysis of spatial correlation describing the positively skewed, while for deeper soils the corre-
variation of the soil property in space; and sponding distributions tended to follow more sym-
3. spatial averaging and variance reduction when metric distributions.
averaging over a volume. The resulting PDF’s are in all cases close to the
normal or lognormal PDF’s. Lacasse & Nadim
Integrated approaches making use of Monte
(1996) reviewed the probability distribution for
Carlo simulation, finite element analysis and
several soil properties (Table 2).
the results of high-level soil variability investiga-
tions have gained interest over recent years. These
approaches allow enhanced modelling of the 3.2 Level 2: Spatial correlation analysis
behavior of geotechnical systems, where spatial
Second-moment statistics alone are unable to
heterogeneity of soil properties invariably plays an
describe the spatial variation of soil properties,
important role.
whether measured in the laboratory or in-situ. Two
Two aspects are described in more detail below:
sets of measurements may have similar second-
the statistical analysis of a random variable
moment statistics (i.e. mean and standard devia-
(level 1) and the analysis of spatially random vari-
tion) and statistical distributions, but could display
ables (level 3). An example is then described for
substantial differences in spatial distribution.
Troll clay offshore Norway.
As part of spatial averaging effect and variance
reduction due to spatial averaging, scale of fluc-
3.1 Level 1: Statistical analysis of random tuation and the spatial coefficient of variation of
variable inherent variability are descriptors of a random
field. Uzielli et al (2006b) provided a review of the
Following the precepts in Ang and Tang (1975),
calculation methods for such parameters.
Figure 1 illustrates the phases leading to the
probabilistic modelling of a random variable. The
descriptive part includes the calculation of sample 3.3 Level 3: Modelling of spatially random
moments and visual inspection of data and histo- variables
grams. The inferential part includes the selection
The description of a random field through a mean,
of a probability density function type and the
standard deviation, a scale of fluctuation and a
distribution parameters, and goodness-of-fit test-
spatial correlation function is useful to characterize
ing. The dashed lines indicate that the results of
a spatially variable soil property (e.g. Vanmarcke
the descriptive analysis can be used in the inferen-
1977, 1983; Elkateb et al. 2003; and Jaksa 2006).
tial analysis; however, inference could also be per-
If spatial variability of soil properties is included
formed without prior statistical description.
in an engineering model, stresses and/or displace-
Different Probability Density Functions (PDF)
ments may change compared to the homogeneous
have been used. The distributions are site- and
case. A design that does not take spatial variability
parameter-specific. Based on cone penetration
into account is biased towards the conservative side
and therefore will lead to more costly solutions. One
of the most important benefits of random field
modelling is the capacity to simulate data series. By
using sets of random field simulations and imple-
menting the variability in non-linear finite element
meshes, the Monte Carlo technique, for example,
can be used to predict reliability of geotechnical
systems with spatially variable properties.
Recent studies have focused on combining
random fields, non-linear finite element analysis
and Monte Carlo simulation to investigate the
reliability of geotechnical systems including the
variability of soil properties. The studies suggest
that:

– When soils are modelled as spatially variable,

the failure mechanisms seem different and
Figure 1. Descriptive and inferential analysis of a ran- more complex than in the case of deterministic
dom variable (Uzielli et al. 2006b). properties.

ISGSR2013.indb 5 10/18/2013 9:36:16 AM

 Figure 2. Probability distributions for Cone Penetration (CPT) and Standard Penetration (SPT) Tests (Popescu et al
1998).

Table 2. Probability distributions for different soil system. Both statistical variability (i.e. second-
properties (adapted from Lacasse & Nadim 1996). moment) and spatial variability (i.e. spatial cor-
relation) of soil properties affect the reliability of
Soil property Soil type PDF geotechnical systems.
Cone resistance Sand, clay N/LN The number of studies making use of random
Undrained shear Clay (triaxial tests) LN field simulation, finite elements and Monte Carlo
strength Clayey silt N simulation is still limited. The importance of the
Normalized undrained Clay N/LN results so far, however, should be a stimulus for the
shear strength transposition of results to practice.
Plastic limit Clay N Popescu et al. (2005) investigated the differen-
Submerged unit weight Clay, silt, sand N tial settlements and bearing capacity of a rigid
Friction angle Sand N strip foundation on an overconsolidated clay layer.
Void ratio, porosity Clay, silt, sand N The undrained strength of the clay was modelled
Overconsolidation ratio Clay N/LN as a non-normal random field. The deformation
modulus was assumed to be perfectly correlated
to undrained shear strength. The settlements (uni-
– There generally exists a critical correlation
form and differential settlements) were computed
distance which corresponds to a minimum
with non-linear finite elements in a Monte Carlo
reliability.
simulation framework. Anisotropy in spatial cor-
– Phenomena governed by highly non-linear
relation was addressed, with the horizontal scale of
behavior laws are affected the most by spatial
fluctuation exceeding the vertical scale of fluctua-
variations.
tion by a factor of 10.
Variance reduction alone cannot convey a Figure 3a shows the contours of maximum
comprehensive picture of the implications of spa- shear strain for a uniform soil deposit with und-
tial variability on the behavior of a geotechnical rained strength of 100 kPa and for a normalized

ISGSR2013.indb 6 10/18/2013 9:36:17 AM

 vertical displacement at the center of foundation Fenton and Griffiths (2005) investigated the
δ/B = 0.1. Different sample realizations of soil reliability of shallow foundations against serv-
properties corresponded to fundamentally different iceability limit state failure, in the form of exces-
failure surfaces. Figure 3b shows one sample reali- sive and differential settlement, both for a single
zation where the spatial distribution of undrained footing and for two footings. Figure 4 (left) shows
strength is not symmetric with respect to the cross-sections through finite element meshes of
foundation. The configuration at failure, shown the single footing and the two footings. Figure 4
in Figure 3c, involves a rotation as well as verti- (right) provides the 3-D finite element mesh for the
cal settlement. The repeated finite-element analysis two-footings case. The elastic modulus of the soil
allows an appreciation of the combined settlement was modelled as a lognormal random field with an
and rotation of the footings, which could not be isotropic correlation structure.
inferred from deterministic bearing capacity calcu- Fenton et al. (2005) investigated the failure and
lations (i.e. neglecting spatial variability). In gen- the reliability of a two-dimensional frictionless wall
eral, the failure surfaces were not variations around retaining a cohesionless drained backfill. Soil fric-
the deterministic failure surface. There was also a tion angle and unit weight are modelled as spatially
significant reduction in the bearing capacity com- variable properties using lognormal random fields
pared to the deterministic case. Figure 3d shows with single exponential-type correlation structures.
that the pressure required to induce a given settle- Figure 5 shows the active earth displacements for
ment is always higher in the deterministic case. two realizations of the finite element mesh. The

Figure 3. Results of investigation on homogeneous and spatially random foundation soil (Popescu et al. 2005).

Figure 4. Left: Single footing and two footings founded on a spatially heterogeneous soil; Right: 3D finite element
mesh of spatially heterogeneous soil volume supporting two footings (Fenton & Griffiths 2005).

ISGSR2013.indb 7 10/18/2013 9:36:19 AM

 Figure 5. Active earth displacements for two realizations with same correlation distance and coefficient of variation
of the random field of soil friction angle (Fenton et al. 2005).

location and shape of the failure surface is strongly (DSS) tests were used (Fig. 6). No outliers were
related to the presence of weaker soil zones (shown evident from visual inspection. The data from
in lighter colors) and is, in both cases, markedly depths 0 to 5 m were excluded for this analysis. The
different from the shapes assumed in earth pres- following model for the total Coefficient of Varia-
sure theory. tion (COVtot) was used:
Griffiths et al. (2013) brings new developments
on soil variability and random finite element Vto2t
COV COV 2ω + COV
Vm2 COV
Vse2 (1)
analysis.
where COVω is the coefficient of variation of
3.4 Application to Troll clay inherent variability, representative of aleatory
uncertainty; COVm is the coefficient of varia-
Uzielli et al. (2006a) did an uncertainty-based
tion of measurement error; and COVSE is the
geotechnical characterization of the Troll clay,
coefficient of variation of statistical estimation
a site offshore Norway for the world’s largest
uncertainty.
gravity structure. Second-moment statistics were
Figures 7 (CAUC data) and 8 (DSS data) and
obtained from laboratory and in situ tests. Bayesian
Table 3 present the second-moment estimates of
updating combined the values of undrained shear
strength resulting from triaxial compression tests
and piezocone tests. Some of the results are pre-
sented herein.
The authors believe that the approach followed
would have been close to what Wilson Tang would
have done himself, if he had been asked to interpret
and calculate the uncertainty in the Troll data.
The characterization consisted of four steps:
(a) visual inspection of data by soil unit and pre-
liminary second-moment data analysis; (b) iden-
tification of a deterministic trend function and
decomposition; (c) identification of a suitable
uncertainty model; and (d) quantification of the
uncertainty (mean, variance, standard deviation or
coefficient of variation). The Kendall’s tau statistic
test (Uzielli et al. 2006b) was run to check whether
or not the data were statistically independent.
A trend function was obtained by regression analy-
sis (Ang & Tang 2007). An uncertainty model was
used to merge the different uncertainty compo-
nents to estimate the total uncertainty.

3.4.1 Laboratory data

The results of anisotropically Consolidated
Undrained Triaxial Compression (CAUC) and Figure 6. Undrained shear strength versus depth from
constant volume (undrained) Direct Simple Shear CAUC and DSS for Troll clay (Unit 1 and Unit 2).

ISGSR2013.indb 8 10/18/2013 9:36:20 AM

 selected trend function. Hence, it is important to
report trends and testing method explicitly when
presenting the results.

3.4.2 Piezocone measurements

Five piezocone soundings (CPTU) were available
for the Unit 1 clay (Uzielli et al. 2006a). Figure 9
indicates the locations of the CPTU soundings
from the 2005 site investigation and Figures 10 and
11 the measured cone resistance and pore pressure.
The water depth was between 305 and 313 m.
The profiles show a considerable regularity and
Figure 7. CAUC undrained shear strength: (a) trends smoothness. Despite the distances between the
and standard deviations; (b) residuals of detrending
(Uzielli et al. 2006a).

Figure 9. Map showing piezocone locations at Troll site

(2005).

Figure 8. DSS undrained shear strength: (a) trends and

standard deviations; (b) residuals of detrending (Uzielli
et al. 2006a).

Table 3. Uncertainty components and total uncertainty

for undrained shear strength (Uzielli et al. 2006a).

Lab test CAUC DSS

Clay unit Unit 1 Unit 2 Unit 1 Unit 2

μt (kPa)* 28.4 80.5 25.8 69.2 Figure 10. Statistics of cone resistance from 5 CPTU
COVω 0.09 0.06 0.21 0.11 tests: (a) measured data. (b) coefficients of variation,
COVm 0.20 0.20 0.20 0.20 (c) mean and standard deviation (Uzielli et al. 2006a).
COVSE 0.02 0.01 0.04 0.02
COVtot 0.22 0.21 0.29 0.23

∗μt is the mean value of the trend.

the uncertainty components and total uncertainty

for the laboratory data. The data show a discon-
tinuity at the interface between the two units, the
higher undrained shear strength being in the upper
unit. This was consistent with the results for the
plasticity index.
The total uncertainty in Unit 2 is smaller than
in Unit 1, due to the smaller aleatory uncertainties.
Uzielli et al. (2006a) suggested that for both tests; Figure 11. Statistics of pore pressure from 5 CPTU
the effect of measurement uncertainty is signifi- tests: (a) measured data, (b) coefficients of variation,
cant. Aleatory uncertainty is directly related to the (c) mean and standard deviation (Uzielli et al. 2006a).

ISGSR2013.indb 9 10/18/2013 9:36:23 AM

 sounding locations, there is considerable overlap area ratio. The cone area ratio for the cone used
between the soundings. was ac = 0.75. Uzielli et al. (2006a) estimated uncer-
The lack of variability suggests homogeneity in tainty in the undrained shear strength from piezo-
the horizontal direction. Spatial correlation in the cone tests with the first-order second moment
horizontal direction could not be investigated reli- (FOSM) approach (Ang & Tang 2007).
ably due to the limited number of soundings and Figure 12 presents the profile of second-moment
the considerable spacing between the soundings undrained shear strength derived from the CPTU
themselves. data. The Coefficient of Variation (COV) varied
At each measurement depth from the common from 21 to 26%, with an average of 24%. The COV
zero-depth, the sample mean and standard devia- is close to the COV from the laboratory CAUC
tion of the measurements from the 5 soundings data of 22%.
were calculated. The classical statistical approach does not allow
The scatter in the data constituting each sam- for the combination of subjective and observed
ple can be ascribed to the inherent variability of data or the merging of data from different sources.
the penetrated soil and to measurement error. However, Bayesian updating can be used to include
Phoon & Kulhawy (1999) suggested a coefficient different sets of data (see also Section 4 below).
of variation of 0.07 for the measurement error in Bayesian updating was done for the 17 CAUC
cone resistance. measurements suCAUC. The second-moment param-
The total coefficient of variation for the cone eters of each CAUC measurement of suCAUC were
resistance qc and pore pressure u2 at each meas- used as prior information. The second-moment
urement depth was obtained from the following parameters of the suCPTU-CAUC values obtained from
model: CPTU data at the same nominal depth of each
CAUC measurement were taken as the likelihood
Vto2t ,D
COV Vξ2,D + COV
COV Vm2 COV 2
VSE , D + COV
Vo2ff function. The updated undrained shear strength
was denoted suB. The details of the analysis are pre-
(2) sented in Uzielli et al. (2006a).
Figure 12 compares the means of the prior, like-
in which COVξ,D is the coefficient of variation of lihood and updated (posterior) undrained shear
the aleatory uncertainty; COVm is the coefficient strength. Table 4 lists the coefficients of variation
of variation of the cone resistance measurement obtained for each. For each data point, the stand-
error; COVSE,D is the coefficient of variation of the ard deviation of the updated date (suB) was always
statistical uncertainty and COVoff is an additional smaller than that of suCAUC and suCPTU-CAUC. The
uncertainty term to account for the artificial offset- COV of the posterior (updated) undrained shear
ting of CPTU soundings at different water depths. strength is much lower than that of the likelihood.
The Harr’s (1987) “rule of thumb” guidelines This is a general and beneficial result of Bayesian
with a value of 0.20 was used for COVoff. While updating.
aleatory uncertainty and statistical uncertainty
were depth-dependent, the measurement error and
the offset-related uncertainty were assumed were
non-depth dependent (subscript ‘D’). The depth
factor was included because inherent variability
is variable with depth and a different number of
measurements may be available at greater depths
(some soundings are deeper than others).
Figures 10 and 11 illustrate, for cone resistance
and pore pressure measured by the CPTU, the
average value at each measurement depth and the
coefficients of variation for each component of
total uncertainty and for the total uncertainty.
The undrained shear strength is usually derived
from cone penetration test through the net cone
resistance (e.g. Lunne et al. 1997) defined as:

qnet = qc + ( − ac )u2 − p0 (3)

in which qc is the measured cone resistance; u2 the

pore pressure measured behind the cone; p0 the Figure 12. Bayesian updating CAUC undrained shear
total vertical overburden stress; and ac the cone strength from CPTU data.

10

ISGSR2013.indb 10 10/18/2013 9:36:24 AM

 Table 4. COV’s for the prior, likelihood and posterior regarding the statistical treatment of data. At the
undrained shear strength. same time, a shift towards an uncertainty-based
perspective is taking place in practice. In these two
Undrained shear respects, the learnings from Wilson Tang, from his
strength, su Range Ave COV17 points
books and papers, are a most useful and effective
Prior COV (suCAUC) 0.19–0.21 0.20 source of information.
Likelihood COV 0.21–0.26 0.24
(suCPTU-CAUC) 4 BAYESIAN UPDATING AND
Posterior COV (suB) 0.15–0.16 0.15 BAYESIAN NETWORKS

Wilson Tang showed a keen interest for Bayesian

The results obtained for the Troll data should
updating, and more so in the latter years of his
not be uncritically exported to data from other
career, with, among others, excellent oral contribu-
sites.
tions in Xian & Oslo in 2008 and the work sum-
The components of uncertainty depend on
marized in Cheung & Tang (2005) and Zhang et al.
trend functions which may be strictly site- or case-
(2009 a; b). Wilson Tang also had a close collabo-
specific. Perhaps most importantly, geotechnical
ration as Kwang-Hua Chair Professor at Tongji
expertise and engineering judgment were found
University in Shanghai. He was important in mov-
to be essential in every phase of the uncertainty
ing the application of Bayesian updating forward.
characterization: in the preliminary examination
Wilson Tang said that the Bayesian method
of data; in the evaluation of second-moment sta-
was “a natural tool for processing geotechni-
tistics of the measured data; in the formulation of
cal information”. He presented applications to
uncertainty models for each parameter and in the
obtain improved estimates of anomaly occurrence
selection of appropriate transformation models to
probability, anomaly size, pile capacity, model
obtain parameters useful for design.
uncertainty, failure probability, liquefaction prob-
ability, slope stability, and even the value of added
3.5 Summary information from additional tests. Bayesian updat-
ing can be assimilated to “the past as a guide-
Wilson Tang published more than 30 papers on
book for the future”, as for instance illustrated by
site characterization, uncertainties in soil prop-
Folayan et al. (1970) for settlement predictions.
erties and spatial variability. Newer design codes
Bayesian work continues in Wilson’s spirit, e.g.
recognize uncertainties in soil properties and engi-
Liu & Nadim (2013) suggest a three-level frame-
neering models, and soil variability then assumes
work for multi-risk assessment, the third level using
an increasingly important role in practice and
the Bayesian approach.
research. Ang & Tang (2007) and Uzielli et al.
Two examples of work that have pursued
(2006b) provided an overview of techniques for
Wilson’s ideas are presented below: the use of
modelling the variability of soils and highlight the
Bayesian networks for (1) the assessment of risk
benefits and limitations of the approaches. A first
for earthquake-triggered landslides and (2) the sta-
step towards an uncertainty-based approach in
bility assessment of talus slopes during road con-
geotechnical practice could be the wider report-
struction works.
ing of data statistics. However, both the simple
and more powerful modelling technique can yield
4.1 Earthquake-triggered landslides
unreliable results if the input data are insufficient
in quantity and quality. Strong earthquakes in mountainous regions usu-
Research is on-going to simplify the use of ally trigger many landslides. Earthquake-triggered
variability-modelling techniques. Research efforts landslides represent some of the most common
focus on advanced simulation techniques, enhanced secondary disasters caused by earthquake in moun-
capabilities of computing tools and use of sophis- tainous areas. In the Wenchuan earthquake of
ticated integrated methodologies to model with May 2008, more than 15,000 landslides were trig-
increasing realism the behavior of complex geotech- gered in the steep mountain slopes (Huang 2008),
nical systems. Geotechnical practice, on the contrary, causing over 20,000 fatalities (Yin et al. 2009) and
still largely relies on deterministic approaches. destroying housing and settlements and irrigation
The gap between geotechnical research and channels (Tang et al. 2011). Landslide dams block-
practice should be narrowed: research should make ing natural rivers create a new hazard that can be
the mathematical techniques more readily usable devastating unless the impounded water may be
and practice should recognize the importance released in a controlled manner.
of addressing uncertainty and variability. There The assessment of the risks associated with
is a necessity to acquire additional competence multi-hazards requires the consideration of

11

ISGSR2013.indb 11 10/18/2013 9:36:25 AM

 the interactions among the hazards and the via the arcs: M-S, D-S, S-L, S-B and L-B. The user
vulnerabilities of the elements at risk. Zhang et al. enters evidence, and the information propagates
(2013) did an assessment of the loss of lives due to through the network. The probabilities in the net-
sequential or concurrent landslides, rock fall and work are updated when new information becomes
debris flows hazards. They proposed approaches available. The posterior probabilities and joint
to estimate the vulnerability factors for loss of life probabilities are calculated on the basis of Bayes’
in multi-risk assessment. For sequential hazards, theorem (Ang & Tang 2007).
the method considers the gradual reduction of the The network in Figure 14 estimates the risk to
elements at risk in earlier hazard events. For con- buildings under an earthquake-triggered land-
current hazards, the method estimates the lower slide. One counts 11 nodes and 16 arcs. Each node
and higher bounds of the vulnerability. The occur- has several discrete states (Table 5). Management
rence of one or two hazards at an early time can includes options of ‘no action’, ‘active’ and ‘pas-
cause redistributions of the elements at risk and sive’ countermeasures, and ‘warning systems’
change the risk profile under subsequent hazards. (a form of passive measure). Active measures,
In earthquake-triggered landslide risk assess- such as retaining walls and drainage result in lower
ment, complex interactions are present between probability of failure and reduced risk. Passive
the earthquake and landslide threats. The vulner- countermeasures, such as rock fall nets or protec-
abilities of the elements at risk can also be corre- tive sheds, reduce the vulnerability.
lated to the threats. To date, the risk assessment
involving multiple hazards neglects possible cas- 4.1.2 Quantifying the network
cade effects of multiple hazards (Marzocchi et al. 4.1.2.1 Seismic hazard
2012; Kappes et al. 2012). Figure 15 shows the distribution of the calculated
A study separating the two hazards (earthquake distances to the seismic source taken as a line
and landslide) as single hazard processes might source. The annual probabilities as a function of
lead to a misjudgment of the risks associated the earthquake magnitude Mw (Fig. 16) used the
with such cascading hazards. The assessment and recurrence relationship from Gutenberg & Richter
mitigation of the risks require a multi-risk analy-
sis approach that can account for the interactions
among the threats and among the vulnerabilities
to these threats.
The risk assessment for earthquake-triggered
landslides using Bayesian network is illustrated
with a sensitivity analysis identifying the most
appropriate risk reduction strategy in a multi-
hazard perspective. Nadim & Liu (2013a) looked
at risk to the buildings exposed to the threat of Figure 14. Decision making Bayesian network for
earthquake-triggered landslides using Bayesian earthquake-triggered landslide risk assessment (after
network. Einstein et al. 2010).

4.1.1 Bayesian network for earthquake-triggered Table 5. Nodes and their possible states in the Bayesian
landslide risk assessment network in Figure 2 (after Nadim & Liu 2013a).
Nadim & Liu (2013a) provided a brief review of # of
Bayesian networks. Figure 13 presents graphically Node states Possible discrete states
a simple Bayesian network with five nodes and five
arcs. The nodes are: Magnitude (M), Distance (D), Magnitude (Mw) 6 4.0–4.5–5.0–5.5–
Seismic severity (S), Landslide severity (L), and 6.0–6.5–7.0
Building damage (B). These nodes are connected Distance (km) 6 22–25–28–31–34–37–40
PGA (g) 6 0–0.08–0.16–0.24–
0.32–0.40–0.48
Landslide 2 Happens; does not happen
Building 3 No damage; some damage;
damage collapse
Alarm 2 On; off
Measure 2 Yes; no
Decision 4 Passive; active; no action;
warning on
Cost measure, ---
Figure 13. Simple Bayesian network (Nadim & Liu cost, utilities
2013a).

12

ISGSR2013.indb 12 10/18/2013 9:36:25 AM

 (1994). The conditional probabilities of the Peak and (3) permanent displacement. The dynamic
Ground Acceleration (PGA), given the magnitude slope performance was modelled with the perma-
and distance to epicenter, were calculated with the nent displacement analysis by Newmark (1965).
ground motion equation from Ambraseys et al. The critical acceleration of a landslide block was:
(2005) and a Monte Carlo simulation. Figure 17
illustrates the joint probabilities of the PGA ac = (FS − 1) ⋅ g ⋅ sinα (4)
inferred from the Bayesian network.
where FS is the static factor of safety; g the
4.1.2.2 Landslide hazard
acceleration of gravity; and α the angle of the
The approaches developed to assess the stability of
sliding surface. For an infinite slope, FS then
slopes during earthquake fall into three analysis cat-
becomes:
egories: (1) pseudo-static, (2) stress-deformation,
FS = c′/(γz ⋅ sinα ⋅ cosa) + (1 − mγw/γ)tanϕ′/tanα (5)

where c′ and ϕ′ are the effective cohesion and fric-

tion angle; z the depth of the failure surface; α
the slope angle; γ the total unit weight of the soil;
and γw the unit weight of water. Table 6 gives the
properties used in the Nadim & Liu (2013a) study.
The probability of slope failure (Pf) as a function
of Newmark displacement (Jibson et al 2000) was
estimated from:

Pf = 0.335 ⋅ [1 − exp(−0.048 ⋅ Dn1.565)] (6)

Figure 15. Discrete probabilities of distance to the where Dn is the Newmark displacement (cm).
seismic source (Nadim & Liu 2013a). The calculated probabilities of slope failure for
different ranges of PGA are listed in Table 7. As
mentioned above, countermeasures made to land-
slide can reduce risk. The probability of slope fail-
ure when active actions were used are also listed
Table 7.
For a building subjected to a multi-hazard situa-
tion involving additive load effects (e.g. earthquake
and landslide), the damage was increased. For the

Table 6. Soil and slope properties (Nadim &

Liu 2013a).

Variable Mean Standard deviation

Figure 16. Discrete probabilities of earthquake magni-
c′ (N/m2) 10,000 2,000
tude (Nadim & Liu 2013a).
ϕ′ (degree) 30 2
z (m) 2.5 0
α (degree) 35 0
γ (N/m3) 27,500 0
γw (N/m3) 10,000 0
M 0.4 0

Table 7. Computed probability of failure (Nadim &

Liu 2013a).

PGA
(10−2 g) 0–8 8–16 16–24 24–32 32–40 40–48

Pf no action 0.124 0.256 0.305 0.328 0.339 0.346

Figure 17. Discrete probabilities of peak ground accel- Pf, active actions 0.025 0.03 0.035 0.04 0.045 0.05
eration (Nadim & Liu 2013a).

13

ISGSR2013.indb 13 10/18/2013 9:36:26 AM

 other nodes, Nadim & Liu (2013a) adopted the expected; otherwise, active measures are preferred,
Einstein et al (2010) probability approach and pre- except for probabilities between 0.15 to 0.25 where
sented the results in tabular form. Table 8 gives an warnings system are slightly preferable to active
example for conditional probabilities of ‘Building measures or no action. This is only an example.
Damage’. The sensitivity of the decision to other factors
needs to be similarly studied.
4.1.3 Results As a further application, one can assume that
Mitigation measures influence the outcome of the average unit rebuilding cost for the “collapse”
multi-risk analyses. The results from the Bayesian damage state is €200,000, and the average repair
network of the entire risk assessment and decision costs for the “yielding” damage state as 50%
are shown in Figure 18 and compared to Einstein (€100,000) of the unit rebuilding cost (Nadim and
et al. (2010). Liu 2013b). Figure 20 presents comparative risk
Different mitigation measures result in different curves with and without the cascade effect. The
utilities. The warning system, showing the lowest mean expected loss increases for the same return
(negative) utility, is the most optimal mitigation period of the hazard(s) when the cascade effects
measure. The expected losses for the four miti- are included.
gation options increase due to the cascade prob- The results are still a preliminary step in fur-
ability triggered by the earthquake. Neglecting the thering the earthquake-triggered landslide risk and
cascade effect could therefore underestimate the multi-hazard risk assessment. The approach fol-
risks. lows the philosophy of the recent work carried out
The parameters in the analysis, e.g. the costs, by Wilson Tang, and the decision-making princi-
the probability of slope failure or the reliability of ples described in the Ang and Tang Part II (1984)
the warning system, can vary. Sensitivity analyses book.
were therefore conducted to assess the effects of
these variations on the results.
4.2 Stability of talus
Figure 19 shows the effect of changing the
probability of landslide occurrence. In this graph, Liu (2011) did a similar Bayesian network study
the best mitigation measure is the one having the as part of his doctoral dissertation for talus
less negative utility. For low failure probabilities
(P[landslide] < 0.15), no action is preferable, as

Table 8. Conditional probabilities of ‘Building Damage’

for PGA = 0–0.08 g (Nadim & Liu 2013a, after Einstein
et al. 2010).

Measure Passive Active

Parent
nodes Landslide Yes No Yes No

Building No damage 0.4 0.1 0.52 0.1 Figure 19. Sensitivity analysis of the risk as a function
damage Some damage 0.3 0.1 0.43 0.1 of the probability of slope failure for different mitigation
actions—horizontal arrows indicate range where type of
Collapse 0.3 0.8 0.05 0.8
mitigation measure is the optimum (after Nadim & Liu
2013a).

Figure 18. Losses for ‘no action’, ‘active measures’,

‘passive measures’ and ‘warning system’ (after Nadim & Figure 20. Example of risk curve with and without
Liu 2013a). cascade effect (Nadim & Liu 2013b).

14

ISGSR2013.indb 14 10/18/2013 9:36:27 AM

 landslides. He investigated the characteristics of states (Table 9). The prior probabilities of the six
talus landslides and the factors affecting talus root nodes (lithology, soil type, gravel content,
stability from existing talus failures in China and vegetation type, slope angle and time of land-
implemented geotechnical engineering risk analy- slide) were quantified from a study of 51 talus
sis with the observed failure mechanism of talus landslide cases. The conditional probability of
slopes. each node could be determined by expert knowl-
Talus landslides are common during the con- edge and interrelationship of each information
struction of highways in mountainous regions and source.
may lead to construction delay and cause fatali- The Bayesian network was constructed using
ties and large economic and environmental losses. logic relationships among triggering factors, vul-
A talus is a slope formed by an accumulation of nerability factors, and consequence factors. The
mainly rock debris or broken rock fragments at the nodes (factors) and arcs (inter-relationships) of
base of mountain cliffs or valley shoulders (talus the network were quantified with historical data,
can also be called “screes”). Talus often have a con- empirical models and experimental results. The
cave upwards shape and the maximum inclination risk value, given the probabilities of the root fac-
corresponds to the angle of repose of the mean tors and vulnerabilities, were calculated based on
debris size. The deformation and failure mecha- the networked interrelationships.
nism of talus slope is different from those of natu- In the application of Bayesian networks along
ral soil and rock slopes. the Shuifu-Maliuwan Highway adjacent to the
The study area is along the Shuifu-Maliuwan Yungui Plateau and Liangshan Mountain in north-
Highway, which is located in the area adjacent to east of Yunnan Province in China, the following
the Yungui Plateau and Liangshan Mountain, in steps were used: (1) the lithology of rock, soil type,
northeast of Yunnan Province in China. The area gravel content, slope angle and vegetation cover
has high mountains, steep gorges prone to heavy
erosion, rapidly moving rivers and saw-cuts. Many
talus slides have occurred along this highway due to
cuts and excavations. Figure 21 provides examples
of some to the structural damage encountered.
The characteristics of and factors affecting fail-
ure of talus were studied from the analysis of typi-
cal talus slides. In addition, to evaluate the input
parameters for the Bayesian network, the com-
position (grain size) and structure of talus mate-
rial were analyzed in the laboratory and by in situ
investigations. Extensive laboratory direct shear
tests were also conducted to study the effects of Figure 22. Prior Bayesian network for assessment of
rock content, rock shape, and soil properties on talus landslide risk (Liu 2011).
the shear resistance of the talus. Model tests inves-
tigated the effect of construction procedure on the
deformation of talus slopes. Table 9. Examples of nodes and their states in the
Building the Bayesian network of talus land- Bayesian network of Figure 21 (Liu 2011).
slide risk is complex. The network was built
Node State 1 State 2
by assembling relevant expert knowledge. The
nodes were divided into three classes: hazard Lithology Sandstone Mudstone
factor node, event node and loss node (Fig. 22). Soil type Silty clay Clay
Each node was characterized by several discrete Gravel content ≤50% >50%
Veget. cover Dry land Mostly woods
Slope angle ≤30º >30º
Travel length ≤60 m >60 m
Volume ≤106 m3 >106 m3
Intensity Weak Strong
Time 06:00∼18:00 18:00∼0:00
Fatalities Badly injured ≤ 3 Badly injured > 3
or death ≤ 1 or death > 1
Economic loss <3% of investment >3% of investment
Figure 21. Structural damage after talus slide: damaged Time overrun ≤30 days >30 days
bridge piers (left) and crack in retaining wall (right) (Liu
2011). *States for marlite (3) and limestone (4) are not shown.

15

ISGSR2013.indb 15 10/18/2013 9:36:29 AM

 types were obtained through a geological survey; the landslides that have occurred. The states of the
(2) to predict the scale of landslides, a stochas- fatalities and losses as well as the corresponding
tic model for generating the talus was developed probabilities are listed in Table 10. Such Bayesian
based on Monte Carlo simulation and realized network could serve as an effective tool to manage
using the AutoCAD VBA program, accounting talus landslide risk, provided that the information
for gravel distribution, shape, position, size and for the prior is available.
content (Fig. 23 left). The finite element software
ABAQUS was used to analyze the most likely slip
surface and travel length (Fig. 23 right). 5 RISK OF TAILINGS DAM BREACH
Using past failure and the spatial model devel-
oped, a decision-making Bayesian network was In mid-career, Wilson Tang published 10 contri-
built to predict the potential economic loss, con- butions on the safety of dams (e.g. Tang & Yen
struction delay and time overrun and fatalities due 1991; Cheng et al. 1993)). Although he did not
to a talus slide. work on this aspect in his later years, his colleagues
Figure 24 presents the posterior probabilities at HKUST distinguished themselves in this area,
of the Bayesian network together with the avail- perhaps also inspired by the work of Wilson (e.g.
able evidence. The zones in red in the monitor win- Xu & Zhang 2009).
dows indicate the parameters that have complete The case study below is a hazard and risk analy-
certainty due to the information acquired. If one sis performed by NGI to estimate the probability of
compares Figures 22 and 24, the risk for losses and non-performance of a tailings management facility
casualties for State 2, after updating, increased as designed for gold mine development in Romania
expected for this specific talus landslide in light of (www.gabrielresources.com/prj-rosia.htm). The
analyses were to establish whether or not the dam
would provide acceptable safety against release of
tailings and toxic water, and whether or not addi-
tional hazard reducing measures were needed. The
project lies within the existing Roşia Montană
mining district north-east of the town of Abrud in
the Apuseni Mountains of Transylvania.
The project should mitigate the consequences
of the historic and future mining operations with
Figure 23. Bayesian network for estimating talus landslide
risk: (left) stochastic model to generate talus; (right) slope
the interception and containment of contaminated
failure calculation with ABAQUS software (Liu 2011). water currently entering the system, treatment of
the contaminated waters and isolation and recovery
of the waste rock piles within the project boundary.
The operation of the project will generate tailings
for approximately 17 years, producing tailings from
the processing of a total of approximately 215 Mt
of ore. The Tailings Management Facility (TMF)
in the valley includes a Starter Dam as a first stage
of the Completed Dam, a Secondary Contain-
ment Dam, a tailings delivery system, a reclaim
water system and a waste rock stockpile (Fig. 25).
The TMF is to provide the required design storage
capacity for the life of the mine, plus an additional
Figure 24. Posterior (updated) Bayesian network for contingency capacity.
estimating specific talus landslide risk (Liu 2011).

Table 10. Results of Bayesian updating of risk associated with talus landslide along
the Shuifu-Maliuwan highway.

Economic loss* Time overrun* Casualties*

State Prior Posterior Prior Posterior Prior Posterior

1 60% 45% 52% 37% 38% 32%

2 40% 55% 48% 63% 62% 68%

*See Table 9 for definition.

16

ISGSR2013.indb 16 10/18/2013 9:36:29 AM

 – Emergency spillway to control any excess water
released.
– Comprehensive geotechnical monitoring system
for safety surveillance.
– Careful control of construction by owner and
contractor/engineer.

Figure 25. Cross-section of tailings dam in Romania 5.2 Event tree analysis
(Corser, P. 2009. Personal comm. MWH Americas Inc.
Bucharest, Romania). To establish whether the dam provides acceptable
safety against “uncontrolled” release of tailings and
To establish whether the dam provides acceptable water during its life, an event tree analysis was done.
safety against “uncontrolled” release of tailings A workshop was organized to develop the event trees
and water during its life, an event tree approach was and reach a consensus when quantifying the hazards.
used to do the hazard analyses. This technique iden- The analysis involved breaking down the complex
tified potential failure mechanisms and followed system into its fundamental components, and deter-
how a series of events leading to non-performance mining the potential “failure” mechanisms leading
of a dam might unfold. The probability of each sce- to non-performance of the dam and the physical
nario, given a triggering event, was quantified. processes that could cause such mechanisms.
The event tree hazard analyses considered the The key factors considered in the analyses
dam at different stages of its life and estimated included: dam configuration (Starter Dam, dam
the probability of non-performance. A non- during construction and Completed Dam), and
satisfactory performance of the dam was defined triggers, including earthquake shaking, extreme
as an uncontrolled release of tailings and water rainfall or snowmelt, natural terrain landslide in
from the dam over a period of time. The release the valley or failure of the waste stockpile into the
could be due to a breach of the dam or overtop- tailings reservoir.
ping without breach of the dam. The analyses Acts of war or sabotage, impact by meteorites
looked at critical scenarios, including all potential or other extreme events of this type were not con-
modes of non-performance under extreme triggers sidered, as they would result in so low probabilities
such as a rare, unusually strong earthquake and of non-performance that they are not realistic to
extreme rainfall in a 24-hour period. consider.
The non-performance modes considered
included:
5.1 Design considerations
1. Foundation failure, due to, e.g. excess pore
The most significant requirements that influenced pressures or weak layer in foundation leading to
the probabilities in the hazard analyses include: cracking, instability and breach of the dam.
– Operational freeboard at all times of one meter 2. Dam slope instability downstream or upstream,
above storage level for maximum reclaim pond due to e.g., construction pore pressure in core
and 2 PMP (probable maximum precipitation); of Starter Dam, excessive pore pressures caused
the requirement leads to a storage volume capac- by static or earthquake loads or instability due
ity of two 1/10,000-yr rainfall within the same to inertia forces.
24 hours. 3. Unravelling of downstream toe and slope, due
– Gentle slopes for the Starter Dam (≈2H:1V to e.g. overtopping or excessive leakage through
upstream and ≈2H:1V downstream). or under the dam. This can be caused by a slide
– Gentle downstream slopes for the Completed into the reservoir, dam crest settlement due to
Dam (3H:1V). deformations of the Starter Dam, piping, inter-
– Good quality rockfill for the Starter Dam con- nal erosion and sinkhole formation, or exces-
struction and the Completed Dam. sive deformations (slumping) of the top vertical
– “Well drained” tailings beach at the upstream part of the Completed Dam during earthquake
face of the dam, where equipment can move shaking.
in for repairs, in case of movement or partial 4. Dam abutment failure followed by breach, due to
breach. e.g. slide close to and/or under part of the dam.
– Secondary Containment Dam (SCD) with about 5. Liquefaction of the tailings.
50,000 m3 containment capacity after 16 years. Figure 26 presents some of the configurations
– Diversion channels along the sides of the valley to and examples of the non-performance modes
divert excess rainfall runoff away from the TMF analyzed. Overtopping without breach of the
pond to minimize the risk of overtopping. dam, including under-capacity or damage of the

17

ISGSR2013.indb 17 10/18/2013 9:36:30 AM

 Figure 26. Examples of non-performance modes.

Secondary Containment Dam was also considered, Table 11. Total probabilities of non-performance.
not as a separate non-performance, but as one of
the events in the sequence of events in the trees. P [non-
Configuration performance]
Different conditions can affect the probabil-
ity of a hazard occurring or severity of a conse- Starter dam (t = 1.5 yr, internal erosion) 1.3 × 10−6 /yr
quence, for example construction deficiencies or Completed dam (t = 16 yrs) 1.3 × 10−6/yr
inadequate response of the field control team at Intermediate stage (t = 4 yrs) 6.5 × 10−7/yr
the site when warning signals may appear. The
Intermediate stage (t = 9–12 yrs) 1.3 × 10−6/yr
analyses also looked into construction deficien-
cies, e.g. inadequate filters leading to uncontrolled
internal erosion, inadequate drainage, very weak
construction layers or zones in the embankment, considered successively. The total probability of
inadequate types of material(s) in the embankment non-performance is the sum of all contributing
fill, or insufficient quality control and unforeseen probabilities to the non-performance for each
construction schedule changes. These conditions of the dam configurations. Table 11 presents the
were also integrated in the event trees as separate total probabilities for each configuration of the
events during the course of the construction of the dam (all triggers included). The probabilities were
Starter Dam and Completed Dam. presented as a function of the release of tailings
and water associated with the non-performance of
the dam. The highest annual probability of non-
5.3 Probability of non-performance
performance was 10−6.
At the event tree workshop, the critical times in the The highest probabilities of non-performance
life of the TMF were defined: during construction were associated with earthquake shaking of the
of the Starter Dam, during the downstream con- completed dam and the static liquefaction of the
struction stages, during the centerline construc- tailings at time 9 to 12 years after the start of
tion of the dam, and/or in the early years after the construction. The non-performance scenarios
Completed Dam is built. A matrix of dam config- would result in some material damage and some
uration versus time was prepared. The modes seen contamination, but only in the vicinity downstream
as most critical and susceptible to lead to the high- of the dam. For the Starter Dam, no reasonable
est probabilities of non-performance were listed. expected scenario lead to a significant release of
As part of the mode screening, the following con- tailings and water because of the limited quantity
siderations were subjected to a consensus decision: of water available and the reserve capacity provided
extreme and critical precipitation (rainfall, flood (2 PMP’s). Internal erosion may cause, with an
and snowmelt), likelihood of failure of the waste annual probability of 10−6, a small escape of tailings
stockpile, critical situations after construction of and water. The escape would cause only modest con-
the dams, and geo-environmental considerations. tamination of the immediate vicinity downstream.
Event trees were developed for each dam con- Essentially all material released could be contained
figuration and trigger, with each non-perform- by the Secondary Containment Dam.
ance mechanism looked at separately. In some The analyses showed (1) no plausible events
cases, two non-performance mechanisms were result in an annual probability of non-performance

18

ISGSR2013.indb 18 10/18/2013 9:36:30 AM

 greater than 10−6. The probabilities are lower than of the dam, gentle downstream slopes for both the
the values considered as acceptable criteria for Starter and the Completed Dam, dam capacity
dams and other containment structures around to store extreme precipitation and/or snowmelt
the world and lower than probabilities of non- events, spillway to release excess water in a control-
performance for most other engineered structures. led manner, the safety monitoring and early warn-
ICOLD (the International Commission on Large ing of early signs of unexpected performance, and
Dams) presented statistics of dam incidents where the proposed preparedness to remediate, given any
the mean probability of failure is between 10−4 and unexpected behavior.
10−5 per year (Londe 1993; ICOLD 1995; Foster et al.
2000; Høeg 2001). Peck (1980), based on work by 5.4 Environmental impact
Baecher et al. (1980a; b) who used the ICOLD data- The physical impact in terms of damage to the
base plus other data, reported that the probability of environment was also studied, if a breach in the
failure of dams in the United States and worldwide, dam should occur. The analysis suggested that
was between 2 and 7 × 10−4 per year. Foster et al. the released tailings’ volume would be limited, and
(2000) reported annual probabilities of an accident would only flow 100 to 200 m (Fig. 28). Studies were
due to downstream slope instability of 1 to 5 × 10−4 also conducted to determine possible pollution of
and an annual probability of failure of 1.5 × 10−5. the river downstream. The levels of pollution may
Historical data are available for embankment be above surface water discharge standards for a
dams that provide failure frequency per dam-year limited period of time and in the immediate neigh-
of operation. Figure 27 shows internal erosion borhood of the tailings dam, but only under the
failure frequencies for US dams. The annual prob- worst case conditions (low flow in the downstream
ability of failure associated with internal erosion river). However, monitoring, early warning and
of earth dams is between 10−4 and 5 × 10−4 per emergency procedures are to be implemented to
year. Internal erosion failures tend to occur more contain damage to a minimum. The weather and
frequently in the first 5 years reflecting first-filling flow conditions for this to occur combined with
failures. The data suggest significantly higher prob- the probability of dam breach occurring at the
abilities of failure than what was computed for the same time resulted in the probability of occurrence
TMF at Roşia Montană. would reduce to 10−7/year (Whitehead, P. 2009.
For tailings dams, the probability of failure is sig- Personal comm. Aquatic Environments Research
nificantly higher than the average annual probability Centre, Univ. of Reading, UK).
of 10−4 and 10−5 reported above. Most of the tailings
dams are dams entirely made of tailings, whereas 5.5 Risk assessment of dams in practice
the Roşia Montană TMF is made up of the Starter
Dam (a regular type rockfill embankment dam), and The example illustrates that the event tree analysis is
when completed to top grade, has a downstream a systematic application of engineering judgment.
slope made of rockfill with gentle inclination of 1:3. Its application does not require the prior existence
The probabilities of failure in Figure 27 are of extensive statistics or the application of com-
higher than the probability of non-performance plex mathematics. The process may provide mean-
computed for the TMF at Roşia Montană. The ingful and systematic estimates and outcomes on
event tree analyses show that the probability of the basis of subjective probabilities (Vick 2002).
non-performance of the TMF is about 100 times With increasing frequency, society demands
lower than the probability of failure of contain- that some form of risk analysis be carried out for
ment dams, based on the performance observed activities involving risks imposed on the public. At
for dams around the world. the same time, society accepts or tolerates risks in
The factors that contribute to the low probabil- terms of human life loss, damage to the environ-
ity of non-performance of the TMF include the use ment and financial losses in a trade-off between
of good quality rock fill for the downstream shell extra safety and enhanced quality of life.
The role of the dam engineering profession
is to explain the uncertainties involved in the

Figure 27. Annual probability of dam failure by inter- Figure 28. Physical impact of TMF dam breach at
nal erosion for different dams in the USA (Von Thun, Roşia Montanǎ (Corser, P. 2009. Personal comm. MWH
1985; Vick 2002). Americas, Inc. Bucharest, Romania).

19

ISGSR2013.indb 19 10/18/2013 9:36:30 AM

 construction and operation of dams and to present retaining wall by relating conventional factors of
the likelihood of incidents and failure in informa- safety (load and resistance factors) to estimated
tive and meaningful terms. The conventional use probabilities of failure. Wilson Tang reviewed and
of a factor of safety just does not do that, and con- commented on the manuscript, and he was really
cepts from probability theory and reliability analy- the first who worked systematically with model
ses should be applied. uncertainty. He wrote comprehensive reports for
The key to making the risk analysis of dams the American Petroleum Institute (e.g. Tang 1988)
effective begins with a detailed overview of all and several papers and discussions (e.g. Tang &
potential failure modes. If shortcuts are taken, the Gilbert 1993b). His efforts were crowned in May
results could be misleading. Once the potential fail- 2013 with the induction in the Hall of Fame of the
ure modes are understood, the screening process paper by Tang et al. (1990) on the performance
will identify the critical modes. A variety of tools reliability of offshore piles.
are available for making the quantitative risk esti- Wilson Tang was always concerned with two
mates. The event tree approach is useful and illus- aspects: (1) the models used to quantify model
trative. It is recognized that risk estimates and risk uncertainty should duplicate as closely as pos-
assessment guidelines are only approximate, but sible the problem situation actually being calcu-
they are useful for choosing among alternatives, lated (Tang & Gilbert 1992), and (2) the profession
comparing risk levels, and making decisions. should improve its ability to use experimental
Høeg (2001) presented the basics of system- results to determine the uncertainties in its engi-
atic risk analysis for dams. He concluded that neering models.
after several years of optimism in the profes- Today, work on this topic is still on-going.
sion with developing and performing meaningful Lacasse et al. (2013 a; b; c) made a contribution
quantitative probabilistic risk analyses for dams, which follows and expands on Wilson Tang’s prin-
there now seems to be a trend towards increased ciple. The ultimate aim of the work was to obtain
use of the qualitative FMECA approach, or the the appropriate factor of safety to use when design-
Failure Modes, Effects and Criticality Analysis ing offshore installations. To illustrate this, Gilbert
(BSI 1991). However, there is increased pressure et al. (2013) present at this conference the case of
on the decision-makers to quantify risk level so three actual offshore structures, presently under
that it can be compared to acceptable or toler- final design, where such calibration of the load
able risk or public protection guidelines. In The and resistance safety factors was done. Only the
Netherlands, the development in this direction is approach and the conclusions are briefly reported
quite advanced and used in the safety evaluation herein.
and upgrading of dikes and storm surge barriers The study was undertaken to document that
(Vrijling 2001). the pile foundations were designed according to
Scott (2011) summarized the practice of risk governing regulations. The goal was to make a rec-
assessment of dam safety of the US Bureau of ommendation on the appropriate resistance factor
Reclamation. Aging infrastructure, population and minimum pile penetration depth to use for the
growth downstream and limited resources render design of the piles on an offshore jacket. The safety
risk assessment of dam safety a reasonable and factors (load and resistance factors) for three case
transparent method for risk management. Key studies were calibrated for a target annual prob-
to making the process effective is starting with a ability of failure, Pf, of 10−4.
detailed analysis of the potential failure modes. The reliability analyses of the axial pile capacity
Scott described a variety of tools available to do methods included seven steps:
quantitative risk estimates. Such estimates and risk
assessment guidelines are only approximate. In – Establish the mean, standard deviation and
each case, it is essential to build the argumentation Probability Density Function (PDF) of the
for the ability of the structure to withstand future soil parameters. Include correlations among
loadings. If done diligently and openly, risk assess- parameters.
ment is a very effective tool for managing the risks – Establish the model uncertainty for the different
associated with containment facilities (Hartford & pile capacity calculation methods used.
Baecher 2004). – Establish the effect of cyclic loading on the axial
pile capacity and determine whether the piles in
compression or in tension govern the design.
6 MODEL UNCERTAINTY AND – Develop a model for the statistics of the static
CALIBRATION OF SAFETY FACTORS (permanent) and environmental loads on the top
of the piles.
In an early paper, Høeg & Murarka (1974) stud- – Do deterministic analysis of the ultimate axial
ied the balanced, yet optimum, design of a gravity pile capacity, Qult.

20

ISGSR2013.indb 20 10/18/2013 9:36:31 AM

 – Do probabilistic analyses of axial pile capacity demonstrate that the annual probability failure
and obtain the PDF of the ultimate capacity, Qult. was less than the target Pf of 10−4/year for the piles
– Calculate the annual probability of failure originally designed with a resistance factor of 1.5.
by combining the loads and the probabilistic It was then possible to use a resistance factor of
description of Qult. 1.3, as for the current API method, instead of the
– Calibrate the load and resistance factors required a priori resistance factor of 1.5 set for the newer
for an annual Pf = 10−4. CPT design methods. A load factor of 1.3 or 1.35
was used.
When doing an axial pile capacity analysis, the The analyses demonstrated the importance of
following aspects should be included: (1) a careful how the characteristic shear strength parameters
selection of the characteristic soil parameters used are defined. Lacasse et al (2013a) recommended
for design; (2) the effect of cyclic loading on the that the characteristic strength be defined in spe-
characteristic shear strength or ultimate pile capac- cific terms, e.g. setting the characteristic shear
ity, for both piles loaded in compression and in ten- strength for the deterministic design of axial pile
sion; (3) the effect of gapping and/or erosion at the capacity at a value equal to the mean minus ½ or
top of the piles on the axial pile capacity; and (4) a one standard deviation.
decision on whether or not to account for the effect The importance of model uncertainty was
of time after pile installation on the axial capacity. pointed out early by Wilson Tang. This confirms
The calibration analyses showed that: the actuality of even his early papers. Here again,
– The calibration of the safety factors demon- Wilson Tang led the way in his study of model
strate that the annual probability of failure vary uncertainty and calibration of safety factors in the
with the axial pile capacity calculation method. early days of his career.
– The values of model uncertainty used in the
analyses have an overwhelming influence on the
resulting annual probability of failure and there- 7 MORE OF WILSON TANG’S LEGACY
fore on the required resistance factor for a target
annual probability of failure. 7.1 Cost-effectiveness of site investigation
– The current state-of-the-art design still relies
heavily on qualified engineering judgment to Wilson Tang worked on the cost-effectiveness of
assess and ensure a consistent safety level. site investigations, a central aspect of our profes-
– The resistance factors calibrated suggest that the sion. His contribution (Tang, 1987), published
newer CPT-methods of pile design are as reli- in the journal Structural Safety, may have passed
able as the current API method. unnoticed.
– The findings on margin of safety and the defi- In general, more extensive site investigations and
nition of the characteristic shear strength have laboratory testing programs reduce the uncertain-
important implications for the design of piles ties in the soil characteristics and design param-
offshore and can result in significant savings. eters. At a certain point however, as Wilson Tang
(1987) pointed out, the benefit obtained from fur-
As illustrated in Gilbert et al. (2013), the pile ther site investigations and testing may not yield
length could be considerably reduced through the sufficient added value (read: increase in the reli-
study of a safety level corresponding to an annual ability of the performance) to the geotechnical sys-
probability of failure of 10−4. Table 12 reproduces tem, and hence may not justify the additional cost
the final results, comparing pile penetration depths. (e.g. Folayan et al. 1970).
The first number is the penetration depth from the Soil investigations, in the way they are planned,
deterministic analyses with a resistance factor of represent a risk-based decision. The complexity of
1.5 on the CPT-methods. The second number is the a soil characterization is based on the level of risk
penetration depth ensuring that the annual prob- of a project. Lacasse & Nadim (1998; 1999) illus-
ability of failure is less than 10−4. trated this graphically. A low risk project involves
The significant reduction in the required pile few hazards and has limited consequences. Simple
penetration depth was possible because one could in situ and laboratory testing and empirical cor-
relations would be selected to document geotech-
nical feasibility. In a moderate risk project, there
Table 12. Pile penetration depth for design (Lacasse are concerns for hazards, and the consequences
et al. 2013c). of non-performance are more serious than in the
former case. Specific in situ tests and good quality
Site A (clay) Site B (sand) Site C (clay and sand) soil samples are generally planned. For a high-risk
90 m to 75 m 51 m to 27 m 45 m to 38 m project involving frequent hazards and potentially
risk to life or substantial material or environmental

21

ISGSR2013.indb 21 10/18/2013 9:36:31 AM

 damage, high quality in situ and laboratory tests turbines (Stuyts et al. 2013). Lacasse & Nadim
are required, and higher costs are involved. (2007) also summarized the applications of sta-
The decision-making process for selecting the tistics, reliability and risk in offshore geotechnical
appropriate soil investigation methods, although engineering based on the original work by Wilson
subconscious, is risk-based. It involves considera- Tang.
tion of requirements, consequences and costs. The methods for assessing hazards offshore can
Uncertainty analysis can help optimize site vary from approximate estimates to more complex
investigations. The uncertainty in a geotechnical calculations. Applications include piled founda-
calculation is often related to the possible presence tions, jack-up structures, gravity foundations
of an anomaly, e.g. boulders, soft clay pockets or and underwater slopes. The applications dem-
drainage layer. Probability approaches can be used onstrate that probabilistic analyses complement
to establish the cost-effectiveness of additional the conventional deterministic safety factor and/
site investigations to detect anomalies. Figure 29 or deformation-based analyses, and contribute to
presents an example where the presence of a achieving a safe and optimum design. The proba-
drainage layer was determinant on the resulting bilistic approach adds value to the results with a
post-construction building settlements. A settle- modest additional effort. Engineering judgment
ment of less than 50 cm would mean an important is still necessary to achieve reliable results in both
reduction in costs. With drainage layer detectabil- hazard and risk assessment.
ity for each boring of 50% or 80% (Fig. 28), and
assuming a given drainage layer extent, 3 to 6 bor-
8 SUMMARY AND CONCLUSIONS
ings were required in this case to establish whether
the drainage layer was present or not.
Wilson Tang’s work was an inspiration to move
forward in the area of statistics, probability and
7.2 Reliability of offshore structures
reliability. He quickly saw the potential of these
Wilson Tang started working with offshore struc- concepts in geotechnical engineering. He published
tures (Høeg & Tang 1977) when he came to NGI his first book, together with Professor A.H-S Ang,
on the Guggenheim research fellowship. Wilson one of the most useful and influential sources of
was very much indebted to the John Simon information on the topic for geotechnical engi-
Guggenheim Memorial Foundation for making neers, only six years after completing his PhD at
possible his research stay first at the Imperial Col- Stanford University. With Ang and Tang’s two vol-
lege of Science and Technology in London and umes, one can find all the essential concepts and
then at NGI in Oslo, Norway. very many applications.
From there on, he continued his research and This paper presented only a few examples of
became a recognized figure in offshore circles, espe- how geotechnical engineers have taken the learn-
cially with respect to model uncertainties and the ings of Wilson Tang and carried on with further
reliability of pile foundations offshore (e.g. Tang & applications in practice. The quantification of the
Gilbert 1992; 1993a). Noteworthy are his studies natural and anthropogenic risks that can affect an
for the American Petroleum Institute, which con- area or engineering structures is today an essential
clusions are still in use today. component of a sustainable environment, land-use
Gilbert et al. (2013), at this conference, give planning, and risk mitigation. To this development,
an overview of the advances in geotechnical risk Wilson Tang was a pioneer and before his time!
and reliability for offshore applications. The les- Wilson Tang was quick to see the importance
sons learned from Wilson Tang are now also used and possible repercussions of using Bayesian
to calculate the reliability of offshore wind energy updating in geotechnical engineering.

Figure 29. Cost reduction with increased number of borings (Lacasse & Nadim 1998 based on Tang 1987).

22

ISGSR2013.indb 22 10/18/2013 9:36:31 AM

 The advantages of the Bayesian approach include: profession over decades of practice, the probabilistic
(1) it is a probabilistic model instead of a determin- toolbox provides a complement to deterministic
istic model. The uncertainties in the parameters analyses, and should be used for several reasons,
and their inter-relationships are represented by including (in Wilson Tang’s own words):
probabilities; (2) a large number of parameters and
their inter-relationships can be considered in a sys- • “(…) Society is demanding more explicit assess-
tematic structure. The probabilities of one param- ment of risk. (…) To work effectively with the
eter can be updated via available information. The public and (…) regulatory agencies, geotechnical
change in one parameter will influence the others in engineers must have some knowledge of prob-
the network through their inter-relationships; and ability theory and probability methodologies, as
(3) physical mechanisms, previous studies, and sta- well as traditional geotechnical expertise.”
tistical data can be accounted for. All three aspects • “Probabilistic methods are useful as a basis for
are key to good geotechnical design. making economic decisions. [For example,] in
The profession only gains by implementing, areas such as dam rehabilitation, landslide haz-
more systematically than before, probabilistic- ards mitigation, environmental remediation, and
based thinking and risk-based methodology. The infrastructure rehabilitation, effective allocation
geotechnical probabilistic approach still has major of funds relies on quantifying the trade-offs
needs, including reducing uncertainty in the calcu- between benefits and risks (…). Probabilistic
lation model by obtaining and analyzing perform- methods provide a quantitative basis through
ance data of high quality, quantifying acceptable which the relative contribution of risk can be
and tolerable hazard and risk levels, and convinc- systemically analyzed and communicated. In
ing stakeholders of the value added in uncertainty- this way, decisions can be made more rationally
based analyses. and justified more logically.”
With the changes in climate and the occurrence • “(…) There is a risk in risk analyses and prob-
of more extreme natural phenomena than before ability analyses, if the analyses are performed
(e.g. storms and precipitation), one cannot only use improperly. This possibility can be minimized by
data from existing experience to evaluate safety, but expanding knowledge of probabilistic methods
one should also include events and triggers that are among geotechnical engineers and by expanding
not covered by e.g. 100- or 1000-year return peri- knowledge of geotechnical engineering practice
ods. Another keyword is the importance of multi- among probability specialists.”
disciplinarity, meaning wider expertise teams than • “(…) With a working knowledge of probabil-
before when evaluating hazard and risk to society, ity theory, geotechnical engineers will be better
and the need to document cost-effectiveness of dif- equipped to deal with the many uncertainties
ferent mitigation measures. that pervade geotechnical engineering practice.”
Bayesian updating and hazard and risk analy- In closing, the authors wish to express their grat-
sis are important and necessary. Hazard and risk itude to Wilson, not only for his competence and
assessment present an opportunity to look at the invaluable scientific contribution, but also for his
bigger picture and seek out designs that meet not friendship, his kindness, thoughtfulness and help
just some arbitrary idea of acceptable/tolerable with articles, discussions, workshops and presenta-
risk but an unknown risk. The engineer should tions, over many years. One example: when NGI
concentrate on exploiting the good features of the decided in the early 80’s to offer an internal edu-
approach. As contributor to the profession’s goals cation program on the practice of statistics and
of documentation, continuity, high-quality and probability in geotechnical engineering, we chose
innovation, and the ever increasing requirement Ang & Tang Part I (1975) as textbook. Hearing
of globalization, hazard and risk assessment and this, Wilson immediately sent NGI his book of
the management of risk serve as communication worked out solutions to all the problems in the
vehicle among geo-specialists and other sectors of book, which turned out to be a godsend. We still
expertise. The authors are convinced that Wilson use this booklet of solved examples!
Tang would have appreciated working on these
emerging aspects and would have contributed with
his usual innovation, elegance and wisdom.
After a special workshop on Reliability Methods REFERENCES
for Risk Mitigation in Geotechnical Engineering at
Ambraseys, N.N., Douglas, J., Sarma, S.K. & Smit, P.M.
Irvine in 1992, Tang & Duncan (1994) concluded 2005. Equations for the estimation of strong ground
that probability methods should be part of the geo- motions from crustal earthquakes using data from
technical engineer’s toolbox. Although every suc- Europe and the Middle East: horizontal peak ground
cessful geotechnical engineer has learned to cope acceleration and spectral acceleration. Bulletin of
with uncertainty by applying lessons learned by the earthquake engineering 3:1–53.

23

ISGSR2013.indb 23 10/18/2013 9:36:31 AM

 Ang, A.H-S. & Tang, W.H. 1975. Probability concepts Safety and Risk. ISGSR 2013. 4–6 Dec. Hong Kong
in engineering planning and design. Volume I Basic Univ. Science & Technology.
principles. John Wiley & Sons. 409 p. Griffiths, D.V., Paiboon, J., Huang, J.S. & Fenton, G.A.
Ang, A.H-S. & Tang, W.H. 1984. Probability concepts in 2013. Homogenization of geomaterials using the
engineering planning and design. Volume II. Decision, random finite element method. Keynote Lecture.
risk and reliability. John Wiley & Sons. 562 p. 4th Intern. Symp. Geotechnical Safety and Risk.
Ang, A.H-S. & Tang, W.H. 2007. Probability concepts ISGSR 2013. 4–6 Dec. Hong Kong Univ. Science &
in engineering. Emphasis on applications to civil & Technology.
environmental engineering. 2nd Ed., John Wiley & Gutenberg, B. & Richter, C.F. 1944. Frequency of earth-
Sons. 406 p. quakes in California. Bulletin of the Seismological
Baecher, G.B. & Christian, J.T. (2003). Reliability and Society of America 34: 185–188.
statistics in geotechnical engineering. Wiley & Sons. Hacking, I. 1975. The emergence of probability.
605 p. Cambridge University Press. 209 p.
Baecher, G.B., Paté, M.-E. & de Neufville, R.1980a. Dam Harr, M.E. 1987. Reliability-based design in civil engi-
Failure in Benefit/Cost Analysis. ASCE Jour. Geot. neering. New York: McGraw-Hill. 290 p.
Eng. Div. 106 (GT1): 101–116. Hartford, D.N.D. & Baecher, G.B. 2004. Risk and
Baecher, G., Paté, M.-E. & de Neufville, R. 1980b. Risk uncertainty in dam safety. Thomas Telford London.
of Dam Failure in Benefit-Cost Analysis. Water 390 p.
Resources Research 16(3): 449–456. Huang, H.W., Su, W.W. & Zhang, J. (2013). Selecting
Benjamin, J.R. & Cornell C.A. 1970. Probability, statis- optimal probability models for geotechnical reliability
tics and decision for civil engineers. McGraw Hill Book analysis. Keynote Lecture. 4th Intern. Symp. Geotech-
Company. New York. 684 p. nical Safety and Risk. ISGSR 2013. 4–6 Dec. Hong
BSI 1991. BS 5760 Part 5: Guide to failure modes, effect Kong Univ. Science & Technology.
and criticality analysis (FMEA and FMECA). British Huang, R.Q. 2008. Preliminary analysis of the develop-
Standards Institution. London. ment, distributions, and mechanisms of the geohaz-
BSI 2001. British Standard Association. BS 5760 Part 5: ards triggered by the Great Wenchuan Earthquake
Guide to failure modes, effects and criticality analysis State Key Laboratory of Geohazards Prevention and
(FMEA and FMECA). London. Geological Environment Protection, Chengdu Univer-
Cheng, S.T., Yen, B.C. & Tang, W.H. 1993. Stochas- sity of Technology. Chengdu, China.
tic Risk Modeling of Dam Overtopping, Reliability Høeg, K. 1996. Performance evaluation, safety assess-
and Uncertainty Analyses in Hydraulic Design. B.C. ment and risk analysis for dams. Int. Journal on
Yen and Y.K. Tung, eds. ASCE Special Publication, Hydropower and Dams 3(6): 51–57.
123–132. Høeg, K. 2001. Embankment-dam engineering, safety
Cheung, R.W.M. & Tang, W.H. 2005. Realistic Assess- evaluation and upgrading. Proc. 15th ICSMGE.
ment of Slope Reliability for Effective Landslide Haz- Istanbul. 4 2491–2504.
ard Management. Geotechnique 55(1): 85–94. Høeg, K. & Murarka R.P. (1974). Probabilistic analysis
Einstein, H.H., Sousa, R.L., Karam, K., Manzella, I. & of a retaining wall. ASCE J. of the Geotechnical engi-
Kveldsvik, V. 2010. Rock slopes from mechanics to neering Division 100 (GT3): 349–370.
decision making. Rock Mechanics in Civil and Envi- Høeg, K. & Tang, W.H. 1977. Probabilistic Consid-
ronmental Engineering, J. Zhao, V. Labiouse, J.P. Dudt erations in the Foundation Engineering for Offshore
and J.F. Mathier (eds). London: CRC Press, 3–13. Structures. Proc. 2nd Intern. Conf. on Structural Safety
Elkateb, T., Chalaturnyk, R. & Robertson, P.K. 2003. and Reliability, Munich, Germany, 267–296.
An overview of soil heterogeneity: quantification and ICOLD 1995. Dam failures—statistical analysis, Bulletin
implications on geotechnical field problems. Canadian No. 99. ICOLD, Paris, France.
Geotechnical Journal 40(1): 1–15. Jaksa, M.B. 2006. Modeling the natural variability of an
Fenton, G.A. & Griffiths, D.V. 2005. Three-dimensional over-consolidated clay in Adelaide, South Australia.
probabilistic foundation settlement. Journal of Geo- 2nd International Workshop on Characterization and
technical and Geoenvironmental Engineering, ASCE Engineering Properties of Natural Soils. Singapore,
131(2): 232–239. pp. 2721–2751.
Fenton, G.A. & Griffiths, D.V. 2008. Risk assessment in Jordaan, I. (2008). Decisions under uncertainty. Proba-
geotechnical engineering. Wiley and Sons. 461p. bilistic analysis for engineering decisions. Cambridge
Fenton, G.A., Griffiths, D.V. & Williams, M.B. 2005. Univ. Press. 672 p.
Reliability of traditional retaining wall design. Geo- Juang, CH., Wang, L. & Atamturktur, S. (2013). Robust
technique 55(1): 55–62. design of geotechnical barriers—A new design per-
Folayan, J.I., Høeg, K. & Benjamin, J.R. 1970. Decision spective. Keynote Lecture. 4th Intern. Symp. Geotech-
theory applied to settlement prediction, ASCE J. of nical Safety and Risk. ISGSR 2013. 4–6 Dec. Hong
the Soil Mechanics and Foundation Division 96 (SM4) Kong Univ. Science & Technology.
1127–1141. Kappes, M.S., Keiler, M., von Elverfeldt, K. & Glade,
Foster, M., Fell, R. & Spannagle, M. 2000. The statistics T. 2012. Challenges of analyzing multi-hazard risk:
of embankment dam failures and accidents. Canadian a review. Natural Hazards 64:1925–1958.
Geotechnical Journal 37(5): 1000–1024. Lacasse, S. & Nadim, F. 1996. Uncertainties in charac-
Gilbert, R.B., Lacasse, S. & Nadim, F. 2013. Advances in terizing soil properties. Uncertainty in the Geologic
geotechnical risk and reliability for offshore applica- Environment: From Theory to Practice, ASCE Geo-
tions. Keynote Lecture. 4th Intern. Symp. Geotechnical technical Special Publication No. 58: 49–75.

24

ISGSR2013.indb 24 10/18/2013 9:36:31 AM

 Lacasse, S. & Nadim, F. 1998. Risk and Reliability in capacity. Probabilistic Engineering Mechanics 20:
Geotechnical Engineering. State-of-the-Art paper. 4th 324–341.
International Conference on Case Histories in Geotech- Popescu, R., Prevost, J.H. & Deodatis, G. 1998. Spatial
nical Engineering. St-Louis MO USA. pp. 1172–1192. variability of soil properties: two case studies. Geo-
Lacasse, S. & Nadim F. 1999. Risk analysis in geo- technical Earthquake Engineering and Soil Dynamics,
engineering. Rocksite 1999 Intern. Conf. on Rock ASCE Geotechnical Special Publication 75: 568–579.
Engineering Techniques for Site Characterization, Scott, G.A. 2011. The practical application of risk assess-
Bangalore, India, pp. 1–15. ment to dam safety. ASCE Geo-Risk 2011, Atlanta.
Lacasse, S. & Nadim. F. 2007. Probabilistic geotechnical Keynote lecture. 129–168.
analyses for offshore facilities. Georisk 1(1): 21–42, Staveren, M.Th. van (2013). Integrated geo risk manage-
Lacasse, S., Nadim, F., Andersen, K.H., Knudsen, S., ment: crossing boundaries. Keynote Lecture. 4th Intern.
Eidsvig, U.K., Yetginer, G., Guttormsen, T.R. & Eide, A. Symp. Geotechnical Safety and Risk. ISGSR 2013.
2013a. Reliability of API, NGI, ICP and Fugro Axial 4–6 Dec. Hong Kong Univ. Science & Technology.
Pile Capacity Calculation Methods. Offshore Technol- Stuyts, B., Gilbert. R. & Cathie, D. 2013. A reliability-
ogy Conference. OTC-24063. Houston, Texas. 22 p. based interpretation framework for pile-supported
Lacasse, S., Nadim, F., Langford, T., Knudsen, S., wind turbines. Offshore Technology Conference. OTC-
Yetginer, G., Guttormsen, T.R. & Eide, A. 2013b. 24011. Houston, Texas. 10 p.
Model Uncertainty in Axial Pile Capacity Design Tang, C., Zhu, J. & Qi, X. 2011. Landslide hazard assess-
Methods. Offshore Technology Conference. OTC- ment of the 2008 Wenchuan earthquake: a case study
24066. Houston, Texas. 10 p. in Beichuan area. Canadian Geotechnical Journal 48:
Lacasse, S., Nadim, F., K.H., Knudsen, S., Eidsvig, U.K., 128–145.
Liu, Z.Q., Yetginer, G. & Guttormsen, T.R. 2013c. Tang, W. & J.M. Duncan 1994. Role of reliability meth-
Reliability of axial pile capacity calculation methods. ods in geotechnical engineering, Geotechnical News.
GeoMontréal 67th Canadian Geotechnical Conference. June. pp. 12–13.
Paper 699. Montréal Canada, 30 Sept–3 Oct 2013. 8 p. Tang, W.H. 1987. Updating anomaly statistics—single
Liu, Z.Q. 2011. Study on stability mechanism and risk anomaly case. Structural Safety 4: 151–163.
decision making for talus slope. PhD dissertation sub- Tang, W.H. 1988. Offshore axial pile design reliability.
mitted to Tongji University. March 2011. 169 p (in Research Report for Project PRAC 86-29B, American
Chinese). Petroleum Institute. Washington DC.
Liu, Z.Q. & Nadim, F. 2013. A three-level framework for Tang, W.H. and Gilbert, R.B. 1992. Offshore piles sys-
multi-risk assessment. 4th Intern. Symp. Geotechnical tem reliability. Research Report Project PRAC 89–29.
Safety and Risk. ISGSR 2013 Dec 4–6. Hong Kong American Petroleum Institute (API).
Univ. Science & Technology. Tang, W.H. & Gilbert, R.B. 1993a. Case study of offshore
Londe, P. 1993. Safety evaluation using reliability pile system reliability. OTC 7196 25th Annual Offshore
analysis. Proc. Workshop on Dam Safety Evaluation. Technology Conference. Houston TX. 677–686.
Grindelwald, Switzerland. 4: 171–188. Tang, W.H. & Gilbert, R.B. 1993b. Discussion Model
Lunne, T., Robertson, P.K. & Powell, J.J.M. 1997. Cone Uncertainty Representation in Geotechnical Reli-
penetration testing in geotechnical practice. London: ability Analyses, J. Geotechnical Engineering ASCE.
Spon Press. 312p. 119(GT11) 1859–1861.
Marzocchi, W., Garcia-Aristizabal, A., Gasparini, P., Tang, W.H., Woodford, D.L. & Pelletier, J.H. 1990. Per-
Mastellone, M.L. & Di Ruocco, A. 2012. Basic prin- formance Reliability of Offshore Piles. OTC 6379 22nd
ciples of multi-risk assessment: a case study in Italy. Annual Offshore Technology Conference. Houston TX.
Natural Hazards 62: 551–573. 299–308.
Nadim, F. & Liu, Z.Q. 2013a. Quantitative risk assess- Tang, W.H. & Yen, B.C. 1991. Dam Safety Inspection
ment for earthquake-triggered landslides using Scheduling, Journal of Hydraulic Engineering, ASCE
Bayesian network. Proc. 18th ICSMGE. Paris. Special publication, 214–229.
Sept 2013. Uzielli, M., Lacasse, S., Nadim, F. & Lunne, T. 2006a.
Nadim, F. & Liu, Z.Q. 2013b. Framework for multi-risk Uncertainty-based characterisation of Troll marine
assessment, Deliverable D5.2. New methodologies clay. 2nd International Workshop on Characterisation
for multi-hazard and multi-risk assessment methods and Engineering Properties of Natural Soils, Singapore.
for Europe. MATRIX FP7 EU-project. With Garcia- 4: 2753–2782.
Aristizabal, A., Woo, G., Aspinall, W., Fleming, K., Uzielli, M., Lacasse, S., Nadim, F. & Phoon, K.K. 2006b.
Vangelsten, B.V., and van Gelder P. Contract No. Soil variability analysis for geotechnical practice.
265138. 2nd International Workshop on Characterisation and
Newmark, N.M. 1965. Effects of earthquake on dams Engineering Properties of Natural Soils, Singapore, 3:
and embankments. Geotechnique 15(2): 139–160. 1653–1752.
Peck, R.B. 1980. Where has all the judgment gone? The Vanmarcke, E.H. 1977. Probabilistic modeling of soil
Fifth Laurits Bjerrum Memorial Lecture. Canadian profiles. J. Geotechnical Engineering Division ASCE
Geotechnical Journal 17(4): 584–590. 103(11): 1227–1246.
Phoon, K.K. & Kulhawy, F.W. 1999. Characterisation of Vanmarcke, E.H. 1983. Random Fields: analysis and syn-
geotechnical variability. Canadian Geotechnical Jour- thesis. Cambridge: MIT Press. 382p.
nal 36: 612–624. Vick, S. 2002. Degrees of belief: Subjective probability
Popescu, R., Deodatis, G. & Nobahar, A. 2005. Effects and engineering judgment. ASCE Press. Reston, VA.
of random heterogeneity of soil properties on bearing 455p.

25

ISGSR2013.indb 25 10/18/2013 9:36:31 AM

 Von Thun, J. 1985. Application of statistical data from Zhang, J., Zhang, L.M. & Tang, W.H. 2009a. Bayesian
dam failure and accidents to risk-based decision analysis framework for characterizing geotechnical model
on existing dams. U.S. Bureau of Reclamation, Engi- uncertainty. J. Geotechnical and Geoenvironmental
neering and Research Center, Denver, Colorado. Engineering, ASCE 135(7): 932–940.
Vrijling, J.K. 2001. Safe dams and dykes—How safe?— Zhang, L.L., Tang, W.H. & Zhang, L.M. 2009b. Bayesian
Proc. Symp. Dams and dykes in development. ICOLD- model calibration using geotechnical centrifuge tests.
ICID. Delft, The Netherlands. March 2001. J. Geotechnical and Geoenvironmental Engineering,
Wong, H.N. (2013). Is landslide risk quantifiable and ASCE 135(2): 291–299.
manageable? Keynote Lecture. 4th Intern. Symp. Zhang, L.M., Nadim F. & Lacasse S. 2013. Multi-risk
Geotechnical Safety and Risk. ISGSR 2013. 4–6 Dec. assessment for landslide hazards. Proc. Pacific Rim
Hong Kong Univ. Science & Technology. Workshop on Innovations in Civil Infrastructure Engi-
Xu, Y. & Zhang, L.M: 2009. Breaching parameters for neering, S.S. Chen, A.H-S. Ang and I.T. Yang (eds.),
earth and rockfill dams. ASCE J. of Geotechnical and National Taiwan University of Science and Technol-
Geoenvironmental Engineering 135(12): 1957–1970. ogy, Taipei, 321–329.
Yin, Y.P., Wang, F.W. & Sun, P. 2009. Landslide hazards
triggered by the 2008 Wenchuan earthquake, Sichuan,
China. Landslides 6: 139–152.

26

ISGSR2013.indb 26 10/18/2013 9:36:31 AM

 2 Keynote lectures

ISGSR2013.indb 27 10/18/2013 9:36:31 AM

This page intentionally left blank
 Geotechnical Safety and Risk IV – Zhang et al. (eds)
© 2014 Taylor & Francis Group, London, ISBN 978-1-138-00163-3

Advances in geotechnical risk and reliability for offshore applications

R.B. Gilbert
The University of Texas at Austin, Austin, Texas, USA

S. Lacasse & F. Nadim

Norwegian Geotechnical Institute (NGI), Oslo, Norway

ABSTRACT: This paper describes recent advances in geotechnical reliability and risk for offshore
applications. The topics addressed include spatial variability, model uncertainty, hazard characterization,
reliability-based design, system reliability and risk management. Conclusions from the evolution of reli-
ability and risk approaches include that practical implementation is key, that assessment is best consid-
ered in the context of decision making, and that collaboration of multiple disciplines and stakeholders is
important to managing risk effectively.

1 INTRODUCTION 5. Considering the reliability of systems as well as

components; and
The offshore oil and gas industry has been a leader 6. Including a wide variety of perspectives, conse-
in considering risk and reliability explicitly in devel- quences and hazards in managing risks.
oping and implementing designs. This industry has
These advances are inevitably motivated by
constantly pushed further the frontiers for design
practical needs in offshore applications. However,
and technology, with facilities being developed at
the advances are general and fundamental, and
present in 3,000 m of water. The consequences of a
therefore relevant to a wide variety of geotechnical
failure can be severe, and the costs associated with
applications. Case histories are presented to illus-
mitigating risks can be enormous. Therefore, there
trate the recent advances. The paper concludes with
is a strong need to avoid both under-conservatism
recommendations for continuing the development
and over-conservatism.
and application of reliability and risk approaches
One of the first reliability-based design guid-
in geotechnical engineering.
ance documents was developed for offshore facili-
ties (API 1993a). A sampling of the work that led
to implementing reliability-based approaches in
2 SPATIAL VARIABILITY
offshore geotechnical practice includes Bea (1983),
Lacasse & Goulois (1989), Wu et al. (1989), Tang
Accounting for spatial variability in geotechnical
et al. (1990), Nadim & Lacasse (1992) and Tang &
properties poses a significant challenge offshore
Gilbert (1993). In addition, the offshore industry
for the following reasons:
has had the opportunity to learn from experi-
ence as the performance of facilities subjected to 1. The locations for offshore developments are not
extreme operation conditions has been observed. readily accessible;
The objective of this paper is to describe recent 2. The cost and time required to conduct offshore
advances in geotechnical reliability and risk for site investigations are orders of magnitude
offshore applications. The following areas are greater than for onshore site investigations1;
highlighted: and
3. The facilities on the seafloor for a single devel-
1. Accounting for spatial variability in geotechni-
opment, including foundations for structures,
cal properties;
wells, manifolds and valves and pipelines, can
2. Characterizing model uncertainty in design
extend many kilometers.
methods;
3. Representing loads and hazards in geotechnical
systems; 1. However, the costs of the geotechnical site investigations
4. Implementing reliability-based design in for offshore installations represent on y a very small
practice; fraction of the total development costs (less than 2%).

29

ISGSR2013.indb 29 10/18/2013 9:36:32 AM

 Figure 2. Horizontal correlation distance2 versus depth
Figure 1. Coefficient of variation versus depth for
for point and depth-averaged values of design undrained
point and depth-averaged values of design undrained
shear strength in random field model (from Cheon &
shear strength in random field model (from Cheon &
Gilbert 2013).
Gilbert 2013).

Consequently, it is not feasible to gather 100 for this model and its calibration are provided in
percent knowledge of the geotechnical proper- Cheon (2011) and Cheon & Gilbert (2013).
ties at the location of or along every foundation The model shows that the influence of the spa-
element. tial variability relative to the mean decreases with
Recent advances have been made in developing depth (Fig. 1), possibly reflecting the increasing
realistic models of spatial variability to account overburden stress damping variations in mineral-
for it in designing foundations and optimizing site ogy or depositional history. The effect of spatial
investigation programs (e.g., Keaveny et al., 1990, averaging in reducing variability for the depth-
Gambino & Gilbert 1999 and Valdez-Llamas et al., averaged strength increases with averaging length
2003). (Fig. 1).
An example of a random field model for the The horizontal correlation distance2 obtained
design capacity of deep foundations is shown in was between 2 and 6 km, and is therefore hundreds
Figures 1 to 4. The geologic setting is normally to of times greater than the vertical correlation dis-
slightly overconsolidated marine clays in 1,500 to tance (Fig. 2). Note that the correlation distance
3,000 m deep water in the Gulf of Mexico. The is much greater for the design undrained shear
random field model represents spatial variations in strength compared to that for individual measure-
the design shear strength. The design strength is the ments of undrained shear strength since the design
strength selected by a designer for the purposes of profile implicitly averages out small-scale varia-
foundation design based on all available laboratory tions (either real or due to measurement methods)
and field test data and geologic information at a and reflects larger-scale variations. Both the hori-
given location. The available data for this geologic zontal and vertical correlation distances are greater
setting included over 100 design profiles of und- for the depth-averaged versus the point strength
rained shear strength from site investigations with (Fig. 2). The horizontal correlation structure is
soil borings, jumbo piston cores, field vane tests best modelled as anisotropic, with a longer hori-
and Cone Penetration Tests (CPT). These design zontal correlation distance moving away from the
profiles are located as close as 100’s of meters to as continental shelf (in the direction of depositional
far as 1,000’s of kilometers from one another. flow) compared to moving along the continental
The three-dimensional random field model con- shelf (Fig. 3).
sists of two cross-correlated models for the design This model of spatial variability can be used to
undrained shear strength: one for the design strength support design decisions. An example application
at a particular depth below the sea floor (to calculate
end bearing) and one for the depth-averaged design
2. Correlation distance was defined here as the separation
strength from the sea floor to that depth (to calcu- distance at which the correlation coefficient is equal to
lated side shear). The model incorporates means 0.37 for an exponentially decreasing correlation coef-
and standard deviations that increase with depth, an ficient with separation distance. This correlation dis-
anisotropic spatial correlation structure, and hori- tance is one-half the scale of fluctuation defined by
zontal correlations that increase with depth. Details Vanmarcke (1983).

30

ISGSR2013.indb 30 10/18/2013 9:36:32 AM

 was conducted to develop a design profile at this
location below a depth of 20 m, then the expected
value of this additional information is a reduction
in required caisson length of about 2 m (obtained
by comparing the curves labeled “Accounting for
Spatial Variability” and “Neglecting Spatial Vari-
ability” in Figure 4). The expected cost savings
can be compared against the cost of obtaining the
additional information and can be determinant for
the decision-making on whether or not to do addi-
tional site investigations.
An important point in Figure 4 is that the added
conservatism required to account for spatial varia-
bility, a reduction in capacity less than ten percent,
Figure 3. Coordinate system describing horizontal is small compared to a typical resistance factor of
distances along and off the continental shelf for hori- 0.8 or material factor of 1.25. Therefore in this geo-
zontally anisotropic correlation model (from Cheon & logic setting, the additional (aleatory) uncertainty
Gilbert 2013). due to not having site-specific geotechnical data is
small compared to the (epistemic) uncertainty in
selecting a design shear strength that represents the
actual strength mobilized when the foundation is
loaded.

3 MODEL UNCERTAINTY

Model uncertainty, which is defined as variations

between the actual performance and that predicted
by a design method, can be one of the largest
sources of uncertainty in offshore geotechnical
design. For example, the coefficient of variation
for model uncertainty in the axial capacity of a pile
foundation is typically greater than 0.2, while the
coefficient of variation due to spatial variability is
less than 0.2 (Fig. 1).
Recent advances have been made in better char-
acterizing model uncertainty for offshore applica-
Figure 4. Factored design axial capacity for 5.5-m tions. One advance has been related to the axial
diameter suction caisson foundation (from Cheon & capacity of driven piles in sand. Based on several
Gilbert 2013). large-scale load testing programs and additional
data, several newly developed design methods
is for the design of a suction caisson that will need could be verified (e.g., Randolph 2003, Jardine
to penetrate below the depth of an available design et al., 2005, Lehane et al., 2005, Clausen et al.,
profile for strength obtained from a jumbo piston 2005, Kolk et al., 2005, Schneider et al., 2008 and
core. Figure 4 shows the factored axial capacity Lacasse et al., 2013c).
(i.e., the nominal design axial capacity reduced by In addition to pile load tests, recent advances
the resistance factor for a Load and Resistance have been made by studying the performance of
Factor Design check). actual offshore structures loaded to or beyond their
The available design profile at this location calculated capacities. Five major hurricanes moved
extends to a depth of 20 m. For a caisson longer through the oil and gas infrastructure in the Gulf
than 20 m, there is additional uncertainty in the of Mexico between 2004 and 2008. Figure 5 shows
axial capacity due to spatial variability. The curve an example of new information on the predicted
labeled “Accounting for Spatial Variability” in versus measured axial capacity of driven piles at
Figure 4 incorporates an additional partial resist- large capacity in normally consolidated clays. The
ance factor to provide the same level of reliability data point with the largest measured capacity is for
as if a design profile were available. If the factored a 1,220 mm diameter by 70 m long pile that failed
design load is 10,000 kN, then the required caisson in tension when a tripod jacket was loaded beyond
length is 27 m. If an additional site investigation its ultimate capacity in Hurricane Ike (2008). It

31

ISGSR2013.indb 31 10/18/2013 9:36:32 AM

 Figure 7. Probability distributions for bias on calcu-
lated ratio of pile system capacity to pile system load
(reserve strength ratio) (from Chen & Gilbert 2013).

Figure 5. Comparison of measured with calculated

axial capacity based on API current guidelines for driven the foundation system, represented by the “Base
piles in normally consolidated clays (adapted from Chen
et al., 2013).
Case” interaction curve in Figure 6. However,
the calculated capacity of the foundation system
is potentially conservative because it assumes a
nominal rather than an average yield strength for
the steel piles; the lateral resistance of the soil was
reduced to account for cyclic loading when the piles
are pushed into undisturbed soil at ultimate failure
of the entire system; and the effect of jacket leg
stubs extending below the mudline was assumed
as negligible. When more realistic assumptions
are used to model the pile system, the calculated
capacity is equal to or greater than the estimated
hurricane load (Fig. 6).
Figure 7 shows how the performance of indi-
vidual platforms (e.g., Figs. 5 and 6) can be used
to update model uncertainty with Bayes’ theorem.
The bias is defined as a multiplicative correction
factor on the calculated ratio of capacity to load,
defined as the reserve strength ratio, where the
Figure 6. Comparison of measured and calculated capacity is calculated using the existing API design
pile system capacity for eight-pile jacket that survived method and the load is calculated using the hur-
Hurricane Ike (adapted from Gilbert et al., 2010). ricane hindcast. Variations between the actual and
calculated reserve strength ratio could occur due
is the largest published failure load to date for a both to errors in the calculated capacity or in the
driven pile in normally consolidated clay. The pile load. The updated probability distribution for this
failed five years after installation under cyclic and ratio is shown in Figure 7 for individual platforms
rapid loading during a hurricane. It is notable that survived or failed in a hurricane. In addition,
because the predicted capacity, when a t-z analysis the results from these individual platforms are
that accounts for strain-softening in side shear and combined together into an overall result, labeled
an axial flexibility of the pile, matches very well “Updated—All Cases,” by assuming independence
with the most likely load at failure based on the between platform performances. The overall result
hurricane hindcast. indicates that while there is possibly a conservative
Figure 6 shows an example of a pile system for bias in the calculated reserve strength ratio, there is
an eight-leg jacket that survived Hurricane Ike. In also considerable uncertainty (Fig. 7). The results
this case, the piles are 920 to 1,070 mm in diam- in Figure 7 should be used with caution because
eter, 52 m long, driven through layers of clay and they are based on a small data set and subsequently
sand, and tipped in sand. The estimated load in treat similarly a variety of different failure mecha-
Hurricane Ike exceeded the calculated capacity of nisms, including lateral and axial pile failures.

32

ISGSR2013.indb 32 10/18/2013 9:36:33 AM

 Figure 8. Comparison of measured axial capacity for
driven piles in clay soils with estimated lower-bound
capacity calculated assuming the side shear equal to the Figure 10. Effect of lower-bound on probability of
remolded undrained shear strength of the clay (adapted failure for TLP foundation (adapted from Gilbert et al.,
from Najjar 2005). 2010).

the mean or standard deviation (Najjar & Gilbert

2010). A lower bound on the capacity is particu-
larly significant to the reliability in cases with
relatively small uncertainty in the load or large fac-
tors of safety. In addition, a lower bound can be
influential even when its exact value is uncertain.
The application of this idea in practice is shown
in Figure 10. The incorporation of a lower bound,
which can be verified with pile driving monitor-
ing during or after installation (i.e., a re-strike
analysis), reduces the probability of failure for this
foundation to within tolerable levels (Fig. 10).

Figure 9. Probability distributions for load and capac- 4 HAZARD CHARACTERIZATION

ity for a Tension Leg Platform foundation (adapted from
Gilbert et al., 2010). The load or hazard is as important as the capacity
in analyzing the reliability of a geotechnical sys-
tem. In many cases, a thoughtful analysis of reli-
Another recent advance has been in refining ability can lead to advances in how the hazard is
models of the left-hand tail of capacity, which is characterized.
the region of interest for reliability. Figure 8 shows Reliability-based design and the decision mak-
an example of establishing a lower-bound on the ing processes in risk management often require
axial capacity of a driven pile in clay based on an assessment of the failure probability during a
the remolded undrained shear strength. This cal- reference time period, e.g., the annual failure prob-
culated lower-bound is less than the measured ability or the failure probability during the lifetime
capacity in every load test. Figure 9 illustrates the of a project. The assessment of this probability
physical significance of such a lower bound on the requires a probabilistic description of the annual
reliability of a Tension Leg Platform (TLP) foun- maximum environmental loads for foundation
dation: the most probable point from a First Order design, or a probabilistic description of frequency
Reliability Method (FORM) analysis, in which a and intensity of trigger(s) for assessment of impact
conventional lognormal distribution is assumed of geohazards on sea floor installations. Using this
for capacity, is well below the lower bound, which information, the probability of foundation fail-
is unreasonable. ure or slope instability can be computed for all
If a lower bound is incorporated into the prob- relevant scenarios and return periods in order to
ability distribution for capacity, then the reliability derive the annual or lifetime failure probability.
can be governed by this lower bound as opposed to However, including all possible scenarios can be

33

ISGSR2013.indb 33 10/18/2013 9:36:33 AM

 acceleration causing slope failure, and ε describes
the variability of the peak ground acceleration at a
given return period.
The probability distribution of Amax was
obtained from the site-specific Probabilistic Seis-
mic Hazard Assessment (PSHA). A Pareto dis-
tribution provided a good fit for Amax with return
periods greater than 100 years. The resistance
parameter Aresist and the variability parameter ε
were respectively assigned lognormal and normal
distributions, and the parameters of the distribu-
tion functions were calibrated to match the condi-
tional failure probabilities for the 3,000-year and
the 10,000 year earthquake events (Fig. 11). With
Figure 11. Results of probabilistic analyses of static this limit state function, the annual probability of
undrained stability, prior to (black), updated (blue) and earthquake-triggered slope failure was computed
after the 3,000-year and 10,000-year earthquake (red) using FORM to be Pf,annual = 4 × 10−4.
(from Nadim 2011). In some situations, such as offshore geohazards
studies, it can be extremely difficult to identify the
trigger(s) for submarine slides and a reference time
frame. One must then rely on the identification and
time-consuming and impractical, and often only dating of recent (in geological sense) slide events
the few scenarios that contribute most to the fail- in the area. The dating results and other relevant
ure probability are needed for a sound assessment. geological evidence can then be used in a Bayesian
Recent advances have been made in developing framework to establish the annual probability of
practical means to calculate the annual probability slope instability (e.g., Nadim 2002). In performing
of earthquake-induced slope failure (Nadim 2002 these analyses, it is very important to consider the
and 2011). This work was supported by a number relevancy of the conditions present in the historical
of joint-industry research projects and offshore record, such as the sea level, to the conditions that
geohazards studies in the North Sea, the Caspian may be present during the reference time period of
Sea, the Black Sea, offshore Indonesia, and the interest.
Gulf of Mexico. The multi-step approach uses Hazard characterization has also provided
FORM, Monte Carlo simulation and Bayesian insight into physical mechanisms. As an example,
updating and is described in detail by Lacasse et al. a recent advance was made in assessing the hazard
(2013d). Nadim (2011) presents an example case for wave-induced mudslides in the Mississippi River
study for a slightly overconsolidated clay slope in delta. For most fixed facilities in shallow water,
a moderately seismic area. Prior analyses showed such as jacket platforms, the loads are governed by
that the earthquake events with return periods the wave height and not the wave period. Therefore,
between 1,000 and 10,000 years contribute most the hazard has conventionally been described by a
to the annual probability of slope failure. The wave height in combination with an associated wave
dynamic response analyses were therefore done for period that corresponds to the strong (right-hand
earthquake events with return periods of 3,000 and in the northern hemisphere) side of a hurricane.
10,000 years. Each of these events was represented However, the wave period is an important
by four sets of properly scaled acceleration time consideration for wave-induced mudslides.
histories. Figure 11 shows the computed and the Figure 12 shows how the factor of safety for a
updated cumulative distribution functions for the slope failure is affected by the wave height and
static safety factor under undrained loading prior the wave period at one location in the Mississippi
to the earthquake, and after the possible impact of River Delta. Wave-induced mudslides occurred
3,000-year and 10,000-year earthquake events. at this location in both Hurricane Ivan (2004)
To estimate the annual probability of slope fail- and Hurricane Katrina (2005). While the maxi-
ure, Nadim (2011) developed a simplified model mum wave height in Ivan was significantly smaller
similar to that suggested by Cornell (1996). The than that during Katrina, the factor of safety was
limit state function for the seismic resistance of smaller in Ivan because of a relatively large wave
the slope was defined as: G = Seismic resistance— period (Fig. 12). The Delta was about 150 km to
Earthquake load = Aresist – ε⋅Amax where Amax is the the left of the eye of Hurricane Ivan, meaning that
annual peak ground acceleration representing the it was on the weak side of the storm. However, the
earthquake load, Aresist is the resistance of the slope wave periods on the weak side were similar to those
to earthquake loading in terms of the peak ground for the much larger wave heights on the strong side

34

ISGSR2013.indb 34 10/18/2013 9:36:34 AM

 Figure 14. Return period for wave-induced mudslides
impacting exiting pipelines in Mississippi River Delta
Figure 12. Factor of safety for slope failure versus wave (from Nodine et al., 2009).
height and period at one location in Mississippi River
Delta (from Gilbert et al., 2010).
a maximum wave height in the Delta is shown in
Figure 13: the most probable combination of wave
height and period represents hurricanes with their
strong side over the Delta, while the other combi-
nations represent hurricanes with their weak side
over the Delta.
An example result from using this hazard charac-
terization in assessing the hazard of wave-induced
mudslides in the Delta is shown in Figure 14. This
map incorporates the wave hazard with the water
depth, bottom slope, geotechnical properties and
pipeline locations.

5 RELIABILITY-BASED DESIGN

A significant benefit of a reliability-based design

Figure 13. Conditional probability distribution for
wave period in the Mississippi River Delta given a maxi- approach is to promote designs that efficiently
mum wave height for a hurricane in the Gulf of Mexico achieve target levels of reliability. Recent advances
(adapted from Nodine et al., 2009). have been made in implementing this principle in
practice.
Lacasse et al. (2013a, 2013b and 2013c) describe
of the storm. Therefore, both large wave heights on a case study concerning the reliability of axially-
the strong side of a storm (i.e., Katrina in Fig. 12) loaded piles in sands. The API RP 2GEO (2011)
and smaller wave heights with longer periods on and ISO 19902 (2007) guidelines included recently
the weak side of a storm (i.e., Ivan in Fig. 12) con- four CPT-methods for calculating the axial capacity
tribute to the risk for wave-induced mudslides. of piles in sands. The design guidelines require that
Based on this experience, an updated hazard if newer methods are to be implemented in design,
representation was developed for wave-induced the same level of safety shall be documented for
mudslides in the Delta. The approach utilized new methods as for existing methods.
the Theorem of Total Probability to account Ensuring adequate reliability under severe load-
for the possibilities that the maximum wave height ing is a necessary consideration, and the calculated
in the Delta corresponds to the strong side of a safety margin depends on the uncertainty in the
storm with the largest waves in the storm or to the parameters used in the analyses and the model
weak side of a storm with larger wave heights out- uncertainty. The design engineer attempts to com-
side of the Delta (Nodine et al., 2009). An example pensate for the uncertainties by introducing appro-
of the conditional probability for wave period given priate (partial) “safety factor(s)” in design.

35

ISGSR2013.indb 35 10/18/2013 9:36:34 AM

 Table 1. Design methods considered in reliability
analysis for axial capacity of driven piles.

Method Methods in clay Methods in sand

API API-RP2 A 20th API-RP2 A

ed.1993 20th ed. 1993
NGI-05 Karlsrud Clausen et al.
et al. 2005 2005
ICP-05 Jardine et al. Jardine et al. 2005;
1996; 2005 API 2011; 2007
Fugro-96/05 Kolk and v.d.Velde Kolk et al. 2005
1996
UWA-05 – Lehane et al. 2005;
Schneider et al. Figure 15. Simplified representation of reliability-based
2008 design calibration process (from Laasse et al., 2013a).

To evaluate the required resistance factor, standard deviation, coefficient of variation and
Lacasse et al. (2013a, 2013b and 2013c) calculated Probability Density Function (PDF). The model
the annual probability of failure for piles on off- uncertainty was obtained by comparing the pre-
shore jackets designed with the API method and dicted to the measured axial pile capacity from
with the newer CPT-based methods. The goal relevant and reliable pile model tests. The NGI
was to make a recommendation on the appropri- database of “super pile” load tests NGI (2000;
ate resistance factor and minimum pile penetra- 2001) was used.
tion depth to use for the design of the piles on an The calibration used (1) the results of the deter-
offshore jacket. Table 1 lists the axial pile capacity ministic analyses giving the ultimate axial pile
methods considered. capacity with the characteristic strength parameters
The reliability analyses of the axial pile capacity (Qult char); (2) the probabilistic analyses giving the
methods included a statistical analysis of the soil PDF of the ultimate axial pile capacity (Qult mean);
parameters; statistical analysis of the model uncer- and (3) the results of the probabilistic analyses giv-
tainty for the different pile capacity calculation ing the annual probability of failure, Pf.
methods used; statistical analysis of the static Figure 15 is a simplification in two dimensions
(permanent) and environmental loads on the top of the overlap of the probabilistic ultimate pile
of the piles; deterministic analysis of the ultimate capacity (Qult) and probabilistic environmental
axial pile capacity, Qult; probabilistic analyses of load (Penv). The probability density function for the
axial pile capacity to obtain the PDF of the ulti- Penv was taken as the same for Pf1 and Pf2 in the cal-
mate capacity, Qult; calculation of the annual prob- culations. The calibration of the resistance factor
ability of failure by combining the statistics of the was coordinated with the definition of characteris-
loads and the probabilistic description of Qult; and tic design load and the characteristic soil strength
calibration of the safety factors (load and resist- profile used for the calculation of axial pile capac-
ance factor) for each pile capacity design method, ity. The calibration details are described in Lacasse
for a target annual probability of failure of 10−4. et al., 2013a.
Three sites, where jackets are currently under Table 2 presents the results of the calibration
design, were analyzed. For Jacket A, the soil con- of the resistance factor for the case study jackets
ditions are characterized by mainly clay layers with to achieve an annual probability of failure of 10–4.
intermittent thin sand and silt layers. For Jacket The resistance factor was obtained based on the
B, the soil consists of mainly dense to very dense axial pile capacity calculated with the characteris-
sand layers, with rather thin clay layers in between. tic undrained shear strength (Qult char). The load fac-
For Jacket C, the soil profile consists of alternat- tors were maintained at the recommended values
ing very dense sand and very stiff clay units. The in the design guidance, although the load factor at
parameters were estimated with statistical analyses the design point was smaller.
of the soil data, combined with well-documented For a given pile length, the calibrated resistance
correlations and experience (bias factors). factor varied with the pile design method. The fac-
An extended study of the model uncertainty tors reflect the varying influence of the uncertainty
was carried out for the different axial pile capac- in the soil parameters and of the model uncer-
ity calculation methods (Lacasse et al., 2013c). The tainties for the different methods. The results are
model uncertainty was expressed as a bias (mean), generally consistent, where the axial pile capacity

36

ISGSR2013.indb 36 10/18/2013 9:36:35 AM

 Table 2. Calibrated resistance factors related to
characteristic ultimate axial capacity.

Site C
Site A (clay) Site B (sand) (clay & sand)
Method 90-m pile 26-m pile 40-m pile

NGI 1.23 1.35 1.20

ICP 1.52 1.45 1.32
Fugro 1.31 1.72 1.55
UWA – – 1.50
API 1.35 2.36 1.93

methods predicting higher axial pile capacity

require a higher resistance factor to ensure that the
annual probability of failure does not exceed 10–4.
The calibrated resistance factors apply to these
case study jackets only, and cannot be transferred
to other sites or structures without site-specific
reliability studies.
These calibrated resistance factors allowed for
a significant reduction in the required pile pen-
etration depth because one could demonstrate
that a target reliabilility could be achieved using
lower resistance factors than the a priori values
in the design guidance. The pile lengths could be
reduced by 15 to 20 percent for Cases A and C and
nearly 50 percent for Case B. A reliability analysis
can therefore have important implications for the
design of piles and result in significant savings.
This reliability study gave insight in the required Figure 16. Interaction curves of pile system capacity
resistance factor for different design methods of exhibiting a robustness check (adapted from Chen et al.,
axial pile capacity to achieve the same annual Pf 2010).
for a given pile penetration depth. The study is not
meant to favor an approach. More case studies are
needed on a variety of soil profiles to enable one
6. The selection of the characteristic parameters
to draw non site-specific recommendations on the
to use in the deterministic analysis is often a
resistance factor for each of the methods. The cali-
source of uncertainty, and can be very subjec-
bration analyses showed that:
tive, varying from one engineer to the other.
1. The calibration of the safety factors demon- Lacasse et al. (2013a) provide recommendations
strates that the annual probability of failure for minimizing this variability.
varies with the axial pile capacity calculation
method.
2. The values of model uncertainty used in the 6 SYSTEM RELIABILITY
analyses have an overwhelming influence on
the resulting annual probability of failure and Design checks are typically conducted on a compo-
therefore on the required resistance factor for a nent by component basis. However, the perform-
target annual probability of failure. ance reliability of the entire system is generally of
3. The current state-of-the-art design still relies greatest interest in managing risk. Recent advances
heavily on qualified engineering judgment to have been made in assessing system reliability for
assess and ensure a consistent safety level. both fixed and floating offshore facilities.
4. The resistance factors calibrated show that the Figure 16 shows a system robustness check for
newer CPT-methods of pile design are as reli- fixed jacket platforms. This idea was motivated by
able as the current API method. the performance of platforms in hurricanes in the
5. The selection of the characteristic shear strength Gulf of Mexico over the past decade. The check
was also a significant parameter that influences involves considering the capacity of the system
the calibrated resistance coefficient. when the lateral or axial capacity of any individual

37

ISGSR2013.indb 37 10/18/2013 9:36:35 AM

 of magnitude smaller than those for the ropes
and chains in the mooring line. In addition, the
probability of failure for individual components
depends on the water depth, with the smallest
probabilities of failure associated with the deepest
water because the uncertain environmental loads
are smaller relative the certain pre-tension loads as
the water depth increases.
Figure 18 shows how the redundancy in this
mooring system is sensitive to whether a semi-taut
or taut3 system is used. The redundancy in the
taut system is greater than in the semi-taut system
because the loads are re-distributed more evenly to
the remaining lines when a single line fails (Fig. 18).
Figure 17. Comparison of probabilities of failure in Therefore, design checks based on single compo-
design life for different components in the most heav- nents in these mooring systems will not necessarily
ily loaded line of a mooring system (from Clukey et al., provide either a consistent or representative reli-
2013). ability with the system. This type of information
is currently being considered in work to update the
design guidance documents for mooring systems.

7 NEW TRENDS IN RISK MANAGEMENT

Disasters like the Maconda Well blowout, which

caused the Deepwater Horizon oil spill in the Gulf
of Mexico in May 2010, can catalyze moments
of change in risk management aims, policy and
practice. The population living along the coastline
who might be affected by offshore accidents are
demanding that their opinions are respected in the
critical risk management decisions.
Quantitatively, risk is the expected consequence
of an adverse event, where the consequences are
obtained from the elements at risk and their vul-
Figure 18. Conditional probability of failure given fail-
ure of the most heavily-loaded line for a mooring system nerabiltiy. Mitigation of risk can be accomplished
in a hurricane (from Clukey et al., 2013). by reducing the probabilty of the adverse event or
by reducing the vulnerability and/or exposure of
the elements at risk, or even by reducing both haz-
pile is reduced. For the three-leg jacket, the system ard and consequence (Fig. 19).
capacity in overturning is essentially proportional Designing participatory processes for stake-
to the axial capacity of the most heavily-loaded holder involvement the risk management deci-
pile (Fig. 16b). For the six-leg jacket (Fig. 16a), sion making process is a new area of research. An
the system capacity in overturning is less sensitive, example of this type of research was provided in
reducing by about 10% for a 30% reduction in the the SafeLand Project (www.safeland-fp7.eu), a
axial capacity of the most heavily-loaded pile. For large collaborative project on landslide risk man-
both cases, the system capacity in shear is much agement within the European Commission’s 7th
less sensitive to the lateral capacity of an individual Framework Programme. The SafeLand project
pile. This proposed design check is to maintain a developed and tested a public communication
minimum system capacity when reducing the axial and participatory process for mitigating the risks
and lateral capacities of individual piles in order to of landslide in the highly at-risk community of
achieve a consistent level of reliability for a wide Nocera Inferiore in southern Italy (SafeLand
variety of pile systems. 2012). The pilot study demonstrated the potential
Figure 17 shows the results from reliability anal-
yses for the mooring system of a floating produc-
tion system located in three different water depths 3. In a semi-taut mooring system there is moderate cat-
(Clukey et al., 2013). The probability of failure for enary in the mooring lines, while in a taut mooring sys-
the suction caisson foundation (anchor) is orders tem there is small catenary in the mooring lines.

38

ISGSR2013.indb 38 10/18/2013 9:36:35 AM

 Figure 19. “Bow tie” diagram illustrating components
in risk management (from Lacasse & Nadim 2009).

and challenges of public participation in decisions

characterized by high personal stakes and intricate Figure 20. Stress testing as a tool to deal with residual
technical, economic and social considerations. It or neglected risk for critical infrastructure (Nadim &
should prove useful in informing similar processes, Sparrevik, 2013).
as stakeholders in Europe increasingly demand a
voice in choosing landslide mitigation measures.
The results of the pilot study in SafeLand and tsunami leading to the Fukushima Dai-ichi
showed that it is feasible to organize an expert- accident. Many aspects of the accident devastating
informed participatory process that respects and the Fukushima Dai-ichi nuclear power plant are
builds on conflicting citizen perspectives and inter- still uncertain. However, the accident highlighted
ests, and demonstrates spheres of policy consen- three areas of potential weakness in the existing
sus as well as policy dissent. Increasingly public safety approaches:
interventions to reduce the risk of landslides and
1. Inadequacy of safety margins in the case of
other hazards are moving from “expert” decisions
extreme external events, especially natural
to include the public and other stakeholders in the
hazards.
decision process. Variations in the role of science
2. Lack of robustness with respect to events that
and scientists, governance structures and interest
exceed the design basis.
groups, legislation, availability of economic and
3. Ineffectiveness of current emergency manage-
political instruments, social learning, facilitation
ment under highly unfavourable conditions.
of communication and trust, media intervention,
access to information, and external pressures and These issues were the focus of the stress tests
shocks were some of the issues identified by the imposed on all nuclear power plants in Europe in
SafeLand research that impact the cognition and 2011 and 2012 (WENRA, 2011).
management of risk practice in a society. A stress test is an examination of the safety of
Another new trend in risk management is stress a system under those particularly unfavourable
testing. Stress testing is a procedure used to deter- scenarios that fall outside the design basis speci-
mine the stability of a system or entity. It involves fied by the regulatory regime, by the operational
testing the said system or entity to beyond its nor- institution or by the stakeholders. A stress test can
mal operational capacity, often to a breaking point, test the system to assess its response to scenarios
in order to observe its performance/reaction to a expected to be in the residual and neglected risk
pre-defined internal or external effects (pressure/ areas (Fig. 20). In this respect, stress testing is not a
force). Stress tests have been used for many years substitute for “conventional” risk or safety assess-
in air traffic safety, in particular for airplanes and ments, but it provides additional valuable insight
helicopters. In recent years, stress testing has often under extreme situations. What stress tests and
been associated with methodologies to assess the “conventional” risk or safety assessments have in
vulnerability of a financial system or specific com- common is that they both rely on a description
ponents of it, such as banks. A number of analyti- of the system of interest, which helps to associate
cal tools have been developed in this area and have the state of the system and a set of consequences
been frequently used since the late 1990’s (e.g., under any given or potential scenario.
Borio et al., 2012). Several multi-national research projects in
More recently, stress testing has been applied to Europe are starting up in 2014 to develop guide-
the comprehensive safety and risk assessment of lines for stress testing of critical infrastructure
nuclear power plants, in particular in the aftermath under the action of natural hazards (Nadim &
of the 11th March 2011 East Japan earthquake Sparrevik, 2013). The premises are that a critical

39

ISGSR2013.indb 39 10/18/2013 9:36:36 AM

 infrastructure is designed to withstand the impact were described to illustrate the practical motivation
of natural hazards according to regulations in for and usefulness of these advances.
codes and standards or specifications from the The following conclusions are drawn from the
owner and/or stakeholders. The regulations are evolution of reliability and risk approaches to their
often set through probabilistic evaluations with the current state for offshore applications:
objective of reducing risk to an acceptable level.
This evaluated risk will be in accordance with what 1. Applying the theory of reliability and risk in prac-
society will tolerate in terms of loss of life, envi- tice is critical to obtaining useful insights from
ronmental damages and the loss of assets through the theory and to developing practical means
the definition of acceptance criteria that are incor- to implement the theory. While the application
porated into regulations. of reliability and risk approaches has matured,
The design rules that result from such regula- each major practical application still involves a
tions implicitly accept that there is a residual significant element of research and development
risk associated with rare, extreme events that is to best suit that particular problem.
neglected because of the (objectively calculated 2. Assessing reliability and risk is most valuable if
or perceived) very low probability of occurrence. it is considered in the context of helping stake-
However, the Fukushima accident showed that as holders make decisions. Opportunities in deci-
a consequence of this neglect, a system that is quite sion making exist both to mitigate risk as well
robust as long as events remain within its design as to reduce the cost required to achieve a target
basis can abruptly shift to complete failure when level of risk
that threshold is passed. Stress tests can help detect 3. Managing risk effectively requires the collabo-
such “cliff-edge effects” and identify ways to intro- ration of multiple disciplines and the involve-
duce some robustness in the system without any ment of stakeholders at all stages of the process,
change in the acceptable level of risk. from assessment to decision making.
Most risk evaluations are based on probability While offshore applications have provided won-
estimates using historical data, observations and/ derful opportunities to advance these approaches,
or experience and engineering judgment, and con- the results of these advances are relevant to a wide
sequence models that try to estimate the impact variety of geotechnical problems.
of unwanted future hazard situations. For natu- The future for reliability and risk approaches is
ral hazards, historical data may in some cases be bright. Public and private stakeholders will always
sparse or highly uncertain. There is also generally welcome, seek and value help in making better
little experience with extreme events, because of and more defensible decisions. The following rec-
their nature. Furthermore, simplified models of ommendations are offered to guide the continued
highly complex situations yield predictions of sys- advancement of these approaches:
tem response that contain significant uncertainty.
The scarcity of data and model uncertainty may 1. Develop means and methods to implement reli-
lead to optimistic evaluations that neglect the ability and risk approaches that are as simple as
risks associated with extreme events. Stress testing possible while still capturing the important char-
provides a framework to address these neglected acteristics that describe hazards, consequences
risks. and the performance of engineered systems.
Stress tests have not yet been applied in off- Simplicity is important both to make imple-
shore projects. However, the safety philosophy and mentation practical and to make the approaches
premises for design of offshore structures are quite as transparent as possible for the stakeholders.
similar to those for onshore critical infrastructure. 2. Encourage the application of reliability and
In the future, stress tests could complement the risk approaches in the earliest stages of project
present risk assessment approaches for many off- development when the greatest opportunities
shore projects. exist to impact decisions and to proactively plan
to acquire valuable data for future decisions.
3. Continuously strive to update knowledge about
hazards, consequences and performance based
8 SUMMARY on historical information. Reliability and risk
approaches provide the link between this infor-
This paper has described recent advances in geotech- mation and the assessment and management of
nical reliability and risk for offshore applications. risk for future applications.
The areas addressed include spatial variability, model 4. Increase awareness and understanding about reli-
uncertainty, hazard characterization, reliability- ability and risk approaches for technical profes-
based design, system reliability and risk manage- sionals and colleagues in the other disciplines such
ment. Case histories from real-world applications as social sciences, as well as the general public.

40

ISGSR2013.indb 40 10/18/2013 9:36:36 AM

 REFERENCES Gilbert, R.B., Murff, J.D. and Clukey, E.C. 2010. Risk
and reliability on the frontier of offshore geotechnics.
API 1993a. Recommended Practice for Planning, Proceedings, International Symposium on Frontiers
Designing and Constructing Fixed Offshore in Offshore Geotechnics, Perth, Australia.
Platforms—Load and Resistance Factor Design. API ISO 19902 2007. Petroleum and natural gas industries—
RP2 A-LRFD, 1st Ed. Washington. Fixed Steel Offshore Structures. 1st ed., Switzerland.
API 1993b. Recommended Practice for Planning, Design- Jardine, R.J. and Chow, F.C. 1996. New design proce-
ing and Constructing Fixed Offshore Platforms— dures for offshore piles. Marine Technology Directo-
Working Stress Design. API RP2 A-WSD, 20th Ed. rate. Publ. 96/103, London, UK.
Washington. Jardine R.J., F.C. Chow, R.F. Overy and J.R. Standing
API 2007. Recommended Practice for Planning, Design- 2005. ICP design methods for driven piles in sands
ing and Construction Fixed Offshore Platforms— and clays. Imperial College, Thomas Telford Publish-
Working Stress Design. API RP 2 A-WSD, 21st Ed. ing, London.
Washington. Karlsrud, K., Clausen, C.J.F. and Aas, P.M. 2005. Bear-
API 2011. Geotechnical and Foundation Design Consid- ing capacity of driven piles in clay, the NGI approach.
erations. ANSI/API RP 2GEO, 1st Ed. Washington. Proceedings, International Symposium on Frontiers
Bea, R.G. 1983. Characterization of the reliability of in Offshore Geotechnics, Perth, Australia.
offshore piles subjected to axial loadings. Structural Keaveny, J.M., Nadim, F. and Lacasse, S. 1990. Auto-
Congress, American Society of Civil Engineering, correlation functions for offshore geotechnical data.
New York, pp. 1–20. Proceedings, International Conference on Struc-
Borio, C., Drehmann, M., and Tsatsaronis, K. 2012. tural Safety and Reliability, 5. ICOSSAR ‘89. San
Stress-testing macro stress testing: Does it live up to Francisco, California, pp. 263–270.
expectations? BIS Working Papers, No 369, January. Kolk, H.J., A.E. Baaijens and M. Senders 2005. Design
Chen, J-Y., Gilbert, R.B., Murff, J.D., Young, A. and criteria for pipe piles in silica sands. Proceedings,
Puskar, F. 2010. Structural factors affecting the sys- International Symposium on Frontiers in Offshore
tem capacity of jacket pile foundations. Proceedings, Geotechnics, Perth, Australia.
International Symposium on Frontiers in Offshore Kolk, H.J. and Velde, E. v.d. 1996. A Reliable Method
Geotechnics, Perth, Australia. to Determine Friction Capacity of Piles Driven into
Chen, J-Y., Gilbert, R.B., Puskar, F. and Verret, S. 2013. Clays. Proceedings, Offshore Technology Conference,
Case study of offshore pile system failure in Hurricane Houston, Texas, OTC 7993.
Ike. Journal of Geotechnical and Geoenvironmental Lacasse, S. and Goulois, A. 1989. Uncertainty in API
Engineering, ASCE, in press. parameters for predictions of axial capacity of
Chen, J-Y. and Gilbert, R.B. 2013. Insights into the driven pile in sand. Proceedings, Offshore Technol-
performance reliability of offshore piles based on ogy Conference, Houston, Texas, Houston, Texas,
experience in hurricanes. Principles and Practices in pp. 353–358.
Geotechnical Engineering, A Geotechnical Special Lacasse, S. and Nadim, F. 2009. Landslide risk
Publication honoring Roy Olson, American Society assessment and mitigation strategy, Chapter 3 in
of Civil Engineers, Reston, Virginia, in press. Landslides—Disaster Risk Reduction, Springer-
Cheon, J.Y. 2011. Analysis of spatial variability in geo- Verlag, Berlin.
technical data for offshore foundations. Ph.D. Dis- Lacasse, S., Nadim, F., Andersen, K.H., Knudsen, S.,
sertation, The University of Texas at Austin, Austin, Eidsvig, U.K., Yetginer, G., Guttormsen, T.R. and
Texas, 2011, p. 212. Eide, A. 2013a. Reliability of API, NGI, ICP and
Cheon, J.Y. and Gilbert, R.B. 2013. Modeling spatial Fugro Axial Pile Capacity Calculation Methods. Pro-
variability in offshore geotechnical data for reliability- ceedings, Offshore Technology Conference, Houston,
based foundation design. Structural Safety, in press. Texas, OTC-24063-MS.
Clausen, C., P.M. Aas and K. Karlsrud 2005. Bearing Lacasse, S., Nadim, F., K.H., Knudsen, S., Eidsvig, U.K.,
Capacity of Driven Piles in Sand, the NGI Approach. Liu, Z.Q., Yetginer, G. and Guttormsen, T.R. 2013b.
Proceedings, International Symposium on Frontiers Reliability of Axial Pile Capacity Calculation Meth-
in Offshore Geotechnics, Perth, Australia. ods. Proceedings, GeoMontréal 67th Canadian Geo-
Clukey, E.C., Gilbert, R.B., Andersen, K.H. and technical Conference. Montréal Canada, Paper 699.
Dahlberg, R. 2012. Reliability of suction caissons for Lacasse, S., Nadim, F., Langford, T., Knudsen, S.,
deep water floating facilities. Foundation Engineering Yetginer, G., Guttormsen, T.R. and Eide, A. 2013c.
in the Face of Uncertainty, Geotechnical Specialty Model uncertainty in axial pile capacity design meth-
Publication Honoring Dr. Fred H. Kulhawy, ASCE. ods, Proceedings, Offshore Technology Conference,
Reston, Virgina. Houston, Texas, OTC 24066.
Cornell, C.A. 1996. Calculating building seismic per- Lacasse, S., Nadim, F., Vanneste, M., L’Heureux, J.S.,
formance reliability: A basis for multi-level design Forsberg, C.F. and Kvalstad, T. 2013d. Case Studies
norms, Proc. 11th World Conf. on Earthquake Eng., of Offshore Slope Stability. Keynote Lecture. Pro-
Acapulco, Mexico. ceedings, ASCE GeoCongress 2013—Stability and
Gambino, S.J. and Gilbert, R.B. 1999. Modeling spatial Performance of Slopes and Embankments III, San
variability in pile capacity for reliability-based design. Diego, California, pp 2379–2418.
Analysis, Design, Construction and Testing of Deep Lehane, B.M., J.A. Schneider and X. Xu 2005. A Review
Foundations, Roesset Ed., ASCE Geotechnical Spe- of Design Methods for Offshore Driven Piles in
cial Publication No. 88, pp. 135–149. Siliceous Sand. UWA Report No. GEO 05358, The

41

ISGSR2013.indb 41 10/18/2013 9:36:36 AM

 University of Western Australia, Perth, Australia, SafeLand 2012. Deliverable D5.7: Design and testing:
Sept. 2005. a risk communication strategy and a deliberative
Nadim, F. 2002. Probabilistic methods for geohazards process for choosing a set of mitigation and preven-
problems: State-of-the-Art. Proceedings, Probabil- tion measures. http://www.safeland-fp7.eu/results/
istics in Geotechnics: Technical and Economic Risk Documents/D5.7.pdf
Estimation, Graz, Austria, September 15–19, p. 18. Schneider, J.A., X. Xu and B.M. Lehane 2008. Database
Nadim, F. 2011. Risk Assessment for Earthquake- assessment of CPT-based design methods for axial
Induced Submarine Slides. Keynote Lecture, 5th capacity of driven piles in siliceous sands. Journal of
International Symposium on Submarine Mass Move- Geotechnical and Geoenvironmental Engineering,
ments and Their Consequences, ISSMMTC, Kyoto, ASCE 34 (9): 1227–1244.
Japan, 24–26 October, Springer. Tang, W.H. and Gilbert, R.B. 1993. Case study of offshore
Nadim, F. and Lacasse, S. 1992. Probabilistic bearing pile system reliability. Proceedings, Offshore Technol-
capacity analysis of jackup structures. Canadian Geo- ogy Conference, Houston, Texas, pp. 677–686.
technical Journal, 29, pp. 580–588. Tang, W.H., Woodford, D.L. and Pelletier, J.H. 1990.
Nadim, F., and Sparrevik, M. 2013. Managing unlikely Performance reliability of offshore piles, Proceedings,
risks posed by natural hazards to critical infrastruc- Offshore Technology Conference, Houston, Texas,
ture. Proceedings, 22nd SRA Europe Conference, OTC 6379.
Trondheim, Norway, 17–19 June. Valdez-Llamas, Y.P., Auvinet, G. and Nunez, J. 2003.
Najjar, S.S. 2005. The importance of lower-bound capac- Spatial variability of the marine soil in the Gulf of
ities in geotechnical reliability assessments. Ph.D. Dis- Mexico. Proceedings, Offshore Technology Confer-
sertation, The University of Texas at Austin, p. 347. ence, Houston, Texas, OTC 15266.
Najjar, S.S., and Gilbert, R.B. 2009. Importance of Vanmarcke, E.H. 1983. Random fields: analysis and syn-
lower-bound capacities in the design of deep founda- thesis. The MIT Press, Boston, Massachusetts.
tions. Journal of Geotechnical and Geoenvironmental WENRA (West European Nuclear Regulators Asso-
Engineering, ASCE 135 (7): 890–900. ciation) 2011. “Stress tests” specifications—
NGI 2000. Bearing Capacity of Driven Piles in Clay. Proposal by the WENRA Task Force, 21 April
Report 525211–1. 23 March 2000. 2011. http://www.oecd-nea.org/nsd/fukushima/
NGI 2001. Bearing Capacity of Driven Piles in Sand. documents/WENRA20110421StressTestsSpecifica-
Report 525211–2. 21 Jan. 2001. tions2011–04–21.pdf.
Nodine, M.C., Gilbert, R.B., Cheon, J.Y., Wright, S.G. Wu, T.H., Tang, W.H., Sangrey, D.A., and Baecher, G.B.
and Ward, E.G. 2009. Risk analysis for hurricane 1989. Reliability of offshore Foundations—state of
wave-induced submarine mudslides. Proceedings, Sub- art. Journal of Geotechnical Engineering, ASCE 115
marine Mass Movements and Their Consequences, (2): 157–178.
4th International Symposium, Springer, pp. 335–352.
Randolph, M.F. 2003. Science and empiricism in pile
foundation design. Geotechnique 44 (3): 427–448.

42

ISGSR2013.indb 42 10/18/2013 9:36:36 AM

 Geotechnical Safety and Risk IV – Zhang et al. (eds)
© 2014 Taylor & Francis Group, London, ISBN 978-1-138-00163-3

Homogenization of geomaterials using the random finite

element method

D.V. Griffiths
Colorado School of Mines, Golden, CO, USA
University of Newcastle, NSW, Australia

Jumpol Paiboon
Colorado School of Mines, Golden, CO, USA

Jinsong Huang
University of Newcastle, NSW, Australia

Gordon A. Fenton
Dalhousie University, Halifax, NS, Canada

ABSTRACT: The homogenized stiffness of geomaterials that are highly variable at the micro-scale has
long been of interest to geotechnical engineers. The purpose of this study is to investigate the influence
of porosity and void size on the homogenized or effective properties of geomaterials. A Random Finite
Element Method (RFEM) has been developed enabling the generation of spatially random voids of given
porosity and size within a block of geomaterial. Following Monte-Carlo simulations, the mean and stand-
ard deviation of the effective property can be estimated leading to a probabilistic interpretation involving
deformations. The probabilistic approach represents a rational methodology for guiding engineers in the
risk management process. The influence of block size and the Representative Volume Elements (RVE) are
discussed, in addition to the influence of anisotropy on the effective Young’s modulus.

1 INTRODUCTION Representative Volume Element (RVE). An RVE

is an element of the heterogenous material that is
The motivation of this work is to investigate the large enough to represent the microstructure but
influence of porosity and void size on the stiffness small enough to achieve computational efficiency
of 3D geomaterials using a statistical approach. (e.g. Liu, 2005; Zeleniakiene et al. 2005).
Even if the expected porosity of a site can be con- Since the concept of the RVE was first introduced
servatively estimated, the location of the voids may by Hill (1963), several theoretical models have been
be largely unknown such as in geological regions proposed for dealing with scale effects. Hazanov &
dominated by karstic deposits. This makes a sta- Huet (1994) derived results involving mixed bound-
tistical approach appealing. The work presented in ary conditions, which locate between the static and
this paper is developed from a study of 2D model kinematic uniform boundary conditions for speci-
homogenization of geomaterials containing voids mens smaller than the size of the RVE. Orthogonal
by random fields and finite elements (Griffiths mixed boundary conditions have also been proposed
et al. 2012) and 3D random finite element methods (e.g. Hazanov & Amieur, 1995; Havanov, 1998;
(Fenton & Griffiths, 2005). The classic problem of Khisaeva & Ostoja-Starzewski, 2006). Numerical
homogenization of heterogeneous materials with methods such as the Finite Element Method (FEM)
variable micro-structure has long been of practi- have also been used to validate the RVE size of ran-
cal interest to engineers. In the current study, the dom heterogeneous materials. Kanit et al. (2003)
influence of voids on effective elastic properties used Monte-Carlo simulations to investigate RVE
is investigated. The goal of homogenization is to and effective properties, while Zohdi & Wriggers
predict the effective property of a heterogeneous (2001) and Ostoja-Starzewski (2006) investigated the
material, where the effective value is defined as the RVE size using a statistical computational approach.
property that would have led to the same response Although there are many models developed to inves-
if the geomaterial had been homogeneous. A use- tigate the effective properties of a material containing
ful concept in this homogenization process is the voids, there is no model that works for all problems

43

ISGSR2013.indb 43 10/18/2013 9:36:36 AM

 (e.g. Böhm, 1998, 2013). See also the reviews pub-
lished by Torquato (2002), Kachanov (2005) and
Klusemann & Svendsen (2009).
In this paper, the Random Finite Element Method
(RFEM) (e.g. Griffiths & Fenton 2007), which com-
bines finite element analysis with random field the-
ory, will be used in conjunction with Monte-Carlo
simulations, to examine the effective elastic proper-
ties of materials with randomly distributed voids.
A 3D cube of material, discretized into a relatively
fine mesh of 8-node hexahedral elements, forms the
basis of the model. Random field theory will be used
to generate a material containing intact material and
voids with controlled porosity and size. The RFEM
can vary the size of the voids through control of
the spatial correlation length and excursion theory
(see e.g. p.141 in Fenton & Griffiths 2008). For each
simulation of the Monte-Carlo process, elements in
the mesh are assigned either an intact stiffness value Figure 1. The 3D finite element model of ideal cubic
or a much lower stiffness value corresponding to a blocks: (a) the solid material, (b) the voids, and (c) the
void. A deterministic analysis follows leading to combined model which show dark and light regions indi-
effective values of the elastic parameters E and υ. cating voids and solid material respectively.
Monte-Carlo analyses are typically repeated numer-
ous times until the output statistics of the effective
elastic properties (mean and SD) stabilize.
The first part of the paper investigates the size
of the RVE for different input void properties. The
second part of the paper investigates the statistics
of the effective Young’s modulus and Poisson’s
ratio in 3D as a function of porosity and void size,
and compares results with numerical and analytical
studies by other investigators. Effective properties
in 3D are also compared with anisotropic results.

2 FINITE ELEMENT MODEL

Examples of the model which combines elastic

material and voids are shown in Figure 1.
The finite element mesh for this study consists of
a cubic block of material of side length L = 50 mod-
Figure 2. Analysis of tied freedom in a “cubic element
eled by 50 × 50 × 50 8-node cubic elements of side test” model with voids. A vertical force is applied on the
length Δx = Δy = Δz = 1.0. Any consistent system of top side. Rollers are fixed at the bottom and two back
units could be combined with the dimensions and sides. The top and the two front sides are tied. The dark
properties described in this paper. Since a mesh such and grey elements represent, respectively, void and intact
as this involves rather large global matrices, equa- solid elastic material.
tion solution in the runs described in this paper will
be performed using a Preconditioned Conjugate
Gradient (PCG) technique with element-by-element are forced to move by the same amount in the
products as described by Smith and Griffiths (2004) analysis. The boundary conditions are such that
which avoids entirely the need to assemble the global the cubic block remains a regular hexahedron after
stiffness matrix. The model in Figure 2 is subjected deformation. Other methods may give similar out-
to a vertical force Q = L × L on the top face lead- comes (see e.g. the effects of tied freedom boundary
ing to an average unit pressure on the top face of condition from Huang et al. 2013). From this idea,
1.0. The boundary conditions of the block involve the effective Young’s modulus and Poisson’s ratio
the use of “tied freedoms” that allow analysis of an easily be back-figured as will be described.
“ideal” block and direct evaluation of the effective In particular, the boundary conditions are such
Young’s modulus and Poisson’s ratio. Tied freedoms that nodes on the base of the block can move only

44

ISGSR2013.indb 44 10/18/2013 9:36:36 AM

 in the x − y plane. The back left and back right faces
are constrained to move only in the y − z and z − x
planes respectively. All z-freedoms on the top plane
are tied, as are the y-freedoms on the front left
plane and the x-freedoms on the front right plane.
A consequence of these constraints is that the top
surface remaining horizontal and the two front
sides remaining vertical following deformation.
These specific boundary conditions enable more
direct comparison to be made with experimental
results, where displacements may be applied with-
out friction on all sides of the specimen. Periodic
boundary conditions have also been used in
homogenization studies of heterogeneous media,
(e.g. Garboczi & Day, 2005). Figure 4. Influence of void element stiffness on the
mean effective Young’s modulus (intact material, E0 = 1).

3 CONTROLLING POROSITY
Thereafter, any element assigned a random field
The random field generator in the RFEM model
value in the range |Z| > zn/2 is treated as intact material
known as the Local Average Subdivision method
with a Young’s modulus and Poisson’s ratio given by
(LAS) (Fenton & Vanmarcke 1990) is used in this
E0 = 1 and υ0 = 0.3, while any element where |Z| ≤ zn/2
paper to model spatially varying voids properties.
is treated as a void element with Young’s modulus
The target mean porosity n is obtained by using the
and Poisson’s ratio given by E0 = 0.01 and υ0 = 0.3
standard normal distribution shown in Figure 3.
(100 times smaller than the surrounding intact
A single value of the random variable Z is initially
material). As can be seen in Figure 4, for the case
assigned to each element of the finite element
when n = 0.2, the results show a small influence of
mesh. Once the standard normal random field
the arbitrarily selected Young’s modulus of the void
values have been assigned, cumulative distribution
elements. In the current work, a void stiffness one
tables Φ (suitably digitized in the software) are then
hundred times less than the surrounding intact mate-
used to estimate the value of the standard normal
rial gave reasonable (and stable) results. The nature
variable zn/2 for which
of random fields is that the mean porosity is under
the user’s control, but the porosity of each individual
Φ ( )
n2 Φ ( 0) n2 (1) simulation processed by the Monte-Carlo method
will vary from one simulation to the next.
where Φ is the cumulative normal distribution
function, and n is the target porosity as shown in
Figure 3. 4 CONTROLLING OF VOID SIZE

As mentioned previously, two materials with the same

average porosity could have quite different void sizes.
One model could have frequent small voids, while
the other could have less frequent larger voids. The
void size in this study is controlled by the random
field spatial correlation length θ which incorporates
a “Markov” spatial correlation structure as follows

ρ ( 2 τ θ) (2)

where ρ = the correlation coefficient; |τ| = absolute

distance between points in the field; and θ = scale
of fluctuation or spatial correlation length. Larger
Figure 3. Target porosity area in standard normal dis- values of θ will lead to larger voids and vice versa.
tribution of random field. Any element assigned a ran- The Markov equation delivers a spatial cor-
dom field value in the range |Z| > zn/2 is treated as intact relation that reduces exponentially with distance.
material a Young’s modulus and Poisson’s ratio given by For example, from Eq. (3), τ < θ, the correlation
E0 = 1 and υ0 = 0.3, respectively. coefficient ρ > 0.13. In the current study, the range

45

ISGSR2013.indb 45 10/18/2013 9:36:37 AM

 Given that L is the side length of cubic block,
and assume stress boundary conditions.

σx 0 0 σ y = 0.0 , σ z = −Q /L2 (4)

δx δy δ
εx = , εy = , εz = z (5)
L L L

hence after substitution into equation (3), the effec-

tive elastic properties can be written as

Q
E= (6)
Lδ z
Figure 5. Typical simulations showing generation of δx
voids at (a) low and (b) high spatial correlation lengths υx = (7)
θ (n = 0.2 in both cases). δz

δy
of ρ varies from 0 to 1. Points close together are υy = (8)
δz
strongly correlated and therefore likely to belong
to the same void. In the limiting case of θ → 0, the
random field value changes rapidly from point to where E = the effective elastic Young’s modulus,
point delivering numerous small voids. At the other Q = stress loading at the top side, υx and υx = the
extreme as θ → ∞, the random on each simulation effective Poisson’s ratios based on the displacement
becomes increasingly uniform with some simula- in the x- and y-directions respectively.
tions representing entirely intact material and In each simulation, the effective Young’s modu-
other consisting entirely of voids. For example as lus is normalized as E/E0 by dividing by the intact
shown in Figure 5, the models show typical simula- Young’s modulus E0. In the current study, fol-
tions of different void clustering for two materials lowing some numerical experiments as shown in
with the same mean porosity. Figure 6, it was decided that 1000 simulations for
each parametric combination would deliver rea-
sonably repeatable results. In this study, we have
5 MONTE-CARLO SIMULATIONS expressed the spatial correlation length in dimen-
sionless form
A “Monte-Carlo” process is combined with the
RFEM and repeated until stable output statistics θ
Θ= (9)
are achieved. The primary outputs from each elas- L
tic analysis are the vertical and horizontal defor-
mations of the block δz, δx and δy. Although all where L is the width of the loaded element
simulation use the same θ and n, the spatial loca- (L = 50).
tion of the voids will different each time. In some
cases, the voids may be located just below the top
of the block leading to a relatively high δz. While
in others, the voids may be buried in the middle of
the block leading to a relatively low δz. Following
each simulation, the computed displacements δz, δx
and δy are converted into the “effective” values of
Young’s modulus and Poisson’s ratio as follows
Based on Hooke’s law,

1
εx (σ x − υ (σ y + σ z ))
E
1
εy (σ y − υ (σ z + σ x )) (3) Figure 6. Sensitivity of the mean effective Young’s
E modulus as a function of the number of simulations for
1 n = 0.2 and Θ = 0.4. It was decided that 1000 simulations
εz (σ z − υ (σ x + σ y ))
E would deliver reasonably repeatability.

46

ISGSR2013.indb 46 10/18/2013 9:36:39 AM

 6 REPRESENTATIVE VOLUME ELEMENT

An RVE is an element of the heterogeneous mate-

rial that is large enough to represent the microstruc-
ture, but small enough to achieve computational
efficiency. The RVE of four cases using the random
field 3D finite element model have been considered
as shown in Table 1.
Figure 7 shows a sequence of five blocks con-
tained within and including the largest block of
50 × 50 × 50 cubic elements. The different block
sizes will indicate the optimal RVE for the given
input conditions. When the RVE is “big enough”,
we expect the standard deviation of the effec-
tive Young’s modulus to be reduced and its mean
essentially constant as shown in Figures 8(a) and
8(b). While the mean values plotted in Figure 8(a)
are fairly constant for different block sizes, it could
be argued that the block size of 20 × 20 × 20 led
to essentially constant values for the low Θ cases
(1 and 3), while a larger block, say 30 × 30 × 30
would be needed for stable mean values with the
larger Θ cases (2 and 4). The standard deviation
shown in Figure 8(b) displays more variability
with block size and tends to zero as the blocks get

Figure 8. Effective Young’s modulus (a) mean and (b)

Table 1. Different input void properties. standard deviation following 1000 simulations for differ-
ent block sizes.
Case Target porosity (n) Θ

1 0.2 0.2
bigger, but at a slower rate for higher values of Θ.
2 0.2 0.7
In both Figures 8, it is noted that the influence of
3 0.7 0.2
Θ on block statistics is greater than that of n. The
4 0.7 0.7
RVE depends more on spatial correlation length
than porosity.

7 RESULTS OF RFEM

Following each set of 1000 Monte-Carlo simula-

tions, the mean and standard deviation of the
normalized effective Young’s modulus were com-
puted for a range of parametric variations of n
and Θ, with results shown in Figures 9 and 10,
respectively.
It can be noted from Figure 9 that the mean nor-
malized effective Young’s modulus drops towards
zero with increasing porosity n and that Θ does
not have much influence. Figure 10 shows that Θ
has more influence on the standard deviation of
the effective Young’s modulus σ E E0 . The standard
deviation values as n → 0 (intact stiffness material)
and n → 1 (very low stiffness material) show low
variance since almost all simulations are the same
and model essentially uniform material. The stand-
Figure 7. Different block sizes for computing the effec- ard deviation was observed to reach a maximum
tive elastic properties of a material with random voids. value at around n ≈ 0.4.

47

ISGSR2013.indb 47 10/18/2013 9:36:41 AM

 Figure 9. μE/E0 vs. n for 0.2 ≤ Θ ≤ 1.0. Figure 12. συ vs. n for 0.2 ≤ Θ ≤ 1.0.

Figure 13. Mean effective values vs. n using Θ = 0.2.

Figure 10. σE/E0 vs. n for 0.2 ≤ Θ ≤ 1.0.

and standard deviation of the effective Poisson’s
ratio. Figure 11 shows that the mean effective
Poisson’s ratio μυ displays a minimum at around
n = 0.5. On the other hand, as shown in Figure 12,
the standard deviation of Poisson’s ratio displays a
maximum at n = 0.7 which is a similar trend to that
observed for Young’s modulus in Figure 10. For
all values of Θ considered however, the standard
deviations were quite small.
Although this paper has focused on Young’s
modulus and Poisson’s ratio, other stiffness mod-
uli may be of interest depending on the context.
Figure 13 combines results from Figures 9 and 11
to show the variation of the mean effective shear
modulus and bulk modulus using Eqns. (10 and 11).
They display a similar trend to that observed for
Figure 11. μυ vs. n for 0.2 ≤ Θ ≤ 1.0.
Young’s modulus.

μE
The result obtained from Equations 7 and 8 for μK = (10)
the effective Poisson’s ratio were in good agreement 3(1 − 2 μυ )
as expected for the range of n and Θ considered.
μE
In the isotropic material model, the two Poisson’s μS = (11)
ratios are essentially identical after Monte-Carlo 2(1 + μυ )
simulation; however the results are based on an
average to account for any small differences. The where μK = the mean effective bulk modulus,
plots shown in Figures 11 and 12 give the mean μS = the mean effective shear modulus.

48

ISGSR2013.indb 48 10/18/2013 9:36:43 AM

 8 COMPARISON OF RFEM AND OTHER
RESULTS

The theoretical results based on the Generalized

Self Consistent Method of Christensen & Lo
(1979) and the numerical results based on the
single-cut GRF model of Roberts & Garboczi
(2002) are compared in Figure 14, with results
from the current study using Θ = 0.6 from Figure 9.
The Generalized Self Consistent Method involved
embedding an inclusion phase directly into an infi-
nite medium. It was demonstrated that the method
could also solve the spherical inclusion problem.
The single-cut GRF model assigns a random
number to each point in space. From Figure 14, it
Figure 14. Comparison of the effective Young’s modu- can be observed that the current method gives sim-
lus obtained from RFEM and other approaches. ilar values of the mean effective Young’s modulus

Figure 15. Analysis of tied freedom in four “cubic element test” models with voids: (a) isotropic model, (b) aniso-
tropic model along x-axis, (c) anisotropic model along y-axis and (d) anisotropic model along z-axis.

49

ISGSR2013.indb 49 10/18/2013 9:36:44 AM

 to those given by the theoretical and numerical voids in 2D (plane strain) are like “tunnels” that
methods for all values of n. continue indefinitely into the 3rd dimension, while
voids in 3D are isotropic, finite in size, and fully
contained within the surrounding material. Thus,
9 COMPARISON OF ISOTROPIC it might be explained that the 2D model is actu-
AND ANISOTROPIC MODELS ally a 3D model with an infinite spatial correlation
length in the 3rd direction.
Anisotropic models are performed with differ-
ent spatial correlation lengths in different direc-
tions. Figure 15(a) shows a model of isotropic 10 CONCLUSIONS
spatial correlation length at Θx = Θy = Θz = 0.6.
The voids tend to disperse in all directions within A 3D RFEM with “tied freedoms” has been used
the material. In Figure 15(b), where Θx = 6 and in this study to investigate the influence of poros-
Θy = Θz = 0.6, and Figure 15(c) where Θy = 6 and ity and void size on homogenized elastic properties
Θx = Θz = 0.6, voids are in horizontally elongated E and υ. It was observed that while porosity had
in the x- and y-directions, respectively. On the a significant effect on both the mean and stand-
other hand, in Figure15(d) where where Θz = 6 ard deviation of E and υ, the void size had little
and Θx = Θy = 0.6, voids are in vertically elongated influence on the mean but more influence on the
in the z-direction. The tied freedom approach standard deviation. The study also investigated
described previously continued to be used in all the RVE needed to capture the essential properties
3D anisotropic models. of a heterogeneous material containing voids. It
Following the Monte-Carlo simulations of the was found that for the same porosity, the larger the
anisotropic models shown in Figure 15, the mean size of the voids, the greater the size of the RVE.
of the effective normalized Young’s modulus was Finally, the paper presented favorable comparisons
compared with isotropic results for a range of n, of the effective elastic properties in 3D with those
as shown in Figure 16. It was noted that similar obtained analytically and numerically by other
results were obtained when the elongated direc- investigators. In addition, the effective Young’s
tion of the anisotropic models was in the x- and modulus of the anisotropic system depends on
y-directions, but the effective Young’s modulus was the direction of voids elongation. The stiffest case
noticeably higher when the elongations were in the was observed when the direction of void elonga-
z-direction (the direction of loading). tion was in the same direction as the loading. The
The isotropic 3D results from the current study RFEM approach to homogenization described
using Θ = 0.6 are also compared in Figure 16 with in this paper shows much promise, and opens
2D (plane strain) for the same spatial correlation the possibility of making probabilistic statements
length as published previously by Griffiths et al. about engineering performance of heterogeneous
2012. The mean normalized effective Young’s geomaterials. The probabilistic aspect has not been
modulus in 3D is obviously higher than in 2D for discussed in the current paper, but remains an area
the same porosity. A direct comparison between of continued research.
2D and 3D may not be justified, however, because

ACKNOWLEDGEMENT

The authors wish to acknowledge the support

of (i) NSF grant CMMI-0970122 on “GOALI:
Probabilistic Geomechanical Analysis in the
Exploitation of Unconventional Resources”, and
(ii) The Royal Thai Government for their support
of the second author.

REFERENCES

Böhm, H.J. 1998, 2013. A short introduction to basic

aspects of comtinuum micromechanics, TU Wien,
Vienna.
Christensen, R.M. and Lo, K.H. 1979. Solutions for
Figure 16. Comparison of the effective Young’s modu- effective shear properties in three phase sphere and
lus obtained from RFEM and other approaches. cylinder models. J. Mech. Phys. Solids 27: 315–330.

50

ISGSR2013.indb 50 10/18/2013 9:36:46 AM

 Fenton, G.A. & Griffiths, D.V. 2005. Three-Dimension Kachanov, M. & Sevostianov, I. 2005. Onquantitative
Probabilistic Foundation Settlement. J. Geotech. characterization of microstructures and effective
Geoenviron. Eng. 131 (2): 232–239. properties. Int. J. Solids Struct. 42: 309–336.
Fenton, G.A. & Griffiths, D.V. 2008. Risk Assessment Kanit, T., Forest, S., Galliet, I., Mounoury, V. & Jeulin, D.
in Geotechnical Engineering, John Wiley & Sons, 2003. Determination of the size of the representative
Hoboken, NJ. volume element for random composites: statistical
Fenton, G.A. & Vanmarcke, E.H. 1990. Simulation of and numerical approach. Int. J. Solids Structure 40
random fields via local average subdivision. J Eng. (13–14): 3647–3679.
Mech. 116 (8): 1733–1749. Khisaeva, Z.F. & Ostoja-Starzewski, M. 2006. On the
Garboczi, E.J. & Day, A.R. 2005. An algorithm for com- size of RVE in finite elasticity of random composites.
puting the effective linear elastic properties of het- Journal of Elasticity 85 (2): 153–173.
erogeneous materials: three-dimensional results for Klusemann, B. & Svendsen, B. 2009. Homogenization
composites with equal phase Poisson ratios. J. Mech. methods for multi-phase elastic composites: Com-
Phys. Solids 43 (9): 1349–1362. parisons and benchmarks. Technische Mechanik 30
Griffiths, D.V. & Fenton, G.A. 2007. Probabilistic Meth- (4): 374–386.
ods in Geotechnical Engineering, CISM Courses and Liu, C. 2005. On the minimum size of representative
Lectures No. 491, Pub. Springer, Wien, New York. volume element: an experimental investigation. Exp.
Griffiths, D.V., Paiboon, J., Huang, J. & Fenton, G.A. Mech. 45 (3): 238–243.
2012. Homogenization of geomaterials containing Ostoja-Starzewski, M. 2006. Material spatial random-
voids by random fields and finite elements. Int. J. Sol- ness: From statistical to representative volume element.
ids Structure 49: 2006–2014. Probabilistic Engineering Mechanics 21 (2): 112–132.
Hazanov, S. & Huet, C. 1994. Order Relationships for Roberts, A.P. & Garboczi, E.J. 2002. Computation of
Boundary-Conditions Effect in Heterogeneous Bod- the linear elastic properties of random porous materi-
ies Smaller Than the Representative Volume. Journal als with a wide variety of microstructure. Proc. Royal
of the Mechanics and Physics of Solids 42 (12): Soc. of London 458 (2021): 1033–1054.
1995–2011. Smith, I.M. & Griffiths, D.V. 2004. Programming the
Hazanov, S. & Amieur, M. 1995. On overall properties finite element method, John Wiley and Sons, Chiches-
of elastic heterogeneous bodies smaller than the repre- ter, New York, 4th edition
sentative volume. International Journal of Engineering Torquato, S. 2002. Random heterogeneous materials:
Science 33 (9): 1289–1301. microstructure and macroscopic properties. Springer-
Hazanov, S. 1998. Hill condition and overall properties Verlag, NY.
of composites. Archive of Applied Mechanics 68 (6): Zeleniakiene, D. Griskevicius, P. & Leisis, V. 2005. The
385–394. comparative analysis of 2D and 3D microstructural
Hill, R. 1963. Elastic properties of reinforced solids: models stresses of porous polymer materials. ISSN
some theoretical principles. J. Mech. Phys. Solids 11: 1392-1207. Mechanika, Nr. 3 (53).
357–372 Zohdi, T.I. & Wriggers, P. 2001. Aspects of the com-
Huang, J.S., Krabbenhoft, K. & Lyamin, A. 2013. Statis- putational testing of the mechanical properties of
tical homogenization of elastic properties of cement microheterogeneous material samples. International
paste based on X-ray microtomography images. Int. J. Journal for Numerical Methods in Engineering 50 (11):
Solids Structure 50: 699–709. 2573–2599.

51

ISGSR2013.indb 51 10/18/2013 9:36:46 AM

This page intentionally left blank
 Geotechnical Safety and Risk IV – Zhang et al. (eds)
© 2014 Taylor & Francis Group, London, ISBN 978-1-138-00163-3

Selecting optimal probability models for geotechnical reliability analysis

H.W. Huang, W.W. Su & J. Zhang

Department of Geotechnical Engineering, Tongji University, Shanghai, China
Key Laboratory of Geotechnical and Underground Engineering of Ministry of Education, Tongji University,
Shanghai, China

ABSTRACT: The cohesion and friction angle of soil are two common random variables encountered
in geotechnical reliability analysis. Its joint distribution can be fully specified with marginal distributions
and a copula function. In this study, a probabilistic framework is suggested to construct, calibrate and
rank probability models for failure probability estimation. It is shown that calibrating the marginal dis-
tributions and copula function may miss the optimal model parameters. The optimal probability model
is problem specific, and the commonly adopted models based on the Gaussian copula function may not
be optimal. When copula function is used to specify the joint distribution of cohesion and friction angle,
many existing algorithms such as the first order reliability method cannot be directly used for failure prob-
ability calculation. Monte Carlo simulation can be applied to general performance functions and hence is
superior to the existing one-dimensional integration method. Both the marginal distribution and copula
function could significantly affect the failure probability calculation. If the probability model is arbitrarily
chosen, the failure probability may be severely overestimated or underestimated. In general, the effect of
probability model on failure probability estimation is more obvious when the failure probability is small.
The effect of probability model on failure probability calculation also depends on the deterministic model
and the available data. Increasing the amount of the measured data is likely to provide more constraint
to the marginal distributions and hence reduce its effect on failure probability estimation. The effect of
copula function on failure probability calculation decreases as there is less correlation between the cohe-
sion and the friction angle.

1 INTRODUCTION angle are often modeled as normal or lognormal

variables. A correlation coefficient is often used
To address the uncertainties in geotechnical to describe the dependence between the cohesion
design, probabilistic methods in which uncertain and the friction angle (e.g., Harr 1987, Cherubini
variables are modeled as random variables have 1997). It is generally believed that the cohe-
been widely employed in geotechnical analysis sion and friction angle are negatively correlated.
(Vanmarcke 1977, Tang 1979, Low & Tang 1997, Nevertheless, a positive correlation between the
Cherubini 2000, Duncan 2000, Whitman 2000, cohesion and the friction angle was also reported
Baecher & Christian 2003, Zhang 2004, Honjo & (e.g., Wolff 1985).
Amatya 2005, Uzielli et al. 2005, Sivakumar Babu Recently, it has been found the marginal distribu-
et al. 2006, Fenton & Griffiths 2008, Huang et al. tions and the correlation coefficient are insufficient
2010, Wang 2011, Luo et al. 2011, Xue & Gavin to characterize a joint distribution (e.g. Durrleman
2007, Zhang et al. 2011, Ching & Phoon 2012, et al. 2000, Cherubini et al. 2004, Nelsen 2006).
Juang & Wang 2013, Zhang et al. 2013). The Rather, a multivariate distribution can be fully
cohesion and friction angle are two commonly specified by the marginal distributions and a cop-
encountered important uncertain variables in ula function. It is increasingly aware in the field of
many geotechnical problems. In many cases, the structural reliability that a joint distribution should
probability model for the shear strength param- be modeled based on the copula theory (e.g. Goda
eters is the fundamental input for a geotechnical 2010, Lebrun & Dutfoy 2009). In geotechnical
reliability analysis. If the probability model for the engineering, Li and his coauthors pioneered the
shear strength parameters is biased, the estimated application of copula theory in a number of geo-
failure probability will also be biased. Lumb technical reliability problems (Li et al. 2013, Tang
(1970) and Chen et al. (2005) studied the marginal et al. 2013). It was found that the estimated failure
distributions of the cohesion and the friction probability could be severely biased if the copula
angle. In practice, the cohesion and the friction function is wrongly chosen.

53

ISGSR2013.indb 53 10/18/2013 9:36:46 AM

 Despite these substantial progresses, several respectively. Let F(c, ϕ) denote the CDF of {c, ϕ}.
important questions still remain in calibrating and It is reasonable to assume c and ϕ are both con-
selecting probability models for the shear strength tinuous variables. Based on Sklar’s theorem (Sklar
parameters. Firstly, the marginal distributions and 1959), F(c, ϕ) can be written as follows
copula functions of the soil strength parameters
were often determined separately, assuming there F (c, ϕ ) C ⎡⎣ F1 (c ) , F2 (ϕ ) , θ ⎤⎦ (1)
is no interaction between the optimal marginal dis-
tributions and the optimal copula function, which
where C(u1, u2, θ) is the copula function, u1 = F1(c),
may not be true. Secondly, many existing methods
u2 = F2(ϕ), and θ is a parameter of the copula
for reliability analysis are implicitly based on prob-
function. Eq. (1) shows that the joint distribution
ability models with the Gaussian copula function,
of c and ϕ can be decomposed into two parts, i.e.
which may not be directly applicable to a probabil-
(1) the marginal distributions, i.e. F1(c) and F2(ϕ),
ity model with a general copula function. Thirdly,
and (2) the copula function that characterize the
when and how the probability model will have sig-
dependent structure of c and ϕ, i.e. C(u1, u2, θ). As
nificant impact on the failure probability estima-
an example, Table 1 shows several commonly used
tion are still not well-understood.
copula models.
The objectives of this paper include: (1) provid-
Based on Eq. (1), the Probability Density Func-
ing a systematical framework to calibrate and rank
tion (PDF) of c and ϕ can be written as
probability models for the shear strength parame-
ters; (2) suggesting a method to evaluate the failure
probability when a general performance function f ( c,ϕ ) c [ F (c ),
) F (ϕ )),θ ] f1 ( c ) f2 ( ϕ ) (2)
is involved; and (3) investigating when and how the
probability model can have significant impact on where c(u1, u2, θ) is the density of the copula func-
failure probability estimation. tion C(u1, u2, θ) defined as follows
This paper is organized as follows. First, the
copula theory is briefly introduced, followed by ∂C ( u1 u2 ,θ )
how various probability models can be calibrated c ( u1 u2 ,θ ) = (3)
∂u1∂u2
and compared for modeling the shear strength
data. Then, how the failure probability can be cal-
culated when a general performance function is Based on the definition of correlation coeffi-
involved is presented. Finally, three examples are cient, the correlation coefficient between c and ϕ
worked out to illustrate the proposed method and can be written as follows
also to investigate the effect of probability model
of soil strength parameters on failure probability
ρ12 =
∫∫ (c − μc )(ϕ − μϕ ) f ( ϕ )dcdϕ
(4)
calculation. σ cσ ϕ

2 PROBABILITY MODELS FOR SHEAR where μc and μϕ are the mean values of c and ϕ,
STRENGTH PARAMETERS respectively, and σc and σϕ are the standard devia-
tions of c and ϕ, respectively.
Let c and ϕ denote the cohesion and friction angle, Comparing Eq. (3) with (4), we can see that the
respectively. Let F1(c) and F2(ϕ) denote the Cumu- correlation coefficient is a function of the copula
lative Distribution Function (CDF) of c and ϕ, model. Previously, the Gaussian copula is often

Table 1. Summary of several commonly used copula functions.

Copula C(u1, u2,θ) c(u1, u2, θ)

Gaussian φG φ −1 (u1 )), φ 1

θ) 1 ⎡ θ 2 ⋅φ 1 2
− 2θ ⋅ 1
( 1) φ
−11
+ θ 2 ⋅ φ −1 ( 2 )2 ⎤
2 exp ⎢ − 1 2
⎥
1−θ 2
⎣ 2(1 − θ 2 ) ⎦
1 1
− −2 −
θ θ θ θ 1 θ θ θ
Clayton ( 1 2 ) ( ))( 1 2) ( 1 2 )
1 ⎡ ( −θ 1 −11)( −θ u − 1) ⎤ −θ ( −θ − )e −θ ( 1 + 2 )
Frank − l ⎢1 + ⎥ −θ
θ ⎣ e −θ − 1 ⎦ [(e − )(e −θ u − 1) + (eθ − 1)]2
FGM u1u2 (1 (1 − u1 )(1 u2 )) 1 θ (1 − 2 1 )(1 − 2 2 )

54

ISGSR2013.indb 54 10/18/2013 9:36:46 AM

 used to construct the joint distribution. As shown the chance to observe D when the values of Θ are
in Table 1, Gaussian copula is only one of the known can be written as
possible options for constructing the probability
models. Different copula functions can result in the n
same correlation coefficient. However, when dif- l ∏f( i
| Θ) (8)
i =1
ferent copula functions are used, the failure prob-
ability could be very different (Tang et al. 2013).
where f(di|Θ) is the chance to observe di. If the den-
The commonly adopted Gaussian copula may not
sity function in Eq. (5) is used, f(di|Θ) can be writ-
always be the optimal copula function to construct
ten as
probability models for shear strength parameters.
Sklar’s theory provides the modeler ample −θ −1
⎡⎡ ⎡ F2 d2i ⎤ ⎤⎥
( ) ( ) ( )
chances to model the real data more realistically. For −θ −θ
f i
F1 d1i ⎤
instance, if one assumes the marginal distributions ⎣ ⎣ ⎦ ⎣ ⎦ ⎦
are normal and that copula function is Clayton, the 1
−1−
PDF of c and ϕ can be written as follows ⎧ ⎫
( ) ( ) ( ) ( )
−θ −θ θ
× ⎨ ⎡ F1 d1i ⎤ + ⎡ F2 d2i ⎤ − 1⎬ f1 d1i f2 d2i
⎩ ⎣ ⎦ ⎣ ⎦ ⎭
−θ −1
f (c,ϕ ) (1 θ ) ⎡⎣ F1 (c ) ⋅ F2 (ϕ ) ⎤⎦
1 (9)

{ }
−θ θ −1−
× ⎡⎣ F1 (c ) ⎦⎤ + ⎣⎡ F2 (ϕ ) ⎤⎦ −1 θ
f1 ( ) f2 ( ) (5) i i
where d1 and d are the first and second elements
2
of di, respectively.
Based on the principle of maximum likelihood,
where f1(c) and F1(c) in this case are respectively the
the optimal values of Θ can be obtained by maxi-
PDF and CDF of a normal variable with a mean
mizing the likelihood function, or equivalently, by
of μc and a standard deviation of σc, and f2(ϕ)
maximizing the logarithm of the likelihood func-
and F2(ϕ) are respectively the PDF and CDF of a
tion as follows
normal variable with a mean of μϕ and a standard
deviation of σϕ. As an example, f1( c) and F1(c) can n
be written as follows L ∑
i =1
f ( i
|Θ ) (10)

1 ⎡ ( x − )2 ⎤
f1 (c ) = p ⎢−
exp c
⎥ (6) If Eq. (10) is used to find the optimal values of
2π σ c ⎢⎣ 2σ c2 ⎥⎦ Θ, the marginal distributions and the copula func-
tion are calibrated simultaneously. Thus, if there is
c any interaction between the optimal marginal dis-
F1 (c ) = ∫ f ( z ) dz (7) tributions and copula function, such interaction
−∞ 1
−
can be readily considered. Nevertheless, estimating
Sklar’s theory implies that the structure of a the values of Θ using Eq. (10) requires solving the
probability model for cohesion and friction angle optimization problem in a five-dimensional space,
can be constructed in two steps: (1) determining which could be non-trivial. To reduce the computa-
the marginal distributions of c and ϕ, and tional challenge involved in the optimization work,
(2) determining the copula function between c and one can assume there is no interaction between
ϕ. After the structure of the probability model is the optimal marginal distributions and the copula
determined, the next step is how to calibrate the function. In such a case, the marginal distributions
model parameters of the probability model, as and copula function can be calibrated separately
described in the following section. (Joe & Xu 1996). For the shear strength probability
model studied here, the optimal values of μc, σc, μϕ,
σϕ, and θ can be obtained by respectively maximiz-
3 MAXIMUM LIKELIHOOD ing the following three likelihood functions
CALIBRATION OF PROBABILITY
MODELS
( )
n
μc c ∏ f1 d1i | μc ,σ c (11)
In the probability model for shear strength param- i=1
eters, the parameters to be calibrated include μc, μϕ,
( )
n
σc, σϕ and θ. Let d = {d1, d2} denote a measurement of l μϕ ϕ ∏ i
f2 d | μϕ , σ ϕ (12)
c and ϕ. Let d1, d2, d3, …, dn denote n measurements i=1

of c and ϕ. For ease of presentation, let Θ = {μc,

( )
n
μϕ, σc, σϕ, θ} and D = {d1, d2, d3, …, dn}. Supposing lθ ∏ f i
| μc* ,σ c* , μϕ* ,σ ϕ* ,θ (13)
the measurements are statistically independent, i=1

55

ISGSR2013.indb 55 10/18/2013 9:36:51 AM

 where in Eq. (13) μc* and σc* are optimal values of values found for Θ based on Eqs. (11)–(13) into
μc and σc determined based on Eq. (11), and μϕ* Eq. (10) to obtain L(Θ*|D), based on which the
and σϕ* are optimal values of μϕ and σϕ determined AIC value can be conveniently calculated.
based on Eq. (12). While no interaction between
the marginal distributions and copula function is
a convenient assumption to reduce the computa- 5 GEOTECHNICAL RELIABILITY
tional work, its accuracy needs to be verified. The ANALYSIS BASED ON COPULA
accuracy of this approximate method will be stud- MODELS
ied later in this paper.
Considering the simple case where there is only one
soil layer involved. Let x = {c, ϕ}. Let g(x) = 0 denote
4 RANKING OF PROBABILITY MODELS the limit state function with g(x) < 0 implying fail-
ure. The probability of failure can be written as
Considering the possible marginal distributions
of cohesion and friction angle as well as the pos-
sible copula function, there could be more than
pf ∫∫ I ⎡⎣ g ( )⎤⎦ f (c ) dcdϕ (17)

one probability models that can be used to fit the

measured data. When multiple probability models where I is the indictor function characterizing the
are available, the probability models can be com- failure domain defined as follows
pared using various information criteria avail-
able in the statistics literature, such as the Akaike ⎧1 g ( x ) < 0
I ⎡⎣ g ( x ) ⎤⎦ = ⎨ (18)
Information Criterion (AIC) (Akaike 1974), the ⎩0 g ( x ) ≥ 0
Bayesian Information Criterion (BIC) (Schwarz
1978), and the Kullback Information Criterion
The First Order Reliability Method (FORM)
(KIC) (Cavanaugh 1999), as listed below:
(Ang & Tang 1984) is widely used in geotechnical
engineering. However, as it is derived for prob-
AIC = −2L( *
| ) + 2k (14) ability models constructed based on the Gaussian
BIC = −2L( *
| ) + k ln n (15) copula function, it is not directly applicable when
other copula functions are used. Tang et al. (2013)
KIC = −2L( *
| ) + 3k (16) illustrated that the double integration in Eq. (17)
can be simplified into a one-dimensional integra-
where Θ* is the point where the likelihood func- tion problem analytically if the performance func-
tion is maximized; k = number of parameters to be tion is relatively simple. However, such a method
calibrated; and n = number of observed data. An may not be applicable when the performance func-
information criterion generally has two compo- tion is complex or when there are several soil layers
nents, i.e. a term related to the value of the maxi- involved. In this study, we used the Monte Carlo
mum likelihood, and a term related to the number simulation to calculate the failure probability. The
of parameters in the model. The term related to advantages of this method include: (1) it does not
the likelihood function measures the degree of fit require that the double integration in Eq. (17)
of the data to the model. The term related to the should be simplified into a single integration prob-
number of parameters measures the complexity lem analytically; and (2) it is applicable even when
of the model. Based on the information criterion, the performance function is complex. Let xi denote
a model should be both accurate and concise to the ith sample of x. The failure probability as indi-
avoid over-fitting. The above criteria have differ- cated in Eq. (17) can be calculated as
ent theoretical background and differ in the weight
1 N ⎡
applied to the complexity of the model. The smaller
the criterion is, the more the model is supported pf =
N i =1 ⎣
( )
∑ I g x i ⎤⎦ (19)
by the data. When the models to be compared have
the same number of parameters, the model rank-
where N = number of samples. For a bivariate dis-
ing result based on different information criteria is
tribution constructed based on the Clayton, Frank
the same. In this study, the models to be compared
and FGM function, the following conditional sam-
all have five parameters, i.e. k = 5. Thus, the model
pling algorithm can be used (Trivedi et al. 2005).
ranking result will not be affected by the informa-
tion criterion adopted. As an illustration, AIC will 1. Draw two independent random variables (v1, v2)
be used for model ranking. Note to obtain the AIC from a uniform distribution bounded between 0
value of the model calibrated based on Eqs. (11)– and 1.
(13), one simply needs to substitute the optimal 2. Set u1 = v1.

56

ISGSR2013.indb 56 10/18/2013 9:36:54 AM

 3. Set t = C2(u2|u1 = v1) = ∂C(u1,u2)/∂u1 = Cu(v). between optimal marginal distributions and cop-
4. Set u2 = C−1u (v, t). ula function is considered. The calibration results
5. Let x1 = F−11(u1), and x2 = F−12(u2). {x1, x2} will be are also summarized in Table 3. From Table 3 the
samples of {c, ϕ}, where F−12(u2) and F−12(u2) are following phenomena can be observed.
the inverse of the CDF of c and ϕ, respectively.
1. When marginal distributions are both normal
For a probability model constructed based on and when the copula function is Clayton, the
the Gaussian copula, the following algorithm can mean and standard deviation of c is 152.3 kPa
be used (Trivedi et al. 2005).
1. Generate two independent distributed samples
from the standard normal distribution which Table 2. Test data from Ankang Hydropower site
are denoted as v1 and v2 here, respectively. (Adapted from Tang et al. 2013).
2. Set y1 = v1.
3. Set y2 = v1 ⋅ θ + v2 ⋅ (1 − θ2)1/2. Test Test
4. Set ui = Φ(yi) for i = 1, 2 (Φ is the cumulative no. c (kPa) ϕ (o) no. c (kPa) ϕ (o)
distribution function of the standard normal
1 165 11.86 14 85 20.81
distribution).
2 127 14.04 15 18 25.64
5. Let x1 = F−11(u1), and x2 = F−12(u2). {x1, x2} is a
3 253 13.50 16 15 22.29
sample of {c, ϕ}.
4 427 10.20 17 78 24.70
5 106 11.31 18 12 26.10
6 242 12.95 19 34 22.78
6 AN ILLUSTRATIVE EXAMPLE 7 209 12.41 20 70 19.80
8 328 13.50 21 20 17.74
The suggested method is used to analyze the data as 9 98 12.95 22 20 20.81
shown in Table 2, which was also studied previously 10 10 15.64 23 217 20.30
in Tang et al. (2013). In this example, 16 possible 11 213 16.17 24 221 20.30
probability models for c and ϕ are considered, as 12 365 17.22 25 254 27.47
shown in Table 3. The 16 models are first calibrated 13 324 20.81
based on Eq. (10) where the possible interaction

Table 3. Calibrated parameters of 16 probability models based on the given data (model parameters calibrated
simultaneously).

Model Marginal
no. Copula distributions* AIC θ μc σc μϕ σϕ

1 Gaussian n,n 468.28 −0.407 156.440 122.229 18.052 5.014

2 l,n 462.58 −0.443 188.762 326.223 18.052 5.013
3 n,l 467.26 −0.421 156.440 122.186 18.058 5.253
4 l,l 461.70 −0.453 188.762 326.223 18.064 5.274
5 Clayton n,n 466.27 −0.439 152.337 124.067 17.848 5.364
6 l,n 463.84 −0.273 196.685 351.868 18.110 4.849
7 n,l 466.34 −0.389 155.700 119.413 18.100 5.613
8 l,l 462.24 −0.299 195.782 351.356 18.070 5.027
9 Frank n,n 468.62 −2.264 152.609 121.861 17.963 4.984
10 l,n 462.15 −2.712 180.607 302.962 18.157 4.973
11 n,l 468.04 −2.233 154.026 121.839 17.981 5.211
12 l,l 461.72 −2.552 183.182 307.764 18.157 5.243
13 FGM** n,n 468.35 −1.000 154.310 122.118 18.042 4.956
14 l,n 462.22 −1.000 184.965 309.786 18.172 4.924
15 n,l 467.71 −1.000 155.206 121.928 18.084 5.222
16 l,l 461.51 −1.000 186.422 311.737 18.201 5.217

*n,n: both c and ϕ follow the normal distribution; l,n: c and ϕ follow the lognormal and normal distributions, respectively;
n,l: c and ϕ follow the normal and lognormal distributions, respectively; l,l: both c and ϕ follow the lognormal distributions.
**For the FGM copula, the value of θ is bounded between −1 and 1 (Trivedi et al. 2005). This constraint is considered
in the maximum likelihood method.

57

ISGSR2013.indb 57 10/18/2013 9:36:57 AM

 Table 4. Calibrated parameters of 16 probability models based on the data shown in Table 2 (model parameters
calibrated separately).

Model Marginal
no. Copula distribution AIC θ μc σc μϕ σϕ

1 Gaussian n,n 468.28 −0.407 156.440 122.229 18.052 5.014

2 l,n 462.58 −0.443 188.762 326.223 18.052 5.014
3 n,l 467.26 −0.423 156.440 122.229 18.064 5.274
4 l,l 461.70 −0.453 188.762 326.223 18.064 5.274
5 Clayton n,n 457.89 −0.500 156.440 122.229 18.052 5.014
6 l,n 463.92 −0.273 188.762 326.223 18.052 5.014
7 n,l 466.50 −0.369 156.440 122.229 18.064 5.274
8 l,l 462.38 −0.298 188.762 326.223 18.064 5.274
9 Frank n,n 468.67 −2.241 156.440 122.229 18.052 5.014
10 l,n 462.18 −2.740 188.762 326.223 18.052 5.014
11 n,l 468.07 −2.227 156.440 122.229 18.064 5.274
12 l,l 461.74 −2.647 188.762 326.223 18.064 5.274
13 FGM n,n 468.37 −1.000 156.440 122.229 18.052 5.014
14 l,n 462.27 −1.000 188.762 326.223 18.052 5.014
15 n,l 467.71 −1.000 156.440 122.229 18.064 5.274
16 l,l 461.57 −1.000 188.762 326.223 18.064 5.274

and 124.1 kPa, respectively. However, when the 4. For a given copula structure, the model assu-ming
marginal distribution of c is changed to be log- c and ϕ are lognormally distributed is most sup-
normal, the mean and standard deviation of c is ported by the data. Comparing models with dif-
changed to be 196.7 kPa and 351.9 kPa, respec- ferent copula functions, Model 16 with the FGM
tively. The optimal model parameters are sensi- copula function has the smallest AIC, and hence
tive to the assumed marginal distribution. is most supported by the data. Models based on
2. When the marginal distributions of cohesion the commonly used Gaussian copula function
and friction angle are both normal distributions are not the optimal probabilistic models.
but when the copula function is assumed to be
Gaussian and Clayton, respectively, the mean
values of c and ϕ are both changed, indicating 7 IMPACT OF PROBABILITY
that the optimal marginal distribution is also MODELS ON FAILURE
affected by the copula function. Hence, it may PROBABILITY CALCULATION
not be appropriate to calibrate the marginal
distribution and copula function separately. To 7.1 Example 1
verify this conclusion, Table 4 shows the cali- To investigate the effect of probability model on
bration results of the 16 models when marginal failure probability calculation, consider an infinite
distributions and the copula function are cali- slope as shown in Figure 1. Its factor of safety can
brated separately based on Eqs. (11)–(13). For be calculated as follows:
the same model the AIC value in Table 4 is
generally larger than that in Table 3, indicating
c γ H s 2 α tan φ
that the model calibrated ignoring the interac- Fs = (20)
tion between marginal distributions and copula γ H i α cos α
function is less supported by the data.
3. When the copula function is Gaussian, the four where H = thickness of soil layer; α = slope angle;
probability models with different marginal dis- γ = unit weight of the soil. In this example, H = 5 m
tributions have different AIC values, indicat- and γ = 17 kN/m3.
ing that the validity of each probability model Monte Carlo simulation is first used to calcu-
is affected by the marginal distributions. Also, late the failure probability of the slope when dif-
when the marginal distributions are the same, ferent probability models are adopted and as the
the AIC values of the model vary with the slope angle varies. In the Monte Carlo simulation,
copula function, indicating the validity of each 1,000,000 samples are generated. The obtained
model is also affected by the copula function. failure probabilities are summarized in Table 5.

58

ISGSR2013.indb 58 10/18/2013 9:36:57 AM

 For comparison, the failure probabilities calculated from models 1 and 3 are very similar, and also the
using the one-dimensional integration method sug- predictions from models 2 and 4 are very similar, as
gested by Tang et al. (2013) are also shown. We can can be found in Table 5. It seems that the calculated
see that the failure probability calculated using failure probability is more sensitive to the marginal
Monte Carlo simulation and that calculated using distribution of c than that of ϕ. Also, there is a
the one-dimensional integration method is practi- trend that the difference between predicted failure
cally the same, thus verifying the validity of the sug- probability increases as the calculated failure prob-
gested method for failure probability calculation. ability decreases, i.e. the small failure probability
Figures 2(a)–2(d) compare the calculated failure is more sensitive to the selection of probability
probability of the slope calculated when the mar- model. Similar phenomena can also be found in
ginal distributions vary but the copula function is Figures 2(c) and (d). Comparing Figures 2(a)–(d),
fixed. In Figure 2(a), although there are four prob- the effect of marginal distribution is most obvious
ability models involved, only two curves can be in Figure 2(b) for the case of α = 20o. In such a
clearly identified. This is because the predictions case, the failure probability calculated based on
Model 5 is almost four orders of magnitude larger
than that calculated based on Model 6.
Figures 3(a)–(d) compare the calculated failure
probability of the slope when the marginal distri-
butions are the same but the copula function is
different. In Figures 3(a) and (c), the effect of copula
function on failure probability calculation is not very
obvious. However, the copula function could sig-
nificantly affect the calculated failure probability, as
shown in Figures 3(b) and (d). Indeed, in Figure 3(b)
the failure probability calculated based on Model 10
is more than two orders of magnitude larger than
that calculated based on Model 6, implying the cop-
ula function can also significantly affect the failure
probability calculation. Comparing Figures 3(b) and
(d), the effect of copula function is also more obvi-
Figure 1. Infinite slope analyzed in example 1. ous for small failure probability problems.

Table 5. Failure probability calculated based on different probability models (example 1).

pf

α = 30o α = 25o α = 20o

Model Marginal 1-D Monte Carlo 1-D Monte Carlo 1-D Monte Carlo
no. Copula distribution integration simulation integration simulation integration simulation

1 Gaussian n,n 0.119 0.120 0.110 0.109 0.098 0.098

2 l,n 0.038 0.038 0.011 0.011 0.001 0.001
3 n,l 0.119 0.119 0.109 0.109 0.098 0.098
4 l,l 0.039 0.039 0.012 0.012 0.002 0.002
5 Clayton n,n 0.119 0.130 0.108 0.112 0.097 0.108
6 l,n 0.048 0.050 0.012 0.013 9.0 × 10−6 7.0 × 10−6
7 n,l 0.121 0.116 0.110 0.105 0.099 0.094
8 l,l 0.048 0.051 0.012 0.014 7.0 × 10−6 7.0 × 10−6
9 Frank n,n 0.121 0.126 0.110 0.116 0.099 0.104
10 l,n 0.044 0.043 0.016 0.015 0.004 0.004
11 n,l 0.118 0.124 0.108 0.113 0.098 0.102
12 l,l 0.046 0.045 0.017 0.017 0.004 0.004
13 FGM n,n 0.120 0.124 0.109 0.114 0.099 0.103
14 l,n 0.049 0.046 0.018 0.017 0.003 0.003
15 n,l 0.121 0.122 0.099 0.112 0.099 0.101
16 l,l 0.051 0.047 0.020 0.018 0.004 0.003

59

ISGSR2013.indb 59 10/18/2013 9:36:57 AM

 Figure 2. Effect of marginal distributions on failure probability calculation (example 1): (a) models 1, 2, 3, 4;
(b) models 5, 6, 7, 8; (c) models 9, 10, 11, 12; (d) models 13, 14, 15, 16.

Figure 3. Effect of copula function on failure probability calculation (example 1) (a) models 1, 5, 9, 13;
(b) models 2, 6, 10, 14; (c) models 3, 7, 11, 15; (d) models 4, 8, 12, 16.

60

ISGSR2013.indb 60 10/18/2013 9:36:58 AM

 To understand how the marginal distribu- but differ in the copula function. We could see that
tions affect the failure probability calculation, the distribution of samples is also affected by the
Figures 4(a)–(d) compare the distribution of sam- copula function. The limit state functions are all
ples together with the limit state function when close to the edge of the region where simulated
Models 1–4 are employed, respectively. In each samples are concentrated. As shown in Figure 3(b),
figure, there are 1,000,000 simulated samples. For the failure probabilities estimated based on these
comparison, the measured data are also shown four models are quite large, implying the tail of the
in these figures. We can see that the assumption joint distribution is sensitive to the copula function
on the marginal distribution could significantly used.
affect the distribution of samples. For instance, As mentioned previously, the FGM copula with
when cohesion is assumed to be normal, a large c being lognormal and ϕ being lognormal is most
portion of samples will lie in the failure region, supported by the data. Currently, the Gaussian
indicating a relatively larger failure probability. copula is often used in practice. As shown in
However, when cohesion is assumed to be lognor- Table 5, when the marginal distributions of c
mal, the number of samples in the failure region is and ϕ are both lognormal, the failure probability
significantly reduced, making the calculated failure calculated from the Gaussian copula function is
probability small. In this example, the distribution generally larger than that obtained based on the
of samples is very sensitive to the assumed mar- FGM copula function. In such a case, the failure
ginal distributions. probability will be overestimated if the Gaussian
To investigate how the copula function affects copula is used.
the failure probability calculation, Figures 5(a)–(d)
compare the distribution of samples together with
7.2 Example 2
the limit state functions when Models 2, 6, 10 and 14
are used for failure probability calculation. All the In Example 1, the performance function is rela-
four models have the same marginal distributions tively simple, which can be solved by both the one-

Figure 4. Effect of marginal distributions on the distribution of samples (example 1): (a) model 1; (b) model 2;
(c) model 3; (d) model 4.

61

ISGSR2013.indb 61 10/18/2013 9:36:58 AM

 Figure 5. Effect of copula function on the distribution of samples (example 1): (a) model 2; (b) model 6;
(c) model 10; (d) model 14.

dimensional integration method and Monte Carlo

simulation. To illustrate the advantage of the
Monte Carlo simulation algorithm, a strip footing
subjected to a load of q (kPa) resting on a horizon-
tal ground is analyzed (Fig. 6).
The bearing capacity of the strip footing Qu
can be calculated using the following equations
(Cherubini 2000)

Qu 0 B ⋅ Nγ c ⋅ Nc 2 D f ⋅ Nq (21)
Figure 6. Strip footing analyzed in example 2.
2
⎛ ⎛ π φ ⎞⎞
Nq eπ φ
⋅ ⎜ tan ⎜ + ⎟ ⎟ (22)
⎝ ⎝ 4 2⎠⎠
bottom. In this example, Df = 0.5 m, and B = 2 m,
Nγ 1.8 ( Nq − 1) ⋅ tan φ (23) γ1 = γ2 = 19.5 kN/m3.
As the performance function here is complex,
the one-dimensional integration method can
Nq − 1 hardly be applied. However, the Monte Carlo sim-
Nc = (24)
tan φ ulation algorithm is still applicable. Table 6 sum-
marizes the failure probability calculated using
where Df = embedded depth of foundation 16 different probability models (N = 1,000,000).
(m); B = foundation width (m); γ1 = unit weight Similar to Figures 2, 3 and Figures 7, 8 show the
of the soil below the foundation bottom and effect of marginal distribution and copula function
γ2 = unit weight of the soil above the foundation on failure probability calculation in this example,

62

ISGSR2013.indb 62 10/18/2013 9:36:59 AM

 Table 6. Failure probability predicted based on different probability models (example 2).

pf
Model Marginal
no. Copula distribution q = 50 kPa q = 80 kPa q = 110 kPa

1 Gaussian n,n 0.091 0.094 0.096

2 l,n 0 0 2.7 × 10−4
3 n,l 0.091 0.093 0.096
4 l,l 0 0 2.7 × 10−4
5 Clayton n,n 0.100 0.103 0.106
6 l,n 0 0 0
7 n,l 0.087 0.090 0.092
8 l,l 0 0 0
9 Frank n,n 0.096 0.099 0.102
10 l,n 3.0 × 10−5 3.2 × 10−4 0.001
11 n,l 0.094 0.097 0.100
12 l,l 2.0 × 10−5 2.5 × 10−4 0.001
13 FGM n,n 0.094 0.096 0.100
14 l,n 1.0 × 10−5 1.0 × 10−4 8.9 × 10−4
15 n,l 0.093 0.095 0.098
16 l,l 0 1.1 × 10−4 9.2 × 10−4

*0 denotes that no failure sample is found during the conditional sampling with 1,000,000 samples.

Figure 7. Effect of marginal distributions on failure probability calculation (example 2): (a) models 1, 2, 3, 4;
(b) models 5, 6, 7, 8; (c) models 9, 10, 11, 12; (d) models 13, 14, 15, 16. (failure probability of 0 in Table 6 is plotted
as 1.0 × 10−6).

respectively. The phenomena observed in Figures 7 Figures 9(a) and (b) show the calculated fail-
and 8 are generally similar to those observed in ure probabilities based on 16 models for Exam-
Figures 2 and 3, i.e. the effect of marginal distribu- ple 1 and Example 2, respectively. The difference
tions and copula functions on failure probability between calculated failure probabilities based on
calculation is more obvious when the failure prob- the 16 probability models in Example 2 is even
ability is small. larger than that observed in Example 1, indicating

63

ISGSR2013.indb 63 10/18/2013 9:37:01 AM

 Figure 8. Effect of copula function on failure probability calculation (example 2): (a) models 1, 5, 9, 13; (b) models
2, 6, 10, 14; (c) models 3, 7, 11, 15; (d) models 4, 8, 12, 16. (failure probability of 0 in Table 6 is plotted as 1.0 × 10−6).

that the effect of probability model on failure prob-

ability calculation is also affected by the determin-
istic model. In other words, even for the same set
of data and for the same probability models con-
sidered, the effect of probability model on failure
probability calculation changes with the reliability
problem considered.

7.3 Example 3
In the above two examples, the probability mod-
els have important effect on the failure probability.
The third example shows a case where such an
effect is less obvious. In this example, the data are
adopted from Xu & Yang (1998) about the shear
strength parameters of clay in Shanghai, as shown
in Table 7. In this example, there are 81 measure-
ments, compared to only 25 measurements in
Example 1. The same 16 models are considered in
this example, and the calibration results are shown
in Table 8.
Among the 16 models considered, Model 8 has
the smallest AIC value and hence is the optimal
one, in which both c and ϕ are assumed to be log-
normal and the copula function is Clayton. The
optimal model in this example is different from
that found in Example 1, indicating the optimal
probability model is problem specific.
Using the calibrated samples, the failure prob-
ability of the infinite slope in Example 1 is calcu-
lated again. Figure 10 shows the effect of marginal
Figure 9. Failure probabilities calculated based on distribution on failure probability calculation. The
16 models: (a) example 1; (b) example 2; (c) example 3. difference between failure probabilities calculated
(Failure probability of 0 in Table 6 is plotted as 1.0 × 10−6). based on different probability models is generally

64

ISGSR2013.indb 64 10/18/2013 9:37:02 AM

 Table 7. Soil data analyzed in example 3 (Adapted from Xu & Yang 1998).

Test Test Test Test

no. c (kPa) ϕ (ο) no. c (kPa) ϕ (ο) no. c (kPa) ϕ (ο ) no. c (kPa) ϕ (ο)

1 30 15.8 22 35 15.6 43 20 15.1 64 13 13.5

2 14 24.8 23 20 10.8 44 15 14.0 65 15 17.0
3 38 20.9 24 29 19.5 45 27 14.2 66 9 15.0
4 19 18.8 25 25 10.9 46 15 13.2 67 13 13.5
5 43 19.5 26 25 18.8 47 26 14.7 68 7 19.0
6 13 20.6 27 30 15.8 48 11 12.7 69 12 13.5
7 13 24.6 28 28 14.6 49 35 12.4 70 13 14.0
8 24 25.5 29 30 9.6 50 15 20.4 71 14 10.0
9 7 25.9 30 32 15.6 51 33 15.4 72 16 22.0
10 27 19.3 31 30 14.7 52 17 10.9 73 19 15.3
11 26 15.1 32 24 19.5 53 25 17.2 74 8 13.6
12 19 22.3 33 34 21.3 54 34 17.8 75 12 18.5
13 22 19.8 34 29 18.8 55 17 12.2 76 16 14.0
14 13 17.7 35 35 13.3 56 32 14.0 77 13 12.5
15 15 19.5 36 21 13.5 57 20 11.6 78 15 19.2
16 17 16.0 37 30 19.8 58 33 15.6 79 12 17.5
17 20 17.2 38 15 17.2 59 20 15.7 80 10 14.0
18 7 14.8 39 18 14.4 60 21 11.5 81 11 14.0
19 19 16.8 40 16 14.5 61 25 21.5
20 18 10.8 41 25 13.0 62 30 17.4
21 10 21.8 42 17 10.4 63 28 12.4

Table 8. Calibrated parameters of 16 probability models based on the data in Table 7.

Model Marginal
no. Copula distribution AIC θ μc σc μϕ σϕ

1 Gaussian n,n 1028.57 −0.005 20.852 8.405 16.236 3.746

2 l,n 1025.20 −0.040 20.966 9.472 16.235 3.747
3 n,l 1022.83 −0.014 20.852 8.405 16.236 3.767
4 l,l 1019.60 −0.009 20.966 9.472 16.232 3.750
5 Clayton n,n 1028.56 −0.030 20.852 8.401 16.237 3.747
6 l,n 1025.04 −0.084 20.945 9.463 16.231 3.750
7 n,l 1022.81 −0.036 20.844 8.399 16.236 3.767
8 l,l 1019.18 −0.088 20.945 9.463 16.236 3.767
9 Frank n,n 1028.48 0.182 20.875 8.405 16.235 3.743
10 l,n 1025.34 0.092 20.966 9.472 16.231 3.747
11 n,l 1022.71 0.226 20.862 8.404 16.232 3.750
12 l,l 1019.58 0.150 20.975 9.501 16.232 3.750
13 FGM n,n 1028.48 0.101 20.877 8.405 16.234 3.744
14 l,n 1025.33 0.049 20.966 9.472 16.225 3.745
15 n,l 1022.70 0.123 20.866 8.405 16.232 3.750
16 l,l 1019.58 0.094 21.008 9.491 16.216 3.746

no more than two orders of magnitude, which are mum difference between failure probabilities calcu-
less than those observed in Example 1, indicating lated based on different copula functions occurs in
the effect of marginal distribution is less obvious in Figure 11(d), which is less than one order of mag-
this example. Figure 11 shows the effect of copula nitude and is also significantly than that observed
function on failure probability calculation. This is in Example 1. Note the copula function measures
possibly because there are more measurements in the dependence relationship between the cohesion
this example, which provide more constraint on and the friction angle. To find out the reason why
the shape of the marginal distributions. The maxi- the effect of copula is less obvious in example, the

65

ISGSR2013.indb 65 10/18/2013 9:37:02 AM

 Figure 10. Effect of marginal distributions on failure probability calculation (example 3): (a) models 1, 2, 3, 4;
(b) models 5, 6, 7, 8; (c) models 9, 10, 11, 12; (d) models 13, 14, 15, 16.

Figure 11. Effect of copula function on failure probability calculation (example 3) (a) models 1, 5, 9, 13; (b) models
2, 6, 10, 14; (c) models 3, 7, 11, 15; (d) models 4, 8, 12, 16.

66

ISGSR2013.indb 66 10/18/2013 9:37:02 AM

 correlation coefficients between c and ϕ in Exam- effect of probability model on failure probability
ple 1 and Example 3 are both calculated, which estimation is more obvious when the failure prob-
are −0.407 and −0.005, respectively. Although the ability is small. The effect of probability model
correlation coefficient is not an accurate measure on failure probability estimation depends on the
of dependence, it does show that the dependence deterministic model and also the data available for
relationship between c and ϕ is much weaker in constructing the probability models. Increasing the
Example 3. Thus, it is reasonable to expect that amount of measured data may provide more con-
in Example 3 the effect of copula function is less straint to the theoretical model and hence reduce
obvious. the effect of marginal distributions on failure
Figure 9(c) shows the calculated failure prob- probability estimation. As the correlation between
ability based on 16 models, which also shows that the cohesion and friction angle decreases, the effect
the effect of probability model on failure probabil- of copula function on failure probability calcula-
ity calculation is less obvious in this example than tion also decreases.
that in Example 1. As analyzed previously, this is
due to the combined effect of more measured data
and less correlation between c and ϕ. When there ACKNOWLEDGEMENTS
are more data for constructing the probability
model, the probability model may have less effect This research was substantially supported by the
on the failure probability calculation. NSFC/ANR Joint Research Scheme (51161130523),
National Basic Research Program of China (973
Program 2011CB013800) and Shanghai Outstand-
8 SUMMARY AND CONCLUSIONS ing Academic Leaders Program (12XD1405100).
The authors also wish to thank Prof. Dianqing Li
The joint distribution of the cohesion and the fric- from Wuhan University for helpful discussions of
tion angle can be fully specified by the marginal the work presented.
distributions and a copula function. The model
parameters can be calibrated simultaneously by
the maximum likelihood method. Calibrating REFERENCES
the model parameters separately can enhance the
computational efficiency, but may miss the opti- Akaike H. 1974. A new look at the statistical model
mal model parameters. Different models could be identification. IEEE Trans Autom Control 19(6):
ranked based on the information criteria available 761–723.
in statistics which are capable of considering the Ang, A.H-S. & Tang, W.H., 1984. Probability concepts
trade-off between model accuracy and model com- in engineering planning and design: design, risk and
reliability, Vol. 2. Wiley, New York.
plexity. The examples studied in this paper indicate Baecher, G.B. & Christian, J.T. 2003. Reliability and Sta-
that the probability models constructed based tistics in Geotechnical Engineering, Wiley, Chichester,
on the Gaussian copula, which are widely used United Kingdom.
in the literature, may not be the optimal probabil- Cavanaugh, J.E. 1999. A large-sample model selection
ity model. The optimal probability model depends criterion based on Kullback’s symmetric divergence.
on the data available and is problem specific. Statistics & Probability Letters 42(4): 333–343.
Many existing reliability methods such as the Chen, L.H., Chen, Z.Y. & Liu, J.M. 2005. Probability
first order reliability method are developed for distribution of soil strength. Rock and Soil Mechanics
probability models constructed based on the 26(1): 37–40, 45. (In Chinese).
Cherubini, C. 1997. Data and considerations on the varia-
Gaussian copula, and can not be directly applied bility of geotechnical properties of soils. In Proceedings
when other copula functions are used. Monte of the International Conference on Safety and Reliabil-
Carlo simulation, however, are generally applicable, ity (ESREL) 97, Lisbon, Vol. 2, pp. 1583–1591.
even when the performance function is complex. Cherubini, C. 2000. Reliability evaluation of shallow
Monte Carlo simulation can be applied to general foundation bearing capacity on c′, ϕ' soils. Canadian
performance function and hence is superior to the Geotechnical Journal 37: 264–269.
existing one-dimensional integration method when Cherubini, U., Luciano, E. & Vecchiato,W. 2004. Copula
the performance function is complex or when there Methods in Finance. Wiley Finance Series. Wiley,
are several pairs of cohesion and friction angle in Chichester.
Ching, J.Y. & Phoon, K.K. 2012. Establishment of generic
the geotechnical reliability analysis. transformations for geotechnical design parameters.
Both the marginal distribution and copula func- Structural Safety 35: 52–62.
tion could significantly affect the failure probabil- Duncan, J.M. 2000. Factors of safety and reliability in
ity calculation. The failure probabilities calculated geotechnical engineering. ASCE Journal of Geo-
based on different probability models could vary technical and Geoenvironmental Engineering 126(4):
by several orders of magnitude. In general, the 307–316.

67

ISGSR2013.indb 67 10/18/2013 9:37:03 AM

 Durrleman, V., Nikeghbali, A., & Roncalli, T. 2000. Sivakumar Babu, G.L., Srivastava, A. & Murthy, D.S.N.
Which copula is the right one? Groupe de Recherche 2006. Reliability analysis of the bearing capacity of a
Opérationnelle, France: Crédit Lyonnais Working shallow foundation resting on cohesive soil. Canadian
paper. Geotechnical Journal 43: 217–223.
Fenton, G.A. & Griffiths, D.V. 2008. Risk Assessment in Sklar, A. 1959. Fonctions de répartition à n dimensions
Geotechnical Engineering. John Wiley & Sons, New et leurs marges. Publications de l’Institut de Statis-
York. tique de l’Université de Paris 8, pp. 229–231.
Goda K. 2010. Statistical modeling of joint probability Tang, W.H. 1979. Probabilistic evaluation of penetra-
distribution using copula: Application to peak and tion resistances. Journal of Geotechnical Engineering
permanent displacement seismic demands. Structural 105(10): 1173–1191.
Safety 32(2): 112–123. Tang, X.S., Li, D.Q., Rong, G., Phoon, K.K. &
Harr, M.E. 1987. Reliability-based design in civil engi- Zhou, C.B. 2013. Impact of copula selection on geo-
neering. McGraw-Hill Book Company, New York. technical reliability under incomplete probability
Honjo. Y. & Amatya, S. 2005. Partial factors calibration information. Computers and Geotechnics 49: 264–278.
based on reliability analyses for square footings on Trivedi, P.K. & Zimmer, D.M. 2005. Copula Modeling:
granular soils. Géotechnique, 55(6): 479–491. An Introduction for Practitioners. Foundations and
Huang, J., Griffiths, D.V. & Fenton, G.A. 2010. System Trends in Econometrics 1(1): 1–111.
reliability of slopes by RFEM. Soils and Foundations Uzielli, M., Vannucchi, G. & Phoon, K.K. 2005. Random
50(3): 343–353. field characterisation of stress-normalised cone pen-
Joe, H. & Xu, J.J. 1996. The estimation method of infer- etration testing parameters. Géotechnique 55(1): 3–20.
ence functions for margins for multivariate mod- Vanmarcke, E.H. 1977. Probabilistic modeling of soil
els. Department of Statistics. University of British profiles. ASCE Journal of Geotechnical Engineering
Columbia, Technical Report, 166. 103(GT11): 1227–1246.
Juang, C.H. & Wang, L. 2013. Reliability-based robust Wang, Y. 2011. Reliability-based design of spread foun-
geotechnical design of spread foundations using dations by Monte Carlo simulations. Géotechnique
multi-objective genetic algorithm. Computers and 61(8): 677–685.
Goetechnics 48: 96–103. Whitman, R.V. 2000. Organizing and evaluating uncer-
Lebrun R. & Dutfoy A. 2009. A generalization of the tainty in geotechnical engineering. ASCE Journal of
Nataf transformation to distributions with ellipti- Geotechnical Engineering 125(6): 583–593.
cal copula. Probabilistic Engineering Mechanics 2009 Wolff, T.H. 1985. Analysis and design of embankment
24(2):172–178. dam slopes: a probabilistic approach. Ph.D. thesis,
Li, D.Q., Tang, X.S., Phoon, K.K., Chen, Y.F. & Purdue University, Lafayette, Ind.
Zhou, C.B. 2013. Bivariate simulation using copula Xu, C. & Yang, L.D. 1998. Test of goodness of fit of
and its application to probabilistic pile settlement random variables and Bayesian estimation of distri-
analysis. International Journal for Numerical and Ana- bution parameters. Journal of Tongji University 26(3):
lytical Methods in Geomechanics 37(6): 597–617. 340–344. (In Chinese).
Low, B.K. & Tang, W.H. 1997. Efficient reliability Xue, J.F. & Gavin, K. 2007. Simultaneous determination
evaluation using spreadsheet. Journal of Engineering of critical slip surface and reliability index for slopes,
Mechanics 123(7): 749–752. ASCE Journal of Geotechnical and Geoenvironmental
Lumb, P. 1970. Safety factors and the probability distri- Engineering 133(7): 878–886.
bution of soil strength. Canadian Geotechnical Journal Zhang, J., Zhang, L.M. &Tang, W.H. 2011. Kriging
7: 225–242. numerical models for geotechnical reliability analysis.
Luo, Z., Atamturktur, H.S., Juang, C.H., Huang, H. & Soils and Foundations 51(6): 1169–1177.
Lin, P.S. 2011. Probability of serviceability failure in Zhang, L.L., Zuo, Z.B., Ye, G.L., Jeng, D.S. & Wang, J.H.
a braced excavation in a spatially random field: Fuzzy 2013. Probabilistic parameter estimation and pre-
finite element approach. Computers and Geotechnics dictive uncertainty based on field measurements for
38: 1031–1040. unsaturated soil slope. Computers and Geotechnics 48:
Nelsen R.B. 2006. An introduction to copulas. 2nd ed. 72–81.
Springer: New York. Zhang, L.M. 2004. Reliability verification using proof
Schwarz, G. 1978. Estimating the dimension of a model. pile load tests. Journal of Geotechnical and Geoenvi-
Ann. Statist. 6: 461–464. ronmental Engineering 130(11): 1203–1213.

68

ISGSR2013.indb 68 10/18/2013 9:37:03 AM

 Geotechnical Safety and Risk IV – Zhang et al. (eds)
© 2014 Taylor & Francis Group, London, ISBN 978-1-138-00163-3

Robust design of geotechnical systems—a new design perspective

C. Hsein Juang, Lei Wang & Sez Atamturktur

Glenn Department of Civil Engineering, Clemson University, Clemson, SC, USA

ABSTRACT: This paper presents a new geotechnical design concept, called Robust Geotechnical Design
(RGD). The new design methodology seeks to achieve design robustness and economics, in addition to
meeting safety requirements. Here, a design is considered robust if the variation in the system response
(i.e., failure probability) is insensitive to the statistical characterization of noise factors, such as uncertain
geotechnical parameters. It can be shown that when safety requirements are met, cost and robustness are
conflicting objectives with a trade-off, meaning that a single best design may be unattainable. In such a
case, a multi-objective optimization considering cost and robustness can be performed to obtain a suite of
best designs within the solution space in which safety is guaranteed. To this end, the suite of best designs
forms the Pareto Front, which is a useful tool to aid in the design decision making within the RGD meth-
odology. The new design methodology is illustrated with examples of spread foundation designs and rock
slope designs, through which the significance of the RGD methodology is demonstrated.

1 INTRODUCTION well the random soil parameters are characterized.

If the knowledge of the statistical distribution of
It is well recognized that uncertainty in geotechnical soil parameters is “perfect,” the results of reliabil-
parameters is usually unavoidable in the geotechni- ity analysis will be sufficiently accurate and the
cal design (Whitman 2000). The uncertainty in the reliability-based design can be easily implemented
geotechnical parameters, as well as the uncertainty by selecting the least cost design constrained with
in the adopted geotechnical model, can lead to the the failure probability requirement.
uncertainty in the predicted response of a system. In a typical project in geotechnical practice, the
In a deterministic approach, the engineer uses statistical characterization of geotechnical param-
an experience-calibrated factor of safety to cope eters is uncertain due to lack of data, measurement
with the uncertainty in the predicted response. Of error, transformation error, etc. The variation
course, the factor of safety adopted in a particular range of geotechnical parameters is usually consid-
design depends not only on the degree of uncer- erably large (Harr 1987; Phoon & Kulhawy 1999
tainties but also on the consequence of failure; in a & b; Lee et al. 2012) and thus, the variation can
other words, it depends on the “calculated risk” be either overestimated or underestimated. Such
(Casagrande 1965). To account for the uncertainty overestimation or underestimation of the variation
in geotechnical parameters explicitly, the probabil- of geotechnical parameters can lead to over-design
istic or reliability-based approaches are often used or under-design (Juang and Wang 2013).
(e.g., Harr 1987; Wu et al. 1989; Tang & Gilbert While reduction of the uncertainty in geotech-
1993; Christian et al. 1994; Lacasse & Nadim 1996; nical parameters is important, which should be
Duncan 2000; Phoon et al. 2003; Chalermyanont pursued whenever it is deemed cost-effective, in
& Benson 2004; Fenton & Griffiths 2008; Najjar & this paper, we focus on a different approach by
Gilbert 2009; Wang et al. 2011; Zhang et al. 2011). achieving robustness in the design without elimi-
In a traditional reliability-based geotechnical nating the sources of uncertainty. Here, a design
design process, multiple candidate designs are first is considered robust if the variation in the system
checked against safety requirements (in terms of response is insensitive to (or robust against) the
probability of failure), and then the acceptable variation of uncertain geotechnical parameters
designs are optimized for cost, which yields the (called noise factors in this paper). The essence of
final design. The reliability-based design is quite a robust design is to select a design that yields a
straightforward if the results of the reliability minimal variation in the system response without
analysis are accurate and precise so that there will eliminating the sources of uncertainty or reducing
be no question whether a given design satisfies the the level of uncertainty.
safety requirement. The accuracy and precision In this paper, a Robust Geotechnical Design
of a reliability analysis, however, depends on how (RGD) methodology is proposed to fulfill the goal

69

ISGSR2013.indb 69 10/18/2013 9:37:03 AM

 of minimizing the effects of the uncertainty of soil
parameters. Robust design concept is first proposed
by Taguchi (1986) for improving product quality
and reliability in Industry Engineering. Early appli-
cations of robust design are closely related to prod-
uct and mechanical design to avoid the effects of
the uncertainty from environmental and operating
conditions (Taguchi 1986; Phadke 1989). The more
recent applications are found in various fields such
as mechanical, structural and aeronautical design
(e.g., Chen et al. 1996; Chen et al. 1999; Lee & Park
2001; Doltsinis et al. 2005; Kang 2005; Park et al.
2006; Brik et al. 2007; Lagaros & Fragiadakis 2007;
Kumar et al. 2008; Marano et al. 2008; Lee et al.
2010). Robust design attempts to adjust design
parameters (i.e., the so-called “easy to control”
factors) so that the system response of the design
is insensitive to, or robust against, the variation of
noise factors (i.e., “hard to control” factors). In a
geotechnical design, the noise factors mainly refer
to the uncertain geotechnical parameters, although
model bias may also be considered. Thus, in a robust
design, regions in the solution space that yield low
variation in the system response are sought.
It should be noted that Robust Geotechnical
Design (RGD) is a design strategy to complement
the traditional methods. With the RGD approach,
the focus is to satisfy three design requirements,
namely safety, cost, and robustness (against the
variation in system response caused by noise
factors). The safety requirement is usually imple- Figure 1. Flowchart for robust geotechnical design
mented through constraints of reliability, and (modified after Juang et al. 2013a).
hence the design becomes a bi-objective optimiza-
tion problem. For the bi-objective problem exam-
ined in this paper, it is very likely that no single best first established, which can be in the form of either
solution exists that is most optimal with respect to an analytical model or a numerical model. For all
both objectives (cost and robustness). In such a sit- the input parameters of the adopted geotechnical
uation, a detailed study of the trade-offs between model, the design parameters and the noise factors
these objectives can lead to a more informed design are then classified.
decision. Step 2 is to characterize the uncertainty in statistics
In this paper, the RGD approach is demon- of noise factors and specify the design domain. For
strated with two design examples: one on spread the design of geotechnical systems, the key uncertain
foundation designs and the other on rock slope geotechnical parameters are usually identified as
designs. The significance and versatility of the noise factors. The uncertainty in the statistics (e.g.,
RGD approach is demonstrated through these coefficient of variation) of the noise factors may be
examples. estimated based on published literature guided by
engineering judgment or the bootstrapping method
based on limited data. The design domain (space) of
2 ROBUST GEOTECHNICAL DESIGN the design parameters are prescribed based on their
METHODOLOGY typical ranges, augmented with local experiences.
The design parameters are usually specified in dis-
In reference to Figure 1, the Robust Geotechnical crete space for construction convenience. Thus, the
Design (RGD) methodology conceptually consists design space will consist of a finite number (M) of
of five steps outlined below (Juang et al. 2013a): designs.
Step 1 is to define the geotechnical design Step 3 is to assess the variation of the sys-
problem of interest. For a given problem, the tem response as a measure of robustness for a
deterministic model (limit state or performance given design. In this paper, the variation of sys-
function) of the intended geotechnical system is tem response (e.g., failure probability) caused by

70

ISGSR2013.indb 70 10/18/2013 9:37:03 AM

 uncertain statistics of noise factors is used as a
measure of robustness, and a smaller variation
of system response indicates a greater robust-
ness. The variation of the failure probability is
determined using the advanced Point Estimate
Method (PEM) developed by Zhao & Ono (2000)
based on the estimated variation in statistics of
noise factors. The PEM evaluates the failure
probability for a given design at selected values
(say N sets of values) of statistical parameters
of noise factors, and then the mean and stand-
ard deviation of the failure probability can be
computed based on the PEM formulation. For
a given set of statistical parameters, the failure
probability is determined using the traditional
reliability method, such as the first-order reliabil- Figure 2. Illustration of Pareto Front (modified after
Wang et al. 2013).
ity method (FORM; Ang & Tang 1984). Thus, in
this step, for a given design in the design space,
the mean and standard deviation in the failure
In this paper, the Non-dominated Sorting
probability is computed by PEM integrated with
Genetic Algorithm version II (NSGA-II), devel-
FORM procedure, as represented by the inner
oped by Deb et al. (2002) is used to establish the
loop shown in Figure 1.
Pareto Front. The obtained Pareto Front can
Step 4 is to repeat the procedure in Step 3 for
then be used for as a design aid for choosing the
each design in the design space. The mean and
most preferred design when a specific cost range
standard deviation of the failure probability for all
or a target robustness level is selected (Juang et al.
the M designs in the design space are obtained by
2013a&b).
M repetitions of the PEM integrated with FORM
It should be noted that when the number of
procedure in Step 3, as represented by the outer
candidate designs in the design space is large, the
loop shown in Figure 1.
process for evaluating the mean and standard
Step 5 is a key step in RGD for choosing the
deviation of failure probabilities for all designs
most preferred design. A multi-objective optimiza-
may be computationally prohibitive. In such cases,
tion is carried out considering three distinct crite-
NSGA-II can be directly applied with PEM and
ria including safety, cost and robustness. Through
FORM as its integrated components. In this sce-
the multi-objective optimization, a Pareto Front is
nario, the NSGA-II will automatically search for
established, which serves as a guide for selecting
the “non-dominated designs” and only a limited
the most preferred design.
number of selected designs in the design space that
In the implementation of optimization, the
has potential to be on the Pareto Front require the
mean failure probability is set as the safety con-
full PEM and FORM procedures, which makes it
straint to screen out the unsatisfactory designs,
more efficient.
and cost (in terms of construction cost for the geo-
technical system) and robustness (in term of stand-
ard deviation of the failure probability) are set as
3 RGD APPLICATION I: DESIGN
two objectives for optimization.
OF SPREAD FOUNDATION
Note that the multi-objective optimization sel-
dom leads to a single best solution (or design).
3.1 Illustrative example
With conflicting objectives, it leads to a set of
non-dominated solutions that collectively form a A design example of a spread foundation sub-
Pareto Front. For a minimization problem with jected to the combined axial compressive loads
two objectives, the solution A dominates solution and moment loads is first used to demonstrate the
B when no objective value of B is less than A, and application of the proposed RGD approach. The
at least one objective value of B is strictly greater axial loads are applied at the center of foundation
than A (Ghosh & Dehuri 2004). By screening with a permanent load component G of 900 kN
designs or solutions in the solution space and elim- and a random variable load component of Q with
inating the solutions that are dominated by others, a mean of 458.7 kN and a Coefficient of Varia-
a set of solutions that are non-dominated by any tion (COV) of 0.15. The moment load is applied
other solution can be obtained, which collectively at the center of foundation in the direction of the
form a Pareto Front as shown in Figure 2 (Wang length of foundation with a deterministic value
et al. 2013). M = 500 kN ⋅ m.

71

ISGSR2013.indb 71 10/18/2013 9:37:04 AM

 The spread foundation is to be installed in a 17%; COV[mv] is assumed to have a mean of 0.30
stiff till with a deterministic total unit weight of and a COV of 17%.
22 kN/m3, a mean effective friction angle φ' of The robustness of a reliability-based design is
36.4° (c' = 0), a mean undrained shear strength achieved if the probability of failure of a design is
cu of 235.3 kPa and a mean coefficient of volume insensitive to the variation of the estimated COVs
compressibility mv of 0.01875 m2/MN. The unit of the input soil parameters. For a given set of
weight of concrete foundation is 24 kN/m3. The fixed COVs of soil parameters (φ', cu and mv in this
foundation is founded at just above the ground example), the failure probability determined from
water table with a foundation depth of D = 0.8 m. FORM procedure (Hasofer & Lind 1974) will be a
The deterministic model for computing the ULS fixed value. If the COVs of these soil parameters
capacity of spread foundation subjected to both are treated as random variables, the resulting fail-
vertical and moment loads is the “effective area” ure probability will also be a random value. Thus,
method (Meyerhoff 1953). For the Serviceability the variation of the failure probability has to be
Limit State (SLS) capacity, the deterministic model evaluated.
defined in Eurocode 7 is adopted in this study
(Orr & Farrell 1999; Orr & Breysse 2008). The
3.3 RGD of spread foundation
ULS failure is said to occur when the computed
bearing capacity is smaller than the total combined In this example, the footing width B and footing
loads. The SLS failure is said to occur if the com- length L are considered as design parameters,
puted total settlement exceeds the maximum allow- the design domain of which should be specified.
able settlement of 25 mm. Here, B and L are assumed to vary from a mini-
mum value of 1.0 m to a maximum value of 4.0
m, with an increment of 0.1 m for construction
3.2 Uncertainty modeling
convenience. Thus, the number of possible designs
For the spread foundation design in a stiff till, the in the design pool (or solution space) is 961. For
uncertain soil parameters are considered as noise a typical rectangular footing, the range of the
factors, including φ', cu and mv. length-to-width ratio (L/B) is between 1 and 10
In geotechnical practice, the sample size (i.e., (Akbas 2007). When screened with this geometry
number of specimens tested) of the geotechnical constraint for the length-to-width ratio, the design
parameters is usually small due to the budget con- pool is reduced to 496 designs.
straints, so the statistics of geotechnical param- For the safety requirements, the reliability
eters derived from a few testing results may requirements defined in Eurocode 7 are adopted
not reflect the statistics of the real geotechnical for this foundation design (Wang 2011). The tar-
properties. Generally, the “population” mean get failure probability for ULS is set as 0.000072
can be adequately estimated based on the sample (corresponding to a reliability index of 3.8) and
mean even with a small sample size (Wu et al. the target failure probability for SLS is set as 0.067
1989). However, the estimation of variation of (corresponding to a reliability index of 1.5). A sen-
“population” is likely to be subjected to error sitivity study indicates that the ULS failure proba-
when the sample size is small. With insufficient bility requirement is always more stringent than the
data, the variation of soil parameters in terms SLS failure probability requirement and the ULS
of Coefficient of Variation (COV) is usually esti- controls the design of spread foundations, which
mated based on the prior knowledge from local is consistent with those reported by other investi-
experiences or published literature (e.g., Phoon & gators (e.g., Wang & Kulhawy 2008; Wang 2011).
Kulhawy 1999a). For example, Orr and Breysse Indeed, in all analyses performed in this study, the
(2008) provided typical ranges for COV of three ULS requirement is more stringent than the SLS
soil parameters φ', cu and mv for this spread requirement for this shallow foundation problem.
foundation example. The COV of φ', denoted as Therefore, in the subsequent analysis only the ULS
COV[φ'], typically ranges from 4% to 11%; the failure probability is considered.
COV of cu, denoted as COV[cu], typically ranges Following the flowchart of the RGD methodol-
from 20% to 40%; and the COV of mv, denoted ogy presented in Figure 1, the mean and standard
as COV[mv], typically ranges from 20% to 40%. deviation of the ULS failure probability, denoted
Thus, the COVs of φ', cu and mv should be treated as μp and σp, can be obtained for all designs in the
as random variables. design pool using PEM integrated with FORM
As an example to illustrate the robustness con- procedure. The cost for each design in the design
cept presented herein, COV[φ'] is assumed to have a pool can be determined using the cost estima-
mean of 0.08 and a COV of 25% (roughly to cover tion procedure for spread foundation proposed
the typical range of COV[φ']). Similarly, COV[mv] by Wang & Kulhawy (2008). The cost of spread
is assumed to have a mean of 0.30 and a COV of foundation construction includes the costs for

72

ISGSR2013.indb 72 10/18/2013 9:37:04 AM

 excavation, formwork, concrete, reinforcement
and compacted backfill, respectively based on U.S.
average unit price for construction of spread foun-
dation (Wang & Kulhawy 2008). In this paper, the
total cost for spread foundation is therefore a func-
tion of design parameters B and L.
Then the multi-objective optimization using
NSGA-II may be achieved considering safety, cost
and robustness in the design. Safety is generally
assured through a constraint that the computed
mean failure probability is less than the target fail-
ure probability. Thus, the design can be optimized
with two objectives, robustness and cost, subjected
to the safety constraint and any additional con- Figure 4. Pareto Front for design of spread
straint such as geometry. This optimization scheme foundation.
can be set up as shown in Figure 3.
The design parameters (B and L in this case)
are generated in the discrete space using NSGA-II.
The population size is arbitrarily set at 100 indi- 1700 USD, the design with the least standard
viduals. For this example, a converged Pareto deviation of failure probability (meaning greatest
Front is obtained at 30th generation, meaning that robustness) within this budget constraint has the
Pareto Front has already reached a “stable” condi- design parameters of B = 2.2 m and L = 2.7 m with
tion and negligible improvement can be made at a construction cost of 1694.6 USD.
further generations. For this spread foundation Although the Pareto Front with a specified trade-
design, 80 “unique” designs are selected into the off relationship offers a valuable tool for decision-
final Pareto Front, as shown in Figure 4. Thus, making, it may be desirable to further refine the
the RGD approach yields a set of optimal designs, decision-making use a more user-friendly index for
called Pareto Front, which describes an obvious robustness assessment. The feasibility robustness
trade-off relationship between cost and robust- is defined as the confidence probability that the
ness. The designer can select a design with greater actual failure probability satisfies the target fail-
robustness (i.e., the failure probability is insensitive ure probability in the face of uncertainty, which is
to the statistical characterization of soil param- expressed as follows (Parkinson et al. 1993; Juang
eters) only at the expense of a higher cost. The et al. 2013a):
least cost design on the Pareto Front is design with
B = 2.1 m and L = 2.5 m with a construction cost Pr[( p f − pT ) [( βT − β )
] = Pr[( ] = Φ(ββ ) > P0
of 1522.2 USD. (1)
The obtained Pareto Front can then be used as
a design aid in decision making for choosing the where pf is the computed failure probability, which
“best” design based on the target cost or robust- is a random variable affected by uncertainty in sta-
ness level. For example, if the allowable budget tistical characterization of noise factors; pT is the
for the spread foundation construction is set at target failure probability; Pr[(pf – pT) < 0] is the con-
fidence probability that the requirement of target
failure probability is satisfied; and P0 is an accept-
able level of the confidence probability specified by
the designer. Thus, the feasibility robustness index
ββ can be used as a measure for feasibility robust-
ness, which corresponds to different confidence
probability P0 that the actual failure probability
meets the requirement of target failure probability
under the uncertainty in statistical characteriza-
tion of noise factors.
The feasibility robustness index ββ for each of
80 designs on the Pareto Front of Figure 4 can
be computed, the results of which are shown in
Figure 5. As expected, a design with higher feasi-
bility robustness (higher ββ) costs more. By select-
Figure 3. Formulation of multi-objective optimization ing a desired target feasibility robustness level, the
for design of spread foundation. least-cost design among all on the Pareto Front can

73

ISGSR2013.indb 73 10/18/2013 9:37:04 AM

 assumed as 2.6 ton/m3. For shear strength proper-
ties of rock discontinuity (along the failure plane),
mean cohesion c is 10 ton/m2 and mean friction
angle φ is 35°.
A deterministic model with single failure mode
developed by Hoek & Bray (1981) is employed
herein for case study of Sau Mau Ping slope in
Hong Kong. The model is a two dimensional limit
equilibrium analytical model for plane failure, in
which the rock slope is assumed with a 1-meter
thick slice through slope. The factor of safety is
expressed as follows (Hoek & Bray 1981; Hoek
2006):
Figure 5. Cost versus feasibility robustness for design cA + [W ( ψ α i ψ ) U − V sin
i ψ ]]tan φ
of spread foundation. FS =
W (si ψ α sψ ) V cosψ
(2)
readily be identified. For example, when the feasi-
bility robustness level is set at ββ = 1, which corre- where FS denotes the factor of safety for rock slope;
sponds to a confidence probability of 84.13%, the c is the cohesion of the rock discontinuity (or joint
least-cost design is B = 2.2 m and L = 2.5 m. The surface) (ton/m2); A is the base area of the wedge
cost for this design is 1583.9 USD. When the feasi- (m); W is the weight of rock wedge resting on the
bility robustness level is set at ββ = 2, the least-cost failure surface (°); ψ is angle of failure surface
design is B = 2.5 m and L = 3.0 m, which corre- measured from horizontal level (°); α is ratio of
sponds to a confidence probability of 97.72%. The horizontal earthquake acceleration to gravitational
cost for this design is 2077.1 USD. The feasibil- acceleration; U is the uplift force due to water pres-
ity robustness Pareto Front offers an easy-to-use sure on the slip surface (ton); V is horizontal force
measure for a more informed decision-making due to water in tension crack (ton); φ is the fric-
considering cost and robustness after satisfying the tion angle of the rock discontinuity (°). Additional
safety requirement. details for computing the factor of safety FS can
be found in Hoek (2006). The slope failure is said
to occur if the factor of safety computed from a
4 RGD APPLICATION III: DESIGN deterministic model (such as Eq. 2) is less than 1.
OF ROCK SLOPE In this example, the slope height H and the slope
angle θ are considered as design parameters for the
The example presented in this section is a sum- rock slope, typical ranges for which should be spec-
mary of the prior work by the authors (Wang et al. ified based on the design situation in hand.
2013), and the reader is referred to this manuscript
for additional details.
4.2 Uncertainty modeling
In a routine practice of rock slope design, only a
4.1 Illustrative example
small sample of data will be available for determin-
The design of rock slope is used to further dem- ing shear properties of the rock discontinuities.
onstrate the proposed RGD methodology. The The mean value of these rock properties can usu-
case history of Sau Mau Ping rock slope in Hong ally be adequately estimated even with a small size
Kong is used to illustrate the RGD of the rock of data. However in often cases, the Coefficients
slope. The rock mass of the Sau Mau Ping slope of Variation (COV) of uncertain rock proper-
is un-weathered granite with sheet joints. The sheet ties and the coefficients of correlation (ρ) among
joints are formed by the exfoliation processes dur- these properties can only be estimated with engi-
ing cooling of granite. Initial study by Hoek (2006) neering judgment, guided by published literature
led to a simplification of Sau Mau Ping slope as (for example, see Hoek 2006 and Lee et al. 2012).
a slope composed by a single unstable block with Hence, these estimates are subjected to error.
a water-filled tension crack, which involves only a In the rock slope design discussed herein, cohe-
single failure mode. Following Hoek (2006), the sion and friction angle of rock discontinuity are
slope before remediation has a height H of 60 m treated as noise factors. Both parameters are
and a slope angle θ of 50°. The potential failure assumed to follow normal distribution as per Hoek
plane is inclined at 35°. The unit weight of rock is (2006). The COVs of c and φ, denoted as COV[c]

74

ISGSR2013.indb 74 10/18/2013 9:37:05 AM

 and COV[φ], as well as the correlation coefficient computed for all possible designs in the design space
(ρ) of c and φ, denoted as ρc,φ, are treated herein using PEM integrated with FORM procedure.
as random variables. In addition, the horizontal Then, a multi-objective optimization approach
distance to tension crack b is assumed to follow can be adopted for identifying the designs opti-
normal distribution with a mean of 15.3 m and a mal to both objectives, robustness (minimizing the
COV of 21%. standard deviation of the failure probability) and
According to Low (2007), the ratio of the depth cost (minimizing the construction cost), while sub-
of water in the tension crack (zw) over the depth of jected to the safety constraint that μp is less than
tension crack (z), denoted as iw, is assumed as an the target failure probability. The scheme of this
exponential distribution with a mean of 0.5, trun- optimization is set up as shown in Figure 6.
cated to [0, 1]. Besides, the ratio of the horizon- The multi-objective optimization is performed
tal earthquake acceleration over the gravitational here with the NSGA-II algorithm, in which the
acceleration, denoted as α, is assumed as an expo- designs optimal to both objectives (cost and robust-
nential distribution with a mean of 0.08, truncated ness) are searched iteratively in the discrete design
to [0, 0.16]. As noted by Low (2007), b is likely to space. With the NSGA-II algorithm, 89 “unique”
have a positive correlation with iw. Thus, in addi- designs are selected into the final Pareto Front as
tion to the variation in rock properties, the correla- shown in Figure 7.
tion of b and iw, denoted as ρb,iw, is also considered In Figure 7, an apparent trade-off relationship
as a random variable. between cost and robustness is evident. In this case,
Based on the typical ranges for the above uncer- greater design robustness can only be attained at
tain statistical parameters reported in literature the expense of a higher cost. The trade-off rela-
(e.g., Low 2007; Lee et al. 2012), COV[c] is assumed tionship in Figure 7 can be used as a design guide
to have a mean of 0.20 and a COV of 17%; COV[φ] for choosing the preferred design when a desired
is assumed to have a mean of 0.14 and a COV of
12%; ρc,φ is assumed to have a mean of −0.50 and
a COV of 25%; ρb,iw is assumed to have a mean of
0.50 and a COV of 25%.

4.3 RGD of rock slope

As noted previously, the slope height H and the
slope angle θ are treated as design parameters and
the design space should first be specified for rock
slope design. The slope height H may typically
range from of 50 m to 60 m, and slope angle θ may
typically range from 44° to 50°. For convenience
in construction, slope height H may be rounded to
nearest 0.2 m and slope angle θ may be rounded to
nearest 0.2°. Thus, H can take 51 discrete values in Figure 6. Formulation of multi-objective optimization
its typical range, while and θ can take 31 discrete for design of rock slope.
values. Thus, 1581 designs (pairs of H and θ) are
possible, which define the entire design space.
For the safety consideration, the rock slope is
designed to meet a target reliability index βT of 2.5
(Low 2008), which corresponds to a target failure
probability pT of 0.0062. A simple cost estimation
method is used in this paper for illustration of the
RGD of rock slope. The cost of a given rock slope
design is simplified as the volume of rock mass
that needs to be excavated (Duzgun et al. 1995). It
should be noted that cost estimation for the rock
slope design is not the focus of this paper, and the
proposed RGD method is not dependent on any
particular cost estimation method. In fact, any rea-
sonable cost estimation method can be adopted.
Following the RGD procedure outlined in
Figure 1, the mean and standard deviation of the Figure 7. Pareto Front at convergence for design of
failure probability, denoted as μp and σp, can be rock slope.

75

ISGSR2013.indb 75 10/18/2013 9:37:05 AM

 Finally, the RGD methodology presented in this
paper is equally applicable to other geotechnical
design problems such as drilled shafts and braced
excavations, and the reader is referred to Juang
et al. (2013a&b) for additional details.

5 CONCLUDING REMARKS

In this paper, a new design methodology, called

Robust Geotechnical Design (RGD), is presented
and illustrated with applications in several geotechni-
cal problems including design of spread foundations
and rock slopes. In a geotechnical reliability-based
Figure 8. Cost versus robustness for design of rock design, the statistics of the noise factors, including
slope.
the Coefficient of Variation (COVs) of the noise
factors (mainly referring to uncertain geotechni-
cal properties herein) and the coefficients of cor-
cost level or a robustness level is specified. For relation between noise factors, are often difficult to
example, when the maximum acceptable cost for ascertain. When these statistics are overestimated or
the designer is 200 units [note: 1 unit = unit cost ($) underestimated, the design obtained from the tradi-
of excavated volume of rock mass in m3/m], then tional reliability based design can be cost-inefficient
the design with the least σp while in the acceptable or unsafe. The proposed RGD approach aims at
cost range will be the best design. In this scenario, reducing the effect of uncertainties associated with
it is a design with H = 50 m and θ = 50°. the estimated statistics of noise factors by carefully
Furthermore, for each of the 89 designs on adjusting the design parameters. Within the RGD
Pareto Front, the feasibility robustness index ββ framework, a multi-objective optimization is per-
can be determined. The ββ value for each of the formed to identify designs optimal in both cost and
89 designs in the Pareto Front is plotted against robustness, while satisfying the safety requirements.
the corresponding cost as shown in Figure 8. As Through this optimization, a Pareto Front is derived,
expected, a design with higher feasibility robustness which typically describes a trade-off relationship
has a higher cost. When a desired feasibility robust- between cost and robustness at a given safety level.
ness level (in terms of ββ) is selected, the least-cost This trade-off relationship offers a design aid that
design among those on the Pareto Front can read- can assist in selecting the most suitable design. The
ily be identified. The feasibility robustness can be effectiveness and significance of the proposed RGD
used as a valuable design aid for decision making. framework is demonstrated through two design
For example, when the target feasibility robustness examples presented in this paper.
level is set at ββ = 1, which corresponds to a confi-
dence probability of 84.13%, the least-cost design
is H = 52.8 m, θ = 50° with a cost of 132.0 units. ACKNOWLEDGMENTS
Similarly, when the target feasibility robustness
level is set at ββ = 2, which corresponds to a confi- The study on which this paper is based has been
dence probability of 97.72%, the least-cost design is supported in part by National Science Founda-
H = 50.4 m and θ = 48.4° with a cost of 247.6 units. tion through Grant CMMI-1200117 and the
The feasibility robustness provides an easy-to-use Glenn Department of Civil Engineering, Clemson
quantitative measure for more informed decision University. The results and opinions expressed in
making from designs on the Pareto Front. this paper do not necessarily reflect the view and
It should be noted that in this paper, the robust- policies of the National Science Foundation.
ness within the proposed RGD framework is meas-
ured by the variation of the system response (in
terms of standard deviation of failure probability). REFERENCES
Of course, this is not the only measure for design
Akbas, S.O. 2007. Deterministic and probabilistic
robustness; alternative measures for robustness
assessment of settlements of shallow foundations in
have been reported in the fields of structural and cohesionless soils. Ph.D. thesis, Cornell Univ., Ithaca,
industry engineering. For example, measures such New York.
as signal-to-noise ratio (Phadke 1989), vulner- Ang, A.H.-S. & Tang, W.H. 1984. Probability concepts
ability function (Brik et al. 2007), and reliability in engineering planning and design: decision, risk, and
sensitivity (Zhang et al. 2005) have been reported. reliability. Vol. 2. New York: Wiley.

76

ISGSR2013.indb 76 10/18/2013 9:37:05 AM

 Brik, B.A., Ghanmi, S., Bouhaddi, N. & Cogan, S. 2007. Juang, C.H., Wang, L., Hsieh, H.S., & Atamturktur, S.
Robust design in structural mechanics. International 2013b. Robust geotechnical design of braced excava-
Journal for Computational Methods in Engineering tions in clays. Structural Safety, doi:http://dx.doi.
Science and Mechanics 8 (1): 39–49. org/10.1016/j.strusafe.2013.05.003.
Casagrande A. 1965. The role of the “calculated risk” in Kang, Z. 2005. Robust design optimization of struc-
earthwork and foundation engineering. Journal of the tures under uncertainties. Aachen, Germany: Shaker
Soil Mechanics and Foundations Division 91 (4):1–40. Verlag.
Chalermyanont, T. & Benson, C. 2004. Reliability-based Kumar, A., Nair, P.B., Keane, A.J. & Shahpar, S. 2008.
design for internal stability of mechanically stabilized Robust design using bayesian monte carlo. Interna-
earth walls. Journal of Geotechnical and Geoenviron- tional Journal for Numerical Methods in Engineering,
mental Engineering 130 (2): 163–173. 73: 1497–1517.
Chen, W., Allen, J.K., Mistree, F. & Tsui, K.L. 1996. A Lacasse, S. & Nadim, F. 1996. Uncertainties in charac-
procedure for robust design: minimizing variations terizing soil properties. In ASCE Uncertainties’96
caused by noise factors and control factors. Journal of Conference Proceedings, C.H. Benson, Ed., Madison,
Mechanical Design 118 (4): 478–485. WI, pp. 49–75.
Chen, W., Wiecek, M.M. & Zhang, J. 1999. Quality utility: Lagaros, N.D. & Fragiadakis, M. 2007. Robust perform-
a compromise programming approach to robust design. ance based design optimization of steel moment
Journal of Mechanical Design 121 (2): 179–187. resisting frames. Journal of Earthquake Engineering 11
Christian, J.T., Ladd, C.C. & Baecher, G.B. 1994. (5): 752–772.
Reliability applied to slope stability analysis. Journal Lee, K.H. & Park, G.J. 2001. Robust optimization con-
of Geotechnical Engineering 120 (12): 2180–2207. sidering tolerances of design variables. Computers and
Deb, K., Pratap, A., Agarwal, S. & Meyarivan, T. 2002. Structures 79 (1): 77–86.
A fast and elitist multi-objective genetic algorithm: Lee, M.C.W., Mikulik, Z., Kelly, D.W., Thomson, R.S.
NSGA-II. IEEE Transactions on Evolutionary & Degenhardt, R. 2010. Robust design-a concept
Computation 6 (2): 182–197. for imperfection insensitive composite structures.
Doltsinis, I., Kang, Z. & Cheng, G. (2005). Robust design Composite Structures 92 (6): 1469–1477.
on non-linear structures using optimization methods. Low, B.K. 2007. Reliability analysis of rock slopes
Computer Methods in Applied Mechanics and Engi- involving correlated nonnormals. International
neering 194: 1779–1795. Journal of Rock Mechanics and Mining Sciences
Duncan, M.J. 2000. Factors of safety and reliability in 44 (6): 922–935.
geotechnical engineering. Journal of Geotechnical and Low, B.K. 2008. Efficient probabilistic algorithm illus-
Geoenvironmental Engineering 126 (4): 307–316. trated for a rock slope. Rock Mechanics and Rock
Duzgun, H.S.B., Pasamehmetoglu, A.G. & Yucemen, M.S. Engineering 41 (5): 715–734.
1995. Plane failure analysis of rock slopes: a reliabil- Marano, G.C., Sgobba, S., Greco, R. & Mezzina, M.
ity approach. International Journal of Surface Mining, 2008. Robust optimum design of tuned mass damp-
Reclamation and Environment 9 (1): 1–6. ers devices in random vibrations mitigation. Journal
Fenton, G.A. & Griffiths, D.V. 2008. Risk assessment in of Sound and Vibration 313 (3–5): 472–492.
geotechnical engineering. New York: Wiley. Meyerhof, G.G. 1953. The bearing capacity of
Ghosh, A. & Dehuri, S. 2004. Evolutionary algorithms foundations under eccentric and inclined loads. In
for multi-criterion optimization: a survey. International Proceedings, Third International Conference on Soil
Journal of Computing and Information Sciences 2 (1): Mechanics and Foundation Engineering, Zurich, Vol.1,
38–57. pp. 440–445.
Harr, M.E. 1987. Probability-based design in civil Najjar, S.S. & Gilbert, R.B. 2009. Importance of
engineering. New York: McGraw-Hill Book Company. lower-bound capacities in the design of deep founda-
Hasofer, A.M., & Lind, N.C. 1974. Exact and invariant tions. Journal of Geotechnical and Geoenvironmental
second-moment code format. Journal of the Engineer- Engineering 135 (7): 890–900.
ing Mechanics Division 100 (1): 111–121. Orr, T.L.L. & Farrell, E.R. 1999. Geotechnical design to
Hoek, E. (2006). Practical rock engineering. Eurocode 7. Berlin: Springer.
Chapter 7: A slope stability problem in Hong Kong; Orr, T.L.L. & Breysse, D. 2008. Eurocode 7 and
and Chapter 8: Factor or safety and probability of reliability-based design. In: Reliability-based design in
failure. In: http:// www.rocscience.com/hoek/Practical geotechnical engineering: computations and applications
RockEngineering.asp. (ed. Phoon KK), ch. 8, pp. 298–343. London: Taylor
Hoek, E. & Bray, J. 1981. Rock slope engineering. 3rd ed. & Francis.
London: Institution of Mining and Metallurgy. Park, G.J., Lee, T.H., Lee, K.H. & Hwang, K.H. 2006.
Juang, C.H. & Wang, L. 2013. Reliability-based robust Robust design: an overview. AIAA Journal 44 (1):
geotechnical design of spread foundations using 181–191.
multi-objective genetic algorithm. Computers and Parkinson, A., Sorensen, C. & Pourhassan, N. 1993.
Geotechnics 48: 96–106. A general approach for robust optimal design. Journal
Juang, C.H., Wang, L., Liu, Z., Ravichandran, N., of Mechanical Design 115 (1): 74–80.
Huang, H. & Zhang, J. 2013a. Robust Geotechnical Phadke, M.S. 1989. Quality engineering using robust
Design of drilled shafts in sand—A new design per- design. Englewood Cliffs: Prentice Hall.
spective, Journal of Geotechnical and Geoenviromental Phoon K.K. & Kulhawy, F.H. 1999a. Characterization
Engineering, doi: http://dx.doi.org/10.1061/(ASCE) of geotechnical variability. Canadian Geotechnical
GT.1943-5606.0000956. Journal 36 (4): 612–624.

77

ISGSR2013.indb 77 10/18/2013 9:37:06 AM

 Phoon, K.K. & Kulhawy, F.H. 1999b. Evaluation Wang, Y., Au, S.K. & Kulhawy, F.H. 2011. Expanded
of geotechnical property variability. Canadian reliability-based design approach for drilled shafts.
Geotechnical Journal 36 (4): 625–639. Journal of Geotechnical and Geoenvironmental
Phoon, K.K., Kulhawy, F.H. & Grigoriu, M.D. 2003. Engineering 137 (2): 140–149.
Development of a reliability-based design framework Whitman, R.V. 2000. Organizing and evaluating
for transmission line structure foundations. Journal of uncertainty in geotechnical engineering. Journal
Geotechnical and Geoenvironmental Engineering 129 of Geotechnical and Geoenvironmental Engineering
(9): 798–806. 126 (7): 583–593.
Taguchi, G. 1986. Introduction to quality engineering: Wu, T.H., Tang, W.H., Sangrey, D.A. & Baecher, G.B.
designing quality into products and processes. New 1989. Reliability of offshore foundations—State-of-
York: Quality Resources. the-art. Journal of Geotechnical Engineering 115 (2):
Tang, W.H. & Gilbert, R.B. 1993. Case study of offshore 157–178.
pile system reliability. Proc., 25th Offshore Technology Zhang, J., Zhang, L.M. & Tang, W.H. 2011.
Conference, Society of Petroleum Engineers, Houston, Reliability-based optimization of geotechnical
Texas, pp. 677–686. systems. Journal of Geotechnical and Geoenvironmental
Wang, L., Hwang, J.H., Juang, C.H. & Atamturktur, S. Engineering 137 (12): 1211–1221.
2013. Reliability-based design of rock slopes—a new Zhang, Y., He, X., Liu, Q. & Wen, B. 2005. Robust
perspective on design robustness. Engineering Geology reliability design of Banjo flange with arbitrary
154: 56–63. distribution parameters. Journal of Pressure Vessel
Wang, Y. & Kulhawy, F.H. 2008. Economic design opti- Technology 127 (4): 408–413.
mization of foundations. Journal of Geotechnical and Zhao, Y.G. & Ono, T. 2000. New point estimates
Geoenvironmental Engineering 134 (8): 1097–1105. for probability moments. Journal of Engineering
Wang, Y. 2011. Reliability-based design of spread foun- Mechanics 126 (4): 433–436.
dations by Monte Carlo Simulations. Géotechnique
61 (8): 677–685.

78

ISGSR2013.indb 78 10/18/2013 9:37:06 AM

 Geotechnical Safety and Risk IV – Zhang et al. (eds)
© 2014 Taylor & Francis Group, London, ISBN 978-1-138-00163-3

Integrated geo risk management: Crossing boundaries

M.Th. van Staveren

VSRM Research and Development, Breda, The Netherlands

ABSTRACT: Any engineering and construction project involves dealing with risk, which has often a
geotechnical origin. Therefore, the objective of applying Geotechnical Risk Management (GeoRM) is
contributing to the engineering, construction, and maintenance of successful projects, despite the inher-
ently uncertain character of its ground conditions. However, for realizing these projects in an effective
and cost-efficient way, GeoRM has to be well-integrated within Project Risk Management (ProjectRM).
As ISSMGE TC304 Task Force 3 revealed, a complete GeoRM application that is integrated with Pro-
jectRM, seems not yet common practice in a lot of project phases, in a lot of projects, in a lot of coun-
tries. This requires crossing often hidden boundaries between risk concepts, disciplines, project phases,
scopes, project types and even boundaries between industries and countries. Finally and perhaps first
of all, we have to cross our own mental boundaries. Within this paper these boundaries are highlighted
and crossed, for catching benefits of learning, effectiveness, and cost-efficiency. The promise of well-
integrated GeoRM in ProjectRM, by crossing the mentioned boundaries, is substantially contributing to
successful construction projects from a geotechnical engineering and construction perspective, with fewer
cost overruns, less delays, less non-conformances of safety and quality, and therefore happier project
participants and stakeholders.

1 INTRODUCTION years, the project circumstances may change dra-

matically. Therefore, effectively dealing with risks
1.1 Risk is pervasive and underlying uncertainties will make the dif-
ference between project success and failure for all
Beck (1992) classifies our modern societies as
parties involved. These include clients, government
risk societies. As mentioned by van Staveren and
agencies, engineering firms, main contractors, sub-
Halman (2013), over the last 25 years several risk
contractors, suppliers, end-users, the public and
research topics were awarded with Nobel Prices,
sometimes social pressure groups.
including Markowitz and Sharpe in 1990 for their
project portfolio theory, Nash in 1994 for his game
theory, Kahneman and Tversky in 2002 for their
1.2 Risk management as solution
prospect theory, and rather recently Sargent and
Sims in 2011 for their cause and effect theory. Well-structured and explicitly dealing with risk,
Especially civil engineering and construction including its key element of risk analysis, is by
projects are inherently risky, due to numerous many researchers and practitioners seen as a mean
uncertainties of many types. Examples are political for effectively and cost-efficiently dealing with
uncertainties, financial uncertainties, market uncer- project uncertainties (e.g. Boothroyd & Emmet
tainties, stakeholder uncertainties, technical uncer- 1996, Chapman & Ward 2011, Edwards 1995,
tainties and, not to forget, ground uncertainties. Edwards & Bowen 2005, Flanagan and Norman
The resulting challenges, such as ever-increasing 1993, Godfrey 1996, Smith 2003, Thompson &
operational complexity and serious safety and Perry 1992, and Weatherhead et al. 2005).
integrity concerns are widely acknowledged (e.g. In many sectors and organizations several types
Bea 2006, Beal 2007, Blockley & Godfrey 2000, of risk management are already in use, includ-
Brandl 2004, Schinzinger & Martin 2000, Sowers ing enterprise risk management, project risk
1993). management, and discipline-based risk manage-
Each type of construction project, either large ment. Examples of this latter type are financial
or small, involves several phases of feasibility, risk management, IT risk management, and also
design, contracting, construction, and finally geotechnical risk management. Several research
maintenance. Often, it takes a lot of years before a results confirm the anticipated effects and benefits
substantial project is actually realized. Over those of risk management. One example is provided by

79

ISGSR2013.indb 79 10/18/2013 9:37:06 AM

 Raz et al. (2002), who discovered a statistically and ground conditions in the overall risk profile of
significant correlation between the application civil engineering and construction projects.
of several project risk management practices and A probably well-known example is that of a
success in meeting schedule and budget objectives. contractual risk that is caused by the occurrence of
Another example gives Cooke-Davies (2002), who differing ground conditions during construction.
derived a correlation between on-time perform- This may generate an expensive and time consum-
ance and organizational risk management charac- ing contractual dispute, due to an apparent poor
teristics, such as company-wide risk management site investigation program in the early project
education, processes for assigning risk ownership, phase. These differing ground conditions can have
using a risk register, and keeping an up-to-date risk several manifestations. Examples are harder rock
management plan. than expected during excavation, which requires
the mobilization of additional equipment to the
site, softer soil layers, which require unforeseen,
1.3 Or failure of risk management application?
expensive, and time consuming ground improve-
Nevertheless, despite these promising research ment, or much more groundwater inflow during
results, this available risk management knowledge tunnel boring activities that demands for addi-
apparently does not prevent a lack of effective risk tional grouting measures. All of these situations
management application. For instance, a longitu- will create additional costs and delay within the
dinal project management analysis over the years project, with one major question: which party is
1998–2003, which has been initiated by the Project going to bear these unfavourable effects, by paying
Management Institute (PMI), shows that organi- the costs? Is it the client, the contractor, the gov-
zations consistently fail to apply risk management ernment or perhaps an insurance company?
across projects (Mullaly, 2006). Researchers and
practitioners, such as Hillson (2002), Hillson and
1.5 Integrated risk management: A better
Murray-Webster (2007), Olsson (2006), and Power
solution?
(2009) confirm similar failures of risk management
application and its effectiveness. This paper aims to take a new risk manage-
This application gap may be a root cause of ment development route, that of integrated risk
failure in a lot of (construction) projects, which management. In particular, this paper will focus
seems to be in line with the assessments of failure on the integration of Geotechnical Risk Man-
costs in the construction industry, typically ten to agement (GeoRM) in Project Risk Management
thirty percent of the total construction costs, as for (ProjectRM), in all types of civil engineering and
instance raised by Barber et al. (2000), Egan (1988), construction projects in the international construc-
and Latham (1994). In conclusion, many managers tion industry. Integrated risk management aims to
and engineers in the construction industry would combine and synthesize these different types of risk
agree that risk management is not always produc- management, in order to minimize the adverse effects
ing the expected results. Or positively stated: there of uncertainties and maximize their potential ben-
are lots of opportunity for further risk manage- efits, for all project participants and stakeholders.
ment development and implementation within the As the Dutch Geo-Impuls program on reduc-
construction industry. ing geotechnical failures teaches, the integration
of GeoRM in ProjectRM seems to be promising.
In this development program over 40 Dutch
1.4 The geotechnical root cause of project risk
organizations, including ministries, city coun-
As mentioned before, many of these adverse effects cils, other public clients, contractors, engineering
of risks and uncertainties in civil engineering and consultants, universities, and research institutes
construction projects have geotechnical risk drivers are involved. Within the Geo-Impuls, over
with causes in the subsoil. Examples are unfore- 200 professionals and managers work together in
seen weak, hard, or organic soil layers, variations 12 Working Groups on the development and imple-
of groundwater tables, boulders, collapsible soils mentation of GeoRM within (project) organiza-
or sinkholes in calcareous rock. These unforeseen tions (Cools 2011, van Staveren et al. 2013a).
differing ground conditions can have a serious Also within the International Society of
impact on the project. For instance, according to Soil Mechanics and Geotechnical Engineering,
the US National Committee on Tunnelling Tech- Technical Committee 403, Task Force 3 (TC304-
nology, poor site investigations result in different TF3), this risk management integration topic has
site conditions claims averaging 28% of the entire been worked out. Geotechnical representatives
contract price (Smith, 1996). A recent survey of ten countries reported the degree of applica-
including ten countries (van Staveren 2013a) once tion and integration of ProjectRM and GeoRM
again confirmed the serious stake of the subsoil in their countries. The participating countries

80

ISGSR2013.indb 80 10/18/2013 9:37:06 AM

 are, in alphabetical order, Austria, China, Czech project types and even boundaries between indus-
Republic, Finland, Germany, Japan, Netherlands, tries and countries. But perhaps we have to cross
Sweden, Switzerland and United Kingdom. first another one: our own mental boundary.
The segregation of Geotechnical Risk Manage- This paper ends with conclusions and recom-
ment (GeoRM) and Project Risk Management mendations for further and deeper integrating
(ProjectRM) within project organizations is likely GeoRM and ProjectRM. This aims to support
to obstruct the catchment of all potential risk civil engineering and construction projects, in each
management benefits, in terms of improved safety project phase, for serving all project participants
and quality, cost and time savings and strengthen- and stakeholders.
ing of reputations. In addition, potential synergies
of scale and learning of risk management remain
unused. This resulted in the following TC304-TF3
2 RISK: WHAT ARE WE TALKING
research focus:
ABOUT?
Contributing to the integration of GeoRM in
2.1 Introduction
ProjectRM, by evaluating existing international
knowledge and lessons. Due to their rather abstract character, the terms
risk and risk management have a lot of mean-
By applying this international knowledge and ings and definitions. For instance, the words
lessons, geotechnical professionals and project uncertainty and risk are often intermingled (Van
managers, contract managers, safety managers, Asselt & Vos 2006, van Staveren 2009b), as are
and quality managers worldwide are facilitated by the words risk and hazard (Waring & Glendon
jointly managing serious project risks with geo- 1998).
technical risk drivers, in a cost-efficient way. In Also the ten countries participating in the
the resulting TC304-TF3 State of the Art Report TC304-TF3 research reveal that there is not yet
(van Staveren 2013a), the status of ProjectRM and uniform and widely accepted risk terminology
GeoRM application and integration has been ana- (van Staveren 2013a). Regarding definitions for
lysed and summarized. These valuable results are project risk management (ProjectRM), in the
included in this paper. majority of seven out of the ten countries (70%),
no specific and widely accepted ProjectRM defi-
nitions are in use. In the remaining three coun-
1.6 Paper set up
tries ProjectRM definitions are slightly different,
This necessarily rather extensive introduction with also different origins. China reports their
resulted in the following paper set up. First, the own ProjectRM definition, Netherlands refers to
main types of risk and risk management are pre- the RISMAN definition (van Well-Stam 2004),
sented in some more detail and with tight defini- and Sweden uses the definition of ISO 73:2009
tions. These definitions are essential for the rather (ISO 2009b).
abstract concept of risk and its management, of Within the TC304-TF3 research, the same situa-
which each individual has his or her own, often tion has been observed for definitions of Geotech-
hidden, perceptions. nical Risk Management (GeoRM). In the majority
A shared language between users of project risk of eight out of the ten participating countries (80%)
management and geotechnical risk management, no specific and unified GeoRM definitions are in
with similar definitions of uncertainty, risk, haz- use. In Sweden GeoRM is not specifically defined,
ard, risk analysis and risk management proves to because it is considered an overall part of risk
be essential for integrating geotechnical risk man- management. The remaining two countries, China
agement in project risk management. In particu- and Netherlands, do have their slightly different
lar the positioning of (geotechnical) risk analysis GeoRM definitions.
within (geotechnical) risk management, as well as Therefore, a starting point for effective risk man-
its difference with risk management, needs to be agement in general, and for effective ProjectRM
clarified. and GeoRM in particular, is to provide clarity
Next, four distinguished degrees of GeoRM about the core definitions. When aiming to inte-
and Project RM application and integration are grate GeoRM in more generic ProjectRM, which is
worked out in more detail. This provides the foun- the purpose of this paper, the specific geotechnical
dation for the next section on exploring and cross- parts of the common risk terminology should be
ing GeoRM boundaries, for providing effective and clarified as well. For this reason, pragmatic defini-
cost-efficient risk management. These boundaries, tions five sets of risk-related terminology are pre-
which seem to be often quite hidden, are between sented. Generic risk terms are presented together
risk concepts, disciplines, project phases, scopes, with their geotechnical counterparts, in order to

81

ISGSR2013.indb 81 10/18/2013 9:37:06 AM

 create a common, uniform, and shared risk man- 2.5 Geotechnical risk
agement language:
By using the mentioned ISO 31000 risk defini-
1. Project uncertainty & geotechnical uncertainty tion, geotechnical risk can be simply defined as
2. Project risk & geotechnical risk the effect of geotechnical uncertainty on realiz-
3. Hazard & geotechnical hazard ing objectives, such as settlements or horizontal
4. Risk analysis & geotechnical risk analysis deformations within pre-set limits. A geotechni-
5. Risk management, ProjectRM & GeoRM. cal risk has a probability of occurrence and one
or more geotechnical causes. Moreover, and very
This should minimize the usual confusion
important for non-geotechnical professionals
when talking about and dealing with risk and is
and managers, if occurring, geotechnical risks
management. The following definitions follow
have usually a number of unfavourable effects.
as much as possible the KISS approach: Keep
These stretch far beyond geotechnical engineer-
It Simple, Stupid. The presented definitions are
ing into the arena of project management, such
selected from the abundant definitions, as found in
as physical damage to properties, injured persons
the literature, and proved their value in the prac-
or even dead victims, cost overruns and delay.
tice of geotechnical engineering and construction
By also including the positive side of risk, geo-
in The Netherlands (van Staveren 2013b).
technical uncertainty may also cause favourable
effects. Examples are the presence of less settle-
2.2 Uncertainty ment sensitive ground layers than expected below
a road foundation under construction, or a lower
According to the pragmatic approach of Ward
ground water inflow than foreseen in the design
and Chapman (2011), uncertainty is simple a lack
phase at the executing a deep excavation. These
of certainty. Within construction projects there is
in fact positive risks or opportunities may result
a lack of certainty in many ways, such as funding
for instance in cost savings for the contractor,
uncertainty in the feasibility phase of a project,
and, depending on the type of contract, also for
price uncertainty of construction materials during
the client. Obviously, these benefits can only be
construction, and uncertainty about the required
materialized if the project is flexible enough to
degree of maintenance in after realization of for
absorb these opportunities in the construction
instance a tunnel or bridge.
phase.

2.3 Geotechnical uncertainty

2.6 Hazard
Building forward on the pragmatic uncer-
tainty definition by Ward and Chapman (2011), When considering the provided definitions on
geotechnical uncertainty is simple a lack of uncertainty, risk and geotechnical risk, what could
geotechnical certainty. It may be helpful for geo- be the added value of a definition for the widely
technical engineering purposes to distinguish used term hazard? Remarkably, while the word
between four types of uncertainty: (1) ran- hazard is deep-lee embedded in our daily language,
domness, (2) fuzziness, (3) incompleteness and it seems not explacidly defined in a lot of stand-
(4) incorrectness (Blockley & Godfrey 2000, van ard literature on risk and its management, includ-
Staveren 2006). ing Bernstein (1996), Ewards & Bowen (2005),
Chapman and Ward (2011), and Hillson (2010).
Amongst the many available definitions, hazards
2.4 Risk can be defined as threats to people and the things
Amongst the many definitions of risk, the one they value (Van Staveren 2006). Therefore, hazards
by the ISO 31000 Risk Management Guideline and risks seem closely coupled terms with a more
(ISO 2009a) proves to be useful in practice: risk or less similar meaning. This seems consistent with
is the effect of uncertainty on realizing objectives. the definition in the on-line version of the Oxford
Within this definition, the effect of uncertainty Dictionary, which defines hazard as danger or risk.
can be negative, which is the traditional mean- Therefore, for the sake of clarity, hazard is here
ing of risk. However, its neutral approach allows defined simply as risk.
also a positive meaning, being the positive effect
of uncertainty or in other words an opportunity.
2.7 Geotechnical hazard
By this ISO definition, fruitless debates about dif-
ferences and similarities between uncertainty and Given the above discussion about a workable defi-
risk, initiated already by Knight (1921) and for nition of a hazard, the most pragmatic definition
instance critically judged by Hubbard (2009), can of a geotechnical hazard seems to be that of a geo-
be omitted. technical risk, as previously defined.

82

ISGSR2013.indb 82 10/18/2013 9:37:06 AM

 2.8 Risk analysis Apart from the ISO 31000 Risk Management,
well-known generic frameworks are for instance
Risk analysis is another term that is widely used in
the COSO Enterprise Risk Management and the
our day to day communication. Not seldom how-
frame-work of the Institute of Risk Management
ever, risk analysis and risk management are inter-
(IRM). Nevertheless, most frameworks are more
mingled (van Staveren 2013b). ISO31000 (ISO
or less similar, with often somewhat different ter-
2009) defines risk analysis as a process in order to
minology. The ISO31000 (ISO 2009a) defines risk
understand the characteristics of a risk to determine
management simply as the coordinated activities to
the risk level. The latter risk level, or risk profile, is
direct and control an organization with regard to its
the assessment of the probability of a risk, together
risks. This straightforward risk management defi-
with its expected consequences. Ideally, these risk
nition will be used in this paper.
consequences, or risk effects, are for construction
projects expressed in key performance indicators,
such as cost, time, quality, safety and reputation. 2.11 Project Risk Management (ProjectRM)
Any risk analysis should include at least risk
Well-known frameworks for project risk manage-
identification and risk classification. Risk clas-
ment are for instance those from the Project Man-
sification, often also called risk assessment, can
agement Institute (PMI) and the International
be subdivided in a qualitative and a quantitative
Project Management Association (IPMA). In the
approach. The latter one is usually preferred by
Netherlands the RISMAN method is widely used
engineers, because it allows them to calculate prob-
in construction projects (van Well-Stam et al. 2004).
abilities and effects.
Within this method, in each project phase the fol-
However, not seldom there are not sufficient
lowing subsequent steps need to be performed at
reliable data available to allow calculations with a
least once:
statistical acceptable accuracy. In these cases nec-
essarily qualitative assessments have to be made, 1. Setting project objectives: what information?
for instance based on expert judgment. Obviously, 2. Identifying risks: what risks?
these qualitative assessments can be transformed 3. Classifying risks: how serious?
into numbers by using expert judgment, for calcu- 4. Remediation risks: what actions?
lation purposes. However, sensitivity analyses are 5. Evaluating risks: effective remediation?
of utmost importance in these cases, to catch an 6. Reporting risks: to the next project phase!
understanding of the effects of the uncertain input
Most risk management literature provides risk
data on calculated results.
management processes with distinct risk manage-
Finally, it is of utmost importance to understand
ment steps that are more or less equal to these six
that risk analysis and risk assessment are an essen-
steps while the names of the steps may be slightly
tial part of, yet not similar to, risk management.
different (van Staveren 2013b). Therefore, are there
any relevant differences between risk management
2.9 Geotechnical risk analysis and project risk management? This question was
By building on the foregoing definition, a geotech- raised during a project risk management lecture
nical risk analysis can be defined as the process in that the author of this paper provided.
order to understand the characteristics of geotechni- A literature scan learned that the majority of
cal risk to determine the project risk level. As men- the literature considers project risk management
tioned before, geotechnical risks, such as excessive implicitly as being risk management for projects
settlements or even collapse of ground embank- (Chapman and Ward 2011, Edwards & Bowen
ments or bored tunnels can have very severe effects 2005, van Well-Stam et al. 2004, Hopkinson 2011).
on the project manager’s key performance indica- Some literature does explicitly distinguish project
tors costs, time, safety, quality, as well as reputa- risk management, from risk management, such as
tion of all parties involved in the project. For this Hopkin (2012).
reason, the outcomes of geotechnical risk analyses However, some specific points of attention for
should be compared with those of the usually more project risk management surfaced, because projects
generic project risk management analyses. This is are one-off and always new to some extent. Also,
already one main reason for integrating GeoRM projects have a start and an end with several project
within ProjectRM, however without losing the spe- phases, while the decisions in earlier phases have
cific degree of detail of geotechnical risk analyses. an irreversible effect on the later phases. Moreover,
many projects have a variety of stakeholders, often
with different and even conflicting interests.
2.10 Risk Management (RM)
Finally, projects need to be realized within the
Worldwide, there are many Risk Management (RM) pre-set limitations of budget and time, with usually
definitions and related frameworks available. specified criteria for safety and quality. All these

83

ISGSR2013.indb 83 10/18/2013 9:37:06 AM

 project characteristics give a specific flavour to risk In summary, geotechnical risk management,
management within projects, when compared to GeoRM, is an explicitly structured and
enterprise-wide risk management in organizations. well-communicated way of dealing with geotechni-
Nevertheless, the correct answer remains to be cal risk that is executed in all construction project
no, there are no significant methodological dif- phases, in order to achieve project objectives effec-
ferences identified between risk management and tively and cost-efficiently. Herewith, GeoRM is in
project risk management. Therefore, the ISO31000 fact a new name for the GeoQ process for geotech-
(ISO 2009a) definition will be used as well for nical risk management (van Staveren 2006).
defining for project risk management: the coordi- The process of geotechnical risk management is
nated activities to direct and control a project organ- similar to the process of project risk management or
ization with regard to its project risks. even enterprise-wide risk management. It involves
the same sequence of risk management steps as
presented for ProjectRM. Therefore, GeoRM fits
2.12 Geotechnical Risk Management (GeoRM)
perfectly well in any sort of ProjectRM. The only
Finally we arrive at geotechnical risk management. main difference is that GeoRM is a more detailed
First, what is it not? GeoRM is not about guar- and in-depth geotechnical approach of project
anteeing with 100% certainty that all geotechni- risk management, for giving geotechnical risk the
cal risks are eliminated. Nor is it about predicting attention it requires in all phases of engineering
the future. Its theory is also not difficult, and the and construction projects (van Staveren 2013b).
GeoRM application is not necessarily expensive,
when executed properly in accordance with the
geotechnical risk profile and geotechnical risk 3 INTEGRATING GeoRM & ProjectRM
acceptance within a project. Furthermore, while it
may certainly be valuable in particular complicated 3.1 Introduction
construction projects, probabilistic calculations or
finite element geotechnical modelling is by far not While considering integrated GeoRM, what is
automatically always required for adequate geo- integration anyway? According to conventional
technical risk management (Van Staveren 2013b). dictionaries and several definitions on the Inter-
So what is geotechnical risk management? Simi- net, the verb integration can be defined as forming,
lar to all the previously presented terminology, the coordinating, or blending elements into a function-
literature presents many definitions, and different ing or unified whole. Related meanings are uniting
people often have a different view on it. For the with something else and incorporating into a larger
sake of simplicity, geotechnical risk management unit. This latter definition seems appropriate for
is defined here as explicitly communicating about integrating GeoRM within ProjectRM.
continuously dealing with geotechnical risk, in a As raised earlier, lot of serious project risks
structured manner. have their origin in ground or groundwater condi-
Explicitly means completely describing all of tions and can be adequately controlled by sound
the aspects of geotechnical risks and underlying geotechnical risk management. However, in the
uncertainties. Communicating means that geotech- day-to-day practice of construction projects the
nical risks are discussed with and reported to the two complementary types of risk management,
persons and parties involved, in order to share dif- GeoRM and ProjectRM, are either (1) not applied
ferent perceptions of the causes, probabilities, and at all, (2) quite separately applied, (3) somewhat
effects of risks. Dealing indicates that risk manage- blurred into each other, or (4) fully integrated. In
ment is not only about reducing or eliminating risk. the next sections these four options are worked out
Also, opportunities may emerge from uncertainty. in some more detail, supported by the results of
Basically, it is about making choices regarding the the TC304-TF3 research.
best way to handle a geotechnical uncertainty in
view of the project objective. This can also mean
3.2 No GeoRM & ProjectRM applied
simply accepting a risk. Continuously means
that risk management steps are cyclic and there- In the introduction has been discussed that while
fore repeated after a predefined time of project risk management is promising in supporting the
phase. This is necessary because of the inherently completion of successful civil engineering and con-
dynamic character of risks. Due to the fact that struction projects, this promise is by far not always
circumstances and perceptions change, risks will realized. Substantial failure costs in the construc-
change over time as well. Finally, structured indi- tion industry demonstrate this situation. One of
cates the importance of using of a defined process the causes may be that risk management, either
of discrete steps, which are presented in the next ProjectRM or GeoRM, is not applied at all in the
section. subsequent phases of construction projects.

84

ISGSR2013.indb 84 10/18/2013 9:37:06 AM

 A real-life project that seems to demonstrate hazard would mean stopping of the works, which
no application of GeoRM and ProjectRM is the motivated all involved parties including construc-
Heathrow Express Line in London. In October tion companies to hiding potential problems. Also
1994, three tunnels of this project collapsed over in the Netherlands applying GeoRM in projects is
a period of three days. There were surprisingly no relatively new.
injuries. Nevertheless, the tunnel collapses gave In Japan it becomes a common view that it is
widespread disruption at Heathrow airport. More- important to study geotechnical risks during the
over, it seriously delayed the project, as well as the construction stage of projects, especially those
Jubilee extension project, on which the same New risks that could lead to some serious troubles or
Austrian Tunnelling Method (NATM) was used. accidents.
According to Clayton (2008), investigation of the In the United Kingdom GeoRM is not applied
collapses revealed a lack of risk awareness within to all projects. It is a generalization, but the evi-
the project. dence suggests that the smaller the financial value
The British newspaper The Guardian published of a project the less likely that GeoRM will be
an article with the title “Managers ‘overlooked applied. Also the Geo-Impuls participants in
risk’ in airport tunnel collapse” (Harper 2000). the Netherlands notice this point regularly. At
According to this article, a report by the health the smaller projects, or in the smaller work pack-
and safety executive concluded that warning signs ages of larger projects, often executed by sub-
had not been recognised. The problem would not contractors like piling firms or dewatering firms,
have occurred if safety had been put before results. the application of ProjectRM and GeoRM seems
According to the executive’s chief inspector of quite underdeveloped.
construction: “The collapses could have been pre- The situation of no application, and therefore
vented but for a cultural mindset which focused also no integration, of ProjectRM and GeoRM
attention on the apparent economies and the need is symbolized in Figure 1 by the empty and sepa-
for production rather than the particular risks.” rated circles, with GeoRM and ProjectRM put in
The direct cause of the collapses was a chain of between brackets.
events, which started with substandard construc-
tion, inadequate repairs, and the construction of
3.3 Separated GeoRM and ProjectRM
another tunnel in failing ground. The collapses
application
have therefore been classified as an organisational
incident. Regarding the impact on the project par- The second option of GeoRM and ProjectRM
ties, the principal contractor was fined £1,2 million application is that both are applied, however still
after pleading guilty to offences under the Health in a separated way. Examples of such a separated
and Safety Act. Austrian advisers got a fine of application are separated risk files for geotechnical
£0.5 million (Harper 2000). The costs of recovery risks and project risks, no communication between
were £150 million, with a project delay of 6 months the project risk manager and geotechnical engi-
(Clayton 2008). neers, geotechnical risk remediation measures that
In the TC304-TF3 research Austria, Czech are not related to quality control, and so on.
Republic, Germany, Switzerland and United A real-life example of separated risk files
Kingdom reported explicitly that no standardized have been encountered on a 0.5 billion euro tun-
ProjectRM is applied in all construction projects. nel project in a city in the Netherlands. The city
Similar findings were reported about GeoRM. The government asked a Review Board to execute an
following examples illustrate this situation. independent risk identification and classification,
Austria, Germany, Switzerland report that in order to reveal the main risks with an impact on
theoretically GeoRM should be applied in every safety and hindrance for the inhabitants of the city.
construction project. Norm DIN4020 (DIN, 2010)
describes the requirements on sharing the geo-
technical risk in construction projects. The risk
of unexpected geotechnical aspects is born by the
owner (Sondermann and Trunk 2008). In prac-
tice however, GeoRM has not been applied in all
projects. The share of projects where it is used is
increasing and GeoRM application is mostly pro-
moted by the construction companies.
In the Czech Republic GeoRM is in general
a new approach. For example, the Czech State
Mining Authority has accepted the concept of risk
just recently. Before that, admitting any risk or Figure 1. No application of GeoRM and ProjectRM.

85

ISGSR2013.indb 85 10/18/2013 9:37:06 AM

 In total sixteen top risks were identified by The GeoRM and ProjectRM, these are somewhat
Review Board, including five with geotechnical or blurred towards each other. The definitions of both
geohydrological causes. are not explicit, as well as the scopes and relation-
These results were compared with two other risk ships. Examples are vague and incomplete descrip-
files of the project, a project management risk file, tions of geotechnical risks in the project risk register,
including the project phases of design, construc- such as simply stated by words as “settlements” or
tion, and use of the tunnel, and a risk register “poor soil conditions”. These poorly described
focussing on contractual risks. Both of these files geotechnical risks are not explicitly analysed, clas-
did also include geotechnical risks. Two main con- sified, and managed. In addition, geotechnical risk
clusions could be drawn from this comparison. remediation measures, such as (additional) risk-
First, two out of the three risk files did present driven site investigations, contractual geotechnical
each risk as independent ones, without providing risk allocation by the Geotechnical Baseline Report
inter-relationships. The relationship between (GBR) (e.g. Essex 1997), or geotechnical monitor-
for instance geotechnical risks and contractual ing programs are not explicitly related to key per-
risks, including remediation measures, was not formance indicators of project risk management,
explicitly made. In other words, GeoRM was sepa- such as costs, budget, quality and safety.
rated from ProjectRM. A real-life example of rather blurred GeoRM
Second, some 50% of the top risks, identified and ProjectRM application has been identified
by the Review Board from a local government per- during a redevelopment project of existing infra-
spective, were not found in the two other risks files. structure in the Netherlands. In the project risk
The latter two were compiled with merely a project register some geotechnical risks were identified.
management and contractual focus. This case learns However, their risk descriptions incomplete, with
that separated application of risk management, fuzzy causes and effects. Also non-geotechnical
from different perspectives, may lead to incomplete risks were often poorly described. Furthermore,
and suboptimal risk management. The good news risk management in this project consisted mainly
is that after the review the project organization was of rather instrumental risk analyses, which stopped
able to combine the three risk registers, in order to at risk management step 4 of taking risk remedia-
strengthen its overall risk management. tion actions. Also step 5 of risk evaluation was not
Figure 2 presents the situation of separated done in a well-structured way. Therefore, any effec-
GeoRM and ProjectRM application. Contrary tiveness and cost-efficiencies of risk remediation
to the previous Figure 1 with the empty circles, in measures, if performed, remained hidden. This
Figure 2 the circles are shaded, but yet separated. obstructed learning from and improving of inte-
This symbolizes separated application of GeoRM grated risk management. Also, a geotechnical risk
and/or ProjectRM. did occur, leakage of groundwater in a construc-
The TC304-TF3 survey did not reveal many exam- tion pit, yet without serious consequences for the
ples of fully separated but applied GeoRM and Pro- project.
jectRM. The participating countries report either no This situation, together with an increase of
application of GeoRM and ProjectRM, or a some- project complexity and disputes with the con-
what blurred version of GeoRM and ProjectRM tractor, triggered to project manager to pay more
integration, which is explained in the next section. attention to improving the risk management on
her project. After a period with no or not so expe-
3.4 Blurred GeoRM and ProjectRM integration rienced risk managers, a well-experienced risk
manager was appointed for 2–3 days a week. His
In the third option of the relationship between the task was improving integrated ProjectRM for con-
two complementary types of risk management, tributing to project success, by making risk man-
agement useful and attractive for the project team
members. He was supported by a risk consultant,
who provided risk management training and guid-
ance to the project team.
Figure 3 illustrates the blurred GeoRM and
ProjectRM integration. Both ways of risk man-
agement are touching each other and GeoRM is
even partly integrated in ProjectRM. However, the
dotted circles of both indicate the vagueness and
incomplete execution of GeoRM and ProjectRM.
The TC304-TF3 survey identified that this type
of blurred GeoRM and ProjectRM integration
Figure 2. Separated GeoRM and ProjectRM. seems to be common practice in at least 80% of

86

ISGSR2013.indb 86 10/18/2013 9:37:07 AM

 motorway expansion projects ever in The
Netherlands. This multi-billion Design, Build,
Finance, and Maintenance (DBFM) project,
includes 60 km of very busy motorways in one of
the most densely populated areas, with soft soil
conditions and high groundwater tables. The con-
struction works have to be executed with a mini-
mum of hindrance to the daily motorway users, for
avoiding unacceptable mega-traffic jams.
For this and other large infrastructure projects,
the client developed project-specific risk manage-
Figure 3. Blurred GeoRM and ProjectRM integration.
ment flyers. The target groups for these flyers are
technical managers, contract managers, and entire
project teams. These flyers summarize the proven
the ten participating countries. For instance, in benefits of well-integrated GeoRM and ProjectRM
Austria, Germany, and Switzerland, ProjectRM for these specific project management roles within
is obliged in the large construction firms, and the project. An example of these benefits for the
GeoRM is required normatively as well. Never- technical manager are the execution of additional
theless, integration of GeoRM and ProjectRM risk-driven site investigations in the mentioned
is not perfect and both are yet too much focused motorway expansion project in The Netherlands,
on only contractual allocation of risk. In China which proved the feasibility of embankments in
and the Netherlands, GeoRM fits theoretically a conventional way and saved a lot of money. An
well in ProjectRM, however more integration in example of GeoRM benefits for the contract man-
practice, by more cooperation between ProjectRM ager is the application of a geotechnical baseline
and GeoRM professionals, may help to reduce too report in a large bridge project, which settled the
general and therefore inefficient and ineffective occurrence of differing site investigation effectively
GeoRM approaches. and efficiently, without delaying the project. The
entire project team gained for instance from the
results of a GeoRiskScan (Bles et al. 2009) in a
3.5 Complete GeoRM and ProjectRM
project for broadening a canal, which revealed just
integration
in time a gap in essential, yet lacking, ground infor-
Finally, there is the fourth option of complete mation which has been filled by hiring specific geo-
GeoRM and ProjectRM integration. In case of technical expertise (van Staveren et al. 2013c).
complete integration the definitions of ProjectRM Figure 4 presents complete GeoRM and Pro-
and GeoRM are explicit and shared, as well as jectRM integration. While GeoRM is absorbed
their respective scopes and relationships. by ProjectRM, the clear lines of both circles sym-
Examples are explicit and complete descriptions bolizes the explicit scopes of GeoRM and Projec-
of geotechnical risks in the project risk register. tRM. While being an integral part of ProjectRM,
Simply stated geotechnical risks by words as “set- GeoRM has keeps its own explicit and structured
tlements” or “poor soil conditions” are avoided. way of dealing with geotechnical risk, with its own
Moreover, well-described geotechnical risks are risk remediation tools.
explicitly analysed, classified, and managed. In In the TC304-TF3 survey only the Scandina-
addition, geotechnical risk remediation measures, vian countries Finland and Sweden, which form
such as (additional) risk-driven site investigations, 20% of the participating countries, indicate almost
contractual geotechnical risk allocation by the full integration of GeoRM and ProjectRM.
Geotechnical Baseline Report (GBR) (e.g. Essex
1997), by geotechnical monitoring programs are
explicitly related to key performance indicators of
project risk management, such as costs, budget,
quality and safety.
Also, there is close cooperation between geotech-
nical professionals, contract managers, quality man-
agers, safety managers, planners, cost managers and
so on, in order to optimize geotechnical risk man-
agement in line with the key performance indicators
of the civil engineering or construction project.
A real-life example of complete GeoRM
and ProjectRM integration is one of the largest Figure 4. Complete GeoRM and ProjectRM integration.

87

ISGSR2013.indb 87 10/18/2013 9:37:07 AM

 Sweden adds that ProjectRM is applied in most successful risk-driven operations in city areas, close
projects, but not always exactly following stand- to tunnels in use, which avoided leakage, deforma-
ards or guidelines such as ISO31000 (ISO 2009a). tions and resulting damage of adjacent structures.
ProjectRM following these standards is applied Also timely identification of site contaminations,
mainly in large projects. efficiently reacting on unexpected geotechnical
The United Kingdom reported that full integra- conditions, and transparent dialogues about geo-
tion of GeoRM and ProjectRM is provided by the technical risk with the public were reported as
UK Highways Agency. GeoRM successes.
In the Czech Republic GeoRM and ProjectRM Motivated by these successes, the contribution
integration depends currently on the construction of GeoRM to ProjectRM has also been identi-
phase and type of project. The Czech Republic fied within the TC304-TF3 survey. From the ten
reports that GeoRM is well-integrated in Pro- country reports in total ten specific types of con-
jectRM during the construction phase for par- tribution of GeoRM to ProjectRM have been
ticularly tunnel projects. However, in the planning derived, by using data triangulation (Patton 1987,
phase of these type of projects, risk management is Yin 2003). This involved comparing and clustering
often still missing. Moreover, GeoRM seems more similar GeoRM contributions, as reported by the
advanced than Project RM in the Czech Republic, ten countries.
because application of GeoRM is mostly driven by GeoRM contributions to ProjectRM, as raised
the experts in geotechnics and was motivated by by 40% of the ten participating countries, were
geotechnical failures. (1) managing the crucial role of geotechnical
uncertainties that have a major influence on con-
struction projects, (2) increasing the safety dur-
3.6 Benefits of GeoRM and ProjectRM
ing the works and of the final constructions and
integration
(3) combining systematic gathering of geotechni-
The previous section presented some examples of cal information with construction performance
completely integrated GeoRM and Project RM. and costs, for improving know-how and fostering
However, what are the benefits of it? learning from past performance.
Within the TC304-TF3 survey ProjectRM The remaining seven GeoRM contributions
success examples were provided by 60% of the were each identified by 20% or 10% of the coun-
participating countries. The reported ProjectRM tries, which means by one or two countries. This
benefits seem mainly realized in large expressway implies that the views on the GeoRM contribu-
and tunnel projects. Benefits include sensible tions vary considerably amongst several coun-
risk allocation, no serious risks and losses hap- tries, which allows learning from each other. One
pening, improved safety, prevention of conflicts, country may adopt one or more GeoRM contri-
finding innovative solutions in unexpected con- butions from other countries for strengthening its
ditions, perseverance of public project accept- ProjectRM.
ance and ease and facilitation of design and For instance, China and the Netherlands reme-
construction. diate geotechnical risk in projects by involving
However, as Sweden noticed, ProjectRM suc- QA/QC professionals, for aligning processes and
cesses are quite difficult to find and not readily achieving economies of scale and learning. This
available. Moreover, the ProjectRM benefits as approach could be valuable for the remaining eight
presented in all examples were qualitatively, rather countries that participated in this survey.
than quantitatively. ProjectRM successes are not Another example, the United Kingdom suggests
expressed in figures of the usual performance indi- that GeoRM may influence the choice of which
cators for project success, such as cost, time, safety, risk responses to adopt. Some responses may be
and quality. able to deal not only with geo-risks, but also with
Also GeoRM success examples were provided by other project risks at the same time, which also
60% of the participating countries by the TC304- may result in economies of scale and learning. This
TF3 survey. As the Czech Republic noticed, severe benefit may be attractive for some other countries
problems may be a trigger to start with GeoRM. as well.
Similar to the ProjectRM examples, also in the Finally, all of the ten types of contribution of
provided GeoRM success the examples, the results GeoRM to ProjectRM are again quite general
are qualitatively, rather than quantitatively. Also and qualitative, rather than proven and quantita-
GeoRM successes are not expressed in figures of tive contributions expressed in money, time, safety
the usual performance indicators for project man- records and so on. Nevertheless, all presented
agement, such as cost, time, safety, and quality. examples of benefits by applying ProjectRM and
The reported GeoRM benefits are derived from GeoRM seem to confirm the value of more well-
mainly tunnel projects and rail projects and include integrated ProjectRM and GeoRM. However, this

88

ISGSR2013.indb 88 10/18/2013 9:37:07 AM

 requires crossing a number of boundaries, which is their valuable geotechnical activities within civil
the topic of the next chapter. engineering and construction projects.

4 CROSSING GeoRM BOUNDARIES 4.2 Crossing boundaries between risk concepts

This first boundary separates conventional rule-
4.1 Introduction
based risk management from more modern
Despite the promising success examples of the pre- principle-based risk management. These are basi-
vious chapter, complete and integrated GeoRM cally the two main routes of doing risk manage-
and ProjectRM seems not yet common practice ment. The first approach is using rules and gives
in a lot of project phases, in a lot of projects, in apparently one best way for managing risk. This
a lot of countries. More complete and integrated is not appropriate for GeoRM, because of the
GeoRM and ProjectRM seems to require the cross- large diversity of engineering and construction
ing of several boundaries. This may support faster projects in complexity, size, location, and ground
learning how to increase GeoRM and ProjectRM conditions. For this reason, the Dutch participants
completeness and integration, which is expected to of the earlier mentioned Geo-Impuls development
pay off in effectiveness and cost-efficiency within program for reducing geotechnical failures and
projects. Within this chapter, the following eight related costs, take the principle-based route for
boundaries are explored and crossed: allowing fit-for-purpose GeoRM. Based on the
eleven generic risk management principles from
1. Boundaries between risk concepts
ISO31000 (2009a), eight specific geotechnical or
2. Boundaries between disciplines
GeoRM principles have been defined by a team of
3. Boundaries between phases
representative of Geo-Impuls participants.
4. Boundaries of scope
By definition, principles are always abstract
5. Boundaries between project types
and need a translation into activities. Such a
6. Boundaries between industries
translation can be done for specific projects, and
7. Boundaries between countries
even for specific project phases, again in a fit-for-
8. Boundaries within our minds.
purpose way. By this approach GeoRM is a mean
By awareness of, and even better, crossing these for realizing project objectives, and not an end in
boundaries, we can avoid to re-inventing the wheel itself. Table 1, derived from van Staveren et al.
in our own discipline of geotechnical engineering (2013b) presents the eight GeoRM principles,
and construction. Outside this discipline there together with a number of examples of related
appear to be abundant readily available concepts, actions. The table can be used as a GeoRM
methods, tools, and approaches that geotechnical checklist, simply by completing the column Done
engineers can use for strengthening the impact of Yes/No.

Table 1. GeoRM principles in a checklist format.

Done

No. GeoRM principles Examples of GeoRM actions Yes No

1 Generate and protect value Make all geotechnical risks in each project phase
explicit, with risk effects remediation measures
2 Participate in decision making Make a geotechnical risk file from the start of the
in all project phases project, use it for decisionmaking
3 Make geotechnical uncertainty Include geotechnical sensitivity analyses with
explicit margins in project reports
4 Work systematically, structured, Include GeoRM explicitly in the project planning
and in time and reserve resources for it
5 Use all available information Work from a general level to a detailed level, from
using geological maps to geotechnical monitoring
6 Work transparently together with Indicate and communicate any dependencies of
all stakeholders geotechnics with other disciplines in the project
7 Include the role of the human Make differences in organizational culture of all
factor involved project parties visible and feasible
8 Use experiences and lessons for Use all available and relevant project evaluations, risk
continuous improvement checklists, and experiences from professionals

89

ISGSR2013.indb 89 10/18/2013 9:37:07 AM

 Another example of using a “one-size-fits-all” management and safety management. The TC304-
rule-based risk management approach: the foun- TF3 survey revealed that in a number of countries,
dation design of a three-story building. Let us such as Switzerland, risk management is quite
assume that it is common practice to execute one safety-driven. Other discipline boundaries to cross
borehole for every fifty square meter for assessing are those between the technical disciplines, such
and remediating the foundation risks of differential as civil engineering and geotechnical engineering,
settlement, independent of the anticipated ground and the managerial disciplines, including project
conditions. By adopting a principle-based risk management, organization sciences, and organiza-
management approach, the number of boreholes tional change management. For instance, because
will be defined in a project-specific way. The site effectively, cost-efficiently and routinely applying
investigations depends on the anticipated ground GeoRM requires and appropriate organizational
conditions, the structural design of the building structure and culture within all of the organiza-
and resulting settlement tolerances, as well as of tions that are involved in construction projects.
the degree of risk tolerance of the client and con- The good news is that an appropriate organiza-
tractor and future inhabitants. tional culture for risk management has a lot in
Therefore, the main results of crossing the common with an effective organizational culture
boundaries of conventional rule-based to more for quality management and Health, Safety and
modern principle-based risk management con- Environment (HSE) management. Therefore, risk
cepts is delivering more project-specific geotech- managers, quality managers and HSE managers
nical risk management, that fits the specific risk can join their forces in order to establish organi-
tolerances of the parties involved, against reason- zational cultures within their projects that support
able costs. Effectiveness and cost-efficiency are risk management, rather than frustrate it.
the main anticipated benefits of crossing the rule- Van Staveren (2013b) derived seven key con-
based-principle-based boundary of geotechnical ditions for an effective geotechnical and project
risk management. risk management culture. These are derived from
The TC304-TF3 survey reveals that in seven out research on implementing risk management in
of the ten countries (70%) principle-based stand- organizations in the construction sector (van
ards, are mentioned. However, in three of these Staveren 2009). These key conditions are (1) shared
seven countries the ISO 31000 (ISO 2009a) Guide- understanding of the geotechnical risks in a project
line for risk management is not yet well known facilitated by speaking the same “risk manage-
(Czech Republic, Germany), or mentioned in only ment language” in the organization, (2) acknowl-
one project (Finland). Principle-based ISO stand- edgement that dealing with geotechnical risk is
ards seems rather well incorporated in Austria, driven by a combination of rationality and (often
Netherlands, Sweden and United Kingdom. unconscious) emotion, specifically with regard to
Finally, principle-based approaches are not the assessments of risk probabilities and effects,
limited to the discipline of risk management. For (3) understanding of differences in risk perception
instance, the principle-based concept of High Reli- by individual engineers and managers, which has
ability Organizations (HRO) (Weick and Sutcliffe a close relationship with the second cultural condi-
2007) has been developed in the discipline of safety tion, (4) understanding that geotechnical risk man-
management. The five HRO principles include agement goes beyond geotechnical engineering and
reluctance to simplification, sensitivity to opera- has close relationships with planning, cost control,
tions, and respect for expertise. Organizations safety, and quality, (5) sharing of geotechnical risk
that work in high risk environments according knowledge is common practice, (6) acknowledge-
to these principles proved to have substantially ment of differences in motivation to apply geotech-
better safety records than similar organizations nical risk management amongst engineers and
that did not adopt these principles in their opera- managers, and finally (7) cooperation on geotechni-
tions. Examples of typical HRO organizations are cal risk management with other disciplines within
nuclear power stations and aircraft carriers. In the and outside the organization.
Netherlands, the HRO approach in winning popu- An example of the latter condition is that one
larity within other types of organizations, includ- party reduces the probability of a risk, while
ing healthcare and construction. This confirms the another partner reduces the risk effects in case
value of crossing borders between disciplines, as of occurrence. For instance, an engineer makes a
explored in the next section. robust geotechnical design for a deep excavation in
a city centre, while the contractor works out a fall-
back scenario in case, despite the robust design,
4.3 Crossing boundaries between disciplines
horizontal deformations exceed pre-set values. For
The second boundary separates several different avoiding confusion, a robust geotechnical design
disciplines. Let us start with the disciplines of risk is here defined as one that reduces the probability

90

ISGSR2013.indb 90 10/18/2013 9:37:07 AM

 of failure or unacceptable deformation to a well- Table 2. Project phases, GeoRM and ProjectRM.
defined and agreed-upon risk level.
An additional way of joining forces by risk man- Project ProjectRM GeoRM
agers, quality managers, and HSE managers is the phases in countries in countries
execution and control of risk remediation measures Feasibility 60% 50%
by quality and safety engineers. For instance, the Final design 70% 50%
risk of the placement of filter materials with inap- Contracting 80% 50%
propriate grain sizes at a dam site can be reduced Construction 90% 90%
by taking regularly bulk samples by the QA/QC Maintenance 10% 10%
team members. Moreover, the risk of slope instabil- Decommissioning 10% 0%
ity during excavation works adjacent to an existing
motor-way can be monitored by HSE engineers,
because this concerns the safety of the workers as
well as of the users of the motorway. tries (if applied at all, which is not yet a routine, as
These are just two simple examples of crossing explained in the previous chapter).
the boundaries of the risk, quality and safety dis- In Table 2 the conventional order of subsequent
ciplines on construction sites. For being effective, project phases is presented, from feasibility to
the delegation of risk remediation measures need and including decommissioning. The table shows
to be formalized within the organizational project that ProjectRM is in particular applied during the
structure. construction phase (in 90% of the participating
Therefore, the main benefits of crossing the countries) and in the contracting phase (in 80% of
boundaries between the disciplines of risk man- the participating countries). If applied, then in a
agement, quality management and health, safety majority of countries ProjectRM is even executed
and environmental management are creating cost during the feasibility phase and design phases.
and time efficiencies, because risk remediation Only the United Kingdom reports that Pro-
measures are combined with quality and safety jectRM is sometimes even executed during main-
controls. tenance and decommissioning. Nevertheless, as
Finally and obviously, integrating GeoRM and explicitly indicated by the United Kingdom, is
ProjectRM requires also crossing the boundaries should not be excluded that ProjectRM is mainly
between discipline-based risk management, such applied in only one or two project phases of indi-
as GeoRM, and ProjectRM, which is in fact part vidual projects, rather than in all project phases,
or project management. Therefore, by crossing as ProjectRM theories do recommend (e.g. van
this border the border between a rather technical Well-Stam 2004).
approach (geotechnical engineering) and manage- From Table 2 it can also be concluded that
rial approach (project management) is also crossed. GeoRM is in particular applied during the con-
Both approaches complement each other and are struction phase, in 90% of the participating coun-
fundamental for realizing successful civil engineer- tries. In 50% of the ten participating countries
ing and construction projects. GeoRM is applied in the feasibility, design and
contracting phases. Nevertheless, as indicated by
the United Kingdom, it is likely that also GeoRM
4.4 Crossing boundaries between phases
is mainly applied in only one or two phases, rather
The third border separates commonly distin- than in all phases as GeoRM theories do recom-
guished project phases in civil engineering and mend (van Staveren 2013b). For instance Aus-
construction projects. Conventional project phases tria, Czech Republic, Germany, and Switzerland
for these type of projects are feasibility, pre-design reported that geotechnical risks are typically not
and final design, contracting, construction, main- well analysed, communicated, and managed during
tenance and finally decommissioning. Obviously, the project planning phase. Therefore, once again,
these phases have slightly different names in dif- Table 2 presents only in which project phases Pro-
ferent countries. Also the sequence of these phases jectRM and GeoRM are applied, if applied at all.
may be different, due to the type of contract. In When compared with ProjectRM, GeoRM
case of Design and Build or Design and Construct seems to be applied to a lesser extent in the dis-
contracts for example, the contracting phase will tinguished project phases than ProjectRM. The
be before the design phase, because design is part exception is the construction phase. In total 90%
of the contract to be executed by a construction of the countries report that ProjectRM as well as
company. GeoRM is executed during construction. However,
Table 2 is derived from the TC304-TF3 survey the presented data does not reveal whether
and shows in which project phases ProjectRM and ProjectRM and GeoRM are both performed in the
GeoRM are applied in the ten participating coun- same projects or not. In other words, it is possible

91

ISGSR2013.indb 91 10/18/2013 9:37:08 AM

 that in some projects only ProjectRM is executed, admission of potential problems. Especially in the
and on other projects only GeoRM. technical field, engineers highly rely on standards
Table 2 and the previous remarks demonstrate and norms and they are not used to analyse pos-
that ProjectRM and GeoRM are not yet cyclic sible deviations from an ideal standardized state or
applied in the entire sequence of project phases, progress. Therefore, risks are typically not commu-
as highly recommended by the risk management nicated to the public in advance. It has been com-
theories (e.g. van Well-Stam 2004). This is simply mon practice to present the project as a safe and
demonstrated by the following example. A deci- certain action. However, this paradigm seems to be
sion to apply the observational method as a risk changing. There has been a rising debate about the
remediation measure during construction should number of sever cost overruns in large construc-
be taken already during design, because it requires tion projects and about the fact, that uncertainty
a certain degree of design flexibility that allows of the cost estimates should not be neglected in the
changes during construction, if the ground proper- planning phase.
ties encountered require changes in order to satisfy In the Czech Republic many serious accidents
the specifications (van Staveren 2006). and cost overruns occurred over the last years.
In conclusion, crossing the boundaries of Now the public is generally suspicious about the
project phases may become an important step in large construction projects. Open communica-
developing the effectiveness of both GeoRM and tion of geotechnical risks in future projects can
ProjectRM. thus cause exaggerated reactions of the affected
inhabitants.
In the Netherlands it is for many clients and
4.5 Crossing boundaries of scope
contractors quite a dilemma to either communicate
The forth border is about scope. In particular, this about geotechnical risk before starting the project
border explores the scope of communication about (which may make the public feeling uncomfortable
geotechnical uncertainty, risk and the potential about the project), or only once the geotechnical
consequences for project parties and external problems occur (for instance damage due to set-
stakeholders. While geotechnical calculations tlements) which would make the public not only
are obviously of key importance for successful feeling uncomfortable, but quite angry as well.
projects, clear communication of the calculation Switzerland demonstrates that public involve-
results, their expected accuracy, as well as the need ment in decisions about the large projects pays off.
for reliable input parameters (to derive from site For example, the Gotthard base tunnel showed
investigations) should also be part of the scope of the benefits of public involvement into the proc-
the geotechnical engineer, once he or she aims to ess. This included three referenda on realization of
work according the principles of GeoRM. the project on for instance its financing, ensured a
For example, many geotechnical engineers com- general acceptance of the project amongst people,
plain from time to time about the lack of interest of in spite of large time and cost overruns and even
clients in site investigations. Who of us is not famil- fatalities during its construction.
iar with a site investigation report for a project that All of these examples invite geotechnical engi-
is too limited with regard to the number of Cone neers to jump over the boundary that separates
Penetration Tests (CPTs), boreholes, and labora- the area of geotechnical calculations with results
tory tests? Apparently, in those cases geotechnical in numbers from the area with convincing stories,
engineers are not able to communicate clear about graphs and other visual presentation means about
the geotechnical risk profile of the project and its the need for sound and risk-driven geotechnical
potential consequences for the client and other engineering, in order to realize successful and safe
project parties and stakeholders. As indicated by engineering and construction projects.
the United Kingdom in the TC304-TF3 report This type of communication should start within
(van Staveren 2013a): the project team inside projects, should expands
towards the public and other stakeholders in the
“Technical people often struggle to present informa- external project environment. An example of with-
tion in a non-technical manner, reverting to techno- in-the-team risk communication that is proposed
speak and jargon that simply turns others off.” in the TC304-TF3 survey is simply including geo-
technical risk communication within all coordina-
The risk language issue is made explicit by tion and common practices inside the project, as
a number of countries that participated in the done in Finland. Another example is presenting
TC304-TF3 research. For instance, Austria, the consequences of identified geotechnical risks in
Germany, Switzerland report that communication terms that can be more widely understood by non-
of risks in general is quite limited. This is mainly geotechnical project team members and managers,
due to cultural specifics that do not motivate open such as their impact to meeting business objectives

92

ISGSR2013.indb 92 10/18/2013 9:37:08 AM

 and the ‘balance’ to be achieved between risk and is applied in underground construction in 40% of
reward, as practiced in the United Kingdom. the participating countries, and in major highway
Examples for geotechnical risk communication projects in 40% of these countries. In the remain-
with the public outside the project are organiz- ing types of projects, GeoRM is applied in 30% or
ing workshops, evening meetings, door knocking, less of the participating countries.
promoting in the local press, having a question & When compared with ProjectRM, GeoRM
answer kiosk or information centre outside the site seems to be applied to a lesser extent than
and manned by trained staff, and Open Doors Days ProjectRM. For example, in underground con-
for the project, as occasionally done in for example struction projects, ProjectRM is used in 70% of the
Czech Republic and in the United Kingdom. In countries, against GeoRM in 40% of the countries.
the Netherlands, information centres for the public Remarkably, for building foundations ProjectRM
are common for large tunnel and railway projects and GeoRM are not or almost not applied.
in city centres, such as bored and cut and cover In conclusion, the TC304-TF3 survey results
tunnels under construction and other works in the confirm that for certain types of project ProjectRM
cities of Amsterdam, Delft and Utrecht. and sometimes also GeoRM is more common than
In conclusion, crossing the boundaries of scope for other types of project. This implies that pro-
by geotechnical engineers concerns in particular fessionals concerned with project types for which
communication from only numbers towards con- risk management is not yet common practice, may
vincing stories in a language that can be easily be able to learn a lot from engineers and manag-
understood by non-geotechnical managers, clients, ers who are already more familiar with applying
and public. ProjectRM and GeoRM in their projects. The
same applies of course for larger projects, like
highways, where risk management is rather com-
4.6 Crossing boundaries between project types
mon practice and smaller road projects, which
The fifth boundary of concern is that between dif- may also benefit from applying risk management
ferent project types. Civil engineering projects can practices. For these reasons, crossing the border of
be divided by several criteria, for instance by func- project types may pay off.
tion or by size. The ten countries participating in
the TC304-TF3 survey were asked in which type
4.7 Crossing boundaries between industries
of projects ProjectRM and GeoRM are applied, if
applied. The results are summarized in Table 3. The sixth boundary is about crossing boundaries
Table 3 shows for instance that ProjectRM is between different industries. The introduction of
applied in underground construction in 70% of the this paper started with the statement that especially
ten participating countries, and in major highway construction projects are inherently risky, due to
projects in 60% of these countries. In the remain- numerous uncertainties of many types. While true,
ing types of projects ProjectRM is applied in a the construction industry is not the only industry
minority of the participating countries. GeoRM facing challenges of high uncertainty. For instance,
projects within the chemical, energy, offshore, oil,
gas, military and space industry do also have their
Table 3. Project types, GeoRM and ProjectRM. specific high uncertainty challenges.
ProjectRM GeoRM
There are a number of striking similarities
Types of projects in countries in countries between the space industry and the under-ground
construction industry, involving tunnels works for
Underground 70% 40% instance. Quality and safety are key issues in the
construction: tunnels, space industry, because once a rocket with a satel-
subway (metro) lite is launched, repair is very difficult, if possible
stations, parkings at all. Moreover, there is time and cost pressure in
Major highways 60% 40% the space industry, because of increasing competi-
Large, multidisciplinary 40% 30% tion and governments cutting budgets.
and complex Similar situations occur in the construction
infrastructure
sector. Especially in Europe, there is severe price
Railways 40% 30%
competition and government budgets are under
Harbours 30% 20%
pressure, due to the financial crises. Moreover,
Bridges 30% 10%
when returning to underground construction, once
Large dams 30% 10%
Dikes & slopes 10% 10%
a tunnel is bored and completed, it is often very
Windfarms 00% 10%
expensive to repair in case of for instance excessive
Building foundations 00% 10% groundwater inflow by cracks in the tunnel lining.
The same applies to deep foundations.

93

ISGSR2013.indb 93 10/18/2013 9:37:08 AM

 Therefore, the need for adequate risk manage- more effective and cost-efficient GeoRM can be
ment in the construction sector may match well found. These suggested industries are presented in
with the need for effective and cost-efficient risk Table 4.
management in the space industry. For instance, All of the suggestions in Table 4 demonstrate
the NASA Risk Management Handbook (NASA that there is probably a wealth of proven concepts,
2011), free to download from the Internet, provides methods, techniques, tools, and approaches wait-
a wealth of risk management knowledge that may ing for adoption in well-integrated ProjectRM
be of particular use for the construction industry. and GeoRM. This may save a lot of time, energy
A learning example is the rigorous relationship and money by avoiding to re-invent the wheel by
between risk and opportunity management that keeping considering ProjectRM and GeoRM as
the NASA applies. Therefore, all uncertainties are separated disciplines, as indicated by the United
perceived from the conventional risk side (the neg- Kingdom in TC304-TF3.
ative down-side), as well as from an opportunity In conclusion, the main benefits of crossing the
perspective (the positive upside). Apart from the boundaries between the different industries are
possibilities of financial gains from this approach, cost-effective organizational learning and conse-
is much more fun for all professionals and manag- quently costs savings. A lot of risk management
ers involved, because they are not only looking for experiences and lessons from other industries are
potential problems. ready to be applied within the construction indus-
Other lessons from space, derived from try, obviously in the earlier discussed principle-
a risk management implementation case in based and project specific way.
The Netherlands, are (1) paying more attention to
interdisciplinary teamwork in all project phases,
4.8 Crossing boundaries between countries
(2) explicitly discussing mistakes made in order to
learn from them in other projects and (3) evaluat- The seventh boundary is about different countries
ing the effectiveness and cost efficiency of project We are living in an ever more globalizing world,
risk management. The latter is done by assessing in which millions of professionals in different
the Total Cost of Risk (TCoR) in projects, with countries are interconnected by different kinds
is the sum of all risk management process costs, of social media. For instance, the professional
the costs of all risk remediation measures and the LinkedIn network has currently globally more than
costs of all risks that occurred, despite the meas- 225 million users (per June 2013). With regard to
ures taken (which again requires a high degree of risk management, there are numerous user-groups
openness and trust within the organization). with participants from different countries, who can
Exploring the boundaries between industries easily learn from each other.
for the benefits of learning about risk manage- When focusing on the construction sector, over
ment was also part of the TC304-TF3 survey. In the last fifteen years Dutch engineering firms and
total five out of the ten participating countries contractors imported for instance successfully
(50%) suggested in total seven sectors outside the soft soil tunnel boring expertise from abroad to
construction industry, where valuable lessons for The Netherlands. In return, Dutch technology for

Table 4. Project and GeoRM lessons from other industries.

Industry Lessons for ProjectRM and GeoRM

Mechanical industry How to increase productivity, despite the unique character of construction
projects and the related risks.
IT-industry How to use simulation tools for optimizing construction processes under risky
conditions.
Space industry How to integrate risk management in systems engineering in construction projects.
How to develop risk management as part of continuous organizational
improvement programs.
Insurance industry How construction insurance claims can be avoided or reduced, by using qualified
and competent people.
Chemical and nuclear industry How to develop clear risk benchmarks
How to combine risk and safety management
Financial services industry How to consider risk in a holistic and integral way by applying Enterprise Risk
Management (ERM).
Consulting industry How to apply proven project management techniques within ProjectRM and
GeoRM processes.

94

ISGSR2013.indb 94 10/18/2013 9:37:08 AM

 flood protection, gained during the Delta Works, is the community, and increasing risk awareness and
exported for tens of years all over the world, from open communication, which would likely increase
northern America to south-east Asia. With regard their acceptance both in society and amongst
to geotechnical risk management, the concept of practitioners.
the Geotechnical Baseline Report (GBR), as origi- The remaining eleven recommendations, were
nally developed in the United States (Essex 1997) reported by 10% or just one of the ten countries.
has been introduced in The Netherlands (van Examples of these recommendations are adopt-
Staveren 2006). The same applies for the Observa- ing new investigation methods or technologies,
tional Method (OM), as developed by Peck (1969) such as GPS and GIS in GeoRM to provide suffi-
in the 1930s. The OM is currently starting up a suc- cient geotechnical information for minimizing the
cessful second live in the Netherlands, in particular uncertainty in ProjectRM (suggested by China),
for major tunnel projects. It proves to be a valuable and making an effort to involve people who are
geotechnical risk control tool that has been made directly involved in risk handling to adopt those
very efficient by using the latest sensor and infor- parts of risk management that they can benefit
mation technology. from (proposed by Sweden). Because these recom-
The ten countries participating in the TC304- mendations are suggested by just one country, they
TF3 survey provided in total 19 different recom- may be of particular interest for other countries,
mendations for (further) integrating GeoRM and where these recommendations may be unknown to
ProjectRM. These recommendations have been date. This makes it worthwhile to consider cross-
derived, by using data triangulation (Patton 1997, ing boundaries of countries, either virtual just by
Yin 2003), which involved comparing and cluster- surfing over the Internet, or real by participating
ing largely similar recommendations, as reported in international conferences and by actually vis-
in the ten country reports. All of the recommenda- iting organizations and construction projects in
tions have been classified in three types: (1) recom- countries.
mendations that are primarily realized by changes Similar to crossing the boundaries of indus-
in the organization structure of (project) organi- tries, the main benefit of crossing the bounda-
zations involved in construction projects, (2) those ries between countries is fostering cost-effective
primarily realized by changing the culture within organizational learning. This is expected to result
these organizations, and (3) recommendations in substantial costs savings, because a lot of risk
primarily involving technical measures. Measures management experiences and lessons from other
for changing organizational structure and culture countries with more or less similar ground condi-
seem often closely coupled. In total 13 out of the tions can be applied, without paying the money of
19 recommendations (68%) have an organizational learning by mistake.
cause, either structural (47%) or cultural (21%). Finally, as mentioned before, six out of the
In total six technical recommendations have been nineteen recommendations (32%) made by the ten
identified (32%). countries for integrating and developing ProjectRM
The top five recommendations for integrating and GeoRM were technical recommendations.
GeoRM and ProjectRM, reported by 60% to 50% However, in the same TC304-TF3 survey also the
of the ten countries, involved education of apply- main hurdles for applying ProjectRM and GeoRM
ing ProjectRM and GeoRM, communication of have been investigated. Of all hurdles investigated,
their effects and successes, as well as learning from in total 5% of the ProjectRM hurdles and 18% of
past performance of projects. All of these recom- the GeoRM hurdles were of technical origin. All
mendations are classified as of an organization remaining hurdles, i.e. 95% for ProjectRM and
structure type, because the education should be 82% for GeoRM seem to have an organizational
structurally organized. In other words, the pro- cause.
posed ProjectRM and GeoRM education should Therefore, it seems that there is a tendency to
not stay voluntary for engineers and managers. solve organizational problems with technical solu-
It should become a formalized part of their indi- tions, rather than with organizational solutions.
vidual professional development. This results into This should be a point of attention, when one is
professional development of organizations, either trying to reduce ProjectRM and GeoRM hurdles
of clients, engineering firms, contractors, knowl- for (further and deeper) integrating both types of
edge institutes and universities in the discipline of risk management. Here we return to the area of
risk management. the human mind, which is the last border to cross.
Other recommendations, reported by 40% to
30% of the countries, involve increasing the inter-
4.9 Crossing our mental boundaries
est of public clients in risk, to provide standards
for risk management processes in public invest- Finally, we may have to cross some of our mental
ments projects that would be broadly accepted by boundaries, in particular the one between different

95

ISGSR2013.indb 95 10/18/2013 9:37:08 AM

 risk perceptions. This requires a few rather philo- and better risk management tools. Practice teaches
sophical and psychological reflections. that it seems often the result of better organizing
Understanding of differences in risk perception, the required expertise for dealing at the right time
the fact that people judge and classify the same and place with geotechnical uncertainties.
factual risk data in a different way, is a key success For instance, research by van Tol (2008) of 40
factor for effective risk management (van Staveren failed excavation pits in the Netherlands proved
2006). Individual risk perception is the result of an that these failures were not caused by a lack of
(often hidden) mixture of ratio and emotion. This geotechnical expertise. For 80% of these cases the
has been thoroughly investigated by for instance geotechnical expertise was readily available in the
the Nobel laureates Kahnemann and Tversky Netherlands, but not organized at the right time at
(1979) in their prospect theory about how people the right place in the project. Therefore, these fail-
make choices under uncertainty. ures seem more of an organizational problem than
Dealing with different risk perceptions is by far an geotechnical problem. This stresses the need to
not simple. Prof. Frans Barends (2005), emeritus pay more attention to (project) organization design
professor at the Delft Technical University in The approaches, including their structures and cultures
Netherlands, concluded in 2005 during his Terza- (van Staveren 2009b).
ghi Oration: Effective design of project organizations
requires another approach than conventional
‘The effects of subjective individual interpretation of technical design. For instance, technical design
facts and data are underestimated.’ of constructions like a bridge or tunnel uses con-
ventional algorithmic technological rules, which
This firm statement was illustrated with a geo- guarantee that a solution will be found. However,
hydrological example about the interpretation of suitable organizational design, necessary for well-
an observed porewater response, caused by chang- integrated and completely applied GeoRM and
ing water levels. This interpretation proved to be ProjectRM, needs so-called heuristic technologi-
highly subjective, as six different geohydrological cal rules (Roozenburg & Eekels 1995). Contrary
models could be applied to interpret the same data, to the algorithmic ones, heuristic technological
obviously with different outcomes. Awareness of rules do not guarantee that a solution can or will
our own subjective expert judgment is a there- be found. A heuristic rule facilitates finding of a
fore a first step towards effective dealing with risk solution. This aligns with a hermeneutic onto-
perceptions. logical view, which considers the world as a social
The next step is trying to understand the risk construct with inherent subjectivities (Gumesson
perceptions of other people: colleagues, geotech- 1988).
nical engineers, other engineers involved in the Accepting these facts for developing GeoRM
project, project managers, technical managers, and ProjectRM implementation is another exam-
contract managers, site managers, procurement ple of crossing a mental boundary. Moreover, it
managers, financial managers, quality and safety helps to keep a realistic and pragmatic view on
managers, our clients, the public, and other exter- developing risk management in civil engineering
nal stakeholders. Trying to understand their risk and construction projects.
perceptions, and comparing these with our own
perceptions, is a prerequisite for effective decision
making under uncertainty. This applies also to geo- 5 CONCLUSIONS
technical risks.
Explicit and open communication is required Based on the previous chapters, and in particular
to identify and discuss any differences in risk per- the results of the TC304-TF3 survey, the following
ception between ourselves and other persons. It is main conclusions are made.
likely to result in a more shared risk understand- First, abundant literature, and the TC304-TF3
ing, which opens doors towards less confusion, survey provide evidence that integrated Geotech-
less misunderstanding, and consequently more nical Risk Management (GeoRM) and Project
successful engineering and construction projects, Risk Management (ProjectRM) may considerably
despite their inherent geotechnical and project contribute to successful civil engineering and con-
risks. Therefore, understanding our own and other struction projects for all involved project parties
risk perceptions can be seen as crossing our men- and stakeholders.
tal boundary from unaware dealing towards well- Second, from a methodological point of view,
aware dealing with risk. there seem to be no objections for (further) inte-
In addition, as mentioned in the previous section, grating GeoRM and ProjectRM in all project
more effective and cost-efficient risk management phases, because their pragmatic definitions and
seems often not so much a matter of only more processes fit well and complement each other.

96

ISGSR2013.indb 96 10/18/2013 9:37:08 AM

 Third, considerable failures and resulting addi- This education should focus on the concepts and
tional costs in civil engineering and construction practices of uncertainty, risk and its management
projects, often with a geotechnical root cause, as in general, and on ProjectRM and GeoRM in par-
found in the literature, has been confirmed by the ticular. These professionals and managers should
TC304-TF3 survey. In at least 80% of the partici- then become more capable of raising risk aware-
pating countries GeoRM and/or ProjectRM are ness and providing acceptable risk solutions for
(1) not applied at all in all phases of civil engineer- their projects and organizations, their clients, the
ing and construction projects, are (2) incompletely public, and other stakeholders of civil engineering
and separately applied, or are (3) applied in a rather and construction projects.
blurred way. Only two out of the ten countries Second, expand the degree multi-disciplinary
(20%) reported (4) rather complete and integrated and inter-disciplinary research and development
application of GeoRM and ProjectRM in most of GeoRM and ProjectRM. This research should
projects, however not always performed according in particular focus on the interfaces with project
to standards and guidelines. Therefore, the degree management, with engineering and construction
of implementation of integrated ProjectRM and management, and with safety and quality manage-
GeoRM varies considerably between and within ment. A perhaps extreme relevant research topic
the participating countries. Similar variations in is measuring the GeoRM and ProjectRM effec-
the degree of GeoRM and ProjectRM application tiveness and cost-efficiency, in relation the project
and integration are identified between different management key performance indicators such as
types of projects, between different project phases, safety, quality, costs, time, and even reputation.
and between different sizes of projects. Case study research may be an effective research
Forth, in total eight boundaries have been methodology. In particular quantitative research
explored. Crossing these boundaries will probably results of risk management effectiveness are highly
improve the degree of GeoRM and ProjectRM welcome, in order to convince sceptical profession-
application and integration within civil engineer- als, clients, and the public about the added value
ing and construction projects. These boundaries of risk management within civil engineering and
concern (1) risk concepts, (2) disciplines, (3) project construction projects. Depending on the research
phases, (4) scopes, (5) project types, (6) industries results, further developing and fine-tuning of
and (7) countries. For learning how to cross these GeoRM and ProjectRM can be realized.
seven boundaries, we may have to cross first some Execution of these recommendations will allow
(8) mental boundaries. us to follow Deming’s widely accepted and applied
Therefore, the final conclusion of this paper Plan-Do-Check-Act (PDCA) cycle, specifically for
could be that in theory there are no objections for risk management improvement in engineering and
a full integration of GeoRM and ProjectRM, that construction. Plan equals providing education, do
the awareness of the need for and potential benefits concerns application of integrated GeoRM and
of such an integration is growing, but that the inte- ProjectRM in projects, check is about evaluation
grated application of both types of risk manage- by research of the risk management performance
ment, for realizing real project success in practice, of projects, and finally act is about further develop-
has still a lot of room for development and improve- ing and fine-tuning risk management theories and
ment. For this reason, the next and final chapter practices. Here emerges an important task for uni-
presents some main recommendations for further versities and other knowledge institutes.
and deeper integrating GeoRM and ProjectRM. The follow up of these recommendations, pref-
erably in close cooperation with many countries,
may realize the promise of well-integrated and
6 RECOMMENDATIONS completely applied GeoRM and ProjectRM. The
ultimate objective remains substantially contribut-
Throughout this paper already many examples ing to successful construction projects, by applying
and recommendations for further and deeper risk-driven geotechnical engineering and construc-
integrating GeoRM and ProjectRM have been tion, which give us less cost overruns, less delays,
presented. Nevertheless, two main recommenda- less non-conformances of safety and quality, and
tions need considerable more attention, because therefore happier project participants, client, pub-
these are considered as key success factors for lic and other stakeholders.
developing complete and well-integrated execution
of GeoRM and ProjectRM, in all project phases
of civil engineering and construction projects. ACKNOWLEDGEMENTS
First, improving the quality and quantity of
education of geotechnical professionals, other The author is grateful to the ISSMGE TC304-TF3
engineers, and managers participating in projects. participants, who prepared their country reports

97

ISGSR2013.indb 97 10/18/2013 9:37:08 AM

 that made it possible to compare and analyse the Chapman, C.B. & Ward, S. 2011. How to Manage Project
degree of GeoRM and ProjectRM application and Opportunity and Risk: Why Uncertainty Management
integration in ten different countries. These profes- can be a Much Better Approach than Risk Manage-
sionals are, in alphabetical order by first names, ment, 3rd Edition. Chichester: John Wiley & Sons.
Clayton, C.R.I. 2008. The Heathrow Tunnel Collapse.
Alexander Rozsypal, Annina Peisa, Hongwei Advanced Course on Risk Management in Civil Engi-
Huang Joost van der Schrier, Lars Olsson, Kati neering LNEC Lisbon November 17–22 2008. URL:
Kaskiala, Leena Korkiala-Tanttu Markku Tavi, http://riskmanagement.lnec.pt/pdf/papers/Nov21_
Makoto Kimura, Martin Srb, Matti Kokkinen, apresentacoes/20_presentation_Clayton.pdf.
Miroslav Matousek, Olga Spackova, Paul Cools, Cooke-Davies, T. 2002. The “real” success factors on
Paul Maliphant, Pentti Salo, Tadashi Hashimoto, projects. International Journal of Project Management
Tobias Nevrly, Qunfang Hu and the members of 20(3): 185–90.
the Swedish Geotechnical Society Committee on Cools, P.M.C.B.M. 2011. The Geo-Impuls Programme
Risk Assessment. reducing geotechnical failure in the Netherlands. In
N. Vogt et al. (eds), Geotechnical Safety and Risk: 191–
Furthermore, the author is grateful to all profes- 198. Karlsruhe: Bundesanstalt fur Wasserbau (BAW).
sionals, engineers and managers, he met personally DIN Deutsches Institut für Normung. 2010. DIN
over the years during lectures, workshops, confer- 4020: Geotechnische Untersuchungen für bautech-
ences, and consultancy activities in many organiza- nische Zwecke [Geotechnical Investigations for Civil
tions, inside and outside the construction industry. Engineering Purposes]. Berlin: Beuth-Verlag.
Without their open sharing of experiences and Egan, J. 1998. Rethinking Construction: The Report of the
insights, this paper could not have been realized. Construction Task Force. London: Department of the
Environment, Transport and the Regions, HMSO.
Edwards, L. 1995. Practical Risk Management in the
Construction Industry. London: Thomas Telford.
REFERENCES Edwards, P.J. & Bowen, P.A. 2005. Risk Management
in Project Organisations. Oxford: Butterworth-
Barber P., Graves A., Sheath D. & Tomkins, C. 2000. Heinemann.
Quality failure costs in civil engineering projects. Inter- Essex R.J. (ed.). (1997). Geotechnical Baseline Reports for
national Journal of Quality & Reliability Management Underground Construction. Technical Committee on
17(4/5): 479–492. Geotechnical Reports of the Underground Technology
Barends, F.B.J. 2005. Associating with advancing insight: Research Council. Danvers: ASCE.
Terzaghi Oration 2005. In: Proceedings 16th Inter- Flanagan, R. & Norman, G. 1993. Risk Management
national Conference on Soil Mechanics and Geotech- and Construction. Oxford: Blackwell Scientific
nical Engineering, 12–16 September, Osaka, Japan, Publications.
pp. 217–48. Rotterdam: Millpress. Godfrey, P.S. 1996. Control of Risk: A Guide to the
Barends, F.B.J. 2009. Over Onzekerheid en Duurzaam- Systematic Management of Risk from Construction.
heid in de Geotechniek [About Uncertainty and Sus- CIRIA Special Publication 125. London: CIRIA.
tainability in Geotechnics]. Geotechniek Special Nov.: Gummesson, E. 1988. Qualitative Methods in Management
28–33. Research. Lund: Studentlitteratur.
Bea, R.G. 2006. Reliability and human factors in geo- Harper, K. 2000. Managers ‘overlooked risk’ in airport
technical engineering. Journal of Geotechnical and tunnel collapse. The Guardian, Thursday 6 July 2000
Geoenvironmental Engineering 132(5): 631–44. 01.46 BST.
Beal, A.N. 2007. CDM Regulations: 12 years of pain but Hillson, D. 2002. Risk Management Maturity Level
little gain. New Civil Engineer International August: 82–8. Development. Risk Management Specific Interest Group.
Beck, U. 1992. Risk Society: Towards a New Modernity. Newton Square, PA: Project Management Institute.
London: Sage. Hillson, D.A. 2010. Exploring Future Uncertainty:
Bernstein, P.L. 1996. Against the Gods: The Remarkable Creating Value From Risk. Farnham: Gower.
Story of Risk. New York: John Wiley & Sons. Hillson, D.A. & Murray-Webster, R. 2007. Understanding
Bles, T.J., van Staveren, M.Th., Litjens, P.P.T. & and Managing Risk Attitude, 2nd edition. Aldershot:
Cools, P.M.C.B.M. 2009. Geo Risk Scan: Getting Gower.
grips on geotechnical risks. In Y. Honjo et al. (eds) Hopkin, P. 2012. Fundamentals of Risk Management:
Geotechnical Safety and Risk: 339–346. Leiden: Understanding, Evaluating and Implementing Effec-
CRC Press. tive Risk Management, 2nd edition. London: Kogan
Blockley, D. & Godfrey, P. 2000. Doing It Differently: Page.
Systems for Rethinking Construction. London: Hopkinson, M. 2011. The Project Risk Maturity
Thomas Telford. Model: Measuring and Improving Risk Management
Boothroyd, C. & Emmet, J. 1996. Risk Management: Capability. Farnham: Gower.
A Practical Guide for Construction Professionals. Hubbard, D.W. 2009. The Failure of Risk Management:
London: Witherby. Why It’s Broken and How to Fix It. New Jersey: John
Brandl, H. 2004. The Civil and Geotechnical Engineer Wiley & Sons.
in Society: Ethical and Philosophical Thoughts, Chal- ISO 2009a. ISO 31000:2009 Risk management—Princi-
lenges and Recommendations. The Deep Foundations ples and guidelines. Geneva: International Organiza-
Institute, Hawthorne. tion for Standardization.

98

ISGSR2013.indb 98 10/18/2013 9:37:08 AM

 ISO 2009b. ISO Guide 73:2009 Risk management— Van Asselt, M.B.A. & Vos, E. 2006. The precautionary
Vocabulary. Geneva: International Organization for principle and the uncertainty paradox. Journal of Risk
Standardization. Research 9(4): 313–36.
Kahneman, D. & Tversky, A. 1979. Prospect theory: van Staveren, M.Th. 2006. Uncertainty and Ground
An analysis of decision under risk, Econometrica 47: Conditions: A Risk Management Approach. Oxford:
263–291. Butterworth Heinemann.
Knight, F.H. 1921. Risk, Uncertainty, and Profit. Boston, van Staveren, M.Th. 2009a. Extending to geotechnical
MA: Houghton Mifflin. risk management. Georisk 3(3): 174–183.
Latham, M. 1994. Constructing the Team: Joint Review van Staveren, M.Th. 2009b. Risk, Innovation & Change:
of Procurement and Contractual Arrangements in the Design Propositions for Implementing Risk Manage-
UK Construction Industry. London: Department of ment in Organizations. Köln: Lambert Academic
the Environment, HMSO. Publishing.
Mullaly, M. 2006. Longitudinal Analysis of Project Man- van Staveren, M.Th. 2013a. International State of the Art
agement Maturity. Newton Square, PA: Project Man- Report on Integration of Geotechnical Risk Manage-
agement Institute. ment and Project Risk Management. Draft version 1
NASA 2011. NASA Risk Management Handbook., of July 2013 (in review process of TC304-TF3 mem-
NASA/SP 2011-3422, Version 1.0 November 2011. bers). ISSMGE TC304-TF3.
Washington: National Aeronautics and Space Admin- van Staveren, M.Th. 2013b. Geotechnics on the move:
istration. URL: http://www.hq.nasa.gov/office/codeq/ guidance for a risk-driven way of working. Georisk
doctree/NHBK_2011_3422.pdf. (accepted, ready for publication).
Olsson, R. 2006. Managing Project Uncertainty by using van Staveren, M.Th. & Halman, J.I.M. 2013. Key con-
and Enhanced Risk Management Process. Ph.D. thesis, ditions for risk management implementation. Risk
Mälardalen University Dissertations No. 34. Västerås: Analysis (paper in preparation).
Mälardalen University Press. van Staveren, M.Th., Litjens, P.P.T. & Cools, P.M.C.B.M.
Patton, M.Q. 1987. How to Use Qualitative Methods in 2013a. Embedding geo risk management: The Geo-
Evaluation. Newbury Park, CA: Sage. Impuls approach. In: Proceedings of the 18th Interna-
Peck, R.B. 1969. Advantages and limitations of the tional Conference on Soil Mechanics and Geotechnical
observational method in applied soil mechanics. Engineering Paris (in press).
Géotechnique 19: 171–87. van Staveren, M.Th., Litjens, P.P.T. & Cools, P.M.C.B.M.
Power, M. 2009. The risk management of nothing. 2013b. Implementing geo risk management in the
Accounting, Organizations and Society 34(6–7): 849–55. construction industry. In: Proceedings of the 4th Inter-
Raz, T., Shenhar, A.J. and Dvir D. 2002. Risk manage- national Symposium on Geotechnical Safety and Risk
ment, project success, and technological uncertainty. ISGSR2013, Hong Kong (in press).
R & D Management 32(2): 101–9. van Staveren, M.Th., Litjens, P.P.T. & Heerema, J.J.
Roozenburg, N.F.M. & Eekels, J. 1995. Product Design: 2013c. Implementing geo risk management in a cli-
Fundamentals and Methods. Chichester: Wiley. ent organization. In: Proceedings of the 4th Interna-
Schinzinger, R. & Martin, M.W. 2000. Introduction to tional Symposium on Geotechnical Safety and Risk
Engineering Ethics. New York: McGraw-Hill. ISGSR2013, Hong Kong (in press).
Smith, N. 2003. Appraisal, Risk and Uncertainty. London: van Tol, A.F. 2008. Schadegevallen bij bouwputten
Thomas Telford. [Failure Cases of Construction Pits] Cement 6: 6–13.
Smith R. 1996. Allocation of risk: the case for manage- van Well-Stam, D., Lindenaar, F., Van Kinderen, S. &
ability. The International Construction Law Review Van den Bunt, B.P. 2004. Project Risk Management:
4:549–69. An Essential Tool for Managing and Controlling
Sondermann, W. & Trunk, U., 2008. Spezialtiefbau und Projects. London: Kogan Page.
risikomanagement: gibt es einen gemeinsamen Weg? Waring A.E. & Glendon, A.I. 1998, Managing Risk, 1st
[Deep constructions and risk management: is there edition. London: International Thomson Business
a joint way?] In: Beiträge Zum 7. Geotechnik-Tag in Press.
München. Risiko Und Sicherheit in Der Geotechnik, Weatherhead, M., Owen, K. & Hall, C. 2005. Integrating
15.02.2008 München (Deutschland). Presented at the Value and Risk in Construction. London: CIRIA.
7th Geotechnik-Tag, München, pp. 3–16. Weick, K.E. & Sutcliffe, K.M. 2007. Managing the
Sowers, G.F. 1993. Human factors in civil and geotechni- Unexpected: Resilient Performance in an Age of
cal engineering failures. Journal of Geotechnical Engi- Uncertainty, 2nd edition. San Fransisco: John
neering 119(2): 238–56. Wiley & Sons.
Thompson, P. & Perry, J. 1992. Engineering Construction Yin, R.K. 2003. Case Study Research: Design and
Risks. London: Thomas Telford. Methods, 3rd edition. Thousands Oaks: Sage.

99

ISGSR2013.indb 99 10/18/2013 9:37:08 AM

This page intentionally left blank
 Geotechnical Safety and Risk IV – Zhang et al. (eds)
© 2014 Taylor & Francis Group, London, ISBN 978-1-138-00163-3

Is landslide risk quantifiable and manageable?

H.N. Wong
Civil Engineering and Development Department, Geotechnical Engineering Office, Government of the
Hong Kong Special Administrative Region, Hong Kong, China

ABSTRACT: Quantitative Risk Assessment (QRA) enables decision-makers to exercise informed and
rational judgement in evaluating and managing risk. It is commonly acknowledged that Hong Kong has
showcased not only the state-of-the-art in, but also the practicability of, quantification and management
of landslide risk. The author would caution that this is a simplistic view. With reference to Hong Kong’s
experience, it is highlighted that there are different categories of landslide problems, each with its own
factors that combine to give rise to risk. As to whether the risk is quantifiable and manageable, different
categories have their particular circumstances and the answers are not the same. It is credible to manage
landslide risk without quantification, although risk quantification can undeniably aid risk management.
Landslide risk quantification may not be practicable in all cases. Even if landslide risk is quantifiable, it is
important not to lose sight of the degree of uncertainty involved.

1 INTRODUCTION intensely with human activities in urban areas.

Landslides pose particularly grave challenges to
Landslide risk management and Quantitative Risk densely populated built-up areas like Hong Kong
Assessment (QRA) have been undergoing rapid because:
development in the past 20 years. Notable recent
developments and applications of landslide QRA – landslide consequences are serious due to the
are described in Fell & Hartford (1997), Wong et al. close proximity to and the high concentration of
(1997), Ho et al. (2000) and Wong (2005). Sepa- population and vulnerable facilities;
rately, under funding from the European Commis- – urban development, when carried out with-
sion, the integrated research project ‘SafeLand’ has out adequate geotechnical input, results in the
been conducted in Europe by 27 institutions from formation of potentially unstable slopes and
13 countries, aiming at developing generic quanti- increases landslide frequency;
tative risk assessment and management tools and – landslide problems in an urbanised setting are
strategies for landslides. aggravated by human factors, including concen-
The quantitative approach in landslide risk trated surface water flow and localised cutting
assessment and management has gained wider and filling; and
use, both in global and site-specific assessments, – relocation of existing facilities to avoid landslide
as illustrated by successful applications in Hong hazards is often not viable, while at the same
Kong and elsewhere. time landslide prevention and mitigation works
This paper aims to discuss whether landslide risk are difficult and costly given space and access
is quantifiable and how landslide risk is being man- constraints.
aged in practice, with and without quantification. In the early years of Hong Kong’s rapid urbani-
Particular reference is made to the practice and sation, a large number of cut and fill slopes and
experience in Hong Kong, which is recognised as a retaining walls were formed to pave the way for
role model in urban landslide risk management. housing and infrastructure developments. Typi-
cally formed with little geotechnical input at the
time, these unengineered man-made slope features
2 LANDSLIDE AND SLOPE SAFETY
continue to pose landslide risks to the community
MANAGEMENT IN HONG KONG
to this day. Against this background, the Landslip
Prevention and Mitigation (LPMit) Programme
2.1 Landslide problems in Hong Kong
(known as Landslip Preventive Measures, LPM,
While landslides are widespread in many parts Programme before 2010) has been implemented
of the world, landslides in Hong Kong interact by the Geotechnical Engineering Office (GEO,

101

ISGSR2013.indb 101 10/18/2013 9:37:08 AM

 formally the Geotechnical Control Office) of the of priority and partnership for reducing landslide
Civil Engineering and Development Department, frequency and consequence, and (ii) to address
with significant headway made in risk reduction. public attitude and tolerability of landslide risk
The severe weather conditions experienced in to avoid unrealistic expectations. The System
Hong Kong and the increasing urban development also adds value to the society through avert-
close to natural hillsides have also called for atten- ing potential fatalities and improving the built
tion to managing the risk of landslides originating environment.
from natural terrain. The Slope Safety System has proven success-
ful in containing landslide risk within an As Low
as Reasonably Practicable (ALARP) level, via:
2.2 Slope safety management system
(i) improving slope safety standards, technology
Following a number of disastrous landslides with and administrative and regulatory frameworks,
serious fatalities in the 1970s in Hong Kong, the (ii) ensuring safety standards of new slopes,
GEO was established in 1977. Over the years, (iii) rectifying substandard Government slopes and
a comprehensive Slope Safety System has been maintaining them, (iv) ensuring that private own-
developed and implemented by the GEO to com- ers take responsibility for slope safety, and (v) pro-
bat landslide problems. moting public awareness in and response to slope
The key components of the Slope Safety Sys- safety. Besides slope safety, the improved aesthet-
tem in Hong Kong together with their functions ics and ecology of engineered slopes have also con-
in a risk management context are summarised tributed to enhancing the built environment.
in Table 1. A range of initiatives are included in Retrofitting of the old unengineered slopes is
the System to manage landslide risk in a holistic carried out in the pre-2010 LPM Programme and
manner. post-2010 LPMit Programme. The Programmes
The goals of the System are: (i) to minimise serve to systematically assess the stability of
landslide risk to the community through a policy old man-made slopes according to their ranked

Table 1. The slope safety system in Hong Kong (based on Malone 1998).

Primary contribution of each component to

Reduce landslide risk Address public

attitude and
Components of slope safety system Reduce likelihood Reduce consequence tolerability

Policing
– Checking new slope works
– Slope maintenance audits
– Recommending safety clearance
of vulnerable squatters and unauthorised
structures threatened by hillsides
– Exercising geotechnical control through
input in land use planning
– Safety screening studies and
recommending statutory repair orders
for private slopes
Works projects
– Retrofitting substandard government
man-made slopes
– Natural terrain landslide mitigation
and boulder stabilisation works
Research and setting standards
Education and information
– Slope maintenance campaigns
– Risk awareness programmes
and personal precaution campaigns
– Information services
– Landslip warning and emergency services

102

ISGSR2013.indb 102 10/18/2013 9:37:08 AM

 flow or slide or avalanche). Proper classification
prevents fundamentally different landslide types
and their related data from being lumped together.
As different fault trees and event trees are involved
in assessing the risks of different types of land-
slides, mixing them together is not conducive to
meaningful analysis and diagnosis of risk patterns
and trends.
Factors other than physical characteristics
could also be critical to landslide classification.
For example, the landslide susceptibility of cut
slopes formed before and after 1977 in Hong Kong
is markedly different due to different levels of geo-
technical input in their design and construction.
Changes in slope design standards and process-
ing systems applied to slopes should also be taken
into account in the classification of slopes and
Figure 1. Historical annual landslide fatalities in Hong landslides.
Kong. In terms of landslide classification, three nota-
ble categories of landslides have contributed to the
landslide risk in Hong Kong: (i) landslides induced
order of priority and to upgrade substandard by construction activities, (ii) landslides triggered
Government slopes to the required standards. by heavy rain on slopes that do not meet engineer-
ing standards, and (iii) landslides under severe
weather conditions, on slopes that meet engineer-
2.3 Risk trend
ing standards. The answers to the question about
A detailed analysis of the landslide risk trend in whether landslide risk is quantifiable and manage-
Hong Kong is given in Wong (2009a). Two types able are different for each of the categories. Their
of landslide risk trend have been assessed: his- respective risk quantification and management
torical and theoretical. They illustrate how the issues for each class of landslides are explored in
holistic landslide risk management framework the following sections.
has contributed to reducing landslide risk in
Hong Kong.
4 LANDSLIDES INDUCED
Since the 1940s, documentary records have been
BY CONSTRUCTION ACTIVITIES
available in Hong Kong for reliably tracing histori-
cal landslide fatalities. Figure 1 shows the annual
4.1 Description of problem
landslide fatality figures. Although historical fatal-
ities reflect the risk that has actually been realised, Construction activities that are improperly carried
they do not necessarily represent the true (or theo- out may cause failure of both temporary and per-
retical) level of landslide risk, because such figures manent slopes. Slopes that are poorly designed or
are affected by the actual rainfall conditions, spatial built may fail during or soon after its formation.
distribution of heavy rain with respect to that of Construction activities such as unauthorised exca-
existing facilities, near-miss events, etc. QRA has vation may also cause failure of the adjacent slopes
therefore been applied in Hong Kong to quantify even if the slopes have been designed to a high safety
the levels of theoretical landslide risk. The histori- standard. Such failures could involve system defi-
cal and theoretical landslide risk trends in Hong ciency (e.g. poor design guidelines and construction
Kong are further described in Section 5.2 below. practice), malpractice (e.g. unauthorised construc-
tion activities), and human error (e.g. design fault).
Heavy rainfall is often one of the aggravating
3 CLASSIFICATION OF LANDSLIDES factors of construction-induced landslides. The
most notable example in Hong Kong is the Po
Proper classification of landslides is important Shan landslide on 18 June 1972 (Fig. 2), which
for risk assessment and management. Common resulted in the collapse of the 15-storey Kotewall
practice is to classify landslides based on either the Court and 67 fatalities. The cause of the landslide
physical characteristics of the slope (e.g. whether was attributed to poorly executed site formation
man-made or natural; whether cut slope or fill works which undermined the stability of the hill-
slope) or the landslide mechanism (e.g. whether side and resulted in a catastrophic failure when hit
sliding, liquefaction or washout; whether debris by heavy rain.

103

ISGSR2013.indb 103 10/18/2013 9:37:08 AM

 control over private developments is also exercised
by the GEO through the statutory powers of the
Buildings Authority. Design submissions made by
developers or owners are subject to approval before
construction commences.
Geotechnical control needs to be supported by
regulatory and administrative systems to ensure
effective enforcement. Also of importance is good
professional ethics and accountability in uphold-
ing good practice in design and construction.
Given the system in place, extensive landslides
triggered by construction activities are no longer
expected in Hong Kong. However, the possibility
of isolated incidents arising from unforeseen cir-
cumstances or exceptional irregularities at individ-
ual sites cannot be entirely ruled out. Also, smaller
failures may still result from localised unauthorised
construction, e.g. in squatter and village areas.

Figure 2. The Po Shan landslide on 18 June 1972. 5 LANDSLIDES TRIGGERED BY HEAVY

RAIN, ON SLOPES THAT DO NOT MEET
ENGINEERING STANDARDS
4.2 Risk quantification
5.1 Description of problem
The risk of construction-induced landslides is very
difficult to quantify. Such landslides are typically On average, approximately 300 landslides are
the result of multiple causes that are highly vari- reported in Hong Kong every year, most of which
able. In many historical cases, the causes of failure involve failure of substandard man-made slopes
are poorly documented and detailed design infor- triggered by rain when the Landslip Warning is in
mation is not available. As construction practice force. Issue of Landslip Warning is decided by the
evolves with time, so do construction problems. GEO based on rainfall forecasts from the Hong
There is often a lack of reliable data that can be Kong Observatory and the landslide frequency-
consistently analysed for landslide frequency rainfall model established by the GEO using histori-
assessment. Risk quantification of this kind of cal data. Normally three to four Landslip Warnings
failures has therefore not been conducted in Hong are issued in Hong Kong every year. The rainfall
Kong, with the narrow exception of a preliminary concerned is heavy, but usually not extreme.
review of trench excavation-induced failures by Slopes not meeting engineering standards refer
Kwong (2003). While construction-induced land- to unengineered man-made slopes formed with-
slides are also commonplace in other parts of the out proper geotechnical design and control. This
world, to the author’s knowledge, there is hardly occurred prior to the establishment of the GEO
any systematic, comprehensive risk quantification in 1977. In those days, slopes in Hong Kong
carried out. were designed and formed empirically by experi-
ence: cut slopes typically constructed to 55° and
fill embankments to 35° were deemed to be sat-
4.3 Risk management
isfactory. A total of 39,000 sizeable unengineered
Despite the difficulty in risk quantification, the slopes (>3 m high) have been registered in the
risk of construction-induced landslides is nonethe- GEO’s catalogue of man-made slopes. After 1977,
less manageable. The primary factors causing such good practice and regulatory system were put in
landslides are reasonably well understood and can place, which virtually eliminated the formation of
be regulated by improved system and practice. new substandard slopes. However, due to the pres-
This has been done in Hong Kong through the ence of the large number of these old unengineered
geotechnical control of Government and private slopes, they are the principal source of landslide
developments. All permanent geotechnical works risk in Hong Kong after 1977.
including man-made slopes and retaining walls A notable example of failure of unengineered
formed under public works developments are sub- slope triggered by heavy rainfall was the col-
mitted to the GEO for checking (ETWB 2002), lapse of a fill embankment at Sau Mau Ping in
and there is also an established independent check- 1976 where landslide debris punched through the
ing system for temporary works. Geotechnical ground floor of Block 9 of Sau Mau Ping Estate

104

ISGSR2013.indb 104 10/18/2013 9:37:09 AM

 landslide data together with detailed inventory of
slopes and the facilities affected in the event of fail-
ure also enable the development and application of
landslide consequence models for quantifying the
consequence of failure (Wong et al. 1997).
Global QRA has regularly been carried out to
quantify the landslides risks for this category of
slopes, as an integral part of landslide risk man-
agement in Hong Kong. The effect of construc-
tion activities is excluded in the quantification of
rainfall-induced landslides for this category of
slopes, given that landslide risk from construction
activities has been largely managed.
Figure 4 shows the rolling 15-year average
Figure 3. Fill slope failure at Sau Mau Ping on values of the annual fatalities, which depict the
25 August 1976. historical landslide risk trend in Hong Kong.
Figure 5 presents diagrammatically the calculated

killing 18 people (Fig. 3). Subsequent investigation

revealed that the loose sandy fill materials in the
surface 3 m of the slope had undergone liquefac-
tion failure when it became saturated in the heavy
rain (Morgenstern 1978).
Natural hillsides that have marginal stability and
are active in landsliding may also fall into this cat-
egory. From the interpretation of the comprehen-
sive inventory of historical aerial photographs, it is
found that an average of about 300 natural terrain
landslides occur on the natural terrain in Hong
Kong each year. However, most of these landslides
do not result in any major consequences and are
not reported to the GEO.

5.2 Risk quantification

The landslide risk associated with this kind of
slopes can be more readily quantified, provided Figure 4. 15-year rolling average of historical annual
that comprehensive data are available on: landslide fatalities in Hong Kong.
i. historical landslides, their locations, time, scale
and consequence; and
ii. the characteristics of the slopes, i.e. their type,
size, gradient, etc.
The above conditions are favourably met in Hong
Kong, which renders risk quantification practica-
ble. Comprehensive and good quality failure data
are available, and all existing, sizeable man-made
slopes have been registered and surveyed. In Hong
Kong, the instability problem of this category of
slopes is principally controlled by the deficiency of
empirical-based design, which displays a probabilis-
tic pattern of failure associated with: (i) presence of
geological weaknesses, and (ii) occurrence of heavy
rain. Historical landslide data, together with local
knowledge and information on ground and rain-
fall conditions, provide a good basis for analysing
both of these factors, and thereby establishing the
frequency of landslides. The numerous historical Figure 5. Landslide risk trends in Hong Kong.

105

ISGSR2013.indb 105 10/18/2013 9:37:09 AM

 landslide risk trend for this category of slopes in On natural terrain, the distribution of risk
Hong Kong, as assessed by global QRA. amongst different types of catchments has been
In 1977, territory-wide geotechnical control assessed, including open slopes, topographic
was introduced by the GEO to ensure that newly depression and channelised debris flow (Wong
formed slopes in Hong Kong are designed and built et al. 2006, Cheng & Ko 2010). These results have
to the required standards of safety. Since then, the been instrumental in formulating the post-2010
increasing risk trend has levelled off, i.e. turned LPMit Programme, which implements risk mitiga-
from line BC to line BD in Figure 5, despite that tion works for 30 natural hillside catchments annu-
an additional more than 20,000 sizeable man-made ally apart from retrofitting old man-made slopes.
slopes have been formed as a result of urban devel- Global QRA has also provided useful figures
opment in Hong Kong. Since 1977, the outcome of on the amount of risk reduction achieved and the
the retrofitting effort under the LPM Programme is effectiveness of risk management effort, which aids
progressive risk reduction along Line BE (Fig. 5). risk communication.
From a global QRA, the calculated overall landslide Site-specific QRA, on the other hand, facili-
risk in year 2000 is 50% of that in 1977 (Cheung & tates a rational consideration of risk tolerability
Shiu 2000), i.e. Point X in Figure 5. It was also and evaluation of risk mitigation strategy at a
found that by year 2010, the retrofitting works particular locality. This is where state-of-the-art
should have brought the overall risk level to within quantified risk management is usefully applied.
25% of that in 1977 (Cheng 2011), i.e. Point E. For example, a natural terrain landslide hazard
The global QRA conducted on the unengineered study has been carried out at Sha Tin Heights,
slopes in Hong Kong has given results with good Hong Kong (Fig. 6) following six landslides in
reliability, as is also illustrated by the good match 1997, three of which developed into debris flows
of the calculated risk figures with the historical that affected residential buildings. QRA was used
landslide risk trend (Fig. 4). to evaluate the acceptability of risk levels (Fig. 7)
Wong (2005) also describes the application of and the cost effectiveness of risk mitigation strat-
QRA at individual sites in Hong Kong, for the egies. Risk mitigation works were completed in
quantification and management of the risk at site- 2004.
specific level. In general, given the use of the same In parallel, besides landslip prevention and miti-
set of data, the reliability of the risk quantifica- gation works, non-works measures such as pub-
tion would decrease from global to site-specific lic education, landslide warning, etc. have helped
QRA. Hence, in practice, site-specific QRA nor- reduce the consequence of landslides. The GEO
mally calls for more detailed information and runs a public education and publicity campaign on
assessment. slope safety to promulgate the importance of slope
maintenance and educate the public on personal
precautionary measures to be taken during heavy
5.3 Risk management
rainstorms. The GEO also provides a 24-hour
Hong Kong’s Slope Safety System is instrumental emergency service to attend to landslip incidents
in its success in managing landslide risks arising and to advise on emergency and follow-up actions,
from old substandard slopes. This has however
not come easily. The process has called for a strong
commitment in resources, time and strategy. In
contrast with forming new slopes that are up to
standard, retrofitting existing substandard ones is
resource-demanding and is subject to many other
engineering, environmental and social constraints.
It takes a long time to achieve notable results.
A system manager, a role taken on by the GEO, is
required to devise good strategy and organisation,
involving the compilation of catalogues, prioritisa-
tion, systematic study and retrofitting, etc.
Global QRA has been successfully used to
identify the scale of the landslide problem and to
facilitate the formulation of strategy and prioriti-
sation. It has formed the basis for resource allo-
cation and assessment of relative priority between
different slope types under the LPM and LPMit
Programmes, as incorporated into the prevailing Figure 6. Residential buildings affected by landslide
risk-based priority ranking system. hazards at Sha Tin Heights (after Wong & Ho 2005).

106

ISGSR2013.indb 106 10/18/2013 9:37:10 AM

 such as building evacuation and road closure, to (Line BE) that accounts only for the effect of slope
minimise the impact of landslides on the public. retrofitting works. Over the past 20 years, the actual
The non-works initiatives form an integral part annual landslide fatalities in Hong Kong have been
of holistic landslide risk management. They have consistently less than the theoretical risk level by at
proven to be effective in reducing the landslide risk least 50%. The overall risk trend suggests that the
to the ALARP level. This is shown diagrammati- contribution of the non-works initiatives to reduc-
cally as Line BF in Figure 5, reflecting a further tion of landslide risk in Hong Kong could be fairly
reduction in landslide risk from the theoretical level significant. As an illustration, in the past 20 years,
the number of Landslip Warnings issued, landslide
incidents reported, buildings and people tempo-
rarily evacuated, and sections of road closed as a
result of the landslide emergency responses of the
GEO and the community are shown in Table 2.

6 LANDSLIDES UNDER SEVERE

WEATHER CONDITIONS,
ON SLOPES THAT MEET
ENGINEERING STANDARDS

6.1 Description of problem

An account of the key lesson learnt from the sys-
tematic landslide investigation in Hong Kong that
engineered slopes still have a finite chance of fail-
ure is given in Wong & Ho (2003). Such failures
Figure 7. Comparison of societal risk in terms of
F-N pairs with societal risk criteria at Sha Tin Heights. are typically related to the progressive degrada-
societal risk was reduced from 5.67 × 10−3 PLL/year to tion of man-made slopes under repeated rainfall
4.44 × 10−3 PLL/year following the implementation of conditions or to geological defects not identified
risk mitigation measures (after Fugro Maunsell Scott in slope design by the current state of knowledge
Wilson JV 2004). and practice.

Table 2. Breakdown of landslides affecting different facilities in Hong Kong since 1994.

No. of buildings* evacuated

No. of landslip No. of reported No. of sections of
Year warning issued landslide incidents Block House Flat/unit road affected/closed

1994 4 436 4 24 207 225

1995 6 295 5 8 170 163
1996 2 153 0 7 58 87
1997 8 491 0 19 176 253
1998 1 216 0 6 43 120
1999 3 402 4 10 128 166
2000 6 322 0 3 15 184
2001 8 214 0 6 48 98
2002 3 138 0 7 5 73
2003 1 201 0 2 3 98
2004 2 69 0 0 3 18
2005 2 481 7 4 46 219
2006 4 193 0 1 8 85
2007 1 83 0 0 4 36
2008 5 849 1 22 80 420
2009 2 124 0 0 2 45
2010 3 266 0 0 7 93
2011 0 99 0 0 0 27

*A ‘block’ is a multi-storey building, which may comprise up to several dozen of flats/units. A ‘house’ is typically within
3 storeys, which comprises several flats/units.

107

ISGSR2013.indb 107 10/18/2013 9:37:11 AM

 Between 1997 and 2002, 106 landslide incidents A total of 370 mm rainfall was recorded at the
involved man-made slopes with past geotechnical site within 24 hours before the landslide, and
engineering input and geotechnical design submis- the 31-day rainfall of 1303 mm exceeded the
sions checked and accepted under the slope safety highest calendar monthly rainfall ever recorded
system, 24 of which were major failures (i.e. ≥50 m3 by the rain gauge at the Hong Kong Observa-
in volume). All the 106 failures were subject to the tory since records began in 1884. The land-
GEO’s detailed landslide investigation to enable slide consisted of a translational failure with
a systematic diagnosis of the probable causes of the detached ground mass sliding on a surface
failure. Of the 106 cases, 53 affected engineered soil dipping gently out of the slope. Large failures
cut slopes, 15 of which were major failures. All these of this type are unusual in Hong Kong. The
53 failures involved unsupported soil cuts, with no post-failure investigation established that the
structural support such as soil nails or earth retain- basal slip surface of the landslide developed
ing structures. The comprehensive analysis showed along a laterally extensive (>50 m) weak layer
such engineered man-made slopes that were formed of kaolinite-rich altered tuff, which was about
with old technology before the 1990s, known as 15 m below the crest of the cut slope and dip-
‘old technology slopes’ in Hong Kong, are not suf- ping out of the slope at about 10° to 25°. The
ficiently robust in withstanding degradation and failure caused one fatality.
could fail under severe rainfall conditions. ii. The adverse effect of significant geological
During the severe rainstorms between May and defects on slopes previously subjected to engi-
July 2008, 28 landslides occurred on engineered neering studies was evident in the Shek Kip Mei
man-made slopes, of which two were major fail- landslide of 25 August 1999 (Fugro Maunsell
ures (Li et al. 2012). The one at Pak Fuk Road Scott Wilson JV 2000). The 4-day rainfall
was an unsupported soil cut slope formed between before the landslide was most severe, with a
1976 and 1980. The one at Tsing Yi Road involved total of 641 mm of rain and was the heaviest
a wedge failure in its rock portion. These failures recorded since slope formation. The slope dis-
on engineered slopes are typical of the residual placed forward by about 1 m at the slope toe.
landslide risk problem associated with old technol- The total volume of the displaced mass was
ogy slopes in Hong Kong. about 6000 m3. A laterally-persistent (over 60 m
Unengineered slopes have a much higher failure long) discontinuity dipping at a shallow angle
frequency than engineered ones, and they tend to out of the slope formed the basal plane of the
show instability under much less intense rainfall. southern part of the landslide. This discontinu-
In contrast, the risk of failure of engineered slopes ity was infilled with polished, slickensided kao-
tends to become evident under severe rainfall con- lin and manganese oxide deposits up to 15 mm
ditions. The dire consequences of such failures can thick only (Fig. 9), which was not mapped in
be illustrated by the following two examples: the past geotechnical studies. As a result of the
landslide, two public housing blocks were per-
i. The 13 August 1995 landslide at Fei Tsui Road
manently evacuated.
(GEO & Knill 1996) involved a large-scale fail-
ure (14,000 m3) of a 27 m high engineered cut
Natural terrain in Hong Kong also exhibits
slope that had no structural support (Fig. 8).
instability that resembles this category of landslide
problem, in that natural terrain tends to respond
vigorously to severe rainfall. Ko (2005) established
from detailed rainfall and natural terrain landslide
correlations that the density of natural terrain
landslides increases exponentially with rainfall
intensity (Fig. 10). Under moderate to heavy rain-
fall conditions, there are generally few failures on
natural terrain in Hong Kong. However, when
rain becomes severe, e.g. when 20% or more of the
average annual rainfall is recorded within 24 hours,
widespread natural terrain landslides tend to occur.
For instance, in the June 2008 rainstorm, Lantau
Island was subjected to a maximum of 623 mm
rainfall in 24 hours, with a statistical return period
of about 200 years. The 4-hour rainfall of 384 mm
in the rainstorm was most severe, which measured
Figure 8. Failure of cut slope at Fei Tsui Road on over 1000-year return period. More than 2400
13 August 1995. natural terrain landslides occurred during the

108

ISGSR2013.indb 108 10/18/2013 9:37:11 AM

 Figure 11. Natural terrain landslides at Tung Chung,
Lantau Island during the June 2008 rainstorm.

6.2 Risk quantification

Quantifying the landslide risk under severe weather
conditions is difficult and is subject to significant
uncertainty. First, only limited data are available
on the landslide frequency in severe rainfall condi-
tions, given the limited observation period for rela-
tively rare events. For instance, the current natural
terrain landslide inventory maintained by the GEO
is primarily based on low-level aerial photographs
which reveal historical natural terrain landslides
over the past 60 years. Caution is needed when
using the historical data as a basis for assessing
extreme, or low-frequency-high-magnitude events
by simple projections. Large and mobile landslides
Figure 9. Failure of cut slope at Shek Kip Mei on are potentially underrepresented.
25 August 1999 (top); laterally persistent discontinuity Second, the quantification of landslides occur-
infilled with slickensided kaolin and manganese oxide ring during severe weather conditions requires
deposits (bottom). extensive extrapolation, which increases uncer-
tainty. There is little information available on
the presence and extent of geological defects
that could subvert slope stability, even for man-
made slopes that have been properly engineered.
The phenomenon and process of slope degradation
are not well understood. There is limited knowl-
edge of how landslide scale and mobility increase
under extreme rainfall conditions.
Third, effects of climate changes could alter
both the frequency of occurrence as well as the
severity of extreme rainfall conditions. However,
this is an area awaiting further study and analy-
sis to ascertain the effects, their potential influence
on the severity and frequency of extreme rainfall,
and how slope instability would respond to such
changes.
In short, risk quantification may be attempted
for landslide risk under severe weather condi-
tions but the findings should be interpreted with
Figure 10. Correlation between rainfall and natural care and with due consideration of the significant
terrain landslide density.
uncertainty involved.

6.3 Risk management

rainstorm (Fig. 11). Furthermore, it was evident
that the scale of the natural terrain landslides and Managing landslide risks that may be brought
the mobility of the landslide debris also increased about by severe weather conditions poses some of
drastically under such extreme rainfall (Wong the greatest challenges for the geotechnical profes-
2009b). sion today. The difficulty is two-fold. Firstly, the

109

ISGSR2013.indb 109 10/18/2013 9:37:12 AM

 uncertainties in risk quantification reflect the lack Proper classification, including consideration of
of understanding of the nature and severity of the both the physical and system aspects of the land-
problem, which hinders the formulation of risk slide hazards, is important in risk quantification
management strategy and deployment of resources. and management. Different types of landslides
Secondly, given the present state of knowledge and should not be lumped together in analysis. The
technology, pragmatic solutions to curb landslides availability of good documentation and data of
in extreme weather conditions may not always be historical landslides and slopes are prerequisites to
available. QRA.
For man-made slopes, robust design solutions The main challenge in risk quantification often
have already been adopted in Hong Kong. This lies with the assessment of landslide frequency
includes the use of structural support such as soil (including the scale and mobility of failure). Con-
nailing and prescriptive measures such as enhanced sequence assessment is generally less difficult,
drainage provisions. Engineered slopes formed although existing consequence models do warrant
with old technology could be vulnerable to failure further review of their applicability and reliability
in severe rainfall conditions. However, it is notable under different circumstances. This is beyond the
that no mobile, high consequence or major failures scope of this paper.
have occurred on soil nailed cut slopes in Hong It is possible to manage landslide risks with-
Kong to date. This indicates the robustness of soil out quantification, but risk quantification can
nailed systems in preventing large scale instabil- undeniably aid risk management. The problem
ity even in the face of severe weather conditions. of construction-induced landslides in Hong Kong
Minor failures do occur, and these may be reduced illustrates how the multiple uncertainties involved
by improved slope surface protection and drainage could render risk quantification difficult, even
measures (Ng et al. 2008). though there are a wealth of data, good under-
For natural terrain landslides, on the other hand, standing of landslide mechanisms and ample expe-
existing technology has limited capability in coping rience in use of QRA.
with the hazards that may arise under severe rain- Where risks cannot be quantified with reason-
fall conditions. There is thus a pressing need for able confidence, it is difficult to apply the quanti-
the development of better mitigation measures and fied risk assessment and management framework.
innovative solutions. Some thoughts on enhanced However, the relevant risk assessment principles
risk management and mitigation strategy for natu- and considerations are still useful in general terms.
ral terrain arising from observations and lessons Applied in a qualitative manner together with
learnt from the June 2008 rainstorm are given in expert judgment, they can still facilitate manage-
Wong (2009b). For instance, in dealing with size- ment of risk.
able debris flows exceeding several thousand cubic In Hong Kong, the risk of rainfall-induced
metres, there may be a need to explore the use of landslide on old, unengineered slopes has been
multiple layers of debris resisting barriers rather reduced to a low level. As the risk level reduces,
than a single layer. risk quantification becomes increasingly sensitive
Non-works options will also continue to feature to the uncertainties involved. In the past, this cat-
prominently in the management of landslide risk egory of landslides was the predominant compo-
under severe weather conditions, including public nent of the overall landslide risk in Hong Kong.
education, landslip warning and emergency prepar- Upon reduction of the risk to a relatively low
edness as discussed earlier. At the same time, closer level, other factors have come into play, e.g. the
collaboration with the meteorological profession is effect of construction activities and risk associated
required in assessing the effects of climate change with extreme weather conditions, which originally
and its implications in slope safety. would not contribute significantly to the overall
landslide risk.
Therefore, it may become more difficult and
7 CONCLUDING REMARKS unreliable to further quantify the actual risk reduc-
ON RISK QUANTIFICATION tion that could be achieved by a certain risk man-
AND MANAGEMENT agement initiative, when the risk is at a low level
already and is sensitive to changes associated with
A range of factors affect the practicability of risk other factors.
quantification. These need to be evaluated before Viewed from another perspective, containing
deciding whether or not it is practicable to perform risk within the ALARP level calls for continual
QRA and whether or not risk quantification can effort to curb various factors that may increase
be expected to give reliable results. Even where it risk. This in turn requires risk management efforts
is possible to quantify the risk, it is important not which may not necessarily achieve a noticeable or
to lose sight of the degree of uncertainty involved. quantifiable risk reduction.

110

ISGSR2013.indb 110 10/18/2013 9:37:13 AM

 ACKNOWLEDGEMENT Li, A.C.O., Lau, J.W.C., Cheung, L.L.K. & Lam, C.L.H.
2012. Review of landslides in 2008. GEO Report No.
This paper is published with the permission of the 274. Hong Kong: Geotechnical Engineering Office.
Director of Civil Engineering and Development, Malone, A.W. 1998. Risk management and slope safety
in Hong Kong. In K.S. Li, J.N. Kay & K.K.S. Ho
Government of the Hong Kong Special Administra- (eds), Slope Engineering in Hong Kong; Proc. 16th
tive Region. Thanks are due to Mr Eugene K.L. Wong Annual Seminar of the HKIE Geotechnical Division,
for his assistance in preparing this paper. Hong Kong, 2 May 1997: 3–17. Rotterdam: Balkema.
Morgenstern, N.R. 1978. Mobile soil and rock flows.
Geotechnical Engineering 9: 123–141.
REFERENCES Ng, A.F.H., Lau, T.M.F., Shum, K.W. & Cheung, W.M.
2008. Review of selected landslides involving soil-
Cheng, P.F.K. 2011. Assessment of landslide risk posed nailed slopes. GEO Report No. 222. Hong Kong:
by man-made slopes as of 2010. GEO Special Project Geotechnical Engineering Office.
Report No. SPR 1/2011. Hong Kong: Geotechnical Wong, H.N. 2005. Landslide risk assessment for indi-
Engineering Office. vidual facilities. Landslide Risk Management; Proc.
Cheng, P.F.K. & Ko, F.W.Y. 2010. An updated assess- intern. conf., Vancouver, 31 May–3 June 2005: 237–296.
ment of landslide risk posed by man-made slopes and London: Taylor & Francis.
natural hillsides in Hong Kong. GEO Report No. 252. Wong, H.N. 2009a. Holistic urban landslide risk
Hong Kong: Geotechnical Engineering Office. management—challenges and practice. Keynote Lec-
Cheung, W.M. & Shiu, Y.K. 2000. Assessment of global ture at the 7th Asian Regional Conference of the Inter-
landslide risk posed by pre-1978 man-made slope fea- national Association for Engineering Geology and the
tures: risk reduction from 1977 to 2000 achieved by Environment, Chengdu, 9–11 September 2009.
the LPM Programme. GEO Report No. 125. Hong Wong, H.N. 2009b. Rising to the challenges of natural
Kong: Geotechnical Engineering Office. terrain landslides. Natural Hillsides: Study and Risk
ETWB 2002. Geotechnical control for slopes and retain- Management Measures; Proc. 29th Annual Seminar of
ing walls. ETWB Technical Circular (Works) No. the HKIE Geotechnical Division, Hong Kong, 17 April
29/2002. Hong Kong: Environment, Transport and 2009: 15–53. Hong Kong: Hong Kong Institution of
Works Bureau. Engineers.
Fell, R. & Hartford, D. 1997. Landslide risk management. Wong, H.N. & Ho, K.K.S. 2003. Lessons learnt from
In D. Cruden & R. Fell (eds), Landslide Risk landslides. Case Histories in Geotechnical Engineer-
Assessment: 51–110. Rotterdam: Balkema. ing in Hong Kong; Proc. 23rd Annual Seminar of the
Fugro Maunsell Scott Wilson JV 2000. Report on the HKIE Geotechnical Division, Hong Kong, 9 May 2003:
Shek Kip Mei Landslide of 25 August 1999. Hong 255–263. Hong Kong: Hong Kong Institution of
Kong: Geotechnical Engineering Office. Engineers.
Fugro Maunsell Scott Wilson JV 2004. Detailed study of Wong, H.N. & Ho, K.K.S. 2005. Landslide risk man-
the hillside area below Sha Tin Heights Road. GEO agement and slope engineering in Hong Kong. Proc.
Report No. 143. Hong Kong: Geotechnical Engineer- State-of-the-Practice of Geotechnical Engineering
ing Office. in Taiwan and Hong Kong, Hong Kong, 20 January
GEO & Knill, J. 1996. Report on the Fei Tsui Road Land- 2005: 97–136. Hong Kong: Hong Kong Institution of
slide of 13 August 1995. Hong Kong: Geotechnical Engineers.
Engineering Office. Wong, H.N., Ho, K.K.S. & Chan, Y.C. 1997. Assessment
Ho, K.K.S., Leroi, E. & Roberds, W. 2000. Quantitative of consequence of landslides. In D. Cruden &
risk assessment—application, myths and future R. Fell (eds), Landslide Risk Assessment; Proc. intern.
direction. Geotechnical and Geological Engineering workshop, Honolulu, 19–21 February 1997: 111–149.
GeoEng2000; Proc. intern. conf., Melbourne, 19–24 Rotterdam: Balkema.
November 2000, vol. 1: 269–312. USA: Technomic. Wong, H.N., Ko, F.W.Y. & Hui, T.H.H. 2006. Assessment
Ko, F.W.Y. 2005. Correlation between rainfall and natu- of landslide risk of natural hillsides in Hong Kong.
ral terrain landslide occurrence in Hong Kong. GEO GEO Report No. 191. Hong Kong: Geotechnical Engi-
Report No. 168. Hong Kong: Geotechnical Engineer- neering Office.
ing Office.
Kwong, J.S.M. 2003. A Study on Past Failures of Trench
Excavations. GEO Report No. 137. Hong Kong:
Geotechnical Engineering Office.

111

ISGSR2013.indb 111 10/18/2013 9:37:13 AM

This page intentionally left blank
 3 Geotechnical uncertainty and variability

ISGSR2013.indb 113 10/18/2013 9:37:13 AM

This page intentionally left blank
 Geotechnical Safety and Risk IV – Zhang et al. (eds)
© 2014 Taylor & Francis Group, London, ISBN 978-1-138-00163-3

Quantifying epistemic uncertainty and aleatory variability

of Newmark displacements under scenario earthquakes

W. Du & G. Wang
Department of Civil and Environmental Engineering, The Hong Kong University of Science and Technology,
Clear Water Bay, Hong Kong SAR, China

ABSTRACT: Earthquake-induced slope displacement is an important parameter for safety evaluation

and design of slope systems. Many Newmark displacement models have been recently developed using
various ground motion databases and different intensity measures as predictors. In this study, the epis-
temic uncertainties among different models and aleatory variability of the predicted Newmark displace-
ments are quantified. The epistemic uncertainty is represented by the variation between different model
predictions. The standard deviation of the uncertainty is approximately 0.5–1 in natural log scale for
scenario earthquakes of Mw 5.5–7.5, which is much larger than that of the ground motion prediction
equations. This indicates further development of the Newmark displacement models is much needed. The
total aleatory variability considering both GMPEs and displacement models is roughly 1.5–2.5 in natural
log scale. The large epistemic uncertainty and aleatory variability imply that it is extremely important to
account for both in seismic slope analysis.

1 INTRODUCTION moment magnitude of the earthquake (M), Peak

Ground Velocity (PGV) etc. Arias intensity (Ia)
Newmark displacement model is commonly used is also an important predictors, as it incorporates
to estimate the seismic performance of slopes dur- cumulative effect of an acceleration time history
ing earthquake (Newmark, 1965). The Newmark and represent a measure of earthquake energy.
displacement model assumes the sliding mass is Arias Intensity is calculated by the following equa-
rigid, and sliding occurs on a predefined interface. tion (Travasarou & Bray, 2003):
The critical acceleration (ac) represents the resist-
π ttot
a (t ) dt
2 g ∫0
ance of the slope against sliding. It can be deter- 2
Ia = (1)
mined by the strength of material and the slope
angle etc. Sliding is initialized when the shaking
acceleration exceeds the critical acceleration, and where g is the acceleration of gravity, a(t) is the
the block displaces plastically along the interface. acceleration time history, and ttot is the total dura-
The permanent displacement D is calculated by tion of the time history.
double integrating the exceeded accelerations Because different database and functional forms
with respect to time (Fig. 1). Although the simple are used in developing these empirical prediction
rigid-plastic model does not consider the defor- models, it is not surprising that these models would
mation of the block itself during shaking, this yield different predicted results. Uncertainty asso-
method has been widely used to evaluate earth- ciated with the Newmark displacement models can
quake-induced displacement for natural slopes be classified as two categories: epistemic uncer-
(Jibson 2007). tainty and aleatory uncertainty. Epistemic uncer-
Empirical equation to estimate the Newmark tainty is due to lack of knowledge. In principal, it
displacement was first proposed by Ambraseys & can be reduced by using sufficient data or improved
Menu (1988) as a function of the critical accel- regression techniques. The epistemic uncertainty
eration ratio (critical acceleration ac over the Peak can be approximately evaluated by the variation of
Ground Acceleration PGA). Throughout years, different model predictions. Aleatory variability,
many researchers (e.g., Jibson, 2007; Saygili & on the other hand, represents inherent random-
Rathje, 2008; Hsieh & Lee, 2011) have proposed ness that can not be reduced. Aleatory variability
various empirical equations using various ground is usually quantified by variation of the observed
motion Intensity Measures (IMs) as predictors, data against the model prediction. Recently,
including Peak Ground Acceleration (PGA), Douglas (2010, 2012) studied the consistency of

115

ISGSR2013.indb 115 10/18/2013 9:37:13 AM

 3. [PGA, PGV] RS08 model (Saygili & Rathje
2008):

2
⎛ a ⎞ ⎛ a ⎞
ln D = −1.56 − 4.58 ⎜ c ⎟ − 20.84 ⎜ c ⎟
⎝ PGA
G ⎠ ⎝ PGA
G ⎠
3 4
⎛ a ⎞ ⎛ a ⎞
+ 44.75 ⎜ c ⎟ − 30.5 ⎜ c ⎟
⎝ PGA
G ⎠ ⎝ PGA
G ⎠
− 0.64 ln( PG ) + 1.55 ln(l ( )
σ lln D = 0.411 + 0.52 ( ac /P
PGA
P ) (4)

4. [PGA, Ia, PGV] RS08 model (Saygili & Rathje

2008):

2
⎛ a ⎞ ⎛ a ⎞
ln( ) = −0.74 − 4.93 ⎜ c ⎟ − 19.91⎜ c ⎟
Figure 1. Illustration of Newmark displacement ⎝ PGA
G ⎠ ⎝ PGA
G ⎠
3 4
with critical acceleration ac = 0.2 g. (A) Earthquake ⎛ a ⎞ ⎛ a ⎞
acceleration-time history. (B) Velocity of sliding block ver- + 43.75 ⎜ c ⎟ − 30.12 ⎜ c ⎟
sus time. (C) Displacement of sliding block versus time.
⎝ PGA
G ⎠ ⎝ PGA
G ⎠
− 1.3 ln( PGA
G ) + 1.04
04 l ( ) + 0.67 ln( Ia )
σ ln D = 0.2 + 0.79 ( ac /PGA )
Ground Motion Prediction Equations (GMPEs) (5)
developed in the past four decades for IMs such
as PGA, PGV, Ia, etc. However, the epistemic 5. [PGA, M] RS09 model (Rathje & Saygili,
and aleatory uncertainties of the Newmark dis- 2009):
placement models have not yet been thoroughly
2
investigated. ⎛ a ⎞ ⎛ a ⎞
This study aims at quantifying the epistemic ln( ) = 4.89 − 4.85 ⎜ c ⎟ − 19.64 ⎜ c ⎟
⎝ PGA
G ⎠ ⎝ PGA
G ⎠
uncertainty and aleatory variability of Newmark 3 4
displacement models. Ten recently developed ⎛ a ⎞ ⎛ a ⎞
+ 42.49 ⎜ c ⎟ − 29.06 ⎜ c ⎟
Newmark displacement models are used in this ⎝ PGA
G ⎠ ⎝ PGA
G ⎠
study and listed as follows: + 0.72 ln( G ) + 0.89( − 7)
w
1. [PGA] RS08 model (Saygili & Rathje, 2008):
7 ( ac /P
PGA ) 0.54 ( ac /P ) 2
σ lln D = 0.73 + 0.79 PGA PGA
P
2 (6)
⎛ a ⎞ ⎛ a ⎞
ln D = 5.52 − 4.43 ⎜ c ⎟ − 20.39 ⎜ c ⎟
⎝ PGA
G ⎠ ⎝ PGA
G ⎠
3 4
6. [PGA] J07 model (Jibson, 2007):
⎛ a ⎞ ⎛ a ⎞
+ 42.61⎜ c ⎟ − 28.74 ⎜ c ⎟
⎝ PGA
G ⎠ ⎝ PGA
G ⎠ ⎛ a ⎞
+ 0.72 ln( PGA
G ) log10 ( D ) .215 + 2.
.215 log10 ⎜1 − c ⎟
⎝ PGAG ⎠
σ lln D = 1.13 (2) ⎛ ac ⎞
− 1. log10 ⎜ ⎟
⎝ PGA
G ⎠
2. [PGA, Ia] RS08 model (Saygili & Rathje, σ log10 D = 0.51 (7)
2008):

2 7. [PGA, M] J07 model (Jibson, 2007):

⎛ a ⎞ ⎛ a ⎞
ln( ) = 2.39 − 5.24 ⎜ c ⎟ − 18.78 ⎜ c ⎟
⎝ PGA
G ⎠ ⎝ PGA
G ⎠
3 4 ⎛ a ⎞
⎛ a ⎞ ⎛ a ⎞ log10 ( D ) .71 + 2. log10 1 − c ⎟
+ 42.01⎜ c ⎟ − 29.15 ⎜ c ⎟ ⎝ PGA G ⎠
⎝ PGA
G ⎠ ⎝ PGA
G ⎠
⎛ ac ⎞
− 1.56 ln( G ) + 1.38 ln(
l ( ) − 1. log10 ⎜ ⎟ + 0.424 Mw
⎝ PGA
G ⎠
5 ( ac /P
σ lln D = 0.466 + 0.56 PGA )
P (3) σ llog10 D = 0.454 (8)

116

ISGSR2013.indb 116 10/18/2013 9:37:14 AM

 8. [Ia] J07 model (Jibson, 2007):

log10 ( D)) .401log10 ( Ia)) .481log10 (ac)

.481 .23
σ llog10 D = 0.656
(9)
9. [PGA, Ia] J07 model (Jibson, 2007):

log10 ( D)) .561log10 ( Ia )

⎛ a ⎞
− 3. log10 ⎜ c ⎟ − 1.474
⎝ PGA
G ⎠
σ llog10 D = 0.616 (10)

10. [Ia] HL11 model (Hsieh and Lee, 2011):

log10 ( D)) .847 log10 ( Ia)

a) .62ac
.62
Figure 2. Median predicted displacements by vari-
+ 6. c log10 ( ) + 1.84 ous Newmark displacement models for (a) Mw 7.5
σ llog10 D = 0.295 (11) (b) Mw 6.5 (c) Mw 5.5 earthquake on a strike-slip fault.
Note: for each scenario, the average of the predicted
median IMs are used as input parameters.
Among these equations, the unit of the
Newmark displacement D is cm; PGA and ac are
in the unit of g; PGV is in the unit of cm/s; and Ia displacements by different Newmark displacement
is in the unit of m/s. [PGA] RS08, [PGA] J07, [Ia] models vary significantly for a given scenario.
J07, [Ia] HL11 models employs only a single IM For example, the estimated median displacement
as the predictor. Therefore, they are called scalar ranges from 4 cm to 24.6 cm for Mw 7.5 at the
models. The other models employ a combination rupture distance of 10 km. As rupture distance
of more than one IMs (PGA, Ia or PGV), and they increases, all the predicted sliding displacement
are called vector models. decreases as expected. Epistemic uncertainty is
expected to be higher if the predicted values among
these models are more scattered. From the figures,
2 EPISTEMIC UNCERTAINTY OF the predicted displacements (in log scale) are more
NEWMARK DISPLACEMENT MODELS scattered as rupture distance increases, especially
for smaller magnitude events. This is mainly due
In this study, epistemic uncertainty of the above to specific functional forms chosen by different
Newmark displacement models will be studied. models. Some models decays much faster as rup-
Earthquake scenarios considered in the analysis ture distance increases. Since very small displace-
are moment magnitude (Mw) 5.5, 6.5 and 7.5 events ment values are not of engineering significance,
on a strike-slip fault. The site condition is assumed it is more rational to focus only on large displace-
to be a stiff soil site with Vs30 = 400 m/s. Four Next ment amplitudes. Figure 3 shows the standard
Generation Attenuation (NGA) GMPEs are used to deviations of the predicted median displacements
predict PGA and PGV (Abrahamson & Silva, 2008; by these models for three scenarios. The standard
Boore & Atkinson, 2008; Campbell & Bozorgnia, deviations among predicted values are in the range
2008; Chiou & Youngs, 2008). The GMPEs pro- of 0.5–1 (in natual log scale) for all three scenarios.
posed by Travasarou & Bray (2003), Foulser- The Mw 6.5 scenario shows the smallest epistemic
Piggott & Stafford (2012) and Campbell & uncertainty and the Mw 5.5 event has the largest
Bozorgnia (2012) are chosen for predicting Ia. uncertainty, which is constantly greater than 1 if
In order to minimize the influence of epistemic all Newmark displacement models are considered.
uncertainty existed in the GMPEs, the predicted This is not unexpected since that the number of
median IMs are averaged as the input to predict larger magnitude events (Mw > 6) is usually domi-
the median Newmark displacements. By this way, nant in the ground motion databases that are used
the epistemic uncertainty and aleatory variability to develop the Newmark displacement models.
in the GMPEs are neglected. In addition, the epistemic uncertainty among six
Figure 2 shows the predicted median sliding vector models are much smaller than that of four
displacements versus rupture distances assum- scalar models for all cases considered, as shown in
ing ac = 0.1 g and Mw equals 7.5, 6.5 and 5.5, Figure 3. In general, the standard deviation of vec-
respectively. It can be observed that the predicted tor models is in the range of 0.4–0.6 (in natural log

117

ISGSR2013.indb 117 10/18/2013 9:37:16 AM

 Figure 3. Standard deviations of the median Newmark Figure 4. Variation of the displacements predicted by
displacements (in natural log scale) predicted by the ten various GMPEs and Newmark displacement models, for
Newmark displacement models: (a) For a Mw 7.5 strike- (a) Mw 7.5 (b) Mw 6.5 (c) Mw 5.5 strike-slip earthquakes.
slip earthquake; (b) For a Mw 6.5 strike-slip earthquake;
(c) For a Mw 5.5 strike-slip earthquake. Same as Figure 2,
the average of the median predicted IMs by various
GMPEs are used as input IMs.

scale) for various scenarios. This reveals that the

vector models can provide more consistent results
by using multiple IMs. The vector IMs represent
different aspects of ground motion properties
and provide more information about the ground
motion characteristics.
Furthermore, epistemic uncertainties of both
GMPEs and Newmark displacement models are
considered in the following analysis. Instead of
Figure 5. Standard deviations of predicted displace-
applying the averaged median IMs to compute slid- ments in natural log scale, considering epistemic uncer-
ing displacements as before, the predicted median tainty of both GPMEs and Nemwark displacement
IMs are chosen from different GMPEs (one from models, or considering only epistemic uncertainty of
four NGA models for PGA, PGV, and one from Newmark displacement models, for (a) Mw 7.5 (b) Mw 6.5
three GMPEs for Ia) to compute the median slid- (c) Mw 5.5 strike-slip earthquakes.
ing displacement using 10 Newmark displacement
model, as shown in Figure 4. All together, there
are 120 (4 × 3 × 10) combinations of displace- be much larger than that of the GMPEs, probably
ments in each subplot. The displacements are due to inherent difficulty to correlate the sliding
hugely scattered, indicating that selection of dif- displacements with IMs using simple function
ferent GMPEs and Newmark displacement mod- forms. It calls for more research efforts to be placed
els would significantly influence the final results. to develop more advanced Newmark displacement
For example, the estimated median displacement models in the future.
ranges from 2 cm to 60 cm for Mw 7.5 at the rup-
ture distance of 10 km. Figure 5 shows the stand-
ard deviations of displacements (in natural log 3 ALEATORY VARIABILITY OF
scale) shown in Figure 4. The standard deviations NEWMARK DISPLACEMENT MODELS
by considering both epistemic uncertainties existed
in GMPEs and Newmark displacement models is The aleatory variability of Newmark displace-
generally 20% larger then the previous analysis ment models is also compared for given earth-
by only considering the epistemic uncertainty of quake scenarios. Although vector models usually
the Newmark displacement models. The standard reported a smaller standard deviation compared
deviation obtained from the latter is generally 20% with the scalar models by incorporating more
larger than the former case. The results are con- predictors, inclusion of additional IMs may also
sistent with recent studies on the epistemic uncer- induce extra variability in the IMs themselves.
tainty of IMs. The reported standard deviations of Therefore, the total aleatory variability of the
the epistemic uncertainty for PGA, PGV and Ia Newmark displacement for a scenario earthquake
are approximately 0.2–0.4 in the natural log scale should count for contributions from aleatory vari-
(Douglas, 2010; 2012). The epistemic uncertainty ability in GMPEs and that in Newmark ground
of the Newmark displacement models appears to motion models.

118

ISGSR2013.indb 118 10/18/2013 9:37:18 AM

 Monte-Carlo simulation is used to evaluate the
total aleatory variability of the Newmark displace-
ment for a given scenario earthquake. First, 100
sets of correlated vector IMs are generated for a
specific scenario. The vector IMs are assumed
to follow multivariate lognormal distribution
with mean and standard deviation specified by
GMPEs. For vector models, the joint occurrence
of multiple IMs is specified by the empirical cor-
relations between them. The correlation coef-
ficients among PGA, PGV and Ia are specified
as ρ(PGA, Ia) = 0.88, ρ(PGA, PGV) = 0.69 and
ρ(Ia, PGV) = 0.74 (Campbell and Bozorgnia,
2012), respectively. Secondly, for each set of vec- Figure 7. Standard deviations considering aleatory var-
tor IMs, 100 Newmark displacement values are iability of GMPEs and Newmark displacement models,
simulated by assuming the displacements follow for (a) Mw 7.5 (b) Mw 6.5 strike-slip earthquakes. Cutoff
a lognormal distribution. The standard devia- displacement value is 1 cm.
tion of the resulted 10000 displacement values is
then calculated to estimate the total aleatory vari- gible displacement by Bray & Travasarou (2007),
ability for each Newmark displacement model. It the total standard deviations versus rupture dis-
is noted that very small displacement values have tances are shown as Figure 7. In this case, the
to be excluded, since they are of little engineering standard deviations are generally smaller com-
importance but appear to be highly scattered in log pared with the values shown in Figure 6. This is
scale. Figure 6 shows the obtained total standard expected since the scatter of larger displacement
deviations versus rupture distances for different values would be inevitably smaller in the log scale.
Newmark displacement models by only consid- The standard deviations decreases as separation
ering displacement values greater than 0.001 cm. distance increases, which are approximately within
Quite similar trend can be observed for both mag- the range of 1–1.5 for most models. The total alea-
nitudes considered. Generally speaking, the total tory variability appears to be rather consistent
standard deviations (in natural log scale) fall in for all vector models and scalar models. This is a
the range of 1.5–2.5 for different models, which favorable result since the aleatory variability is the
implies that the displacement distribution is sig- inherent randomness. In principle, it should not
nificantly scattered for each scenario. For example, change significantly from model to model.
although the reported standard deviation is 0.295
in log10 scale (0.68 in natural log scale) for the [Ia]
HL11 model, the total standard deviation of dis- 4 CONCLUSIONS
placements is as high as 1.5 in natural log scale, due
to large standard deviation of Ia (σlnIa ≈ 1). In this paper, epistemic uncertainty and aleatory
If the cutoff displacement value is set as 1 cm, variability of Newmark displacement models are
which is the recommended upper bound for negli- quantified for given scenario earthquakes. Since
the Newmark displacement models are based on
IMs (e.g., PGA, Ia) as predictors, the uncertain-
ties to predict these IMs have to be considered in
the analysis. The standard deviation of the epis-
temic uncertainty of the Newmark displacement
models is within the range of 0.5–1 using the pre-
dicted median IMs as input parameters. In general,
the standard deviation of epistemic uncertainty
becomes 20% larger if both of the epistemic uncer-
tainties of GMPEs and Newmark displacement
models are considered. The epistemic uncertainty
in the Newmark displacement is significant larger
compared with that of GMPEs. It implies that it is
more efforts should be placed to develop advanced
Figure 6. Standard deviations considering aleatory var- Newmark displacement models to reduce the epis-
iability of GMPEs and Newmark displacement models, temic uncertainty in the future.
for (a) Mw 7.5 (b) Mw 6.5 strike-slip earthquakes. Cutoff The total aleatory variability of the Newmark
displacement value is 0.001 cm. displacement model is also studied for given

119

ISGSR2013.indb 119 10/18/2013 9:37:19 AM

 scenario earthquakes. Choosing different cutoff Ambraseys N.N. & Menu J.M. 1988. Earthquake-induced
displacement values would result in different total ground displacements. Earthquake Engineering and
aleatory varialibty. Considering both aleatory Structural Dynamics 16: 985–1006.
varialibilities of GMPEs and the Newmark dis- Boore D.M. & Atkinson G.M. 2008. Ground-motion
prediction equations for the average horizontal com-
placement models, the total standard deviation of ponent of PGA, PGV and 5%-damped PSA at spec-
predicted displacement are 1.5–2.5 in natural log tral periods between 0.01s and 10s. Earthquake Spectra
scale if the cutoff value is chosen as 0.001 cm. The 24(1): 99–138.
total aleatory variability will be reduced if the cut- Bray J.D. & Travasarou T. 2007. Simplified procedure for
off value is chosen as 1 cm. The total aleatory vari- estimating earthquake-induced deviatoric slope dis-
abilities are rather consistent for scalar models and placements. Journal of Geotechnical and Geoenviron-
vector models considered. mental Engineering 133(4): 381–392.
It is important to emphasize that vector models do Campbell K.W. & Bozorgnia Y. 2008. NGA ground
not significantly reduce the total aleatory variability motion model for the geometric mean horizontal com-
ponent of PGA, PGV, PGD and 5% damped linear
of the predicted displacements, due to additional elastic response spectra for periods ranging from 0.1
sources of variability introduced by incorporating to 10 s. Earthquake Spectra 24(1): 139–171.
additional IMs. However, the epistemic uncertainty Campbell K.W. & Bozorgnia Y. 2012. A comparison of
for vector models is generally smaller than that of ground motion prediction equations for Arias inten-
the scalar models. The epistemic uncertainty of vec- sity and cumulative absolute velocity developed using
tor models tends to be consistent for all scenarios a consistent database and functional form. Earthquake
considered, since incorporating multiple IMs can Spectra 28(2): 931–941.
better satisfy the sufficiency criterion (Saygili & Chiou B. & Youngs R.R. 2008. An NGA model for
Rathje, 2008), such that the model is not biased for the average horizontal component of peak ground
motion and response spectra. Earthquake spectra
different magnitude and distances. 24(1): 173–215.
In current practice, seismic slope analysis is Douglas J. 2010. Consistency of ground-motion predic-
a two-step process. First, the ground motion tions from the past four decades. Bulletin of Earth-
IMs (e.g., PGA, PGV or Ia) are estimated using quake Engineering 8: 1515–1526.
GMPEs; Secondly, the estimated IMs are used as Douglas J. 2012. Consistency of ground-motion predic-
input parameters to the Newmark displacement tions from the past four decades: peak ground velocity
models. Both of the epistemic uncertainties and and displacement, Arias intensity and relative sig-
aleatory variabilities in GEMPs and Newmark dis- nificant duration. Bulletin of Earthquake Engineering
placement models should be well considered in seis- 10(5): 1339–1356.
Foulser-Piggott R.F. & Stafford P.J. 2012. A predictive
mic hazard analysis of slopes. A reasonable logic model for Arias intensity at multiple sites and consid-
tree analysis should be employed to account for eration of spatial correlations. Earthquake Engineer-
epistemic uncertainties in GMPEs and Newmark ing and Structural Dynamics 41(3): 431–451.
displacement models. It is also envisioned that Hsieh S.Y. & Lee C.T. 2011. Empirical estimation of the
developing Newmark displacement models directly Newmark displacement from the Arias intensity and
based on earthquake scenarios (magnitude and dis- critical acceleration. Engineering Geology 122: 34–42.
tance etc.) would reduce the inherent complexity Jibson R.W. 2007. Regression models for estimating
and uncertainties in current two-step approach. coseismic landslide displacement. Engineering Geology
91: 209–218.
Newmark N.M. 1965. Effects of earthquakes on dams
ACKNOWLEDGEMENTS and embankments. Geotechnique 15(2): 139–160.
Rathje E.M. & Saygili G. 2008. Probabilistic seismic haz-
ard analysis for the sliding displacement of slopes:
The authors acknowledge financial support from scalar and vector approaches. Journal of Geotechnical
Hong Kong Research Grants Council Grant and Geoenvironmental Engineering 134(6): 804–814.
620311, and DAG08/09.EG13, DAG11EG03G Rathje E.M. & Saygili G. 2009. Probabilistically based
to conduct this study. Any opinions, findings, and seismic landslide hazard maps: An application in
conclusions or recommendations expressed in this Southern California. Engineering Geology 109:
material are ours and do not necessarily reflect 183–194.
those of the sponsor. Saygili G. & Rathje E.M. 2008. Empirical predictive
models for earthquake-induced sliding displacements
of slopes. Journal of Geotechnical and Geoenvironmen-
tal Engineering 134(6): 790–803.
REFERENCES Travasarou T. & Bray J.D. 2003. Empirical attenuation
relationship for Arias Intensity. Earthquake Engineer-
Abrahamson N.A. & Silva W.J. 2008. Summary of the ing and Structural Dynamics 32: 1133–1155.
Abrahamson & Silva NGA ground-motion relations.
Earthquake Spectra 24(1): 67–97.

120

ISGSR2013.indb 120 10/18/2013 9:37:19 AM

 Geotechnical Safety and Risk IV – Zhang et al. (eds)
© 2014 Taylor & Francis Group, London, ISBN 978-1-138-00163-3

Reliability analysis of 640 m long soil retaining wall

for an embedded highway construction

Y. Honjo, Y. Otake, T. Kusano & T. Hara

Gifu University, Gifu, Japan

ABSTRACT: A reliability analysis of 640 m long retaining wall for an embedded highway has been
carried out. The soil profile at the site mainly consists of alternation of three cohesive and cohesionless
soil layers. The soil investigations have been conducted by the borings (SPT N-values) of about 100 m
interval. It was found that the most dominant source of uncertainty to control the performance of the
retaining wall is the thickness of the soil layers which are different from a section to another. The layers
thicknesses are modeled as correlated random processes whose thicknesses are known at the location of
the borings. The thickness of a layer at a certain location is estimated by Co-Kriging method and the
conditional simulation based on it with quantified uncertainty. The design calculation model error is
also evaluated by comparing the calculated results with laboratory model experiment results and actual
site measurements. The response surfaces of the performance of the retaining wall as a function of layer
thicknesses and soil property variations are obtained by extensive design calculations of the retaining
wall. Finally, the reliability evaluation of the retaining wall along the excavation has been carried out by
Monte Carlo Simulation (MCS). The final result takes into account the spatial variability of soil proper-
ties both soil mechanical property variation and the layer thicknesses, the statistical estimation error due
to sparse soil investigations, and design calculation model errors. The quantitative contributions of each
component are also presented.

1 INTRODUCTION statistical estimation error of soil properties and

layer thicknesses, and the calculation model error.
A reliability analysis of 640 m long retaining wall Finally, the reliability of the structure at the
for an embedded highway has been carried out by various locations are evaluated by combining these
a reliability analysis procedure proposed by the information and employing Monte Carlo Simula-
authors (Honjo, Otake and Kato, 2012; Otake and tion (MCS).
Honjo, 2012, 2013; Honjo and Otake, 2012). The
main focus of this study is to introduce a method to
take into account the uncertainty due to the layer 2 DESCRIPTION OF THE SITE
thicknesses on top of uncertainties induced by spa-
tial variability and statistical estimation error of 2.1 The retaining wall
soil properties and design calculation model error. The structure under study is a retaining wall of an
The procedure consists of three part, namely 1) embedded highway construction which runs in an
geotechnical analysis, 2) uncertainty analysis and urban populated area. The total length of the excava-
3) reliability evaluation. tion is 640 m where underground space of 13 m deep
In geotechnical analysis, equations are devel- and 23 m wide should be retained for the highway.
oped, so called Response Surfaces (RS), which The retaining wall is a permanent structure.
relate the inputs and the outputs in the design. The retaining wall is 14 to 20 m long, where the
In the context of the retaining wall design, the excavation depth is 11 to 13 m. The type of the
inputs are soil properties, layer configurations (i.e. retaining wall is a cast-in-site diaphragm wall by
layer thicknesses), dimensions and stiffness of the the Soil Mixing Wall (SMW) method.
retaining wall etc. The outputs can be the maxi-
mum bending moment or displacement induced in
2.2 Geotechnical conditions
the wall during the construction.
The uncertainties caused by various sources are The soil profile at the site is presented in Figure 1.
quantified in the uncertainty analysis. The sources The first layer is landfill soil (B), followed by a
include spatial variability of soil properties, diluvial sand gravel layer (Dg) whose underneath

121

ISGSR2013.indb 121 10/18/2013 9:37:20 AM

 Figure 1. Soil profile at the section.

the deluvial sandy silt layer (Dc). The ground water

level is 2.4 m below the ground surface. The thick-
ness of the layer changes within the construction
location.

2.3 Performance requirements

and performance criteria
The first performance requirement for the retaining
wall is the stability during and after the excavation.
There is also strict requirements for the movement
of the surrounding ground so as to secure the
serviceability of the adjacent building and other
embedded structures.
For the stability of the retaining wall, sufficient Figure 2. The section at Br.5.
embedded depth of the wall should be secured. Also
the maximum bending moment in the wall should Table 1. Excavation steps.
be below the yielding limit. The maximum horizon-
tal deformation of the wall should be limited so as to Step Excavation depth (m) Position of strut (m)
secure the serviceability of the adjacent structures.
The three performance criteria set were the 1 2.0 –
stability of wall for the turnover, the maximum 2 5.5 1.0
bending moment of the wall and the maximum 3 9.5 4.5
horizontal displacement of the wall. The first cri- 4 12.0 8.5
terion was satisfied for all cases if 3 m embedded
depth is secured, which is the minimum required
embedded depth by the code. The maximum bend- a distributed load, whereas the passive side ground
ing moment is limited by the yielding strength of is modeled by linear elastic spring bed which yields
the wall. The criterion for the horizontal displace- at its passive earth pressure. The strata are model
ment is limited to 1% of the excavation depth as elastic springs.
according to the past experiences. Figure 2 is the section at Br.5, where H steel
pile of H-450 × 200 × 9 × 14 is used. The stepwise
excavation depth and the position of struts are pre-
3 GEOTECHNICAL ANALYSIS sented in Table 1. This section is one of the critical
sections in this construction because relatively soft
3.1 Elasto-plastic method Dg layer is thickly deposited.
Elasto-plastic method first proposed by Yamagata Figure 3 presents the distribution of the hori-
et al (1969) is the most commonly used retain- zontal displacement and the bending moment in
ing wall design method in Japan. The method is the wall at various excavation stages. The maxi-
placed as the standard method to design retaining mum horizontal displacement appears at the last
wall for highways (JRA, 1999). In this method, stage of the excavation near the excavation bed,
the retaining wall is model as a beam on a elasto- whereas the maximum moment at one stage before
plastic body. The active earth pressure is applied as the last (3rd step) also at the same location.

122

ISGSR2013.indb 122 10/18/2013 9:37:20 AM

 3. Model uncertainty concerning the elasto-plastic
method of retaining wall design.

4.1 Spatial variability and statistical

estimation error of SPT N-value
Due to the limitation of the space, no further expla-
nation on how the spatial variability and the statis-
tical estimation error of SPT N-values are treated
are not presented in this paper. We have developed
the methodology to treat these uncertainties which
has been applied to this study (Honjo, Otake and
Kato, 2012; Honjo and Otake, 2012, 2013; Otake,
2012). The results of the evaluation are presented
Figure 3. Result of he design calculation at Br.5. in Table 3.

4.2 Layer thickness uncertainty

Table 2. Results of verification.
The layer thickness changes within the site as illus-
Calculated Allowable trated in Figure 1. The RS functions obtained also
Unit (C) (A) C/A indicate that the layer thicknesses have consider-
able influence on the design of the retaining wall.
Plastic rate % 62 90 0.69 Thus, it is important to take the uncertainty of
Bending moment kN ⋅ m 338 341 0.99 layer thicknesses into account.
Displacement mm 52 120 0.43 One of the distinguished characteristics of the
layer thicknesses at the site is the negative corre-
Table 2 summarizes the relationship between the lation between the layer thicknesses of B and Dg
maximum calculated values (C) and the allowable layer, which can be observed from the scattergram,
values (A). C/A for the bending moment is 0.99, Figure 4. The correlation coefficient between the
and it is clear that the section is determined by this two quantities is −0.89.
item. Other items have some extra margin of safety Co-Krigin is used in this study to estimate the
for the limiting values. layer thicknesses of B and Dg layers which should
be interpolated from the observed points together
3.2 Building response surfaces with the statistical estimation error. To evaluate the
The Response Surfaces (RS) for the maximum
Table 3. Uncertainty associated with SPT N-value.
horizontal displacement and the bending moment
have been constructed. RS is a function of basic V θV
variables which indicates the performance of the Layer n σ (m) (m) Γ2(V/θ) ΛG2 σV
structure whether it is exceeding the limit state or
not. A RS function can be given as follows: B 3 12.20 3.0 0.5 0.28 0.40 10.03
Dg 5 4.51 5.5 0.5 0.17 0.24 2.86
y f ( x1 xn ) + ε (1) Dc 6 3.01 6.5 0.5 0.14 0.20 1.76

where xi (i n ) are n basic variables, y is the

response and ε residual error.
Due to the limitation of the space, no further
description on the RSs is presented. We have suc-
cessfully built the RSs in this study.

4 UNCERTAINTY ANALYSIS

Three types of uncertainties are analyzed in this

section. Namely,
1. Spatial variation of the local average of SPT
N-value, and the statistical estimation error
associated in evaluating this local average value.
2. Uncertainty of the layer thickness. Figure 4. Scattergram of B and Dg layer thicknesses.

123

CH09_3-2.indd 123 10/18/2013 6:06:39 PM

 influence of the layer thicknesses to the performance 4.2.2 The conditional simulation
of the retaining wall, the conditional simulation The two correlated random filed can be generated
technique is used. In the conditional simulation, based on the fact that m dimensional normal distri-
necessary number of random filed samples are bution can be decomposed to m independent nor-
generated that take the observed thicknesses at the mal distributions by the eigenvalue decomposition
observation points. Based on the generated sample of the covariance matrix.
layer thicknesses, the RS functions can be used to In order to apply this procedure to the present
evaluate the reliability of the retaining wall. problem, the correlation among the layer thick-
nesses, Z1(xi), Z2(xi), need to be assumed, where
4.2.1 Co-Kriging xi(i = 1, … m) that indicates the coordinate of the
Let the thicknesses of the two layers be location where the layer thicknesses be generated.
Z(x) = (Z1(x), Z2(x)), where x is the coordi- The covariance can be calculated as follows:
nate to show the position corresponds to “Dis-
tance” in Fig. 1). The layer thicknesses at Zk(x) is Cov (Zk ( xi ), Zl ( x j ))
interpolated from 2n observed layer thicknesses at
x1, x2, …, xn. Note that suffix k indicates the layer ⎧⎪σ Z σ Z ρ (| xi x j |) (i k l )
=⎨ k l (5)
⎩⎪−0.89 σ Zk σ Zl ρ ( i
B (k = 1) or Dg (k = 2). | x − x j |) (i k l )
The basic assumptions of Co-Kriging is not dif-
ferent from those of Ordinary Kriging (Journel and
Huijbrechts, 1978) Under the three basic assump- where, (i, j = 1, 2, … m; k, l = 1, 2), σ Zk is the SD of
tions, namely the linear estimator, the un-biasness layer thickness for layer k, and ρ is the autocorrela-
and the minimum estimation variance estimator, tion function of the exponential function type. It is
Co-Kriging problem results to solve the following assumed that the autocorrelation distance for the
simultaneous equation to obtain the weights: two layers are the same.
Finally, the procedure of the conditional simula-
⎡ K11 K12 1 0 ⎤ ⎡ w ⎤ ⎡ k1 ⎤ tion is as follows:
1
⎢K 0 1 ⎥⎥ ⎢ w 2 ⎥ ⎢ k 2 ⎥
⎢ 21 K22 Step 1 Generate layer thickness sk(xj) (j = 1, … m;
⎢ ⎥=⎢ ⎥ (2)
⎢ 1T 0T 0 0 ⎥ ⎢ λ1 ⎥ ⎢ l ⎥ k = 1, 2) by the procedure shown above.
⎢ ⎥⎢ ⎥ ⎢ ⎥ These generated values fulfill the sta-
⎢⎣ 0T 1T 0 0 ⎥⎦ ⎣ λ2 ⎦ ⎣ l ⎦ tistical properties assumed for the layer
thicknesses.
where, Step 2 Interpolate the layer thicknesses at
Kjj ′: Covariance matrix of Zj = {Zj(xi)} and xi(i = 1, … m) by Co-Kriging based just
Zj ′ = {Zj ′(xi)} where ( j, j ′ = 1, 2) on the generated layer thicknesses where
kk: Covariance vector of Zj = {Zj(xi)} and Zk(x) observations are originally made. The inter-
where (i = 1, … n; j = 1, 2; k = 1, 2) polated results are denoted by s k ( xj ) .
wj: Weights in the estimator wj = {wij} (i = 1, …, n; Step 3 Let the results of Co-Kriging based on the
j = 1, 2) original observations be z k ( xj ). A sam-
λj: Lagrange multiplier (j = 1, 2) ple by the conditional simulation can be
1: n − D column vector with 1 obtained by the equation below:
0: n − D column vector with 0
l: = 1 (if j = k) zk ( x j ) = z k ( x j ) + ( sk ( x j ) s k ( x j )) (6)
= 0 (if j ≠ k)
i: Sampling point number (i = 1, … n) The obtained set of layer thicknesses, zk(xj)
j: Layer number (j = 1, 2) (j = 1, … m; k = 1, 2), satisfies the statistical
k: Estimating layer number (k = 1 or 2) properties assumed and also pass through
The estimator is given by the following linear all the observed layer thicknesses at the
equation where wij are the weight assigned to each observed points.
observation that are obtained from Eq. (2): Step 4 Repeat step 1 to step 3 as necessary to
n 2
generate required sets of sample layer
 k (x ) =
Z ∑ ∑ wij Z j ( xi ) (3) thicknesses.
i =1 j =1
The statistics estimated from the observed
values of the layer thicknesses are summarized in
The estimation error can be also evaluated:
Table 4. The autocorrelation distance of the layer
T thicknesses, θH, is not clearly estimated. Thus, two
⎡ w ⎤ ⎡K K12 ⎤ ⎡ w1 ⎤
2
σ EV = ⎢ 1 ⎥ ⎢ 11 ⎥⎢ ⎥ (4) cases, namely θH = 50 and 100 m, are set to com-
w K
⎣ 2 ⎦ ⎣ 21 K 22 ⎦ ⎣ w 2 ⎦ pare the results.

124

ISGSR2013.indb 124 10/18/2013 9:37:22 AM

 4.3 Model error 5 RELIABILITY EVALUATION
Otake (2012) has done an extensive study on the
5.1 Conditions of the evaluation
model error of the elasto-plastic method. Due to
the limitation of the space, only the results are The sections are set 5 m interval for the 640 m long
summarized here. construction site. The reliability evaluations are
Based on the collected model experiment results made for all these sections. The RS functions can
and actual construction case histories, the elasoto- be described as follows by the basic variables:
plastic method has been tested for its accuracy.
The results are presented in Table 5. D = fD(NB, NDg, NDc, HB, HDg, HDc) ⋅ δME (7)
The model error is defined by (observed M = fM(NB, NDg, NDc, HB, HDg, HDc) ⋅ δME (8)
value)/(calculated value) so that when the error
is multiplied to the calculated value, the value be The mean and COV of each basic variable are
transformed to the true value that is represented listed in Table 6.
here by the observed value.
The model error employed in the reliability eval- 5.2 Results
uation in this study adopted the results based on The reliability evaluation results for the maxi-
the model experiment. mum horizontal displacement are presented in
Figure 5, whereas those for the maximum bend-
Table 4. Estimated statistics for B and Dg layer
thickness. ing moment in Figure 6. In (a) in the both figures,
the exceedance probabilities of the limiting values,
θH (m) Pf, are are superposed for θh = 50 (m) and 100 (m)
for comparison. Furthermore in (b) in the both
Layer μ (m) σ (m) Case A Case B

B 3.0 2.10 50 100 Table 6. Statistics of the basic variables.

Dg 5.5 3.94 50 100
μ σ COV

Table 5. Model error in the elasto-plastic method. Statistical estimation NB 15.0 6.43 0.43
Error Γ2 NDg 19.0 1.83 0.10
Item Mean COV n (Normal) NDc 12.0 1.13 0.09
Spatial variability NB 15.0 7.69 0.51
Case histories Max disp. 0.70 0.29 58 Λ2 NDg 19.0 2.20 0.12
Max BM 0.69 0.43 42 (Normal) NDc 12.0 1.35 0.11
Model Max disp. 1.08 0.67 38 Model error Disp. 1.08 0.67 0.62
experiments Max BM 0.73 0.44 24
δME BM 0.73 0.44 0.60
(log normal)
(note 1 error = (observed)/(calculated); (note 2 BM:
bending moment, disp.: displacement).

Figure 5. Results of MCS for the maximum horizontal displacement.

125

ISGSR2013.indb 125 10/18/2013 9:37:24 AM

 Figure 6. Results of MCS for the maximum bending moment.

figures, contribution of each uncertain source are properties are quite limited on the reliability of
illustrated for case θh = 100 (m). the retaining wall.
The following observations can be made for
these cases.
REFERENCES
1. Pf for both the maximum horizontal displace-
ment and the bending moment takes the maxi- Honjo, Y.,Y. Otake & Kato H. 2012. A simpli-fied scheme
mum value between distance 300 to 450 (m). In to evaluate spatial variability and statistical estima-
this section, the thickness of Dg layer is thicker, tion error of local average of geotechnical parameters
which makes the earth pressure lager. in reliability analysis. J. Geotechnical Engineering
2. The accuracy of the layer thickness prediction (JSCE) 68(1):41–55. (in Japanese).
Honjo, Y. & Otake Y. 2012. Verification of statistical esti-
is less for θh = 50 (m) than for 100 (m). It is mation error evaluation theory of local averages of
also less in the area that are distanced from the geotechnical parameters. J. Geotechnical Engineering
observations are made. However, the difference (JSCE) 68(3):475–490. (in Japanese).
is rather small for θh = 50 (m) and for 100 (m). Honjo, Y. & Otake Y. General and local estimation of
3. The uncertainty sources that are critical are the local average of geotechnical parameters in reliabil-
model error and the statistical estimation error of ity analysis. Proc. APSSRA, Phoon, K.K., Beer, M.,
the layer thicknesses. The spatial variability and Quek, S.T., Pang, S.D. (editors), 236–241.
the statistical estimation error of SPT N-value Honjo, Y. & Otake Y. 2013. A Simple Method to Assess
of each layer have only negligible impacts. the Effects of Soil Spatial Variability on the Perform-
ance of a Shallow Foundation. ASCE Geotechnical
Special Publication No. 229, Foundation Engineering in
the Face of Uncertainty (honoring Fred H. Kulhawy),
6 CONCLUSION J.L. Whithiam, K.K. Phoon and M.H. Hussein (eds.),
385–404.
The sources of uncertainty concerning retaining Jounel, A.G. & Huijbregts C.J. 1978. Mining Geostatis-
wall design have been identified and quantified tak- tics, Academic Press.
ing 640 m long embedded highway retaining wall as Otake, Y. & Honjo Y. 2012. Reliability based design on long
an example. The reliability analysis scheme proposed irrigation channel considering the soil investigation
by Honjo, Otake and Kato (2012) has been applied locations, Proceedings of Geo Congres, 2836–2845.
Otake, Y. 2012. Development of geotechnicl reliability
to the reliability evaluation of the retaining wall. based design and chllenges in geotechnical design,
The result of the evaluation has shown that the PhD dissertation submitted to Gifu university, 215 pp.
two major sources of uncertainty are the statisti- (in Japanese).
cal estimation error of the layer thicknesses and Vanmarcke, E.H. 1977. Probabilistic modeling of soil
the model error of the elasto-plastic method for profiles. J. of Geotechnical Engineering, ASCE 103
retaining wall design. The impacts of the spatial (GT11): 1227–1246.
variability of soil properties within each layer,
and the statistical estimation error of the soil

126

ISGSR2013.indb 126 10/18/2013 9:37:25 AM

 Geotechnical Safety and Risk IV – Zhang et al. (eds)
© 2014 Taylor & Francis Group, London, ISBN 978-1-138-00163-3

CPT-based Bayesian identification of underground soil stratigraphy

K. Huang
Department of Civil and Architectural Engineering, Shenzhen Research Institute, City University of Hong Kong,
Hong Kong, China

Z. Cao
State Key Laboratory of Water Resources and Hydropower Engineering Science, Wuhan University,
Wuhan, P.R. China

Y. Wang
Department of Civil and Architectural Engineering, Shenzhen Research Institute, City University of Hong Kong,
Hong Kong, China

ABSTRACT: This paper aims to identify underground soil stratum and classify the soil types using
Bayesian approaches and Cone Penetration Tests (CPT). The uncertainty in the CPT-based soil classifica-
tion using the Robertson chart is modeled explicitly in the Bayesian approaches. The proposed Bayesian
approaches contain two major components: a Bayesian model class selection approach to identify the
most probable number of underground soil layers and a Bayesian system identification approach to esti-
mate the most probable layer thicknesses and determine the soil types simultaneously. Equations are
derived for the Bayesian approaches, and the proposed approaches are validated using a set of simulated
CPT data. It has been shown that the proposed approaches correctly identify the underground soil strati-
fication and classify the soil type of each layer.

1 INTRODUCTION

Determining the underground soil stratigraphy

(i.e. the number of soil types/layers and their
thicknesses/boundaries) is one key aspect in geo-
technical site characterization, during which Cone
Penetration Tests (CPTs) are widely used around
the world. CPT enjoys many advantages (Mayne
2007, Robertson 2009). However, no soil sample
is retrieved for visual inspection to assist in soil
classification.
When performing soil classification from CPT
data, the CPT reading (e.g. tip resistance and
friction ratio) is linked directly to the soil types,
and the Soil Behavior Type (SBT) classification
is determined using soil classification charts (e.g.
Robertson 1990, Jefferies & Davies 1993). Consider,
for example, the soil classification chart, which is
shown in Figure 1 and is frequently referred to as
the Robertson chart (Robertson 1990). The soil
type is determined based on two parameters (i.e.
two axes in Fig. 1): the normalized friction ratio,
FR = 100fs/(qt − σv0), and the normalized tip resist-
ance, Qt = (qt − σv0)/σ′v0, where fs, qt, σv0,σ′v0 are the
sleeve friction, corrected tip resistance, vertical total
stress, and vertical effective stress, respectively. The Figure 1. The Robertson soil classification chart based
nine areas in the chart correspond to nine different on CPT data (After Robertson (1990)).

127

ISGSR2013.indb 127 10/18/2013 9:37:25 AM

 Table 1. Description of soil types in the Robertson’s 2 PROBABILISTIC FRAMEWORK FOR
soil classification chart (After Robertson (1990)). SOIL STRATUM IDENTIFICATION
Area Soil description As shown in Figure 2, identifying the underground
1 Sensitive, fine-grained soil strata based on CPT data is to determine the
2 Organic soils (peats) layer thickness h = [h1,h2, …, hn…, hN] in a soil pro-
3 Clays (clay to silty clay) file containing N soil layers, where hn is the thickness
4 Silt mixtures (clayey silt to silty clay) of the n-th layer. The number N in the first part of
5 Sand mixtures (silty sand to sandy silt) this paper (i.e. before Section 5 “The Most Probable
6 Sands (clean sand to silty sand) Number of Soil Strata”) is treated deterministic but
7 Gravelly sand to sand unknown. It is then determined by a Bayesian model
8 Very stiff sand to clayey sand class selection approach
pp p proposed in Section 5.
9 Very stiff, fine-grained Let ξ ξ 1,ξ 2 ξ ] be a set of ln(FR)
N
and ln(Qt) data obtained from a CPT test
in a soil profile with N soil layers, in which
ξ ξ n (d1 ), ξ n 2 2 ξ nkk (d kn )], n = 1, 2, ..., N ,
n n
soil types, as described in Table 1. Based on which is a set of ln(FR) and ln(Qt) measured at the kn
area the measured (FR,Qt) combination is located depths d1 d2 , ..., d kn within the n-th soil layer. For
in, the soil type is determined accordingly. Note a given layer number N of soil layers, the probabil-
that the Robertson chart was developed based on ity P (ξ N ) that all data points in the n-th layer
n
past observations and engineering experience, and belong to the same soil type can be calculated as:
various uncertainties are involved inevitably, such
as observation scatterness, measurement error, and 9
transformation uncertainty. Such uncertainties can P (ξ N )
n ∑ PSTT (ξ n N )
J
n 1, 2, , N (1)
be properly dealt with by integrating the Robertson J =1
chart with probabilistic methods.
Several probabilistic soil classification where PSTTJ (ξ N ) is the probability that all data
n
approaches have been proposed (e.g. Zhang & points in the n-th layer belong to the same given
Tumay 1999, Kurup & Griffin 2006). However, soil type J. The PSTTJ (ξ N ) can be expressed as:
n
these approaches mainly focused on determining
kn
the soil type from a particular CPT data point.
How to stratify the underground soil profile PSTTJ (ξ N )
n ∏ PST (ξ n,i N )
J
for J = 1, 2, ..., 9 (2)
i =1
from a large number of nearly continuous CPT
data points remains unsolved. Engineering judg-
ment is frequently used to handle such a problem.
However, inconsistencies often arise due to differ-
ent judgment from different engineers. The key
issue to this problem is that the uncertainty in the
CPT-based soil classification and the spatial distri-
bution of the CPT data (e.g. Wang et al. 2010) is
not considered explicitly, but treated implicitly and
vaguely by using engineering judgment.
This paper integrates the Robertson chart with
Bayesian approaches to explicitly and properly
consider the uncertainty in the CPT-based soil
classification and spatial distribution of CPT data.
Bayesian system identification and model class
selection approaches are used to identify the most
probable thicknesses and number of soil layers in
a systematic and rational manner. The paper starts
with the probabilistic framework for soil stratum
identification and probabilistic soil classifica-
tion based on the Robertson chart, followed by
the proposed Bayesian system identification and
model class selection approaches. Then, the appli-
cation procedure for the Bayesian approaches is
described. The proposed approaches are validated Figure 2. An illustration of the underground soil strati-
using a set of simulated CPT data. graphy.

128

ISGSR2013.indb 128 10/18/2013 9:37:25 AM

 where PSTTJ (ξ N ) is the probability that a data
n ,i
point i in the n-th layer belongs to the soil type J.
The value of PSTTJ (ξ n N ) is calculated repeatedly
for J varying from 1 to 9. Then, the probability
P (ξ n N ) is calculated using Equation 1, and the
soil type J with the maximum value of PSTTJ (ξ n N )
among the nine soil types is taken as the soil type
for the n-th soil layer. The probability PSTTJ (ξ n,i N )
is the key input variable in Equations 1 and 2, and
a probabilistic model is developed in the next sec-
tion to estimate the PSTTJ (ξ N ) value.
n ,i

3 PROBABILISTIC SOIL CLASSIFICATION

BASED ON THE ROBERTSON CHART

Consider, for example, a CPT data point i with

a combination of [ln( Ri ), ) ln(Qti )] value meas-
ured at a given depth during g the test. Because the
Robertson chart and [ln( Ri ), ) ln(Qti )] are not per- Figure 3. An illustration of the probability contour and
fect knowledge or information, there exists a plau- two-dimensional joint PDF for soil classification based
on a given CPT data point.
sibility (or occurrence probability) that the soil at
this given depth, in fact, should be classified as any
of the nine types of soil. Such a probability can be
quantified through a Probability Density Function h = [h1,h2, …, hn …, hN], but are not ΩN. Dividing
(PDF) (Wang et al. 2013), as shown in Figure 3, CPT data ξ into different soil layers requires infor-
and be expressed as: mation on the boundaries that separate various soil
layers. Such information is unknown and needs to
1 be determined in site characterization. The next
P (ln( FR ) ln(Qt ) σ ) = section presents a Bayesian system identification
2πσ FFRσ Q
Qt approach to determine the thicknesses and soil
⎧ [ln( FR ) l ( FRi )]2 [ln( t ) ln(
l ( i 2
t )]
⎫ types for a soil profile with a given number N of
× exp ⎨− − ⎬ soil layers.
⎩⎪ 2σ FFR 2
2σ Q
Qt
2
⎭⎪
(3)
4 BAYESIAN SYSTEM IDENTIFICATION
where σ = [σFR,σQt], σFR and σQt are the standard OF THE THICKNESSES OF SOIL
deviation of the joint Gaussian distribution along STRATA
the ln(FR) and ln(Qt) axes, respectively.
Then, the probability PSTTJ that the data point i Within the Bayesian framework, the updated rule
with the measured values belongs to the soil type J for the estimation of the model parameters ΩN is
is expressed as: given by (Cao & Wang 2013, Wang et al. 2013):

PSTTJ ∫∫ P(ln( FR ) l (Qt ) σ )d l ( FR )dd Qt ) P( N ξ, N ) K N P (ξ Ω N N )P ( N N) (5)

J
for J = 1, 2, ..., 9 (4)
where KN is a constant; P(ξ|ΩN,N) is the likelihood
function; P(ΩN|N) is the prior distribution. The
The two dimensional integration in Equation 4 P(ξ|ΩN,N) is expressed as (Wang et al. 2013):
is performed repeatedly for each soil type and each
data point to obtain the PSTTJ (ξ N ) . Note that, N
n ,i
σ = [σFR,σQt] are model parameters in each soil P (ξ N
,N ) ∏ P(ξ n σ n , N ) (6)
layer. As illustrated in Figure 2, for a soil pro- n =1
file containing N soil layers, the N sets of model
parameters are defined as ΩN = [σ1, σ2, …, σΝ], in where P(ξ n|σn,N) is the likelihood function for the
–
which σn = [σFR,n,σQt,n], n = 1,2, …, N. n-th layer and calculated using Equations 1–4.
It is also worthwhile point out that the param- Note that, as the boundaries of soil layers (i.e.,
eters of real interest in the identification of under- layer thicknesses) change, the division of CPT data
ground stratification are the thickness vector also changes. Therefore, ξn is a function of layer

129

ISGSR2013.indb 129 10/18/2013 9:37:28 AM

 thicknesses and P(ξ n|σn,N) is also a function of maximum possible number of soil layers within
– the depth of which CPT is performed. Then, the
layer thicknesses.
Similarly, the prior distribution is expressed as: model class number k is a positive integer varying g
from 1 to Nmax. The most probable model class M k*
N is the model class with the maximum occurrence
P( N N ) = ∏ P( n N) (7) probability among all candidate model classes,
n =1 given that a set of CPT data ξ is observed. The
most probable layer number k*, therefore, can be
where P(σn|N) is the prior distribution of the determined by selecting the one with the maximum
model parameters σn for the n-th soil layer, which value of P(Mk|ξ ) for all candidate model classes.
is calculated using Equation 8: –
According to Bayes’ Theorem (Beck & Yuen
2004, Cao & Wang 2013), P(Mk|ξ) is written as:
P( n N)
⎧⎪ 1 1 Mk ξ )
P (M P (ξ M k )P (M k ) / P (ξ ), k = 1, 2, ..., N max
× fo σ FR ∈[ 0,1.33], [ 0,1.99]
= ⎨1.33 1.99 Qt
Q

⎩⎪0 for other

es (9)
(8) where P(ξ ) is a constant; P(ξ |Mk) is frequently
– –
referred to as the ‘evidence’ for the model class Mk
Interested readers are referred to Wang et al. provided by the CPT data ξ ; P(Mk) is the prior
(2013) for detailed discussions on the prior distri- –
probability of the model class Mk. If no prevail-
bution in Equation 8. ing prior knowledge on the number of soil layers is
Then, the most probable thicknesses of the soil available, the prior probability for each of the Nmax
layers are identified using an asymptotic tech- candidate model classes is the same, and therefore,
nique to approximate the posterior PDF of the P (M k ) can be taken as 1/Nmax. Then, according to
model parameters in Equation 5. The asymptotic Equation 9, P(Mk|ξ ) is proportional to P(ξ |Mk),
technique involves approximating the posterior – –
and the most probable number of soil layer can be
PDF as a Gaussian PDF (Cao & Wang 2013, selected by comparing the value of P(ξ |Mk) among
Wang et al. 2013). Then, the posterior PDF for the –
the Nmax candidate model classes. The most prob-
model parameters is a joint Gaussian PDF with able number of soil layers is the one with the maxi-
the mean value equal to the Most Probable Values mum value of P(ξ |Mk). Details of calculating the
(MPV) of the posterior PDF. The MPV, denoted –
evidence P(ξ |Mk) for each model class are referred
by Ω*N , maximizes the posterior PDF. Under this –
to Wang et al. (2013).
approximation, maximizing P(ΩN|ξ ,N), or for The P(ξ|Mk) is calculated repeatedly for
–
numerical convenience, minimizing an objective k = 1,2, …, Nmax, and the most probable number k*
function f = −ln[P(ΩN|ξ ,N)], leads to the posterior of soil layers are determined by selecting the model
–
mean for the model parameters. class with the maximum P(ξ|Mk) value. Note that
Note that, because the likelihood function in the most probable thicknesses h*k have been deter-
Equation 6 is a function of thicknesses of soil layers, mined in the previous section. Therefore, the deter-
both the posterior PDF and the objective function mination of the most probable number k* of soil
are functions of soil layer thicknesses. Maximizing layers simultaneously leads to the determination
the posterior distribution, i.e. minimizing the objec- of the most probable thicknesses h*k* .
tive function, provides not only the MPV of ΩN but
also the MPV of layer thicknesses.
6 APPLICATION PROCEDURE
5 THE MOST PROBABLE NUMBER
Seven steps are involved in the application proce-
OF SOIL STRATA
dure of the proposed approach. Details of each
step and their associated equations are summa-
Starting from this section, the number N of soil
rized as follows:
layers is considered as a variable k, and a Bayesian
model class selection approach (Beck & Yuen 1. Obtain a set of CPT data and convert them to
2004, Cao & Wang 2013) is used to determine the the normalized terms;
most probable value k* among a pool of candidate 2. Choose an appropriate maximum number Nmax
model classes. A model class herein is referred to of soil layers for the CPT data;
a family of stratification models that share the 3. Define the prior distribution for model param-
same number (e.g. k) of soil layers but have dif- eters using Equations 7 and 8;
ferent model parameters (e.g., layer thickness hN 4. Compute the likelihood function using Equa-
and model parameters ΩN). Let Nmax denote the tions 1–4;

130

ISGSR2013.indb 130 10/18/2013 9:37:32 AM

 5. Construct the objective function, and minimize
the objective function (e.g., by a MATLAB
function “fminsearch”) and determine the most
probable thicknesses h*k for the k-th model class.
Then, compare PSTTJ (ξ n
, N ) J , 2 ..., 9 for
n
the n-th layer and determine the soil type for the
soil layer;
6. Calculate the conditional probability P (ξ M k ) ;
7. Repeat steps 3–6 Nmax times to calculate h*k and
P (ξ M k ) for the Nmax candidate model classes.
The model class with the maximum value of
P (ξ M k ) is selected as the most probable model
class M k* and the corresponding h*k are the most
probable layer thicknesses/boundaries.
Although the proposed Bayesian approach Figure 4. A set of simulated CPT data.
involves quite a number of equations and seems
mathematically complicated, it is rather straight-
forward to program it as a user function or toolbox
in computer software. Geotechnical practitioners
only need to provide prior knowledge, if any, and
project-specific CPT test data as input, and the
user function or toolbox will return the under-
ground soil classification and stratification. This
significantly improves the practicality of the pro-
posed approach.
The proposed Bayesian approaches have been
shown to perform well for some real-life examples
(Wang et al. 2013). In the next section, the pro-
posed approaches will be further validated using a
set of simulated CPT data.

7 SIMULATED EXAMPLE

The CPT data is simulated from a three-layer soil

profile, with respective thicknesses of 3 m, 4 m,
and 5 m and a depth interval of 0.1 m. The dis- Figure 5. The distribution of the simulated CPT data
tribution of the ln(FR) and ln(Qt) of the simulated points on the Robertson soil classification chart.
CPT data is shown in Figure 4. Figure 5 plots the
ln(FR) and ln(Qt) data pairs in the Robertson chart,
7.1 The most probable number of soil strata
and the data points are mainly located within the
areas 1, 3 and 5 (see Table 1 for the description of Table 2 shows the logarithm of evidence (i.e.
each soil type), with some scattered data points in ln[P(ξ|Mk)]) in the second Column for the four can-
the areas of 2, 4 and 6. Note that in practice the didate model classes. It increases from −218.9 to
true boundaries of soil layers are unknown, and −98.4 as k increases from 1 to 3 and it then decreases
they are estimated from project-specific test results from −98.4 to −102.1 as k further increases from 3
and prior knowledge. to 4. The model class with three soil layers, has the
Consider, for example, four candidate model largest value of evidence among all the four model
classes. For all model classes, the prior distribu- classes. Therefore, the most probable number of
tions of model parameters are all estimated using soil layers is three. The true number of soil layers is
Equations 7 and 8. Using the prior knowledge and identified correctly.
the CPT data shown in Figure 4, the proposed
Bayesian system identification and model class
7.2 The most probable thicknesses or boundaries
selection approaches are used to determine the
most probable number of soil layers and identify The most probable thicknesses (i.e. h*k ) of soil layers
the most probable thicknesses and soil types of the for the four candidate model classes are also sum-
layers, as discussed in the following two subsec- marized in Table 2. The most probable boundaries
tions, respectively. for the four model classes and the true boundaries

131

ISGSR2013.indb 131 10/18/2013 9:37:35 AM

 Table 2. Results of the Bayesian model class selection estimate the most probable layer thicknesses and
approach in the simulated case. classify the soil types simultaneously.
Equations were derived for the Bayesian
The most probable approaches, and the proposed approaches were
thicknesses h*k (m)
Model validated using a set of simulated CPT data. It
class ln[P(ξ |Mk)] h*1 h*2 h*3 h*4 has been shown that the proposed approaches cor-
– rectly identify the underground soil stratification
M1 −218.9 12 – – – and classify the soil type of each layer.
M2 −142.1 7.0 5.0 – –
M*3 −98.4 3.2 3.9 4.9 –
M4 −102.1 3.2 2.4 1.3 5.1 ACKNOWLEDGEMENT

The work described in this paper was supported

by a grant from the Research Grants Council of
the Hong Kong Special Administrative Region,
China (Project No. 9041550 (CityU 110210)), a
grant from City University of Hong Kong (Project
No. 7002695), and a grant from National Natu-
ral Science Foundation of China (Project No.
51208446). The financial supports are gratefully
acknowledged.

REFERENCES

Beck, J.L. & Yuen, K.V. 2004. Model selection using

response measurements: Bayesian probabilistic
approach. Journal of Engineering Mechanics 130(2):
192–203.
Figure 6. The most probable boundaries of soil layers Cao, Z. & Wang, Y. 2013. Bayesian approach for proba-
for different model classes. bilistic site characterization using cone penetration
tests. Journal of Geotechnical and Geoenvironmental
Engineering 139(2): 267–276.
of the three soil layers defined in the original data Jefferies, M.G. & Davies, M.P. 1993. Use of CPTU to
are delineated in Figure 6 by dashed and solid lines, estimate equivalent SPT N60. Geotechnical Testing
respectively. For the most probable model class M*3, Journal 16(4): 458–468.
the most probable thicknesses of these three lay- Kurup, P.U. & Griffin, E.P. 2006. Prediction of soil
ers are 3.2 m, 3.9 m and 4.9 m, respectively. They composition from CPT data using general regression
are consistent with the ones defined in the original neural network. Journal of Computing in Civil Engi-
neering, 20(4): 281–289.
data. In addition, the soil types for the three lay-
Mayne, P.W. 2007.Cone penetration testing: A synthesis of
ers are classified as soil type 1, soil type 3 and soil highway practice. NCHRP synthesis 368. Washington,
type 5, respectively. They agree well with the true D.C: Transportation Research Board.
soil types defined in the original data. Robertson, P.K. 1990. Soil classification using the cone
penetration test. Canadian Geotechnical Journal 27(1):
151–158.
8 SUMMARY AND CONCLUSIONS Robertson, P.K. 2009. Interpretation of Cone Penetra-
tion Tests-a unified approach. Canadian Geotechnical
Journal 46(11): 1337–1355.
This paper aimed to identify underground soil
Wang, Y., Au, S.K. & Cao, Z. 2010. Bayesian approach
stratum and classify the soil types using Bayesian for probabilistic characterization of sand friction
approaches and Cone Penetration Tests (CPT). angles. Engineering Geology 114(3–4): 354–363.
It integrated the Robertson chart with Bayesian Wang, Y., Huang, K. & Cao, Z. 2013. Probabilistic iden-
approaches to explicitly and properly consider the tification of underground soil stratification using
uncertainty in the CPT-based soil classification cone penetration tests. Canadian Geotechnical Journal
and the spatial distribution of the CPT data. The (accepted).
proposed Bayesian framework contains two major Zhang, Z.J. & Tumay, M.T. 1999. Statistical to fuzzy
components: a model class selection approach to approach toward CPT soil classifications. Journal
of Geotechnical and Geoenvironmental Engineering
identify the most probable number of underground
125(3): 179–186.
soil layers and a system identification approach to

132

ISGSR2013.indb 132 10/18/2013 9:37:38 AM

 Geotechnical Safety and Risk IV – Zhang et al. (eds)
© 2014 Taylor & Francis Group, London, ISBN 978-1-138-00163-3

Characterization of three-dimensional random crack network

in soil using CT test

J.H. Li
Centre for Offshore Foundation Systems, The University of Western Australia, Perth, Australia
Harbin Institute of Technology Shenzhen Graduate School, Shenzhen, China

L.M. Zhang
The Hong Kong University of Science and Technology, Hong Kong, China

Y. Zhao
China Merchants Shekou Industrial Zone Co., Ltd., Shenzhen, China

ABSTRACT: Cracks in soils are three-dimensional (3D) and provide important preferential path-
ways for rainfall infiltration. The geometrical properties of 3D cracks are crucial parameters for analysis
of seepage in discontinuous soils or rocks. However 3D crack planes in soils are still largely unknown
because the cracks in soils are prone to disturbance and sensitive to moisture content. This paper presents
a method to characterize 3D crack planes in soils by employing a nondestructive Computer Tomography
(CT) test. The traces of cracks are first obtained from sections along three directions in the CT test. Then
three criteria are proposed to determine a crack plane based on these crack traces. Finally the random
crack network in the soil is established by assembling the planes. The probability distribution and statisti-
cal parameters can be obtained based on the obtained crack planes.

1 INTRODUCTION was then proposed to reconstruct the crack plane

in three dimensions. Both the two- and three-
Cracks in soils provide preferential flow pathways dimensional crack geometries were obtained.
for rainfall infiltration into soil slopes, covers
and clay liners. The infiltration in a cracked soil
2 FIELD BLOCK SAMPLE AND CT TESTS
is mainly determined by the geometry of crack
networks (Li et al. 2009). Many researchers stud-
A block soil sample was carefully retrieved in an
ied the pattern and geometrical parameters of a
expansive soil slope in Zhenjiang, China. The
crack network on soil surface (Li & Zhang 2010,
physical properties of the soil are shown in Table 1.
2011; Nahlawi & Kodikara 2006). However cracks
The soil was cut into a cube with dimensions of
normally penetrate the soil and form a three-
0.25 × 0.25 × 0.25 m. The soil sample experienced
dimensional random crack network (Li 2009).
several drying-wetting cycles. The expansive soil
Based on assumed crack shapes and distributions
shrinks during drying and swells during wetting.
some researchers have attempted to establish a three
Therefore many cracks were present within the soil
dimensional network for fractured rock. For exam-
cube as shown in Figure 1a.
ple, a disc-shaped fracture model was proposed by
Baecher (1977) and Long et al. (1985) and a polyg-
onal fracture model was presented by Dershowitz Table 1. Properties of the field soil sample.
(1985). The pattern and geometry of cracks in
soils are however still largely unknown because the Soil property Value
cracks in soils are more easily destroyed and sensi-
Dry unit weight (kN/m3) 16.7
tive to changes in moisture content. The objective
Saturated water content (%) 30.0
of this paper is to present a method to characterize
Liquid limit (%) 40.4
a 3D random crack network in an expansive soil Plastic limit (%) 15.6
mass. A Computer Tomography (CT) test was used Shrinkage index 13.0
to investigate the crack traces in cross sections of a Expansive index 56.0
soil sample without disturbing the soil. A method

133

ISGSR2013.indb 133 10/18/2013 9:37:38 AM

 the y- and z-directions, there are 38 and
50 sections, respectively. Therefore the spac-
ing between the sections along the x direction is
6.4 mm. The spacing between the sections along
the y direction is 6.6 mm. The spacing between the
sections along the z direction is 5 mm.

3 TRACES IN THREE-DIMENSIONAL
SPACE

When a crack plane intersected with a section there

will be an intersection line in the section, which is
a trace of a crack plane. A section in the CT test
shows the traces of different crack planes (see
Fig. 2). The crack traces show a darker color in

Figure 2. A section containing crack traces.

Figure 1. (a) Soil sample; (b) CT test; (c) sections of the

soil sample in the CT test.

After the cracks reached a steady state it

was brought to the laboratory for CT tests, as
shown in Figure 1b. The cross sections of the
soil sample were obtained along the x-, y-, and
z-directions, respectively, as shown in Figure 1c. Figure 3. The crack traces on one section in Figure 2
Along the x-direction there are 36 sections. Along (unit: cm).

134

ISGSR2013.indb 134 10/18/2013 9:37:38 AM

 Figure 4. Crack traces in 3D space (unit: cm).

the section than that of the soil matrix. Assuming

the 3D crack is a plane, a crack trace that develops
along a certain direction can then be modeled as a
line. The coordinates of the two ends of the crack
are extracted from the section. The accuracy of
the data is 0.3 mm. The crack traces in a certain
section can then be obtained. Figure 3 shows the
traces of the section in Figure 2. In this way the
crack traces in every section can be obtained and
put into a 3D space as shown in Figure 4.

4 CRITERIA TO DETERMINE THE

CRACK TRACES THAT BELONG
TO A PLANE

Figure 4 shows all the traces in different sections

in the CT test. However difficulty lies in how to
determine the 3D crack plane from the traces. In
this section three criteria are proposed to deter-
mine the traces that belong to a crack plane:
1. The traces in the sections along the z-direction
must be parallel to each other. Figure 5a shows a
plane in the 3D space. If this plane is intersected
Figure 5. (a) A plane; (2) three sections along the
by three sections along the z-direction (Fig. 5b), z-direction; (c) traces between the plane and the three
the traces of the plane in the sections along the sections.
z-direction are shown in Figure 5c. The three
traces are parallel to each other. Therefore if
the traces in the sections along the z-direction If the traces are parallel to each other but the
belong to a crack plane they should parallel to spacing is different, these traces could be in
each other. different planes. Figure 6 shows two parallel
2. The spacing between the adjacent traces in sec- planes that intersect with three sections along
tions along the z-direction is equal. As the spac- the z-direction. The traces of the two crack
ing between two adjacent sections is equal (i.e., planes are parallel to each other. However
5 mm along the z-direction) in the CT test the the spacing between the traces from different
spacing of the two adjacent traces from a crack planes is not equal (i.e., d1 ≠ d2), as shown in
plane should be the same (as shown in Fig. 5c). Figure 6(b).

135

ISGSR2013.indb 135 10/18/2013 9:37:41 AM

 Figure 6. (a) Two parallel planes that intersect with
three sections; (b) traces of the two planes.

3. The traces that satisfy the first and second

criteria must be interconnected by the traces in
the sections along the x- or y-direction.
There is a possibility that the traces from two
different planes are parallel to each other and
the spacing between them is equal, as shown in Figure 7. (a) Two planes that intersect with three
Figure 7. Figure 7a shows two different planes sections; (b) traces of the two planes in the sections along
(CP1 and CP2) that are intersected by three sec- the z-direction; (c) traces of plane CP1 in the sections
tions. The traces of the two planes (as shown in along the z-direction intersected by the traces in the
Fig. 7b) are parallel to each other and have an sections along the x- and y-direction.
equal spacing. However, the traces shown in solid
line belong to plane CP1 and those shown in dash three criteria are used to determine the traces that
line belong to CP2. The traces belong to the same belong to a certain crack plane.
plane should be intersected by the traces in the
sections along the x- or y-direction. Figure 7c
shows the traces of plane CP1 along the x- and 5 IDENTIFICATION OF CRACK PLANES
y-direction (the dash lines), which intersect with
the traces in the sectios along the z-direction According to the three criteria the traces that
(the solid lines). The traces of plane CP2 do not belong to a certain crack plane can be identified.
intersect with the traces of plane in the sections First, the traces in the sections along the z-direction
along the y-direction. In the following analysis the are projected to the x − y plane. Figure 8a shows the

136

ISGSR2013.indb 136 10/18/2013 9:37:43 AM

 Figure 9. All of crack planes in 3D space (unit: cm).

they satisfy the third criterion. Then there will be a

large possibility that the crack plane is developed
according to these traces. Figure 8c shows clearly
that there is a crack plane crossing these traces.
Following this method, the crack planes in the
soil sample are all identified. The shape of the
crack planes are further investigated by Zhao
(2012). Finally the crack planes with different
shapes are obtained and presented in Figure 9.
There are totally 424 crack planes. The average
depth of the crack plane is about 21 mm. These
crack planes form a random crack network in
the 3D space. The probability distribution and the
statistical parameters can be obtained based on the
obtained crack planes (Zhao 2012).

6 SUMMARY AND CONCLUSIONS

This paper presents a method to characterize a

three-dimensional random crack network in soil
using a CT test. A field block soil sample subjected
to several drying-wetting cycles was retrieved to a
Figure 8. (a) Traces that projected on the x − y plane; laboratory for a CT test. The traces of the cracks in
(b) traces on the sections along the z-direction intersect- the soil sample were obtained in the sections along
ing by the traces in the sections along the x- or y-direc- the x-, y-, and z-directions. Three criteria were
tion; (c) traces in the same crack plane (unit: cm). proposed to determine a crack plane using these
traces of the cracks. Finally many crack planes
were established following the proposed method.
traces of the soil sample in the sections along the It is found that the crack planes intersect with each
z-direction which are projected on the x − y plane. other and form a random crack network. The aver-
The green lines, red lines, and blue lines represent age depth of the crack planes is about 21 mm. The
the crack traces on the sections of z = 25, z = 24.5, probability distribution and the statistical param-
and z = 24, respectively. The black lines represent eters for the cracks can be obtained based on the
the traces on the section along the x- or y-direction. obtained crack planes.
The traces those are parallel to each other and
have similar spacing in the three sections are found
(e.g., the three bold lines in Fig. 8a and 8b). This ACKNOWLEDGEMENTS
means the three traces satisfy the first and second
criterion. The three lines are also intersected by the This research was substantially supported
black lines (as shown in Fig. 8b), which means that by the Natural Science Foundation of China

137

ISGSR2013.indb 137 10/18/2013 9:37:45 AM

 (Grants No. 51379053 and No. 51009043), the Li, J.H., Zhang, L.M., Wang, Y., & Fredlund, D.G. 2009.
Doctoral Program of Higher Education of China Permeability tensor and representative elementary
(No. 20102302120070), and the Science, Industry, volume of saturated cracked soil. Canadian
Trade and Information Technology Commission Geotechnical Journal 46: 928–942.
Li, J.H. & Zhang, L.M. 2010. Geometric parameters
of Shenzhen Municipality. and REV of a crack network in soil. Computers and
Geotechnics 37: 466–475.
Li, J.H. & Zhang, L.M. 2011. Study of desiccation
REFERENCES crack initiation and development at ground surface.
Engineering Geology 123: 347–358.
Baecher, G.B. 1977. Statistical description of rock prop- Long, J.C.S. Gilmour, P. & Witherpoon, P.A.A. 1985.
erties and sampling. Proceedings of the 18th U.S. Model for steady fluid flow in random three-
symposium on rock mechanics. Colorado: Colorado dimensional networks of disc-shaped fractures. Water
School of Mines Press. Resources Research 21(8): 1105–1115.
Dershowitz, W.S. 1985. A new three dimensional model Nahlawi, N. & Kodikara J.K. 2006. Laboratory experiments
for flow in fractured rock. Int. Assoc. Hydrogeol. on desiccation cracking of thin soil layers. Geotechnical
17(7): 441–448. and Geological Engineering 24: 1641–1664.
Li, J.H. 2009. Field experimental study and numerical Zhao, Y. 2012. Reconstruction and numerical modelling of
simulation of seepage in saturated/unsaturated cracked three-dimensional random crack network in soils. MPhil
soil. PhD thesis. The Hong Kong University of Sci- thesis, Harbin Institute of Technology, Shenzhen,
ence and Technology, Hong Kong. China.

138

ISGSR2013.indb 138 10/18/2013 9:37:47 AM

 Geotechnical Safety and Risk IV – Zhang et al. (eds)
© 2014 Taylor & Francis Group, London, ISBN 978-1-138-00163-3

Linear regression and statistical analysis of shear strength parameters

X. Li, L.H. Chen & N. Zhang

School of Civil Engineering, Beijing Jiaotong University, Beijing, China

ABSTRACT: Linear regression is often used to obtain the sample mean value that can be regarded as
the population mean. A common misuse of linear regression is on the standard error of the mean, which
decreases with sample number and has nothing to do with the real variance of regression parameters.
Derivations are reported to calculate the “real” variance and numerical tests are used to verify these
derivations. Further, linear regression is used to analyze the tri-axial test data. The results demonstrate
that an equation transformed from traditional mean stress relation can lead to mean and variance values
with higher accuracy in linear regression, especially under the condition of small sample size (smaller than
1000).

1 INTRODUCTION method for the determination of the mean and

variance of shear strength parameters.
One of the classic job and essential ability of geo-
technical engineers is to evaluate reasonable values
of soil parameters based on results of in situ and 2 SIMPLE LINEAR REGRESSION
laboratory tests. Both the mean value and the vari- METHOD
ance of soil properties are concerned, either for
determining the design values of soil properties, or A Simple Linear Regression (SLR) equation for var-
using directly in reliability analysis. iables Y and X with n pairs of values of xi and yi is
Linear Regression (LR) is often used to calculate
the mean and variance of soil properties. However, Y a + bX + ε (1)
the linear regression all too often is used incor-
rectly in geotechnical engineering and the theoreti- where a and b are regression coefficients, ε is a ran-
cal assumptions of the method are often violated dom variable. The following quantities are defined
(Mann 1987). For example in shear strength esti- as,
mation based on triaxial test data or direct shear
test data, many engineers use the standard devia- μX = ∑ xi /nn
tion of linear regression parameters as the “real”
standard deviation of shear strength and use it in μY = ∑ yi /nn
design. However, this is not correct. The standard
deviation of linear regression parameters only rep-
SXX
X ∑ ( xi − x )2 (2)
resents the variance of the estimated mean value SXY ∑ ( xi − x )( yi − y )
from the linear regression analysis. In fact, the 1
values of standard deviation of linear regression s02 =
n−2
∑ ( yi a − bxi )2
parameters will decrease with the increase of sam-
ple size (Chen 2005, Weisberg 2005) and approach Assume (1) the residual ε has a mean of 0 and
zero when the sample size approaches infinite. a small variance of var(ε), and (2) X is the inde-
Special variance functions should be used to calcu- pendent variable, Y is the dependent variable. The
late the variances of shear strength parameters. regression coefficients a and b can be estimated as,
In this paper, derivations are reported to calcu-
late the “real” standard deviation. Numerical tests b μb SXY /SXX
X (3)
are also carried out to prove the misusing of stand-
ard deviation of linear regression parameters and a μa μY b μX (4)
to verify the special equation used to calculate the
variance. Further, linear regression is used to ana- 2 s02
Se (b ) = (5)
lyze the tri-axial test data, searching for the best SXX
X

139

ISGSR2013.indb 139 10/18/2013 9:37:47 AM

 ⎛ 2 ⎞ In numerical test 1, a, b are taken as normal
s02 ⎜ 1 + μX ⎟
2
Se (a ) (6) distributions and X as a uniform distribution with
⎝ n SXX ⎠
input settings: a = 10, σa = 2, b = 2, σb = 0.4, xmin = 0,
μX 2 xmax = 100. The test procedure for numerical test 1
ρa,b = − s02 (7) is as follows:
SXX
X
1. Giving a, σa and b, σb for a, b.
where a, b are the sample means of a, b; Se(a) is 2. Generating n pairs of random numbers a, b
the standard error of a, Se(b) is the standard error which follow normal distributions N(a, σa2) and
of b, ρa,b is correlation coefficient of a and b; s0 is N(b, σb2).
often called as standard error of regression. 3. Generating n random numbers for X, which fol-
low a uniform distribution on [xmin, xmax].
4. For each xi, selecting one pair a, b and using
3 LIMITATION OF LINEAR Eq. 1 to calculate yi.
REGRESSION METHOD 5. Finally, n pair xi and yi are obtained. Calculating
the sample mean and standard error using
Giving the mean and variance of X, SXX increases Eqs. 3∼6.
with the size of X (referring to Eq. 2). Giving a, b, ε, 6. Comparing the sample mean and standard error
the standard error of regression s0 can be regarded with the input value a, σa and b, σb.
as a constant. If a, b, ε are independent and follow The results of numerical test 1 are listed in
normal distribution, s02 is equal to var(ε) (Yan & Table 2. Referring to Table 2, the sample means cal-
Su 2009). Obviously, Se2(a) and Se2(b) are inversely culated by Eqs. 3 and 4 have some difference with
proportional to the sample size number. the input values and become closer to the input
In fact, the Standard Errors of the Mean (SEM), values with the increase of sample size. Referring
Se(a) and Se(b) are the standard deviation of the to Figure 1, Se2(a) and Se2(b) are found to be
sample-mean’s estimate of a population mean. It inversely proportional to the sample size and has
can also be viewed as the standard deviation of the nothing to do with the input var(a) and var(b).
error in the sample mean relative to the true mean,
since the sample mean is an unbiased estimator.
Clearly, the SEMs calculated by Eqs. 5 and 6 are Table 2. Simple linear regression with uniform X (test 1).
not the real standard deviations of a and b; they
just indicate the accuracy of the mean value calcu- Sample size a b Se2(a) Se2(b)
lated by Eqs. 3 and 4. n (Eq. 4) (Eq. 3) (Eq. 6) (Eq. 5)

10 18.75 1.742 266 0.124

3.1 Numerical test of simple linear regression 50 9.37 2.02 46.83 0.01
method 200 5.40 2.12 10.12 0.003
1000 9.96 2.00 2.21 6.5E-4
Considering Eq. 1, numerical tests can be car- 5000 9.38 2.02 0.42 1.3E-4
ried out to check the validity of simple linear 20000 9.77 2.01 0.1 3.2E-5
regression. The settings of numerical tests are 100000 9.96 2.00 0.02 6.5E-6
listed in Table 1.

Table 1. Summary of numerical test settings.

Test Eq. Table

no. a b X no. no.

1 Ν (10, 4) N (2, 0.16) U (0, 100) 1 2

2 Ν (10, 4) N (2, 0.16) I (0, 100) 1 3
3 Ν (10, 4) N (2, 0.16) U (0, 100) 8 4
4 Ν (10, 4) N (2, 0.16) U (0, 100) 9 5
5 Ν (10, 4) N (2, 0.16) U (0, 100) 12 5
6 Ν (10, 4) N (2, 0.16) I (0, 100) 10 6
7 Ν (10, 4) N (2, 0.16) I (0, 100) 12 6

*Ν(μ,σ2) denotes a normal distribution; U(Xmin, Xmax)

denotes a uniform distribution; I(Xmin, Xmax) denotes m
value with the same spacing in interval [Xmin, Xmax]. Figure 1. Standard error vs. sample size.

140

ISGSR2013.indb 140 10/18/2013 9:37:48 AM

 Sometimes, X may not follow a uniform Table 4. Simple linear regression with uniform X (test 3).
distribution. For example, X may be the confining
pressure in triaxial tests, which normally is taken as Sample a b Se2(a) Se2(b)
50 kPa, 100 kPa, 200 kPa, 400 kPa, etc. To check size (Eq. 3) (Eq. 4) (Eq. 5) (Eq. 6)
the influence of X on simple linear regression, m n = 10 4.2 2.37 3.42 2.7E-2
groups of repeated xi are used in Numerical test 2 n = 50 7.96 2.06 7.8E-2 3.6E-3
(referring to Table 1). The procedure is as follows:
n = 200 10.32 1.98 2.7E-2 1.4E-3
1. Giving a, σa and b, σb for a, b. n = 1000 10.19 2.00 3.9E-3 2.5E-4
2. Generating n*m pairs of random numbers a, b n = 5000 9.33 2.06 1.1E-4 7.9E-4
which follow normal distributions N(a, σa2) and n = 20000 10.55 1.97 1.4E-4 2.3E-4
N(b, σb2). n = 100000 12.66 1.74 1.4E-5 3.0E-4
3. Generating m numbers for X with the same
spacing in the interval [xmin, xmax].
4. For each xi, selecting n pairs of a, b and using
Eq. 1 to calculate yi. values listed Table 4, if the sample size is smaller
5. Finally, n*m pairs of xi and yi are obtained. than 1000, Eq. 8 can lead to better accuracy of a, b
Calculating the sample mean and standard comparing to Eq. 1. However, if the sample size is
error using Eq. 3∼6. larger than 1000, Eq. 1 will lead to better accuracy
6. Comparing the sample mean and standard error of a, b. Such phenomenon is consistent with the
with the input value a, σa and b, σb. values of standard error Se(a) and Se(b), i.e. the
indicator of accuracy of sample mean.
With input settings: a = 10, σa = 2, b = 2, σb = 0.4,
the results of numerical test are listed in Table 3.
The distribution of X does not have significant 3.2 Summary
influence on the results calculated by Eqs. 3∼6.
Again, the accuracy of a, b calculated by Eq. 3 and In short, the numerical test results prove that:
4 increases with the sample size. Se2(a) and Se2(b) 1. Se2(a) and Se2(b) are inversely proportional
are inversely proportional to the sample size and to the sample number; having nothing to do
show no relation with var(a) and var(b). Another with the real var(a) and var(b) but indicating
thing to note is that Se(a)2 is much larger than the accuracy of the mean value calculated by
Se2(b) in the upper two cases. Referring to Eq. 6, Eqs. 3 and 4.
because Se2(a) is about μX2, i.e. 2500 in case 1, times 2. Both Eq. 1 and Eq. 8 can be used to estimate the
of Se2(b). That is to say, the b value calculated by sample mean a, b. The selection of Eq. 1 or Eq.
Eq. 3 is of much higher accuracy comparing the a 8 can be referring to the standard errors Se2(a)
value calculated by Eq. 4. and Se2(b) calculated in two conditions.
Considering another form of Eq. 1 3. In geotechnical engineering, the accuracy of 1%
is normally acceptable for soil parameters. So
Y a according to Tables 2∼4, at least dozens of data
= + b+ ε (8)
X X are required for the estimation of the mean of
regression parameter a, b.
where 1/X can be regarded as independent variable
and Y/X can be regarded as dependent variable.
Applying SLR to Eq. 8 with the same data source
4 DETERMINATION OF VARIANCE
used in numerical test 1, the results of the numeri-
cal test are listed in Table 4. Referring to the a, b
4.1 Theory
A common misuse of linear regression is taken
Table 3. SLR with m group repeated X (test 2). the SEMs calculated by Eqs. 5 and 6 as the “real”
standard deviation of a, b. As SEMs decrease
a b Se2(a) Se2(b)
with the sample number, if the sample number
Sample size (Eq. 4) (Eq. 3) (Eq. 6) (Eq. 5)
is large, such mistake will severely underestimate
n = 5, m = 2 17.12 1.89 1763 0.28 the standard deviation of a, b. On the contrary,
n = 20, m = 2 10.32 1.99 282 4.5E-2 if the sample number is small, such mistake will
n = 20, m = 10 15.24 1.87 15.92 4.1E-3 significantly overestimate the standard deviation
n = 100, m = 2 18.29 1.88 48.85 7.8E-3 of a, b.
n = 1000, m = 5 10.21 1.99 0.78 1.8E-4 As discussed in section 3, the mean of a, b can
n = 5000, m = 10 9.92 2.00 0.06 1.5E-5 be calculated correctly in SLR. The problem is to
estimate the variances σa2 and σb2. Referring Eq. 1,

141

ISGSR2013.indb 141 10/18/2013 9:37:50 AM

 if a, b, ε are independent and follow normal distri- If a, b are assumed to be dependent, Eq. 7 can
bution, the following relation is valid, be written as (if σε is negligible),

Var( ) = μX 2σ b 2 + σ a2 + σ ε2
Var( ) b2 Var( (9) Var( ) = μX 2σ b 2 + σ a2 + 2 ρaabσ aσ b μX
Var( ) b2 Var(
(11)
where b is the mean of b calculated by Eq. 3, which
is considered to be of high accuracy; σε is normally Eq. 11 can be solved by quadratic regression,
regarded as a negligible value. So there are two taking μX as independent variable and [var(Y) −
unknowns σa and σb. b2var(X)] as dependent variable.
If the sample is divided into n groups, and To solve for var(a) and var(b), the following
Eq. 9 is applied to each group, n equations can be equation can be also used in LR instead of Eq. 9
obtained and used to solve the two unknowns σa
and σb. Var( ) b2 Var( ) 1
Obviously, the error of var(X) will affect the = σ b2 + σ a2 (12)
μX 2 μX 2
accuracy of the solved σa and σb. To reduce
the value of var(X), n groups of data should be
The choice of Eq. 9 or 12 depends on the accu-
selected according to the X value, but not selected
racy indicated by the standard errors while apply-
randomly. For example, σa and σb can be solved by
ing these two equations.
using the following procedure:
1. dividing the sample into m groups accord- 4.2 Numerical test of the estimation of variance
ing to the X value with an interval of
[max(X) – min(X)]/m. Taking a, b as normal distributions and X as a uni-
form distribution with input settings (Tests 4 & 5
–2 each group data, a pair of [μX , var(Y) −
2
2. For
b var(X)] can be obtained and a total m pairs of in Table 1): a = 10, σa = 2, b = 2, σb = 0.4, xmin = 0,
data can be obtained. xmax = 100, the results of numerical test are listed
3. Taking μX2 as independent variable and [var(Y) − in Table 5. Referring to Table 5, Eq. 9 can lead to
– accurate var(b), but unreliable var(a), which has
b 2var(X)] as dependent variable, σa2 and σb2 can
be solved using linear regression (Eqs. 3 and 4). great difference with the input value and is of high
standard error. Whereas, Eq. 12 can lead to values
The above procedure is valid for general cases, with higher accuracy for var(a) and similar accu-
such as for X following a uniform distribution. If X racy for var(b) comparing to Eq. 9.
is m groups of repeated data and m groups of data If X is taking as m groups of repeated data,
are selected according to X value, var(X) is 0 and the results of numerical tests are listed in Table 6.
Eq. 9 can be simplified as (if σε is negligible), According to Table 6, Eq. 12 can lead to values
with higher accuracy comparing to Eq. 10. Using
Var(Y ) = μX 2 b
2
+ σ a2 (10) X as m groups of repeated data instead of X with

Table 5. Estimation of variance with X of uniform distribution.

SLR using Eq. 9 (test 4) SLR using Eq. 12 (test 5)

σa2 σb2 σa2 σb2

n* m♥ μ (Eq. 4) Se2 (Eq. 6) μ (Eq. 3) Se2 (Eq. 5) μ (Eq. 3) Se2 (Eq. 5) μ (Eq. 4) Se2 (Eq. 6)

15 3 −288 18391 0.235 6.8E-3 −547 21658 0.311 7.5E-3

60 3 −195 10007 0.215 5.8E-4 25 1034 0.175 6.1E-3
100 5 153 20646 0.046 1.3E-3 6.24 217 0.124 1.9E-3
500 10 −13 6692 0.167 3.2E-4 4.12 1.38 0.164 2.1E-4
1000 10 32 1027 0.143 4.9E-5 3.69 0.39 0.156 6.7E-5
1000 50 56.7 1241 0.132 6.6E-5 4.53 2.9E-3 0.155 7.4E-5
10000 10 −2.14 132 0.163 6.7E-6 5.46 4.4E-3 0.161 6.7E-6
10000 100 0.24 296 0.162 1.5E-5 3.59 6.2E-5 0.158 8.3E-6
100000 10 8.24 7.0 0.158 3.5E-7 5.65 5.2E-3 0.159 8.9E-7
100000 100 0.161 24.8 0.161 1.2E-7 4.08 4.4E-6 0.159 7.2E-7

*Total size of sample; ♥number of groups.

142

ISGSR2013.indb 142 10/18/2013 9:37:50 AM

 Table 6. Estimation of variance with X of m group repeated values.

SLR using Eq. 10 (test 6) SLR using Eq. 12 (test 7)

σa2
σb
2
σa2 σb2

n* m♥ μ (Eq. 4) Se2 (Eq. 6) μ (Eq. 3) Se2 (Eq. 5) μ (Eq. 3) Se2 (Eq. 5) μ (Eq. 4) Se2 (Eq. 6)

15 3 602 1.1E6 0.106 5.9E-2 −108 2.4E5 0.451 1.1E-1

60 3 −121 4.5E5 0.232 2.5E-3 24.6 1.0E3 0.161 4.5E-3
100 5 −94.9 2.5E5 0.230 1.3E-3 2.66 2.6E2 0.187 6.6E-4
500 10 −0.74 3.2E3 0.155 1.6E-4 4.90 9.1E-1 0.158 1.5E-4
1000 10 −11 3.8E3 0.163 1.9E-4 4.25 3.2E-1 0.155 5.2E-5
1000 20 −3.3 1.5E3 0.159 7.5E-5 5.46 2.8E-2 0.155 3.7E-5
10000 10 −8.5 9.4E2 0.168 4.7E-6 3.90 3.9E-2 0.161 6.3E-6
10000 20 23.7 3.5E2 0.151 1.8E-6 4.08 1.0E-2 0.161 1.4E-5
100000 10 5.77 2.8E2 0.159 1.4E-6 3.91 1.9E-3 0.160 3.1E-7
100000 25 5.67 2.7E2 0.160 1.3E-6 3.84 1.9E-4 0.160 4.9E-7

*Total size of sample; ♥number of groups.

a uniform distribution, the accuracy of var(a) ⎧ b −1

and var(b) also increases slightly under the same ⎪⎪ μ f ≈ 2 b
sample size. ⎨ (16)
⎪μ ≈ a
⎪⎩ c 2 b
4.3 Summary
According to Tables 5 and 6, Eqs. 9∼12 can be used The standard deviation of c and f can be calcu-
to solve the variance of the regression parameters lated as:
and Eq. 12 shows better performance in the numer-
ical tests, in which μX is relatively large comparing
to a and b. At least hundreds of data are required 1 1 2 −3
σ c = σ a2 + σ b2 b 0 25 ρababb −2σ aσ b
to obtain a meaningful variance that can be used in 4b 16
geotechnical engineering practice. 1
σ f = σ b (b 05
b −1 5 )
4
5 APPLICATION OF THE STATISTICAL (17)
ANALYSIS OF SHEAR STRENGTH
In reliability analysis, the correlation between c
In geotechnical engineering, a core concern is the and f is often ignored to simplify the problem. In
shear strength, which is often simply character- numerical test, c and f are assumed to be independ-
ized by the parameters of cohesion c and friction ent and used to generate triaxial test data, as listed
angle φ. There are several equations in the SLR for in Table 7. Numerical tests 8∼10 refer to fine soil;
statistical analysis of shear strength based on triax- numerical tests 11 and 12 refer to coarse soil; and
ial test data as follows: numerical tests 13 and 14 refer to well compacted
widely-graded soil. The numerical test results are
2c cos φ (σ 1 + σ 3 )sin
) i φ listed in Table 8 and demonstrate that:
( 1 3) = (13)
1. For the same series of triaxial test data, Eq. 8
σ1 2 (φ 2 + π 4) σ 3 tan2 (φ 2 π 4) (14) and Eq. 15 can lead to the values of highest
accurate among these equations.
2. If the soil cohesion c is not small, hundreds data
σ1 1
= 2c tan (φ 2 π 4) tan2 (φ 2 π 4) (15) is enough to obtain meaningful statistical char-
σ3 σ3 acteristic for shear strength, i.e. μc, μf, var(c),
and var(f). If c is small, the variance of c is
where σ1 is the major effective principal stress; also of low accuracy even when the sample size
σ3 is the minor effective principal stress. Denoting goes up to thousands. In this condition, var(c)
f = tanφ, λ = tan(φ/2 + π/4), 2cλ as a, and λ2 as b, is taken as an empirical value, e.g. 0.01∼0.1μc2,
the mean of c, f can be calculated as, instead of the var(c) obtained in LR.

143

ISGSR2013.indb 143 10/18/2013 9:37:52 AM

 Table 7. Summary of numerical test settings.

Test no. n* c tanφ σ3 a = 2cλ b = λ2

8 40 N (50, 100) N (0.36, 0.002) 50, 100, 150, 200 ∼N (142, 842) ∼N (2.03, 0.029)
9 400 N (50, 100) N (0.36, 0.002) 50, 100, 150, 200 ∼N (142, 842) ∼N (2.03, 0.029)
10 4000 N (50, 100) N (0.36, 0.002) 50, 100, 150, 200 ∼N (142, 842) ∼N (2.03, 0.029)
11 800 N (10, 4) N (0.73, 0.006) 100, 200, 300, 400 ∼N (39.4, 68.2) ∼N (3.89, 0.236)
12 8000 N (10, 4) N (0.73, 0.006) 100, 200, 300, 400 ∼N (39.4, 68.2) ∼N (3.89, 0.236)
13 800 N (100, 400) N (0.73, 0.006) 100, ..., 2000 ∼N (394, 6820) ∼N (3.89, 0.236)
14 8000 N (100, 400) N (0.73, 0.006) 100, ..., 2000 ∼N (394, 6820) ∼N (3.89, 0.236)

Denote λ = tan(φ/2 + π/4); *total size of sample.

Table 8. Results of statistical analysis of shear strength (numerical test).

Eq. 1 and Eq. 13 Eq. 1 and Eq. 14 Eq. 8 and Eq. 15

Test
no. 2ctanφ sinφ a = 2cλ b = λ2 a = 2cλ b = λ2

8 N (86.2, 575) N (0.35, 0.003) N (151, 21.8) N (1.94, 0.105) N (144, 523) N (2.02, 0.066)
9 N (81.7, 606) N (0.36, 0.0021) N (143, 818) N (2.01, 0.044) N (143, 838) N (2.01, 0.030)
10 N (82.5, 617) N (0.36, 0.0020) N (144, 883) N (2.02, 0.041) N (144, 864) N (2.02, 0.043)
11 N (−14.1, −39.6) N (0.615, 0.0040) N (41.5, −834) N (3.88, 0.265) N (41.3, −30.4) N (3.88, 0.250)
12 N (−12.8, 216) N (0.614, 0.0039) N (42.7, 356) N (3.88, 0.242) N (42.4, 270) N (3.88, 0.242)
13 N (82.2, 7490) N (0.606, 0.0018) N (427, −12840) N (3.85, 0.268) N (399, 9104) N (3.89, 0.256)
14 N(69.4, 1458) N(0.609, 0.0021) N (389, 25607) N (3.90, 0.236) N (390, 8864) N (3.90, 0.259)

6 CONCLUSIONS ACKNOWLEDGEMENTS

In this paper, theoretical and numerical efforts This study is supported by the State Key
are conducted to study the capacity of the lin- Development Program of Basic Research of China
ear regression method. The study demonstrates (Project No. 2012CB026104), a Chinese National
that: Key Technology R&D Program of the Ministry of
Science and Technology Project (2012BAH10B01),
1. A common misuse of linear regression is to
and the National Natural Science Foundation of
take the standard errors of the mean (SEM,
China (Grant No. 51109003).
Eqs. 5 and 6) as the “real” standard deviations
of regression parameters a, b. In fact, SEMs
are inversely proportional to the sample size REFERENCES
number, have nothing to do with the real var(a)
and var(b), but indicate the accuracy of the Chen, L.H. & Chen, Z.Y. 2007. Effect of nonlinear
mean value a, b (Eqs. 3 and 4). strength of rockfill on slope stability of high earth-rock
2. If a is relatively small comparing to the mean dam. Rock and Soil Mechanics 9(28): 1807–1810.
of independent variable, μX, Eq. 8 can improve Mann, C.J. 1987. Misuses of linear regression in earth
the accuracy of regression analysis comparing sciences. In W.B. Size (ed.), Use and Abuse of Statisti-
to Eq. 1, especially under the condition of small cal Methods in the Earth Sciences: 74–106. New York:
Oxford University Press.
sample size (smaller than 1000). Weisberg S. 2005. Applied linear regression. Hoboken:
3. In the statistical analysis of shear strength based John Wiley & Sons, Inc.
on triaxial test data, Eqs. 8 and 15 are recom- Yan, X. & Su, X.G. 2009. Linear regression analysis-
mended because they lead to mean and variance theory and computing. Singapore: World Scientific
values with higher accuracy. Publishing Co. Pte. Ltd.

144

ISGSR2013.indb 144 10/18/2013 9:37:54 AM

 Geotechnical Safety and Risk IV – Zhang et al. (eds)
© 2014 Taylor & Francis Group, London, ISBN 978-1-138-00163-3

Uncertainty quantification by polynomial chaos expansion

for geotechnical reliability-based analysis

S.H. Marques, A.T. Gomes & A.A. Henriques

Department of Civil Engineering, Faculty of Engineering, University of Porto, Portugal

ABSTRACT: Uncertainty quantification is the process of determining the effect of input uncertainties
on response metrics of interest. Many approximate methods have been developed so far for the purpose,
and among these methods the polynomial chaos expansion is considered a technique with strong math-
ematical basis and ability to produce functional representation of stochastic variability. The approach has
proven to be an efficient methodology to study several stochastic problems, considered the original form
where optimal convergence for orthogonal Hermite polynomials is only achieved for gaussian stochastic
processes. Regarding the developments in the stochastic response surface methodology and some heuris-
tic and optimisation concepts, results for a design example with correlated nonnormal random variables
are presented in the light of the quality of the approximate metamodels. Considered truncated full and
sparse polynomial chaos expansions, the efficiency and accuracy provided by different schemes of experi-
mental design are analysed and the convergence process is lastly discussed.

1 INTRODUCTION response. As a consequence of this sparse repre-

sentation, a rather small number of polynomial
Metamodelling is a very interesting research area chaos terms is eventually retained, which may be
for the replacement of simulation models when obtained at a reduced computational cost com-
the trade-off between efficiency and accuracy is pared to the classical truncated full polynomial
appropriate. If a model is a simplified and abstract chaos approximation. In fact, the number of terms
representation of reality, a metamodel is a further to be computed grows dramatically with the size of
abstraction, a model of a model, which is typically the input random vector, which makes the compu-
easier to generate and evaluate than the model itself. tational cost of the classical solution schemes unaf-
In fact, the use of metamodels is now established fordable when the model is expensive to evaluate.
to reduce the need to run expensive simulations, For the purpose, Blatman & Sudret (2010b)
as stated by Can & Heavey (2011). The selection present also a methodology for the efficient com-
of a suitable approach is not straightforward and putation of global sensitivity indices by using a
should take into account the final purpose. sparse polynomial chaos expansion. Global sen-
Considered the literature, Huang et al. (2009) sitivity analysis is related to the quantification of
present an Excel add-in implementation for the sto- the output uncertainty due to changes of the input
chastic response surface methodology, developed parameters over their entire domain of variation.
for practitioners and used to analyse illustrative In this context, variance-based methods which rely
geotechnical examples for reliability analysis. In upon the decomposition of the response variance
turn, Li et al. (2011) present an improved stochastic as a sum of the contributions of each input vari-
response surface methodology for reliability analy- able are of major interest to quantify the contribu-
sis of rock slopes involving correlated nonnormal tion of each parameter to the output variability.
random variables. Huang et al. (2009) and Li et al. Thus, soft computing technologies have
(2011) accordingly argue that there appears to be acquired increasing importance in engineering and
no simple method to manage the proliferation of several techniques are nowadays available. In this
terms for the higher random dimensions, noted line of thought, Armani et al. (2011) present a rela-
that it is possible that the order can be kept at four tively new metamodelling building technique that
or below for most practical problems. is able to generate explicit mathematical expres-
To address such problems described in recent sions describing the relationship between the input
work, Blatman & Sudret (2010a) present an adap- variables and the modelled response, the so called
tive algorithm that builds a sparse polynomial genetic programming. In turn, Hofwing et al. (2011)
chaos expansion to represent the random model present a genetic algorithm used to find the terms

145

ISGSR2013.indb 145 10/18/2013 9:37:54 AM

 to be included in a proposed optimal polynomial sional Hermite polynomials; and U = vector of
regression model. In practice, the genetic algo- standard normal random variables.
rithm generates the optimal set of exponents in a Such technique involves the calculation of chaos
polynomial regression model with improved per- expansion coefficients based on a set of response
formance compared to the traditional regression function evaluations. The regression approach uses
model, in that for a full expanded regression model a least-squares solution to find the complete set
a large number of design sampling points is really of coefficients that best match a set of response
required. The determination of the higher order values obtained from a number of design sam-
terms that yields the most accurate regression pling points. A stochastic response surface is then
model is indeed an optimisation problem. produced, describing the relationship between
Therefore, considered truncated full and sparse the input variables and the output, which repre-
polynomial chaos expansions, the efficiency and sents the metamodel. A different order for the
accuracy provided by different schemes of experi- response surface can produce a robust estimation
mental design are analysed in the next sections with of the reliability index for a performance function,
a design example for a concrete gravity retaining and usually the accuracy of the response surface
structure, considered correlated nonnormal ran- increases as the order of the surface response
dom variables. increases. Despite the excellent convergence prop-
erties reported for general probabilistic analy-
sis problems, there is no universal base for every
2 MULTI-DIMENSIONAL HERMITE problem and convergence difficulties were found
POLYNOMIAL CHAOS EXPANSION when highly nonlinear behaviour was considered.
In fact, even if the basic input random variables
Despite being an attractive technique available to are modelled as gaussian, the response may present
the community of researchers and practitioners, highly nongaussian behaviour due to the nonlinear
the application of the stochastic response surface relationship established between the basic input
methodology to geotechnical reliability problems random variables and the system response.
has not been extensively investigated. However, the polynomial chaos expansion suf-
The methodology involves the propagation of fers yet from a set of other drawbacks. In the pres-
input uncertainties through a model to express ence of a particular computational budget for a
a random output, expanded for instance in the problem with a particular dimensionality, the pri-
standard normal space of uncorrelated random mary weakness of the regression approach is the
variables by a multi-dimensional Hermite polyno- common need for a large number of sample points
mial chaos, summarised by Equations 1, 2 and 3: in the design of experiments as the order of the
multi-dimensional Hermite polynomials and the
Y (
a0 + ∑ i1=1 ai1
n
(Ui ))
number of random variables increase. Thus, the use
of a multi-dimensional Hermite polynomial chaos
+∑ ∑
n
i1=1
i1
i 2 =1 (ai i 2 2 (Ui Ui )) expansion for the generation of the response sur-
face has been hindered by the eventual high com-
+∑ ∑
n
i1 1
i1
i2 1 ∑
i2
i3 1 (ai i 2i (Ui1 Ui )
U i 3) + … putational complexity that the technique shows.
Therefore, there is great interest in presenting
(1) metamodelling techniques that are able to provide
⎛ 1⎞ T high quality metamodels at a reasonable computa-
⋅U
U U
Γ o (U U io ) = ( −1) ⋅ e ⎝ 2⎠ tional complexity, introduced some heuristic and
o
optimisation concepts. A novel approach should
⎛ 1⎞ T take into account that truncated full polynomial
∂o −
⎝ 2⎠
⋅U ⋅U
⋅ e (2) chaos expansions are quite far from being simpler
∂U i ∂U io to visualise and evaluate than the model itself, and
so the replacement is not always worthwhile in
⎡U1 ⎤ that Monte Carlo simulation based on counting
⎢U ⎥ remains expensive. Another interesting issue is the
U=⎢ ⎥
2
(3) pursuit of accuracy, considered truncated full and
⎢  ⎥
⎢ ⎥ sparse polynomial chaos expansions.
⎣U n ⎦ The flowchart of procedures adopted for uncer-
tainty quantification by polynomial chaos expan-
where Y = model output; a0, ai1, ai1i2, ai1i2i3, ... = unknown sion is then presented in Figure 1. For the purpose, a
coefficients to be determinated; n = number of simplified approach suggested by Kiureghian & Liu
standard normal random variables; o = order of (1986) and described by Haldar & Mahadevan (2000)
multi-dimensional Hermite polynomial chaos may be considered for the equivalent standard nor-
expansion; Γo(Ui1, ..., Uio) = order o multi-dimen- mal correlation matrix computation. The trans-

146

ISGSR2013.indb 146 10/18/2013 9:37:54 AM

 Table 1. Transformations for representation of
random variables xi* with different distributions as a
function of standard normal random variables yi*.

Distribution xi* = f(yi*)

Normal xi* = μ + σ ⋅ yi*

Lognormal xi* = e(μ1+σ1⋅yi*)
Weibull xi = sc ⋅ (− ln(1 − Φ(yi*)))(1/sh)
Distribution f(xi*) = yi*

Normal (xi* − μ)/σ = yi*

Lognormal (ln(xi*) − μ1)/σ1 = yi*
Weibull Φ−1(1 − e(−(xi*/sc)sh)) = yi*

μ1-log mean; σ1-log standard deviation; sh-shape;

sc-scale.
Φ-cumulative standard normal distribution function.
Φ−1-inverse cumulative standard normal distribution
function.

Figure 1. Flowchart of procedures adopted for uncer-

tainty quantification by polynomial chaos expansion.

formations xi* = f(yi*) and f(xi*) = yi* derived by

Equation 4 are further considered:
Figure 2. Concrete gravity retaining structure.
xi = FXi ⎡
*
⎣ ( yii )⎤⎦
* 1⎡
⎣ ( )
FXi xi ⎤ = yi
*
⎦
*
(4)
c-ϕ soil shown in Figure 2, wherein groundwater
level is away. Considered the inclined eccentric
where xi* = random variable in the original space; loading problem and the calculation model for
yi* = random variable in the standard normal space; bearing capacity, the performance function may be
FXi = cumulative nonnormal distribution function; described by the simplified Equation 5:
FXi −1 = inverse cumulative nonnormal distribu-
tion function; Φ = cumulative standard normal
distribution function; and Φ−1 = inverse cumulative
M (
f B B H H c w w cf f f q ) (5)
standard normal distribution function.
Considered normal, lognormal and weibull dis- where the sum of B1 and B2 is the foundation width
tributions, the described transformations for rep- B; H1 is the wall height; H2 is the foundation height;
resentation of random variables xi* with different γc is the unit concrete weight; ϕw is the friction angle
distributions as a function of standard normal ran- of the soil on the active and passive sides of the
dom variables yi* are further detailed in Table 1. wall; γw is the unit soil weight on the active and pas-
sive sides of the wall; cf is the cohesion of the foun-
dation soil; ϕf is the friction angle of the foundation
3 DESIGN EXAMPLE soil; γf is the unit weight of the foundation soil;
and q is the variable load at ground surface. Other
The design example is referred to the concrete grav- considered parameters, namely for the earth pres-
ity retaining structure on a relatively homogeneous sure coefficients, are the soil-wall interface friction

147

ISGSR2013.indb 147 10/18/2013 9:37:55 AM

 angle on the active side of the wall δwa = 2/3ϕw; the brief analysis, the correlation between the cohesion
soil-wall interface friction angle on the passive side and the friction angle of the foundation soil depends
of the wall δwp = 0; and the soil-foundation inter- particularly on the material, and the question
face friction angle δf = ϕf; considered either a 50% whether the two parameters are correlated is still not
reduction for passive earth pressures. clearly decided in the literature. References for nega-
Table 2 summarises the description of basic input tive correlation are common but regarding a sensitiv-
variables, with different distribution types. The con- ity analysis for the design example, the influence of
sidered coefficients of correlation between the ran- negative correlation on reliability is favourable. Thus,
dom variables are either presented in Table 3, noted the hypothesis of uncorrelated random variables is
that the correlation matrices are idealised with the probably cautious, noted that the characteristic val-
main purpose of creating high correlation between ues of soil properties are often calculated separately,
some of the random variables. The strength of the neglecting the effects of correlation.
association is measured by a correlation coefficient
in the range [−1,1], for the case that both variables 4 RESULTS AND DISCUSSION
vary inversely or together exactly, respectively. In a
The published results should be referred to the
Table 2. Summary description of basic input variables. considered coefficient of variation for the fric-
tion angle of the foundation soil, one of the most
Basic input important parameters concerning reliability, stated
variables Distribution mv cv Statistics that the increased nonnormal behaviour of the
bearing capacity response correlates well with the
B1 (m) Deterministic – – 1.75
higher values of the parameter.
B2 (m) Deterministic – – 1.00
Considered the implementation of the foremost
H1 (m) Deterministic – – 7.00
H2 (m) Deterministic – – 1.00
detailed methodology, Table 4 presents the sum-
mary description of results for the bearing capac-
γc (kN/m3) Deterministic – – 24.00
ity model, included the reliability index obtained
ϕw (º) Lognormal 33.00 0.10 μ1-3.4915
σ1-0.0998 by the first order reliability method and the Monte
γw (kN/m3) Normal 18.80 0.05 – Carlo simulation, in conjunction with the descrip-
cf (kN/m2) Lognormal 14.00 0.40 μ1-2.5648
tion of some important statistical parameters
σ1-0.3853 as the output mean, standard deviation, skew-
ϕf (º) Lognormal 32.00 0.05 μ1-3.4645 ness and kurtosis. It is noted that the differences
σ1-0.0500 between the reliability index obtained from both
γf (kN/m3) Normal 17.80 0.05 – methodologies are somewhat notorious due to the
q (kN/m2) Weibull – – sh-3.25 important nonlinear behaviour.
sc-10.00 The bearing capacity metamodels are built
according to the schemes of experimental design
mv-mean value; cv-coefficient of variation. drawn in Figure 3. They are derived from a number
μ1-log mean; σ1-log standard deviation; sh-shape; of 1000 design sampling points selected at random
sc-scale. from the original space of random variables, noted
that smaller samples are always designed from
Table 3. Coefficients of correlation between the ran- larger samples.
dom variables. Table 5 summarises the number of terms in multi-
Correlation matrix for the bearing capacity model
dimensional Hermite polynomial chaos expansion

ρϕwϕw ρϕwγw ρϕwϕf ρϕwq ρϕwcf ρϕwγf Table 4. Summary description of results for the bearing
ργwϕw ργwγw ργwϕf ργwq ργwcf ργwγf capacity model.
ρϕfϕw ρϕfγw ρϕfϕf ρϕfq ρϕfcf ρϕfγf
ρqϕw ρqγw ρqϕf ρqq ρqcf ρqγf Reliability Reliability
ρcfϕw ρcfγw ρcfϕf ρcfq ρcfcf ρcfγf index index
βFORM βMCS
ργfϕw ργfγw ργfϕf ργfq ργfcf ργfγf
1.0 0.5 0.9 0.0 0.0 0.5 2.484 2.529
0.5 1.0 0.5 0.0 0.0 0.9 Output Output standard
0.9 0.5 1.0 0.0 0.0 0.5 mean deviation Output Ouput
0.0 0.0 0.0 1.0 0.0 0.0 (kN/m) (kN/m) skewness kurtosis
0.0 0.0 0.0 0.0 1.0 0.0
0.5 0.9 0.5 0.0 0.0 1.0 1267 899.1 1.802 6.391

ρ-coefficient of correlation. MCS and statistics results from 106 simulation steps.

148

ISGSR2013.indb 148 10/18/2013 9:37:58 AM

 Table 6. Summary description of results for the bearing
capacity metamodels.

Polynomial Relative Relative

chaos error (%) error (%)
Case* DoE o expansion βFORM βMCS

1 1000 4 Truncated −3.027 −4.040

full
2 150 4 Sparse 0.354 0.572
3 250 4 Sparse 0.274 3.238
4 500 4 Sparse 0.636 2.516
5 1000 4 Sparse −0.209 0.903
6 150 5 Sparse −1.377 −1.805
7 250 5 Sparse −1.578 −3.192
Figure 3. Schemes of experimental design. 8 500 5 Sparse 0.958 −1.459
9 1000 5 Sparse 1.361 −0.391
Table 5. Number of terms in multi-dimensional
Hermite polynomial chaos expansion. DoE-design of experiments; o-order.
MCS results from 106 simulation steps.
Number of terms according *R2-coefficient of determination, for the different cases
n o to (n + o)!/n! × o! 1–9: 1-R2 = 0.993579; 2-R2 = 0.999991; 3-R2 = 0.999958;
4-R2 = 0.999921; 5-R2 = 0.999902; 6-R2 = 0.999981;
6 1 7 7-R2 = 0.999961; 8-R2 = 0.999902; 9-R2 = 0.999897.
6 2 28
6 3 84
6 4 210
6 5 462
6 6 924

n-number of standard normal random variables.

o-order of multi-dimensional.
Hermite polynomial chaos expansion.

for the case of a number of 6 standard normal

random variables and different orders of multi-
dimensional Hermite polynomial chaos expansion. Figure 4. Model deviation for mean and standard
Considered the preliminary work, a fourth order deviation of the bearing capacity metamodels, cases 1–9,
is considered for the truncated full polynomial chaos statistics results from 106 simulation steps.
expansion, which include a number of 210 terms. The
sparse polynomial chaos expansions link a selected
number of 100 terms derived from a sensitivity function; M = ith fitted value of the performance
analysis for importance classification and a heuris- function; and M = mean of the observed values of
tic process of trial and error for a selected group of the performance function.
terms, considered a fourth and a fifth order. The relative errors obtained by different reliabil-
For the purpose, the coefficient of determina- ity techniques are then presented in conjunction
tion evaluated on the data set for the metamodels with the description of some important statisti-
is based on the variances of the fitted values and cal parameters. For the purpose, Table 6 presents
observed values of the dependent variable per- the summary description of results for the bearing
formance function, and is expressed by Equation 6, capacity metamodels, included the relative errors
considered either a target value of 0.999 for the obtained in full analysis by the first order reliabil-
sparse polynomial chaos expansions: ity method and the Monte Carlo simulation. In
addition, Figures 4 and 5 further present the model
∑ i =1(M i )
1 2
M deviation of the main statistical parameters of the
2
R = 1 (6) bearing capacity metamodels.
∑ i =1(Mi )
2
M It is noted that the convergence is verified con-
sidering the quality of the regressions, the stability
where R2 = coefficient of determination; l = number of the solutions for different design of experiments
of observed values of the performance function; and further for different orders. The trun-
Mi = ith observed value of the performance cated full polynomial chaos expansion presents

149

ISGSR2013.indb 149 10/18/2013 9:37:58 AM

 statistically representative of the performance
function. Compared the sparse polynomial chaos
expansions versus the corresponding truncated
full, the key issues gather the stability of the
solutions, even as the number of required design
sampling points and the accuracy shown by the
derived metamodels. It is further noted that for a
selected group of terms, additional data may not
be required for the accuracy improvement, in that
the minimum number of design sampling points
may be enough to match the best solutions in the
field of reliability if considered a reduced model
Figure 5. Model deviation for skewness and kurtosis deviation of the main statistical parameters of the
of the bearing capacity metamodels, cases 1–9, statistics derived metamodels.
results from 106 simulation steps.

REFERENCES
relative errors considered quite acceptable, but
the sparse polynomial chaos expansions may Armani, U., Khatir, Z., Khan, A., Toropov, V.V.,
present improved solutions. So, the fourth order Polynkin, A., Thompson, H., Kapur, N. & Noakes C.J.
presents higher accuracy for the first order reli- 2011. Control of physical consistency in metamodel
ability method, and the fifth order may improve building by genetic programming. In: Proc. of the
in general the results from Monte Carlo simula- Second International Conference on Soft Computing
tion. The statistical description of the output for Technology in Civil, Structural and Environmental
the bearing capacity metamodels is quite accurate Engineering, Chania, Crete, Greece: Civil-Comp Pro-
considered the reference model, with exception to ceedings 97, Paper 43, 1–18. Civil-Comp Press.
the case studies for a number of 150 design sam- Blatman, G. & Sudret, B. 2010a. An adaptive algorithm
to build up sparse polynomial chaos expansions for
pling points, where some differences in kurtosis stochastic finite element analysis. Probabilistic Engi-
are more important. neering Mechanics 25(2): 183–197.
Blatman, G. & Sudret, B. 2010b. Efficient computation
5 CONCLUSION of global sensitivity indices using sparse polynomial
chaos expansions. Reliability Engineering and System
Safety 95(11): 1216–1229.
In the field of abstraction, metamodelling is the Can, B. & Heavey, C. 2011. Comparison of experimental
construction of a cluster of concepts within a cer- designs for simulation-based symbolic regression
tain domain, highlighting properties of the model of manufacturing systems. Computers & Industrial
itself. Due to the computational complexity of Engineering 61(3): 447–462.
the analysis of current systems, metamodelling Haldar, A. & Mahadevan, S. 2000. Reliability assessment
has become an indispensable tool for exploring using stochastic finite element analysis. John Wiley &
the design space and performing sensitivity analy- Sons.
sis, particularly useful for the implementation of Hofwing, M., Strömberg N. & Tapankov M. 2011.
Optimal polynomial regression models by using a
numerical-based reliability techniques, to be used genetic algorithm. In: Proc. of the Second International
for design purposes and risk analysis. Therefore, Conference on Soft Computing Technology in Civil,
there is great interest in the development of tech- Structural and Environmental Engineering, Chania,
niques that facilitate the construction of approxi- Crete, Greece: Civil-Comp Proceedings 97, Paper 39,
mation models by using criteria of efficiency and 1–20. Civil-Comp Press.
accuracy. Huang, S.P., Liang B. & Phoon, K.K. 2009. Geotechnical
Regarding these issues, results for a design probabilistic analysis by collocation-based stochastic
example with correlated nonnormal random varia- response surface method: an Excel add-in implemen-
bles are presented. For the purpose it is considered tation. Georisk: Assessment and Management of Risk
for Engineered Systems and Geohazards 3(2): 75–86.
an inclined eccentric loading problem referred to Kiureghian, A.D. & Liu, P.L. 1986. Structural reliability
a bearing capacity nonlinear calculation model for under incomplete probability information. Journal of
a concrete gravity retaining structure, moreover a Engineering Mechanics 112(1): 85–104.
geotechnical reliability-based analysis problem. Li, D., Chen, Y., Lu, W. & Zhou, C. 2011. Stochastic
Considered the important nonlinear behaviour response surface method for reliability analysis of
of the geotechnical bearing capacity model, results rock slopes involving correlated non-normal variables.
show that the bearing capacity metamodels built Computers and Geotechnics 38(1): 58–68.
by multi-dimensional Hermite polynomial chaos
expansion may provide a general approximation

150

ISGSR2013.indb 150 10/18/2013 9:37:59 AM

 4 Influence of spatial variability on geotechnical design decisions

ISGSR2013.indb 151 10/18/2013 9:38:00 AM

This page intentionally left blank
 Geotechnical Safety and Risk IV – Zhang et al. (eds)
© 2014 Taylor & Francis Group, London, ISBN 978-1-138-00163-3

Mobilized shear strength of soils with constrained slip curves

Y. Hu & J. Ching
Department of Civil Engineering, National Taiwan University, Taipei, Taiwan

ABSTRACT: A previous study by Ching et al. (2013) showed that the mobilized shear strength of
a spatially variable soil is governed by two factors: (a) line averaging effect along potential slip curves;
and (b) the emergent feature of the critical slip curve, i.e., the location of the critical slip curve depends
on the random field realization. In the study done by Ching et al. (2013), the latter factor plays a major
role as the slip curves are allowed to propagate freely without constraints. For scenarios where slip
curves are constrained, e.g., the slip curve in a vertical retaining wall typically passes through the toe,
it is not clear whether this latter factor will be fully suppressed. It is therefore the goal of this study to
understand the behavior of the mobilized shear strength for spatially variable soils when slip curves are
constrained.

1 INTRODUCTION 2 SIMULATION OF ACTIVE LATERAL

FORCES
This study is complementary to a previous study
performed by the second author (Ching et al. The shear strength at a point in soil mass A is
2013). In this study, the mobilized shear strength denoted by τf(x,z), where x and z are the horizontal
of a spatially variable soil in compression and pure and vertical coordinates, respectively. The friction
shear tests was simulated. Although the inclina- angle φ is taken to be 0° in this study for simplic-
tion angle of the critical slip curve is dominated by ity, i.e., τf(x,z) = c(x,z). The shear strength τf(x,z) is
mechanics, it vertical position is not constrained simulated as a stationary Gaussian random field
and varies significantly, depending on the weak with inherent mean E(τf) = μ and inherent standard
zone in the simulated random field. As a result, deviation [Var(τf)]0.5 = σ. The Coefficient Of Varia-
the spatial variability strongly influences the mobi- tion (COV) of this random field is equal to σ/μ. To
lized shear strength; the critical slip curve may define the correlation structure of τf(x,z) between
seek the weak zone in the random field. However, two locations with horizontal distance = Δx and
in the present study, the slip curve is constrained: vertical distance = Δz, an auto-correlation model
for a retaining wall problem, the slip curve must is considered in this study: the single exponential
pass through the toe (see Fig. 1). Therefore, it is model (SExp) (Vanmaqrcke 1977, 1984). And the
expected that the spatial variability should have two-dimensional (2D) stationary Gaussian ran-
a weaker influence on the simulated active lateral dom field τf(x,z) can be readily simulated by the
force (Pa). An important goal of this study is to Fourier series method (Jha & Ching 2013).
understand how to quantify this weaker influence
if it still exists. 2.1 Limit equilibrium method
The process of simulating a random sample of
Pa using the Limit Equilibrium Method (LEM) is
illustrated in Figure 2. Figure 2a shows a realiza-
tion of the τf(x,z) random field, where sections A,
B, and C represent the locations of three poten-
tial slip planes (xA, xB, and xC in Figure 2a char-
acterize the horizontal daylight positions of the
three potential slip planes). It is assumed that the
potential slip curves are planes (straight lines).
This assumption is in principle non-conservative
Figure 1. Finite element model for a retaining wall with because the actual slip curve may not be a plane.
spatially variable soil. However, in a later comparison with Random

153

ISGSR2013.indb 153 10/18/2013 9:38:00 AM

 where β and Lp are the inclination angle and length
of the potential slip, respectively; W is the weight
of the wedge; N is the normal force between the
wedge and potential slip plane; H is the wall height;
γ is the soil unit weight; and x is the daylight posi-
tion of the potential slip plane (x is also the length
of the upper side of the wedge, i.e. x = H/tanβ).
Occasionally, the F is negative, which indicates
that the wedge can stand on its own without the
support from the wall. Thus, F will be zero in this
case. Each potential slip plane corresponds to a
value of F. This F is plotted as a function of x in
Figure 2d. The Pa is then the maximum value in the
F(x) function:

⎛1 x2 H2 ⎞
Figure 2. Process of simulating a random sample of Pa Pa ax F ( x ) ma γ H 2 − τ LLA
f ⋅ , 0⎟
using the LEM.
x ⎝2 x ⎠
(2)

Finite Element Analysis (RFEA) results, it will The x position at which F(x) is maximized is
be clear that the actual effect of this assumption denoted by x*, which is the daylight position of the
is minimal. It is also assumed that the potential critical slip plane.
slip planes pass through the toe. This is barely an Active lateral forces also can be simulated by
assumption because the actual slip curves obtained RFEA. The first step of the RFEA is a geostatic
in RFEA always pass through the toe as well. For step that builds up the in situ stress field, and in the
clarity, only three of these planes are shown. second step a horizontal displacement is applied to
The sectioned τf profiles along the three potential the wall (away from the soil mass) to reach an active
slip planes are plotted in Figure 2b. The average of failure state. Then, the lateral force is calculated as
each profile (the dashed lines in Fig. 2b) is equal to the integration of the soil-wall contact forces in the
the line average for each section. Let us denote the FEA. Figure 3 shows a comparison between the
line average of τf(x,z) along a potential slip plane Pa samples simulated by RFEA and the LEM for
as τfLA. This line average can be directly simulated 100 cases. These 100 cases use wall heights in the
by the Fourier series method developed in Jha & range of [5 m, 20 m]. The random field realizations
Ching (2013) (Eq. 19 in Jha & Ching (2013)). for these 100 cases have fairly wide ranges of μ, σ,
Let us denote the line averages over the three δx, and δz (δx and δz are the horizontal and vertical
potential slip planes shown in Figure 2a by τfLA(xA), Scale Of Fluctuation (SOF), respectively). Each
τfLA(xB), and τfLA(xC). There are an infinite number case is analyzed by both RFEA and the LEM to
of potential slip planes. Each potential slip plane has determine the Pa values under the same random
a line average τfLA, and the resulting continuous line field realizations. It is clear that the Pa values simu-
average process τfLA(x) forms a One-Dimensional lated by both methods generally follow the 1:1 line.
(1D) random field. This 1D random field is illus- As mentioned earlier, although the assumption in
trated in Figure 2c. For each potential slip plane with
daylight position x, force equilibrium can be used to
derive the lateral force F required to support the fail-
ing soil mass. The force equilibrium of the wedge
formed by the potential slip plane can be written as

W N ⋅ cos( β ) + τ LLA
f ⋅ L p ⋅ sin( β )
F τ LA
f Lp ⋅ cos( β ) = N ⋅ sin( β )
1 x
H x ⋅γ = N ⋅ + τ LLA
f ⋅H
2 Lp
H
F τ LA
f x=N⋅
Lp
2
1 Lp 1 x2 + H 2
F γ H 2 − τ LLA
f ⋅ = γ H 2 − τ LLA
f ⋅ (1) Figure 3. Comparison between the Pa values simulated
2 x 2 x by RFEA and the LEM.

154

ISGSR2013.indb 154 10/18/2013 9:38:01 AM

 LEM that the slip curve is a plane is in principle the actual critical slip plane is a prescribed 45° line
unconservative, its effect is minimal, judging from passing through the toe is generally incorrect. The
the fact that RFEA and LEM produce consistent lower plots in Figure 4 show the histograms of the
Pa estimates (note: RFEA does not assume a slip daylight positions (x*) of the critical slip planes
“plane”). This is probably because the actual slip (N = 1,000). It is clear that x* exhibits a certain
curves obtained in RFEA rarely significantly devi- variability. Thus, the critical slip plane is not a pre-
ate from planes. scribed line but varies around the 45° line. When
the COV is large, x* has a large variability, possibly
because there is more opportunity for the critical
3 OBSERVED BEHAVIORS slip plane to find a weak path that deviates from
IN THE SIMULATED PA VALUES the 45°. If a “weak path” is found, the τfLA(x*) has a
mean value less than μ. Then, Pa histogram signifi-
3.1 Effect of the COV cantly shifts to the right of Pa,d.
The statistics of the LEM Pa samples depend on the
COV of the τf(x,z) random field. Consider the ref- 3.2 Effect of SOF
erence case (γ = 20 kN/m3, H = 5 m, μ = 20 kN/m2)
with δx = δz = 1 m. The upper plots in Figure 4 dem- The COV is not the only factor that affects
onstrate the Pa histograms (sample size N = 1,000). the statistics of Pa. Consider the reference case
The solid vertical lines show the deterministic solu- with COV = 0.3 and δx = δz ∈ (0.2 m, 1 m, 10 m,
tion, denoted by Pa,d. It is clear that the Pa sam- 1,000 m). The upper plots in Figure 5 present the
ples are more variable for larger COV’s. But it is LEM Pa histograms. The histograms with white
interesting to note that the Pa samples shift to the faces show the negative Pa samples. These samples
right of Pa,d when the COV is large. As mentioned should be zero because a negative F implies that
earlier, the slip plane is highly constrained because the wedge can stand on its own. It is clear that
it must pass through the toe—it cannot evolve the Pa histograms for δx = δz ∈ (0.2 m, 1 m) are
freely. It is tempting to hypothesize that the actual significantly shifted to the right of Pa,d, whereas
critical slip plane is a prescribed 45° line passing the Pa histograms for δx = δz ∈ (10 m, 1,000 m)
through the toe, in which case Pa is simply a ran- are not. This observation is consistent with those
dom variable that depends on the average shear made in Fenton et al. (2005), where it was found
strength along the prescribed 45° slip plane. The that there is a critical SOF at which the Pa behav-
mean value of τfLA is the same as μ because line ior becomes more complicated. The lower plots
averaging along any prescribed line will not alter in Figure 6 show the x* histograms: whenever
the mean. The consequence is that the mean value the x* variability is large, the Pa samples signifi-
of Pa should be equal to Pa,d. However, this result is cantly shift to the right of Pa,d, and vice versa.
contradictory to the fact that the Pa samples shift This example shows that the variability of x*
to the right of Pa,d. As a result, the hypothesis that depends on not only the COV but also the SOF.
The τfLA(x*) will later be denoted by the “mobi-
lized” shear strength (τfm) because it is the shear

Figure 5. (Upper plots) LEM Pa histograms; (lower

Figure 4. (Upper plots) LEM Pa histograms; (lower plots) histograms of the daylight positions (x*) of the
plots) histograms of the daylight positions (x*) of the critical slip planes [reference case with COV = 0.3 and
critical slip planes (reference case with δx = δz = 1 m). δx = δz ∈ (0.2 m, 1 m, 10 m, 1,000 m)].

155

ISGSR2013.indb 155 10/18/2013 9:38:02 AM

 Figure 7. Relationship between σx*/H and Δ/μ.
Figure 6. Variation of σx*/H and Δ/μ with respect to
δ/H (reference case with δx = δz). Figure 6a Variation of
σx*/H Figure 6b Variation of Δ/μ.

strength that is actually experienced by the

retaining wall.

3.3 More insight into the variability of x*

It is interest to gain insight into the variability
in x*. Consider 15 reference cases with COV ∈
(0.1, 0.2, 0.3) and δx = δz ∈ (0.2 m, 1 m, 10 m,
100 m, 1,000 m). For each case, 1,000 sets of
LEM Pa samples and x* samples are simulated.
The variability of these x* samples is quantified
by its sample standard deviation, denoted by σx*. Figure 8. Relationships between Δ/σ and (δx/H, δz/H).
Figure 6a shows how σx*/H varies with the COV
and δ/H (δ = δx = δz). It is reasonable to observe
that σx* increases with increasing COV. It is 3.3.2 Quantify σx* through Δ
also reasonable to see that σx* is small for large It is expected that σx* is positively correlated with Δ.
SOFs. Note that σx* and Δ have different units: the former
is in meters, and the latter is in kN/m2. Therefore,
it is more sensible to compare σx*/H with Δ/μ.
3.3.1 Connection between σx* and the local Figure 7 shows that σx*/H and Δ/μ are positively
roughness of τfLA(x) correlated for 100 cases in a wide range of values
Recall that σx* is large whenever the critical slip for H, μ, COV, δx, and δz. The result suggests that it
plane is able to find the weak zone. If τfLA(x) is is possible to quantify σx*/H through Δ/μ.
nearly a constant function, then there is no weak Δ/σ can be determined given the information
zone, and x* will be nearly deterministic (≈H) on (δx/H, δz/H) from Figure 8. It is interest-
and σx* ≈ 0. If τfLA(x) is a rough function, there ing to note that the gray line in Figure 8 has a
are significant weak zones, and x* will be more similar shape to the three σx*/H versus δ/H curves
random, whereas σx* will be large. It is expected in Figure 6a. Figure 6b re-plots the (scaled)
that σx* will be correlated with the roughness of Δ/μ versus δ/H relationships (δ = δx = δz) for
τfLA(x). Moreover, the local roughness of τfLA(x) comparison.
near the point x = H is more relevant because
most x* samples cluster around H. Let us define
xL = H × tan(40°) and xR = H × tan(50°), and [xL, 4 NUMBER OF INDEPENDENT
xR] represents the local daylight region of inter- POTENTIAL SLIP PLANES
est. Let us define Δ to be the standard deviation
of [τfLA(xR) − τfLA(xL)]. It is clear that when τfLA(x) Although the critical slip plane is constrained to
is a constant function, Δ will be close to zero. pass through the toe, it is still able to seek the weak
When τfLA(x) is a rough function, Δ will be large. zone and vary locally around the prescribed 45° line.
Therefore, Δ quantifies the local roughness of And the σx*/H can be quantified by Δ/μ. One can
τfLA(x) near x = H. imagine that there are a number of “independent”

156

ISGSR2013.indb 156 10/18/2013 9:38:03 AM

 potential slip planes, and the actual critical slip
plane will be among these independent potential
planes. A large Δ/μ (σx*/H) implies that there may
be many independent potential slip planes from
which to seek. The number of independent poten-
tial slip planes quantifies the tendency of seeking
the weak zone. This number also depends on the
normalized correlation length of τfLA(x), denoted
by θ/H. If θ/H is large, the number of independent
potential slip planes should be small. The notion
of “correlation length” is herein adopted in regard
to τfLA(x) because the SOF of τfLA(x) is infinity.
The correlation length is defined to be the inter-
val length at which the correlation coefficient is
exactly exp(-1) (Li & Der Kiureghian 1993). If the Figure 10. Illustration of the concept of n.
correlation coefficient between τfLA(xL) and τfLA(xR)
is exactly equal to exp(-1), the correlation length is
simply the interval length |xL − xR|. xL and xR should If n ≈ 0, the daylight range covers nearly zero
be centered at x = H because x* occurs around this correlation length, implying that τfLA(x) is nearly
location. In this study, we adopt xL = H/r and constant within the daylight range. Thus, there
xR = H × r (r > 1). The ratio r is first solved, and should be a single independent potential slip
θ is simply |H/r − H × r|. Figure 9 shows a unique plane. As a result, the critical slip plane cannot
mathematical relationship between θ/H and (δx/H, find a weaker zone, and the mean value of the
δz/H). resulting τfm = τfLA(x*) will be equal to μ; thus, the
resulting Pa samples will not shift to the right of
Pa,d. If n is large, the daylight range covers many
4.1 Quantify the number of independent
correlation lengths, implying that τfLA(x) oscillates
potential slip planes by n = (Δ/μ)/(θ/H)
drastically within the daylight range. Thus, there
Conceptually, n = (Δ/μ)/(θ/H) quantifies the should be many independent potential slip planes
number of independent potential slip planes. It is from which to choose. As a result, the critical slip
found that the n being ‘roughly proportional’ to plane is more likely to find a weaker zone, and the
the number of independent potential slip planes. mean value of the resulting τfm will be less than μ;
This concept is illustrated in Figure 10: Δ/μ quanti- the resulting Pa samples will shift to the right of
fies the (normalized) daylight range for the criti- Pa,d. This result can be observed in Figures 4 and
cal slip plane, and θ/H quantifies the (normalized) 5, where the n values are tabulated in the Pa his-
correlation length in τfLA(x). Thus, n = (Δ/μ)/(θ/H) tograms for all cases. It is clear that whenever n is
quantifies the number of (normalized) correlation large, Pa significantly shifts to the right of Pa,d.
lengths of τfLA(x) covered by the daylight range of
the critical slip plane.
4.2 Observed number of independent potential
slip planes
The previous section discusses how to conceptually
quantify the number of independent potential slip
planes by n = (Δ/μ)/(θ/H). Nonetheless, it is pos-
sible to estimate the actual number of independent
potential slip planes based on the LEM results. The
factor n will be referred as the “conceptual number
of independent potential slip planes”. Addition-
ally, the number of planes estimated from the LEM
results will be referred as the “observed number of
independent potential slip planes”, denoted by no.
Corresponding to each LEM Pa sample, there is
a simulated τfm (= minxτfLA(x)). Each τfLA is normally
distributed with a mean value of μ and a variance
of σ2 × Γ2, where Γ2 is the variance reduction factor.
The Probability Density Function (PDF) of τfm can
Figure 9. Relationships between θ/H and (δx/H, δz/H). be derived by differentiating F(τ) with respect to τ.

157

ISGSR2013.indb 157 10/18/2013 9:38:03 AM

 Given the LEM simulation results of for a sample likelihood method. The factor n for each case is
size N (τf,1m, τf,2m, …, τf,Nm). The estimation of (no, also calculated based on its COV, δx, and δz. It is
α) can be achieved through maximizing the follow- found that there is an approximately unique rela-
ing likelihood function L(no,α): tionship between no and n. And there is also a rough
relationship between α and n. It is evident that the
⎛ ⎞ following equations provide a satisfactory fit:
no
⎜ ⎟
⎜ ( )
⎜ α × σ 2 × Γ2 n
p
⎟
⎟
no 1 + 12 × n α ≈ 1 + n (0.2 + 4 × n ) (4)
no −1
⎜ ⎡ ⎛ ⎞⎤ ⎟
N ⎜ τ mf i μ ⎟
L ( no ,α ) = ∏ ⎜ × ⎢⎢1 − Φ ⎜ ⎟⎥ ⎟
⎟⎥ ⎟
i =1 ⎜
⎜ ⎣
⎢
⎜ α × σ 2 × Γ2 n
⎝ ( )
p ⎠ ⎥⎦
⎟
5 SIMULATING Pa SAMPLES USING
A SIMPLIFIED PROCEDURE
⎜ ⎛ τ mf i − μ
⎞ ⎟
⎜ ×ϕ ⎜ ⎟ ⎟ The following simplified procedure can be used to
⎜ ⎜ α × σ 2 × Γ2
⎝ ⎝ ( ) ⎟
⎠
⎟
⎠
simulate Pa samples (for the φ = 0° condition) with-
out actually employing the LEM or RFEA. This
(3) simplified procedure requires knowledge of H, γ,
μ, COV, δx, and δz.
where ϕ is the PDF of the standard normal random
variable. The parameter np characterizes the aver- 1. Determine Δ/σ from Figure 8 based on H,
aging effect along the potential slip plane, no char- δx, δz.
acterizes the number of independent potential slip 2. Determine θ/H from Figure 9 based on H,
planes, and α compensates for the continuous reality. δx, δz.
Figure 11 shows the (no, α) values for 15 reference 3. Calculate n = (Δ/μ)/(θ/H) = (COV × Δ/σ)/(θ/H),
cases with COV ∈ (0.1, 0.2, 0.3) and δx = δz ∈ (0.2 m, and estimate (no, α) using Eq. 4.
1 m, 10 m, 100 m, 1,000 m). For each case, N = 1,000 4. Simulate a random sample U from a uniform
samples of τfm are simulated and used to estimate distribution with the range [0, 1].
(no, α) using the maximum likelihood method. 5. A sample of τfm can be obtained as F−1(U),
Figure 12 shows the no versus n plot for 200 cases where F−1 is the inverse of the CDF F:
with a wide range of H, μ, COV, δx, and δz. For
each case, N = 1,000 samples of τfm are simulated ⎡ ⎛ 1 ⎞⎤

and used to estimate (no, α) using the maximum ( )

τ mf = μ + α × σ 2 × Γ 2 n p 1
⎢1 ( U )⎝⎜ nO ⎠⎟ ⎥
⎣ ⎦
(5)

6. A sample of Pa can be obtained as

⎛ 1 ⎞
Pa H 2 − 2τ mf × H , 0 ⎟ (6)
⎝ 2 ⎠

Repeat the above steps N times to obtain N

samples.
The performance of the above procedure is illus-
Figure 11. Variations of (no, α) with respect to δ trated for a case (Fig. 13). The top row shows the
(reference case with δx = δz). histograms of the LEM Pa samples (N = 1,000).
The middle row shows the histograms of the Pa
samples (N = 10,000) simulated by the simplified
procedure.
The bottom row shows the histograms of
the Pa samples (N = 10,000) simulated based on the
hypothesis that the actual critical slip plane is the
prescribed 45° line passing through the toe (x* is
fixed at H). In general, the Pa samples simulated by
the hypothesis x* = H behave very differently from
LEM when the SOF is small. When δx = δz = 0.2 m,
Figure 12. no-n and α-n relationships for the SExp the LEM samples are clearly not centered at Pa,d but
model. shift to a positive value of approximately 45 kN/m.

158

ISGSR2013.indb 158 10/18/2013 9:38:04 AM

 6 CONCLUDING REMARKS

In this study, the mobilized shear strength of soils

with constrained slip curves is explored. It is found
that although the actual critical slip curve is highly
constrained (must pass through the toe), it is still
able to seek weak zones locally when the SOFs are
smaller than the height of the wall.
The main contribution of this study is to dis-
cover that the tendency of seeking weak zones can
be characterized by the number of independent
potential slip planes. Moreover, this study derives
a set of equations and charts to facilitate the deter-
mination of such a number based on the charac-
teristics of the random field and wall height. The
determination does not require sophisticated anal-
yses, such as the finite element method; instead,
Figure 13. Histograms of the simulated Pa samples for it requires only simple algebraic calculations and
the case in which H = 10 m, COV = 0.3, μ = 50 kN/m2,
δx = δz ∈ (0.2 m, 1 m, 10 m, 100 m, 1,000 m), and the
chart checking. Finally, a simplified procedure is
SExp model is used. (Top row) LEM samples; (middle proposed to simulate the Pa samples without the
row) samples from the simplified procedure; (bottom finite element method or LEM, and the simulated
row) samples based on the hypothesis that x* = H. Pa samples are observed to behave similarly to the
Pa samples simulated by the LEM.

The shift is also quite obvious for δx = δz = 1 m.

However, the samples in the bottom row are always REFERENCES
centered at Pa,d; a major discrepancy exists between
the LEM samples and those based on the hypothe- Ching, J. Phoon, K.K. & Kao, P.H. 2013. Mean and vari-
sis that x* = H when the SOF is small. But for larger ance for mobilized shear strengths of spatially variable
SOFs [δx = δz ∈ (10 m, 100 m, 1,000 m)], there is a soils under uniform stress states. Journal of Engineer-
ing Mechanics, ASCE (accepted w/minor revision).
less noticeable shift. The above observation implies Fenton, G.A., Griffiths, D.V. & Williams, M.B. 2005.
that the hypothesis x* = H may be acceptable when Reliability of traditional retaining wall design.
the SOF is large but is not acceptable when the Geotechnique 55(1): 55–62.
SOF is small. When the SOF is smaller than H, Jha, S.K. & Ching, J. 2013. Simulating spatial averages of
weak zones are more pronounced, so x* starts to stationary random field using Fourier series method.
deviate from H and σx* increases. As a result of Journal of Engineering Mechanics, ASCE 139(5):
these changes, n (no) increases and τfm should be 594–605.
less than μ; thus, the samples shift to the right Li, C. & Der Kiureghian, A. 1993. Optimal discretization
of Pa,d. Nonetheless, the Pa histograms simulated of random fields. Journal of Engineering Mechanics,
ASCE 119(6): 1136–1154.
by the simplified procedure (middle row) behave Vanmarcke, E.H. 1977. Probabilistic modeling of soil
similarly to the LEM Pa histograms, even when the profiles. Journal of Geotechnical Engineering Division,
SOF is small. ASCE 103(11): 1227–1246.
Vanmarcke, E.H. 1984. Random Fields: Analysis and
Synthesis, MIT Press, Cambridge, Mass.

159

ISGSR2013.indb 159 10/18/2013 9:38:06 AM

This page intentionally left blank
 Geotechnical Safety and Risk IV – Zhang et al. (eds)
© 2014 Taylor & Francis Group, London, ISBN 978-1-138-00163-3

Reliability of heterogeneous slopes with cross-correlated shear

strength parameters

T.M.H. Le
Department of Civil and Transport Engineering, Norwegian University of Science and Technology—NTNU,
Norway

ABSTRACT: Spatial variabilities of the soil shear strength parameters (friction angle and cohesion
coefficient) have long been recognised as an important factor influencing the reliability of geo-structures
including slopes. However, these two parameters are frequently considered separately in research studies
even though, in natural soils, both parameters are likely to vary simultaneously with existence of cross-
correlation between them. This study stochastically investigates the reliability of a slope constructed in
soil having spatially varying both friction angle and cohesion coefficient, and compares that with the
scenarios where each soil parameter varies individually. The finite element method is merged with the ran-
dom field theory to probabilistically evaluate the factor of safety and probability of failure of the slope via
Monte-Carlo simulations. A simple procedure to create perfect cross-correlation is discussed. The results
show that the variabilities of both friction angle and cohesion coefficient even without cross-correlation
can elevate the probability of failure relative to the cases where each of them varies individually. If a per-
fectly positive cross-correlation exists, the probability of failure increases significantly due to increasing
occurrences of local failures.

1 INTRODUCTION inherent characteristic that measurement of a soil

property at a given location tends to be similar (i.e.
Probabilistic modelling has been quite frequently correlate with) those of the same property at close
employed in the geotechnical engineering dis- locations but different from those at far locations
cipline to deal with soil variability. Particularly, (e.g. Mostyn & Soo 1992, Hicks & Samy 2002,
slope stability is one of a few engineering prob- Babu & Mukesh 2004, Griffiths & Fenton 2004).
lems that is more often analysed using probabi- This characteristic is usually quantified in stochas-
listic approaches (Alonso 1976, Mostyn & Soo tic modelling by a correlation length (θ). The value
1992, Mostyn & Li 1993, Hicks & Samy 2002, of soil property at a given location also tends to
Babu & Mukesh 2004, Griffiths & Fenton 2004, correlate with the values of other soil properties at
El-Ramly et al. 2005, Hicks & Onisiphorou 2005, the same location or at close locations. This char-
Popescu et al. 2005, Sejnoha et al. 2007, Griffiths acteristic is termed “cross-correlation”.
et al. 2009, Griffiths et al. 2011). Among these Cross-correlation between various soil proper-
studies, a number have investigated the influence ties has been reported by a few authors (Cherubini
of spatial variability of soil properties on Fac- 2000, Baecher & Christian 2003). When consider-
tor of Safety (FoS) and probability of failure (pf) ing problems related to failure, the influence of
of slopes. Monte-Carlo simulations were used in heterogeneity of friction angle (φ′) and cohesion
many of these studies due to its conceptual simplic- coefficient (c′) and the cross-correlation between
ity and its ability to handle complicated geometry, them are likely to play an important role. Cherubini
loading sequence and variability patterns without (2000) suggested that the cross-correlation between
over-simplified assumptions. In a Monte-Carlo φ′ and c′ might be negative (i.e. a large value of φ′
simulation, random samples generated from the occurs with a small value of c′ at a given location or
same input statistical parameters are repeatedly vice versa). It is however intuitively expected that a
analysed. Each analysis is usually referred to as a stochastic slope with negatively correlated φ′ and c′
“realizations”. This method is especially attractive is likely to be more stable than a slope with uncor-
for complex problems without closed form solu- related or positively correlated φ′ and c′. The risk
tions such as slope stability. associated with having no correlation or ignoring
In describing spatial variability, a number an existing positive correlation between these two
of slope stability studies took into account the shear strength parameters should therefore not be

161

ISGSR2013.indb 161 10/18/2013 9:38:07 AM

 ignored. Only few studies have considered the effect sudden increases in displacements of one or more
of cross-correlation between different soil proper- of these points occur during the application of the
ties on slope stability (e.g. Popescu et al. 2005, shear strength reduction procedure. The points
Griffiths et al. 2009, Griffiths et al. 2011). Griffiths are approximately spaced between 2 to 4 m apart,
et al. (2009) briefly investigated a drained slope with which is sufficient to identify the majority of glo-
a cross-correlation between soil c′ and φ′ and con- bal and local failures in the current study.
cluded that a positive cross-correlation decreased In each realisation, spatial variability of φ′
the “critical” value of Coefficient Of Variation and/or c′ is generated by mapping each element of
(COV) (i.e. the value at which ignoring spatial cor- the finite element mesh with a random value of φ′
relation will lead to unconservative estimates). and/or c′ selected from a random field. The values
This study presents a simple procedure to gen- of tanφ′ and c′ are assumed to follow a log-normal
erate cross-correlation between soil shear strength distribution with mean values equal to 0.375 and
parameters modelled by random fields. The proce- 15 kPa, respectively, and a range of COV from 0.1
dure facilitates comparison between the effects of to 1.6. The vertical correlation length (θv) is inves-
variability of single and multiple soil properties on tigated for the range from 1 to 16 m while the hori-
both qualitative characteristics of slip surfaces and zontal correlation length (θh) is assumed to be equal
quantitative values of FoS and pf. This procedure to two times θv. This range of θv is larger than the
is then applied to probabilistically investigate the usual range of θv reported in the literature which is
stability of a slope with spatially varying friction from 0.5 to 4 m (see, for example, Spencer (2007)).
angles and cohesion coefficients with either zero or The larger range is studied to emphasise the influ-
perfectly positive cross-correlation between them. ence of correlation length (if any) on the obtained
results. The Local Average Subdivision (LAS)
algorithm is selected to generate all random fields
2 METHOD in this study because it has been suggested as being
suitable for soil properties (Fenton & Vanmarcke
2.1 Numerical model 1990; Fenton & Griffiths 2008). The Markov cor-
relation function is assumed to model the spatial
A 1:2 slope model is used for every realization in
correlation of random values.
the current study (Fig. 1). The soil behaviour is
assumed to be elastic perfectly plastic, with a Mohr-
2.2 Cross-correlation between soil properties
Coulomb yield/failure criterion. Elastic matrix is
calculated from Young’s modulus and Poisson’s A popular method for generating cross-correlated
ratio through the classical linear isotropic elastic- random fields is the covariance matrix decomposi-
ity theory. This study adopts a non-associated flow tion method. This method is capable of producing
rule with zero dillatancy. Stresses are updated by fields which have a defined degree of correlation
means of a viscoplastic algorithm. The model is (ρ) varying between −1 and 1. This study adopts
numerically analysed with the finite element pro- a simpler approach to generate cross-correlated
gram CODE_BRIGHT (Olivella et al. 1996). random fields of tanφ′ and c′, by assuming a sim-
The failure mechanism is artificially mobi- ple arithmetic transformation between these two
lised by the shear strength reduction technique parameters. This approach has an advantage of
(Zienkiewicz et al. 1975). The displacements of simplifying the conceptual understanding of the
ten points (located along the face, on the crest effect of cross-correlated soil properties, which
and at the base of the slope) are monitored to eases the explanation of failure mechanisms asso-
detect failure (Fig. 1). The FoS is estimated when ciated with the variability of each parameter.
The stability of slopes with either φ′ or c′ vary-
ing singly are compared with those having both
parameters varying simultaneously with zero
cross-correlation (ρ = 0) and with perfectly posi-
tive cross-correlation (ρ = 1). To facilitate this com-
parison, the values of μ(tanφ′) and μ(c′) are kept
constant in every analysis at 0.375 and 15 kPa,
respectively. Four closely related scenarios are
designed as follows:
a. Random tanφ′ and uniform c′
A random field of tanφ′ is generated by the LAS
with μ(tanφ′) = 0.375 and mapped onto the mesh,
Figure 1. Slope dimensions, geometry and boundary while the value of c′ is kept spatially uniform
conditions (scale in meter). at 15 kPa. Figure 2 conceptually illustrates the

162

ISGSR2013.indb 162 10/18/2013 9:38:07 AM

 Figure 2. Variation of failure envelope corresponding
to scenario (a) Random tanφ′ and uniform c′.
Figure 3. Variation of failure envelope corresponding
to scenario (b) Random c′ and uniform tanφ′.

variation of the Mohr-Coulomb failure envelope

relative to the reference envelope which is corre-
sponding to the failure state with c′ = μ(c′) and
tanφ′ = μ(tanφ′). In this scenario, the failure enve-
lopes of different soil elements vary in gradient but
always intersect the vertical axis at c′.
b. Random c′ and uniform tanφ′
The same field of tanφ′ generated in scenario (a)
is used to create a random field of c′ by multiply-
ing each random value of tanφ′ with a constant
k = μ(c′)/μ(tanφ′) = 15/0.375 = 40 kPa. By simple
arithmetics, it can be shown that this transforma-
tion leads to consistency in the variability charac-
teristics of the tanφ′ random field in (a) and the
c′ random field in (b) (i.e. COV(c′) = COV(tanφ′),
θ(c′) = θ(tanφ′) and μ(c′) = k.μ(tanφ′) = 15 kPa).
The random c′ values are mapped onto the mesh Figure 4. Variation of failure envelope corresponding
while the value of tanφ′ is kept spatially uniform at to scenario (c) Random tanφ′ and random c′ with zero
0.375 in this scenario. Therefore, the failure enve- cross-correlation.
lope corresponding to individual soil element shifts
its position up and down but is always parallel to
the reference envelope due to uniform friction assigned with a random value of tanφ′ and a ran-
angle (Fig. 3). dom value of c′. The failure envelope of each soil
c. Random tanφ′ and c′ with zero cross-correlation element changes both gradient and position ran-
(ρ = 0) domly with respect to the reference envelope in this
In this scenario, the random field of tanφ′ gener- scenario (Fig. 4).
ated in scenario (a) is first mapped onto the mesh. d. Random tanφ′ and c′ with perfect cross-correlation
A second random field of tanφ′ (having the same (ρ = 1)
statistical parameters as the first one) is generated The same random field of tanφ′ from scenario (a)
using the LAS generator. This second random field above is used to generate a random field of c′ in the
is then used to create a random field of c′ in a simi- same fashion as described in scenario (b). Both the
lar fashion to the method described in scenario (b) random field of tanφ′ from (a) and the currently
above. The tanφ′ and c′ random fields employed in generated random field of c′ are used to map ran-
this scenario are therefore independently generated dom values of φ′ and c′ onto the mesh. This leads
but they have similar variability characteristics (i.e. to perfectly positive correlation of tanφ′ and c′ at
the same COV and θ while μ(tanφ′) = 0.375 and every location. Because both parameters are var-
μ(c′) = 15 kPa). The c′ random field is also mapped ied and c′ = ktanφ′, it can be shown through trigo-
onto the mesh and hence each finite element is nometry that the failure envelope of soil element

163

ISGSR2013.indb 163 10/18/2013 9:38:07 AM

 Figure 5. Variation of failure envelope corresponding
to scenario (d) Random tanφ′ and random c′ with per-
fectly positive cross-correlation.

varies its gradient (due to the variability of tanφ′)

but always intersects the horizontal axis at the same
point k similar to the reference envelope (Fig. 5).
The use of the same random field of tanφ′ in
all four scenarios, and hence the same configura-
tion of shear strength variability, as described
above, facilitates direct comparison between these
scenarios. The reason is that, for a single realisa-
tion, it eliminates differences between scenarios
caused by the discrepancy in spatial distribution of
shear strength parameters. Also, it becomes pos-
sible to qualitatively compare the effect of vari-
abilities of tanφ′ and c′ on the characteristic of the
slip surface for each individual realization as will
be shown in the subsequent section.

3 RESULT AND DISCUSSION

3.1 Failure mechanism Figure 6. Typical displacement contours and possi-

ble global (continuous line) and local (broken line) slip
Figure 6 illustrates typical examples of ‘local’ and surfaces of the four scenarios: (a) random tanφ′ and uni-
‘global’ slip surfaces for the four scenarios intro- form c′ (b) random c′ and uniform tanφ′ (c) random tanφ′
duced in the previous section. These contour plots and c′ with zero cross-correlation (d) random tanφ′ and c′
and the slip surfaces correspond to slopes having with perfectly positive cross-correlation (μ(tanφ′) = 0.375,
the same random configuration of tanφ′ except μ(c′) = 15 kPa, COV = 0.8, θ = 0.8). Contour scales are
for scenario (c) in which a second random field different between scenarios.
is required to generate the variable configuration
of c′. Figures 2 and 3 show that, at high confin-
ing stresses, the failure shear strength (τf) corre- band and mobilise a slip surface. Between the two
sponds to scenario (a) with random tanφ′ varies first scenarios, even though the slope in scenario
over a broader range compared with scenario (a) has random tanφ′ and that in scenario (b) has
(b) with random c′. Conversely, at low confining random c′, the two scenarios have the same ran-
stresses, the value of τf corresponds to scenario dom configuration due to the assumed perfect
(b) varies over a broader range than to scenario correlation of tanφ′ and c′ used to generate the
(a). When the values of τf spread over a broad random field of c′ as described earlier. Therefore,
range, it becomes ‘easier’ to find a series of weak in deep soil regions (corresponding to regions of
soil elements (i.e. having low τf) to form a shear high confining stresses), the ‘weakest’ path is likely

164

ISGSR2013.indb 164 10/18/2013 9:38:07 AM

 to be weaker in scenario (a) than in scenario (b). parameters (i.e. COV) are compared between the
Conversely, in shallow soil regions (corresponding four scenarios and for different values of θ (Figs. 7
to regions of low confining stresses), the weak- and 8). The horizontal axes in these figures are
est path is likely to have lower shear strength in labelled with a generic term, COV(Shear strength),
scenario (b) than in scenario (a). Consequently, which is equal to COV(tanφ′) in (a), to COV(c′) in
mobilisation of deep failure mechanism is more (b) and equal to both COV(tanφ′) and COV(c′) in
commonly observed in the case of varying tanφ′ (c) and (d). Griffiths et al. (2009) concluded that
(Fig. 6a) compared to varying c′ (Fig. 6b). For the if the input COV(Shear strength) exceeds a cer-
example shown in Figure 6a, both global (i.e. deep tain critical value, ignoring spatial variability can
and large in extent) and local (shallow and limited lead to unconservative estimates of the pf of het-
in extent) slip surfaces are indeed mobilised, but erogeneous slopes. This study therefore considers a
the global slip surface corresponds to a lower FoS large range of values of the input COV (0.1–1.6),
and hence dominates the failure mechanism in this though the typical values of COV(Shear strength)
case. The corresponding slipping mass of the slope are normally less than 0.5.
having random c′ (with the same random configu- The most noticeable trends in Figure 7 are the
ration) tends to be local (Fig. 6b). This difference decreasing value of μ(FoS) and increasing value
in failure mechanisms between the two scenarios, of COV(FoS) as the degree of soil variability
where shear strength parameter is varied singly, is increases, which leads to an apparent increase in
observed for numerous realizations in the current pf with COV(Shear strength) for all cases (Fig. 8).
study. The FoS in scenario (a) is also frequently Note that the logarithm of pf and the log-scale ver-
smaller than that in scenario (b). tical axis have been adopted to facilitate the presen-
For a given realisation in scenario (c), if the tation of results in Figure 8, and hence the upward
simultaneous but not cross-correlated variabilities
of φ′ and c′ lead to a dominance of combinations
of low tanφ′ with low c′ in the weakest path, the sta-
bility of the slope will tend to be lower compared
with the two previous scenarios (e.g. Fig. 6c). On
the other hand, if these variabilities lead to a domi-
nance of combinations of low tanφ′ with high c′,
the stability of the slope is likely to be improved
and a higher FoS compared with other scenarios
is likely to be observed. The failure mechanism can
also vary widely in this scenario, from shallow local
slipping mass if the weakest path is dominated by
low c′, to deep global slipping mass if the weakest
path is dominated by low tanφ′.
Finally, when both tanφ′ and c′ vary with per-
fectly positive cross-correlation, in the region
below the reference envelope, the failure envelope
in scenario (d) is always lower than in scenarios (a)
and (b) because an element with low φ′ will also
certainly have low c′ (Fig. 5). As a result, the weak-
est path is weaker in scenario (d) than in scenarios
(a) or (b) and hence slopes with perfectly positive
cross-correlation have lower FoS than those with
varying only φ′ or only c′. Also, a local slip surface
concentrating in a small soil region is frequently
observed in scenario (d) such as in the example
shown in Figure 6d. This is because the combina-
tion of low values of both tanφ′ and c′ leads to
regions of soil with so low shear strength that the
local slip surface is mobilised much earlier than the
global one.

3.2 Factor of safety and probability of failure Figure 7. Variations of the μ(FoS) and COV(FoS)
with COV(Shear strength) for four scenarios
The variations of μ(FoS), COV(FoS) and pf (μ(tanφ′) = 0.375, μ(c′) = 15 kPa, COV(Shear
with the degree of variability of the strength strength) = COV(tanφ′) = COV(c′), θv = θ, θh = 2θ).

165

ISGSR2013.indb 165 10/18/2013 9:38:09 AM

 observed (Fig. 8). The reason is that, in the case
of positive cross-correlation, all the soil elements
with low value of tanφ′ also have corresponding
low value of c′, while an element with low tanφ′
might have low or high c′ in the absence of cross-
correlation.
The value of μ(FoS) changes insignificantly
with the value of θ, while a very consistent increase
in the value of COV(FoS) with increasing θ is
observed (Fig. 7). When both tanφ′ and c′ vary-
ing, the value of pf increases much more signifi-
cantly with COV(Shear strength) for shorter θ than
for longer ones. Therefore, slopes with shorter
θ appear to pose lower risk of failure than those
with longer θ at low values of COV(Shear strength)
but an opposite trend is observed for COV(Shear
strength) = 1.6 in scenarios (c) and (d) (Fig. 8). The
reason is that the shear strength of a number of
soil elements becomes very low when both tanφ′
and c′ are varied, together with the reduction in
the size of region of ‘correlated’ soil property (with
smaller θ) lead to the dominance of local failure
mechanism occurring within a small region of soil
over global failure mechanism (e.g. Fig. 6d). Long
Figure 8. Variation of the pf with COV(Shear strength) θ in this case causes larger and less fragmented
at different θ for four scenarios (μ(tanφ′) = 0.375, weak soil regions compared with short θ, thus low-
μ(c′) = 15 kPa, COV(Shear strength) = COV(tanφ′) = CO ering the probability of occurrence of these local
V(c′), θv = θ, θh = 2 θ). slip surfaces.

direction denotes the increasing probability of fail- 4 CONCLUSIONS

ure and vice versa. The increasing value of pf is the
result of the increasing probability of soil elements This paper has discussed some interesting effects
with very low tanφ′ and/or c′ with larger degree of cross-correlation between soil shear strength
of variability in shear strength, therefore raising parameters on the failure mechanism and the sta-
the possibility of shearing a sufficient number of bility of a stochastic slope using the finite element
weak soil elements to form a slip surface. It seems method and the random field theory. A simple pro-
that, for the slope considered, the contribution of cedure to generate perfect cross-correlated random
tanφ′ to the stability of the slope is more significant shear strengths is demonstrated, which eases the
than the contribution of c′ because the value of conceptual understanding of the influence of soil
μ(FoS) is always larger for scenario (a) with vary- parameter variability on the slope performance.
ing tanφ′ than for scenario (b) with varying c′. In The results show that, compared with slopes hav-
addition, the scenario (b) results in smaller values ing only one heterogeneous shear strength param-
of COV(FoS) than scenario (a) (Fig. 7). eter, the probability of failure increases when both
In the scenarios with both shear strength friction angle and cohesion coefficient are varied,
parameters varying (either with or without even with zero cross-correlation between them. If
cross-correlation), lower values of μ(FoS) and the two shear strength parameters are perfectly
larger values of COV(FoS) are observed compared positive cross-correlated, the probability of failure
with scenarios with only one parameter varying increases significantly relative to the case where no
due to the possibility of a low tanφ′ occurring with cross-correlation exists, due to dominant occur-
a low c′ in a soil element (Fig. 7). Consequently, rences of local failure mechanism over global fail-
slopes with heterogeneity of both parameters pose ure mechanism.
higher risk of failing than those with only one soil For the slope investigated, an increase in soil var-
parameter being varied. If tanφ′ and c′ are posi- iability decreases the mean however significantly
tively correlated, the value of μ(FoS) decreases and increases the coefficient of variation of the factor
the value of COV(FoS) increases compared slopes of safety, thus resulting in an increasing probabil-
having zero cross-correlation between these two ity of failure in all cases considered. If multiple
parameters (Fig. 7), and hence an increase in pf is shear strength parameters are varied with zero

166

ISGSR2013.indb 166 10/18/2013 9:38:10 AM

 to positive cross-correlation, the increase in the Hicks, M.A. & Onisiphorou, C. 2005 Stochastic
probability of failure with degree of soil variabil- evaluation of static liquefaction in a predominantly
ity (i.e. coefficient of variation of shear strength) is dilative sand fill. Géotechnique 55(2): 123–133.
more significant at small correlation lengths than Hicks, M.A. & Samy, K. 2002 Reliability-based charac-
teristic values: a stochastic approach to Eurocode 7.
at large ones. This can lead to decreases in prob- Ground Engineering 35: 30–34.
ability of failure with increasing correlation length Mostyn, G.R. & Li, K.S. 1993 Probabilistic slope
at sufficiently large degrees of soil variability. analysis: state-of-play. Proceedings of the conference
on probabilistic methods in geotechnical engineering.
Canberra, Australia.
REFERENCES Mostyn, G.R. & Soo, S. 1992 The effect of autocorrela-
tion on the probability of failure of slopes. Proceedings
Alonso, E.E. 1976 Risk analysis of slopes and its applica- of 6th Australia, New Zealand Conference on Geome-
tion to Canadian sensitive clays. Géotechnique 26(3): chanics: Geotechnical Risk.
453–472. Olivella, S., Gens, A., Carrera, J. & Alonso, E. 1996
Babu, G.L.S. & Mukesh, M.D. 2004 Effect of soil vari- Numerical formulation for a simulator (CODE-
ability on reliability of soil slopes. Géotechnique 54(5): BRIGHT) for the coupled analysis of saline media.
335–337. Engineering Computations 13(7): 87–112.
Baecher, G.B. & Christian, J.T. 2003 Reliability and sta- Popescu, R., Deodatis, G. & Nobahar, A. 2005 Effects
tistics in geotechnical engineering, Chichester, United of random heterogeneity of soil properties on bear-
Kingdom, Wiley. ing capacity. Probabilistic Engineering Mechanics 20:
Cherubini, C. 2000 Reliability evaluation of shallow 324–341.
foundation bearing capacity on c′, phi′ soils. Canadian Sejnoha, M., Sejnoha, J., Kalousková, M. & Zeman, J.
Geotechnical Journal 37: 264–269. 2007 Stochastic analysis of failure of earth structures.
El-Ramly, H., Morgenstern, N.R. & Cruden, D.M. 2005 Probabilistic Engineering Mechanics 22(2): 206–218.
Probabilistic assessment of stability of a cut slope in Spencer, W.A. 2007 Parallel stochastic and finite ele-
residual soil. Géotechnique 55(1): 77–84. ment modelling of clay slope stability in 3D. Faculty
Fenton, G.A. & Vanmarcke, E. 1990 Simulation of ran- of Engineering and Physical Sciences. Manchester,
dom field via Local Average Subdivision. ASCE Jour- University of Manchester.
nal Engineering Mechanics 116(8): 1733–1749. Zienkiewicz, O.C., Humpheson, C. & Lewis, R.W. 1975
Griffiths, D.V., Huang, J. & Fenton, G. A. 2011 Probabil- Associated and non-associated visco-plasticity and
istic infinite slope analysis. Computers and Geotechnics plasticity in soil mechanics. Géotechnique 25(4):
38(4): 577–584. 671–689.
Griffiths, D.V., Huang, J. S. & Fenton, G. A. 2009 Influ-
ence of Spatial Variability on Slope Reliability Using
2-D Random Fields. Journal of Geotechnical and
Geoenvironmental Engineering 135(10): 1367–1378.

167

ISGSR2013.indb 167 10/18/2013 9:38:11 AM

This page intentionally left blank
 Geotechnical Safety and Risk IV – Zhang et al. (eds)
© 2014 Taylor & Francis Group, London, ISBN 978-1-138-00163-3

Finite element analysis of compression behavior of ground

improvement with spatial variability

T. Namikawa
Shibaura Institute of Technology, Tokyo, Japan

ABSTRACT: The strength of cement-treated soil varies greatly owing to variability of the in-situ soil,
variability of mixing effectiveness, and other factors. This paper presents the numerical experiments that
investigate the effect of the spatial correlation on the behavior of cement-treated soil columns with dif-
ferent sizes. Finite element analyses were performed to simulate the unconfined compression behavior of
the columns with three different heights. The spatial distribution of the material properties was provided
by means of Monte Carlo simulation. The simulations were carried out with the samples in the presence
of different degrees of spatial autocorrelation. The numerical results provide good understanding of the
effects of the column height on the compression behavior of the cement-treated soil columns with spatial
variability.

1 INTRODUCTION behavior of full scale cement-treated columns

with spatial variability. That study revealed that
Ground improvement by deep mixing method has the spatial autocorrelation in the material prop-
been widely used for structure foundations, retain- erties affects significantly the overall compressive
ing walls and liquefaction mitigation method. In strength of the full scale column. However, since
this method, it is well known that the strength the column size was held constant in that study,
of cement-treated soil varies greatly, even at a the effect of spatial correlation on the behavior of
single column, owing to variability of the in-situ cement-treated columns with different sizes has
soil, variability of mixing effectiveness, and other not been studied.
factors. Therefore, to design adequately the ground This paper presents the numerical experiments
improvement, it is necessary to evaluate appropri- that investigate the effect of the spatial correlation
ately the influence of the variability of cement- on the behavior of cement-treated soil columns
treated soil strength on the behavior of ground with different sizes. FE-analysis with MCS was
improvement. performed to simulate the unconfined compres-
The most popular statistical parameter express- sion behavior of the columns with three different
ing variability is the standard deviation. In practi- heights. The spatial distribution of the material
cal design procedures, the mean and the standard properties was provided by means of MCS. The
deviation of the unconfined compressive strength simulations were carried out with the samples
of cement-treated soils have been used for deter- in the presence of different degrees of spatial
mining the design strength and assessing the qual- autocorrelation. The numerical results provide
ity of ground improvement (e.g., Matsuo 2002, good understanding of the effects of the column
CDIT 2002). In these design procedure, the design height on the compression behavior of the cement-
strength is reduced with the increase in the standard treated soil columns with spatial variability.
deviation of the strength of cement-treated soils.
However, the standard deviation is not sufficient
in characterizing the variability of the strength 2 GENERATING STOCHASTIC FIELD
because the spatial variation of the strength is
generally correlated with each other. Therefore, it The spatial distribution of the strength in the sam-
is necessary to investigate the influence of the spa- ple is provided by means of Monte Carlo simula-
tial correlation of the strength on the behavior of tion (MCS). In the simulations, the probability
the entire ground improvement by cement-mixing. characterization of cement-treated soil strength
Namikawa & Koseki (2013) conducted the Finite is required for generating of random fields. The
Element analysis (FE-analysis) with Monte Carlo type of probability distribution of the cement-
Simulations (MCS) to investigate the compression treated soil strength is required in the generating

169

ISGSR2013.indb 169 10/18/2013 9:38:11 AM

 random field. In this study, the variability of the spatial variability was discussed in the previous
unconfined compressive strength of the cement- paper by the author (Namikawa & Koseki 2013).
treated soil qu is assumed to be characterized by
the normal distribution. Moreover the stationary
random field with an exponential type autocorrela- 3 FINITE ELEMENT ANALYSIS
tion function is assumed for the spatial variability
of qu in a cement-treated column. The exponential 3.1 Numerical condition
type autocorrelation function is defined as: Three-dimensional finite element analyses were
performed using the finite element program devel-
⎛ d ⎞ oped by Shiomi et al. (1993). The unconfined com-
ρ (d ) = exp ⎜ − ⎟ (1) pression behavior of different size cement-treated
⎝ θ qu ⎠ soil columns is examined in this study. The finite
element meshes for the three different size columns
where ρ(d) is the correlation coefficient of qu, d is are shown in Figure 1. The samples are modeled as
distance between two points considered and θqu is a isolated column of 1 m diameter. The height of
the autocorrelation distance of qu, In this study, sample H is set to be 2, 3 and 4 m.
the spatial autocorrelation is assumed to be iso- A mesh consisting of eight-node isoparametric
topic for simplicity. elements was used. The majority of the elements
The stochastic field in the presence of spa- are cubic with a length of 100 mm. The element
tial autocorrelation was generated by covariance size corresponds approximately to the standard
matrix decomposition (Griffiths & Fenton 2007) size of cored samples. Therefore the local average
in this study. Covariance matrix decomposition is process (Griffiths & Fenton 2007) has not been
a direct method of producing a random field with carried out in the analyses. The boundary condi-
prescribed covariance structure. In this method, a tion for a sample of H = 2 m is shown in Figure 2.
random field vector X(r) with the mean value μ is The boundary conditions at the top and bottom
derived as: surfaces are smooth. The loading process consists
of applying uniform vertical displacement at the
() μ+ ⋅ u (r) (2) upper surface of the sample.
The variable unconfined compressive strength
value of each element qui is given by:
where L is the lower triangular matrix of covari-
ance matrix Σ, μ is a vector consisting of μ and
u(r) is a white noise vector with a standard normal qui μqu ( + xV
xiV ) qu xi qu (i Ne ) (4)
distribution. Σ consists of ρ(d) calculated from d by
Eq. (1). L satisfies where μqu is the mean of the strength, Vc is the coef-
ficient of variation of qu, xi is the random variable
LT = Σ .
LL (3) and Ne is the number of elements. The value of xi
generated by the method mentioned in the previous
section is assigned to each element in the sample in
L is obtained using Cholesky decomposition.
the FE-analysis. The value of mean and standard
Using the Box-Muller method (Fushimi 1992),
deviation of xi are 0 and 1 respectively. Here xi is
the normal random numbers (white noise with
classified into the discrete classes whose interval is
the standard normal distribution) were calculated
from uniform random numbers generated by the
M-sequences method (Fushimi 1992). The ran-
dom fields with μ = 0 were generated by adopting
the aforementioned method. 50 realizations of the
Monte Carlo process were performed for several
autocorrelation distances (θqu = 0.2, 0.5, 1.0 and
2.0 m). The number of realizations is not large in
this study. However, since the three-dimensional
nonlinear analysis is very time consuming, the
number of realizations selected was considered
sufficient to investigate the effects of the column
height on the compression behavior of cement-
treated soil columns with spatial variability. The
influence of the number of realizations on the
calculated probability parameters of the over-
all strength of cement-treated soil columns with Figure 1. Mesh for three sizes of column.

170

ISGSR2013.indb 170 10/18/2013 9:38:11 AM

 field and a large value of θqu provides a smoothly
varying field.
An elasto-plastic model developed for cement-
treated soils (Namikawa & Mihira 2007) was used
in these analyses. This model can describe appro-
priately tensile and shear strain-softening behaviors
in the strain localization zone after the peak stress
state. In this model, the smeared crack band con-
cept (Pietruszczak & Mróz 1981) is used to avoid
mesh size dependency due to the strain localization.
Thus the FE-analysis with the elasto-plastic model
can simulate reasonably the post-peak behavior of
cement-treated soils under various boundary con-
ditions. The performance of the used model has
been verified by the simulations for several labo-
ratory tests of cement-treated soils (Namikawa &
Mihira 2007, Namikawa & Koseki 2007) and an
Figure 2. Boundary condition for a sample of H = 2 m. in-situ full scale test of a cement-treated soil col-
umn (Namikawa et al. 2008).
A list of elasto-plastic model parameters used in
the numerical modeling and their values are shown
in Table 1. These parameters for μqu = 1.7 MPa were
determined from the results of various laboratory
tests in which cement-treated sand specimens with
the unconfined compressive strength of 1.7 MPa
were used (Namikawa & Koseki 2006, Namikawa &
Mihira 2007). As indicated in Table 1, the elastic
modulus E, the cohesion c, the tensile strength
Tf, and the fracture energy Gf are regarded as the
stochastic parameters. With the constant friction
angle φ, the value of cohesion c is determined
from the value of qui. The values of E, Tf and Gf
are assumed to be proportional to the value of qui.
Other parameters, Poisson’s ratio ν, the hardening
parameters α and ey, the softening parameter er,
the dilatancy coefficient Dc, the localization size ts0,
Figure 3. Typical samples for FE-analysis (H = 2 m). and the characteristics length lc, are assumed not
The darker regions indicate stronger element. to vary with qui. The detailed descriptions of the
material parameters and the numerical examples
that can illustrate the applicability of this model
set to be 0.1. The classified xi ranges from −3.1 to are available elsewhere (Namikawa & Mihira 2007,
3.1, and the value exceeding the upper and lower
limits of this range is classified into the highest and
lowest classes, respectively. Table 1. Material parameters for FE-analysis.
The mean strength μqu is held constant at
1.7 MPa and Vc is held constant at 0.3. The Elastic modulus E 3000 Mpa Stochastic
autocorrelation distance θqu is set to be 0.2, 0.5,
Poisson’s ratio ν 0.167 Deterministic
1.0 and 2.0 m. Typical samples of H = 2 m are
Friction angle φ 30 degree Deterministic
shown in Figure 3. The samples consist of 1760
Cohesion c 490 kPa Stochastic
elements. Dark and light regions depict “strong”
Tensile strength Tf 380 kPa Stochastic
and “weak” cement-treated soil, respectively.
Hardening parameter α 1.05 Deterministic
These spatial distributions of the strength are
Hardening parameter ey 0.0002 Deterministic
derived from the same probability distribution,
Fracture energy Gf 9.0 N/m Stochastic
that is to say, the normal distribution with the Softening parameter er 0.4 Deterministic
same parameters. The difference in spatial distri- Dilatancy coefficient Dc −0.4 Deterministic
bution is brought about by the difference of the Localization size ts0 0.6 mm Deterministic
autocorrelation distance θqu. Figure 3 shows that Characteristics length lc 100 mm Deterministic
a small value of θqu provides a rapidly varying

171

ISGSR2013.indb 171 10/18/2013 9:38:13 AM

 Namikawa & Koseki 2007, Namikawa & Koseki calculated from the total vertical load at the upper
2013). surface and the strain is calculated from the vertical
displacement of the upper surface. The stress-strain
relationship of the uniform (deterministic) sample
3.2 Numerical results
with μqu = 1.7 MPa is also plotted in these figures.
Five realizations of the overall stress-strain rela- For all the sample heights H, the stress-strain
tionship obtained from the FE-analysis with MCS relationships of the samples with θqu = 1.0 m
are shown in Figure 4 for the case with θqu = 1.0 m. vary widely before the peak stress state. It is also
In these stress-strain relationships, the stress is observed that most of the peak stresses are lower
than the peak stress of the uniform sample. It
can be seen that the axial strain at the peak stress
decreases as H increases. This is because that since
the size of the failure region may not depend on H,
the axial displacement at the peak stress does not
increase as H increases.
The overall strength Qu is defined as the peak
stress in the calculated overall stress-strain
relationship in this study. The sample mean of
the resulting 50 overall strengths mQu is shown in
Figure 5. It can be seen that mQu varies with θqu.
For all H, a minimum value of mQu is observed for
θqu = 0.5 m, implying that a particular value of θqu
which may depend on the sample diameter provide
the minimum value of mQu. Figure 5 also shows that
mQu decreases as H increases for the same θqu value.
Since the possibility that the sample involves weak
regions increases with the sample volume, larger
values of H could lead to smaller values of mQu.
The sample standard deviation of the resulting
50 overall strengths sQu is shown in Figure 6.
sQu is observed to be positively correlated with θqu
for all H.
This is because, as θqu tends to become large, the
field of strength tends to become uniform in a sam-
ple and the difference in mQu between the samples
becomes large. From a theoretical point of view, it
could be speculated that, as θqu becomes infinite,
sQu returns to the σqu set for the elements. Moreo-
ver sQu decreases as H increases for the same θqu
value, indicating that the standard deviation of Qu

Figure 4. Typical stress-strain relationships of realiza-

tions (θqu = 1 m). Figure 5. Sample mean of overall strength mQu.

172

ISGSR2013.indb 172 10/18/2013 9:38:13 AM

 Figure 6. Sample standard deviation of overall
strength sQu.

Figure 7. Overall strength for a 1% probability of

depends on the volume of sample. The standard failure Qu(pf = 1%).
deviation of the mean strength of the elements
composing the sample decreases with increasing
the number of elements Ne. Since Ne is propor- 4 CONCLUSIONS
tional to H, smaller values of sQu are obtained for
larger values of H. The finite element analyses with Monte Carlo
Simulations (MCS) were performed to simulate the
3.3 Probability of failure unconfined compression behavior of the columns
with different heights. The simulations with the
In a design procedure, it is important to evaluate samples in the presence of different degrees of
the probability of failure pf of a cement-treated soil spatial autocorrelation indicate that the sample
column. The type of the probability distribution mean of the resulting 50 overall strengths mQu
of Qu is required to evaluate pf. Namikawa (2012) varies with θqu. For all the height of sample H, a
investigated the probability distribution of Qu minimum value of mQu is observed for θqu = 0.5 m.
calculated from the FE-analysis with MCS. That The sample standard deviation of the resulting 50
study indicates that the probability distribution of overall strengths sQu is observed to decreases as H
Qu corresponds to the underlying distribution of qui. increases. The overall strength Qu(pf = 1%) for a 1%
Therefore, assuming the normal distribution as a probability of failure is evaluated based on mQu and
reasonable fit to the probability distribution of the sQu. The sample height does not affect significantly
computed Qu, pf defined as occurring when exceeds Qu(pf = 1%), indicating that the numerical results
a specific strength is calculated based on the com- for the sample of Η = 2 m or 3 m could be used to
puted mQu and sQu. The ISO 2394 requires pf = 1% evaluate the overall strength of longer columns.
for some consequence of a failure and moderate
relative costs of safety measures. Here the overall
strength Qu(pf = 1%) for a 1% probability of failure ACKNOWLEDGMENTS
is evaluated based on the computed mQu and sQu.
The evaluated Qu(pf = 1%) is shown in Figure 7. The author is grateful to Takenaka Corporation
For all H, Qu(pf = 1%) deceases significantly as θqu for their cooperation in the numerical analysis. The
increases. Since sQu increases as θqu increases (see author acknowledges the support of the Ministry
Fig. 6), the low variability leads to the high value of of Science and Education of Japan (Grant No.
Qu(pf = 1%) for the low value of θqu. Moreover, for 22560505).
the same reason, the value of Qu(pf = 1%) for Η = 4
m is larger than those for Η = 2 m and 3 m. For all
θqu, the minimum values of Qu(pf = 1%) are observed REFERENCES
when Η = 3 m. However Qu(pf = 1%) is not very sen-
sitive to Η, indicating that the numerical results for CDIT (Costal development Institute of Technology).
the sample of Η = 2 m or 3 m could be used to eval- 2002. The deep mixing method: principle, design and
uate the overall strength of longer columns. construction. Balkema: Tokyo.

173

ISGSR2013.indb 173 10/18/2013 9:38:14 AM

 Fushimi, M. 1992. Analysis of Scientific Data Namikawa, T. & Koseki, J. 2013. Effects spatial
(Basic Statistics 3), Chapter 11. Statistics Section, correlation on compression behavior of cement-treated
Department of Social Science, College of Arts and column. Journal of Geotechnical and Geoenvironmental
Science, University of Tokyo. University of Tokyo Engineering ASCE 139(8): 1346–1359.
Press: 307–332. (in Japanese). Namikawa, T, Koseki, K. & Suzuki, Y. 2008. Finite
Griffiths, D.V. & Fenton G.A. 2007. Probabilistic element analysis of a full scale bending test of cement
methods in geotechnical engineering. Springer Wien. treated soil column. 12th International Conference of
New York. the International Association for Computer Methods
International Organization for Standardization. and Advances in Geomechanics: 3635–3641.
1998. General principle on reliability for structures Namikawa, T. & Mihira, S. 2007. Elasto-plastic model
(ISO2394). for cement-treated sand. International Journal for
Matsuo, O. 2002. Determination of design parameters Numerical and Analytical Methods in Geomechanics
for deep mixing. Tokyo Workshop 2002 on Deep 31(1): 71–107.
Mixing: 75–79. Pietruszczak, S.T. & Mróz, Z. 1981. Finite element
Namikawa, T. 2012. Numerical study of compression analysis of deformation of strain-softening materials.
behavior of cement-treated columns with strength fields International Journal for Numerical Methods in
following normal and lognormal distributions. Ground Engineering 17: 327–334.
Improvement and Ground Control 2: 1193–1200. Shiomi, T., Shigeno, Y. & Zienkiewicz, O.C. 1993.
Namikawa, T. & Koseki, J. 2007. Evaluation of tensile Numerical prediction for model No.1. Arulanandan &
strength of cement-treated sand based on several Scott (ed.) Verifications of Numerical Procedures for
types of laboratory tests. Soils and Foundations 47(4): the Analysis of Soil Liquefaction Problems. Balkema:
657–674. 213–219.

174

ISGSR2013.indb 174 10/18/2013 9:38:14 AM

 Geotechnical Safety and Risk IV – Zhang et al. (eds)
© 2014 Taylor & Francis Group, London, ISBN 978-1-138-00163-3

Stochastic finite element analysis for ground improvement by vertical

drains of spatially variable soils

M. Wasiul Bari
Department of Civil Engineering, Curtin University, Perth, WA, Australia

M.A. Shams
Housing and Building Research Center, Cairo, Egypt

M.A. Shahin
Department of Civil Engineering, Curtin University, Perth, WA, Australia

ABSTRACT: A stochastic approach that investigates the effects of soil spatial variability on stabili-
zation of soft clay via Prefabricated Vertical Drains (PVDs) is presented and discussed. The approach
integrates the local average subdivision of random field theory with the Monte Carlo finite element tech-
nique. A special feature of the current study is the investigation of impact of spatial variability of soil
permeability and volume compressibility in the smear zone as compared to that of the undisturbed zone,
in conjunction with uncoupled 3D finite element analysis. The results of this study indicate that the spatial
variability of soil properties has a significant impact on soil consolidation by PVDs; however, it is also
found that the spatial variability of soil properties in the smear zone has a dominating impact on soil
consolidation by PVDs over that of the undisturbed zone.

1 INTRODUCTION The formulation and solution of stochastic

problems are often very complicated. The review
Soils are highly variable from one point to of relevant literature has indicated that although
another in the ground. This inherent variation the significance of soil spatial variability in rela-
of soils with respect to spatial location is known tion to ground improvement by PVDs has long
as soil spatial variability and is due to the une- been realized, little research has been made in this
ven soil micro fabric, complex characteristics of area. However, given the analytical and numerical
geological deposition and stress history. Despite complexity of the problem, available research
the fact that the impact of spatial variation of into the consolidation of highly variable soils
soil properties on soil consolidation has long has been limited to the following two categories:
been recognized by many researchers (e.g., Pyrah (i) one-dimensional consolidation of vertical drain-
1996; Rowe 1972), the design of soil consolida- age, i.e., no PVDs, for either 1D or 2D geometries
tion via Prefabricated Vertical Drains (PVDs) (e.g., Badaoui et al. 2007; Freeze 1977; Huang et al.
has been traditionally carried out determin- 2010; Hwang & Witczak 1984); and (ii) soil con-
istically and thus can be misleading due to the solidation by PVDs considering only the variabil-
ignorance of the uncertainty associated with the ity due to the testing errors in measuring the soil
inherent spatial variation of soil properties. In properties, while the inherent spatial variability
general, acknowledging and quantifying the soil of soil properties has not been taken into account
spatial variability in geotechnical engineering (e.g., Hong & Shang 1998; Zhou et al. 1999). More
has been usually considered using probabilistic recently, preliminary studies have been carried out
modeling techniques that treat the soil properties by the authors (e.g., Bari et al. 2012a, b; Shahin &
as random variables resulting in more realistic Bari 2012) on soil spatial variability for consoli-
solutions. Unlike deterministic analyses, which dation of soft clays by PVDs and have shown
are based on single best estimate (average or valuable insights into the impact of soil spatial
characteristic) values of soil properties, the prob- variability on soil consolidation and enhanced
abilistic analyses explicitly take into account the conceptual understanding about the soil consoli-
variable nature of soil properties, based on their dation problem. However, the above mentioned
statistical characteristics. works have notable limitation of either ignoring

175

ISGSR2013.indb 175 10/18/2013 9:38:14 AM

 the smear effect or considering smear effect with 2. Incorporate the generated soil profile into a
reference to permeability changes alone and vol- finite element modeling scheme of soil consoli-
ume compressibility have been ignored. It has to be dation by PVDs; and
noted that, the smear effect that develops as a con- 3. Repeat Steps 1 and 2 many times using the
sequence of mandrel installation not only reduces Monte Carlo technique so that a series of
soil permeability, k, but also increases soil volume consolidation responses can be obtained from
compressibility, mv. The combined effect of reduced which the statistical distribution parameters
permeability and increased volume compressibility and probability of achieving a target degree of
within the smear zone brings different behavior consolidation can be estimated and analyzed.
from that of the undisturbed soil. Hence, for more
Details of the steps used, as well as the numeri-
accurate prediction of the behavior of stabilized
cal procedures, are described below.
soil with PVDs, the changes of both k and mv in
the smear zone as well as undisturbed zone need to
be considered. However, due to the non-uniform
2.1 Generation of virtual soil profiles
spatial distribution of soil disturbance (which
decreases with the increase of distance from the As mentioned earlier, k and mv are considered to
center of the drain), the variability characteristics be random variables in the present study and are
of the smeared soil may be significantly different characterized in terms of their Probability Den-
from those of undisturbed soil. In addition, as sity Function (PDF) i.e., the mean, μ, standard
expelled water must pass through the smear zone, deviation, σ (the standard deviation can also be
the implication of variability parameters in this represented by variance, σ2, or coefficient of varia-
zone on the overall consolidation behavior may be tion, υ, where, υ = σ/μ), and correlation length, θ.
different from that of the variability parameters in It should be noted that the spatial variation of
the undisturbed zone. In this paper, a parametric soil properties is not entirely random and spatial
study is carried out to investigate the relative sig- dependencies also exist (Fenton & Vanmarcke
nificance of the spatial variability of soil properties 1990; Jaksa et al. 1997; Vanmarcke 1977). That
in the smear zone over undisturbed zone, where the is, a soil property at two separate spatial locations
coefficient of permeability, k, and coefficient of could be similar or otherwise, depending on the
volume compressibility, mv, are separately treated distance they are located apart and this is known
as random variables. as spatial correlation. Vanmarcke (1977) pointed
out that adequate characterization of spatially
variable soil properties requires consideration
2 STOCHASTIC MODELLING OF SOIL
(incorporation) of such spatial correlation. The
CONSOLIDATION BY PVDS
mean and standard deviation are the point statisti-
cal measures with no consideration of the spatial
Among several methods of modeling stochastic
correlation structure of soil properties. Therefore,
problems, the use of deterministic finite element
the correlation length, θ (also known as Scale of
analysis with random input soil parameters in a
Fluctuation, SOF), is introduced as an additional
Monte Carlo framework has gained much popu-
statistic to consider the spatial correlation of soil
larity in recent years (Elkateb et al. 2003). Similar
properties. Generally speaking, a large value of
approach is adopted in the present work to inves-
θ indicates smooth spatial variation of soil prop-
tigate the effects of soil spatial variability on the
erty of interest, whereas a small value of θ implies
behavior of soil consolidation by PVDs. The
erratic variation. In this study, the variability of
approach merges the Local Average Subdivision
both k and mv is characterized by following a log-
(LAS) method (to generate random permeability
normal distribution and assumed as 3D random
fields) and finite element modeling (to calculate
fields. In selecting the probability distribution of
soil consolidation by PVDs) into a Monte Carlo
k and mv, the authors reviewed a broad range of
framework. For a certain problem of ground
literature (e.g., Badaoui et al. 2007; Freeze 1977;
improvement by PVDs, the proposed approach
Huang et al. 2010) and concluded that it is reason-
can be applied using the following steps:
able to assume lognormal probability distribution
1. Create a virtual soil profile for the problem in for both k and mv. Since the same approach is used
hand which comprises a grid of elements that to generate random field of both k and mv, only
is assigned random values of soil properties dif- the procedure to generate the random field of k is
ferent from one element to another across the summarized herein.
grid. The virtual soil profile allows arbitrary In the process of simulating the lognormally
distributions of soil properties to be realistically distributed random field of k, correlated local
and economically modeled according to their averages standard normal random field G(x) are
statistical characteristics; first generated with zero mean, unit variance and

176

ISGSR2013.indb 176 10/18/2013 9:38:14 AM

 a spatial correlation function using 3D LAS tech- where: υk = σk/μk is the permeability Coefficient of
nique (Fenton & Vanmarcke 1990). The correlation Variation (COV). It should be noted that the ran-
coefficient between k measured at a point x1 and a dom fields of both k and mv are generated using
second point x2 is specified by a correlation func- the 3D free access LAS computer code available
tion, ρ(τ), where τ = |x1 − x2| is the absolute distance online at http://www.engmath.dal.ca/rfem/.
between the two points. An isotropic (i.e., the spa-
tial correlation lengths in the horizontal and verti-
2.2 Finite-element modeling incorporating
cal directions are taken to be equal) exponentially
soil spatial variability
decaying (Markovian) spatial correlation function
is used in the current study, as follows (Fenton & With the complete subsurface profile having been
Griffiths 2008): simulated in the previous step, the spatial variability
of k and mv is now known and can be employed as
⎛ 2τ ⎞ input in a Finite Element (FE) consolidation mod-
ρ (τ ) = exp ⎜ − (1) eling of soil improvement by PVDs. In this study,
⎝ θ k ⎟⎠ all numerical analyses are carried out using a modi-
fied version of the finite element computer program
It should be noted that the spatial correlation ‘‘p86’’ from the book by Smith & Griffiths (2004) in
function in Equation 1 is assumed to be statisti- which soil consolidation is treated as a 3D uncou-
cally isotropic, i.e., the scales of fluctuation in the pled (i.e., no displacement degrees of freedom only
horizontal (x), normal to the plane of paper (y) pore pressure degrees of freedom) problem. Origi-
and vertical directions (z) are assumed to be the nally program ‘‘p86’’ was for general two (plane)
same (i.e., θx = θy = θz = θ). Although the correla- or three dimensional analyses of the uncoupled
tion structures in any spatial direction are usually consolidation equation using implicit time integra-
different (i.e., anisotropic), the reason for assuming tion with the ‘‘theta’’ method. The authors modi-
an isotropic correlation structure for both kh and fied the source code of ‘‘p86’’ to allow for input
mv is because the scale of fluctuation is a difficult of the volume compressibility, axisymmetric and
parameter to estimate in practice and assuming an repetitive Monte-Carlo analyses. Since a single-
isotropic condition with smaller scale of fluctuation drain analysis is often enough to investigate the
will provide slightly conservative results (Fenton & soil consolidation behavior, the effect of soil spa-
Griffiths 2008). It is worthy to note that the spatial tial variability is examined using a unit cell of soil
correlation length is estimated with respect to the around a single drain. The consolidation problem
underlying normally distributed random field. considered in this study implies a unit cell (axisym-
Since k is assumed to be characterized statisti- metric) of actual 3D geometry of PVD system (see
cally by a lognormal distribution, the correlated Fig. 1a): L = 1.0 m, re = 0.8463 m, rs = 0.2821 m and
standard normal random field, G(x), generated rw = 0.0637 m, where L is the maximum vertical
using the LAS method is then transformed into a drainage distance; re is the radius of equivalent soil
lognormal distribution by the following transfor- cylinder with impermeable perimeter or the radius
mation function (Fenton & Griffiths 2008): of zone of influence; and rw is the equivalent radius
of the drain. However for the finite-element analy-
ki { k kG (i )} (2)
ses, the circular influence area of the cylindrical
unit cell is transformed into an equivalent square
influence area of side length S, such that S = √πre2
where: G(i) and Xi are, respectively, the local (arith- (i.e., S = 1.5 m). The selection of square influence
metic) average of a standard Gaussian random area instead of the equivalent circular influence
field G(x) over the domain of the i’th element and area is to avoid the unfavorable mesh shape as
the soil property value assigned to that element; μlnk the LAS method requires square (or rectangular)
and σlnk are the mean and standard deviation of elements to accurately compute locally averaged
the underlying normal distribution; μlnk and σlnk are values of kh and mv for each element across the
obtained from the specified permeability μk and σk grid. For the same reason, square shaped smear
using the following lognormal distribution trans- zone of side length Ss = √πrs2 and PVD of side
formation functions (Fenton & Griffiths 2008): length Sw = πrw/2 are employed. It should be noted
that, for simplicity, the well resistance factor which
1 2 may affect the rate of consolidation is not con-
μ ln k l μ k − σ ln k (3)
2 sidered in the FE analysis. This is due to the fact
that the discharge capacities of most PVDs avail-
able in the market are relatively high, and hence
⎛ σ2 ⎞
σ ln k = ln l + υk
l 1 + k2 ⎟ = ln
⎝ μk ⎠
( ) (4) the well resistance effect can be ignored in most
practical cases (Abuel-Naga et al. 2012; Chu 2004).

177

ISGSR2013.indb 177 10/18/2013 9:38:14 AM

 the specified statistical parameters (μ, σ and θ) of
each zone. Both random fields are then mapped
onto the corresponding grid in the finite element
mesh. As the accuracy of the finite element analy-
sis is dependent on the mesh density, a sensitivity
analysis for the problem under consideration is
carried out for both the deterministic and stochas-
tic solutions on various mesh dimensions to ensure
reasonable refinement with minimal discretization
error. The sensitivity analysis indicated that a mesh
with element size of 0.1 m × 0.1 m × 0.1 m gives
reasonable precision. Therefore, it was decided to
discretize the soil domain into a mesh with an ele-
ment size of 0.1 m × 0.1 m × 0.1 m (see Fig. 1b).
The 3D mesh consists of 2250 eight node first
order hexahedral elements.
To simulate reduced permeability condition in
the smear zone during the FE analysis, the mean
values of k in the undisturbed and smear zones
are taken to be equal to μ ku = 0.03 m/year and
μ ks = 0.015 m/year, respectively, which means that
μku /μks = 2.0. Walker (2006) indicates that the value
of the smear zone compressibility could increase
by about 20% from that of the undisturbed zone.
Therefore, to consider increased compressibil-
ity condition in the smear zone, the mean value
of mv in the undisturbed and smear zones are
taken to be equal to μmvu = 8.0 × 10−4 m2/kN and
μmvs = 9.6 × 10−4 m2/kN, respectively, which means
that μ vs /μmv = 1.2. The effect of spatially variable
s
of k and mv on the stochastic behavior of soil con-
solidation by PVDs is investigated over a range of
Figure 1. Schematic diagram of soil consolidation different combinations of standard deviation, σ,
with prefabricated vertical drain: (a) cylindrical unit cell; and scale of fluctuation, θ. It should be noted that
(b) equivalent square geometry with FE mesh σ is presented herein in a normalized form as υ
discretization. (i.e., coefficient of variation). The following values
of υ and θ are considered for the parametric study
presented in this paper:
In soil stabilization by PVDs, soil consolidation
• υk (%) (for both smear and undisturbed
takes place by combined vertical and horizontal
zones) = 50, 100, 200;
(radial) drainage of water. However, in practical
• υ mv (%) (for both smear and undisturbed
sense, soil consolidation due to vertical drainage
zones) = 10, 20, 30 and
is insignificant (due to large drainage length and
• θ (m) (for both k and mv, and smear and undis-
lower permeability in the vertical direction) com-
turbed zones) = 0.25, 0.5, 1.0.
pared to that of the horizontal drainage, thus,
only the component of the overall consolidation It can be noticed that, the selected range of COV
resulted from horizontal drainage is considered to of mv is much less than that of the range selected
be random in the current study. To simulate such for COV of k. This is due to the fact that soil per-
condition, the permeability in the vertical (z) direc- meability is considered to be the most significant
tion, kz is set as to zero in the FE analysis. Since the spatially random soil property affecting soil con-
permeability variance of even one of the directions solidation with high COV of up to 300%, while mv
is rarely known with any accuracy, the two com- can possess a spatial variability of up to 30% (e.g.,
ponents of the horizontal permeability (i.e., kx Beacher & Christian 2003; Kulhawy et al. 1991; Lee
and ky) are assumed as isotropic (i.e., kx = ky). In et al. 1983). However, the range of θ is assumed
order to take the smear effect into consideration, to be the same for both k and mv. This assump-
two independent random fields of both k and mv tion is reasonable because, if one thinks that the
are generated separately (one for the smear zone spatial correlation structure of a soil is caused by
and another for the undisturbed zone) employing changes in the constitutive nature of the soil over

178

ISGSR2013.indb 178 10/18/2013 9:38:16 AM

 the ground, then both k and mv would have similar the 2000 realizations is generated with the same
correlation lengths. Since little is currently known mean, standard deviation and scale of fluctuation,
about the typical COVs and SOFs of soils in the the spatial distribution of k and mv varies from one
smear zone, the same range of υ and θ are selected realization to the next. The above repetitive proc-
for both smear and undisturbed zones. It should be ess is performed for each combination of υ and θ.
noted that no cross-correlation between k and mv The obtained outputs from the suite of 2000 reali-
is assumed in this study and this is left for future zations of the Monte Carlo simulation are collated
refinement. In order to identify the statistical and statistically analyzed to produce estimates of
parameters in the smear and undisturbed zones, the mean and standard deviation of the degree of
υ and θ of k and mv are denoted with appropriate consolidation. In this study, at any given time t,
subscripts “s” and “u” depending on whether they the mean of the degree of consolidation based on
are specified for smear zone or undisturbed zone, the excess pore water pressure, μU, is estimated by
where s refers to the smear zone while u refers to utilizing the geometric average (considered as the
undisturbed zone. An initial pore water pressure of representative mean) of ū(t), as follows:
100 kPa dissipates in a single drain is considered in
all FE analyses. A single generation of a random ⎡ 1 nsim i
⎛ u (t ) ⎞ ⎤
field and the subsequent finite-element analysis of μU = 1 − exp ⎢ ∑ ln ⎜ ⎥
⎝ u0 ⎟⎠ i ⎥⎦
(6)
that field are termed “realization”. For an individ- ⎢⎣ nsim
i i =1
ual realization, the degree of consolidation, U(t),
at any certain consolidation time, t, is calculated The standard deviation of the average degree
with the help of the following expression: of consolidation at any time t defined by the pore
water pressure, σU, is estimated as follows:
u (t )
U (t ) = 1 − (5) nsim
u0 1 i
2
σU
nsim
∑ ⎡(U t )i
−1 ⎣
μU ⎤⎦ (7)
i i =1
where: u0 = initial pore pressure; and ū(t) = average
pore pressures at any time of the consolidation
where: nsim is the number of Monte Carlo simu-
process. It has to be emphasized that ū(t) of the
lations; (ū(t)/u0)i and (U(t))i are, respectively, the
consolidation process is calculated by numerically
ratio of the average excess pore pressure to the
integrating the pore pressure across the volume
initial excess pore water pressure and the degree
of each element at a particular time, summing the
of consolidation at any time t for the ith simula-
contribution of each element and dividing by the
tion (see Eq. 5). The use of the geometric aver-
total mesh volume (element volume are also calcu-
age of ū(t) in computing μU is due to the fact that,
lated by numerical integration).
in a 2D or 3D space, compared to the 1D space,
the flow of water has more freedom to avoid low
2.3 Repetition of process based permeability zones by detouring around them and
on the Monte Carlo technique therefore, the geometric average may be a better
estimator (e.g., Dagan 1989) for computing the
Following the procedures of the Monte Carlo representative mean of the average excess pore
technique, the process of generating random fields water pressures. For the same reason, Huang et al.
of kh and mv and performing the finite element (2010) also used geometric average in determining
analysis is repeated numerous times. The accuracy equivalent coefficient of consolidation for a 2D
of the estimated statistics of the output quantities system.
of interest is dependent on the number of reali-
zations required in the Monte Carlo procedures.
Therefore, to maintain accuracy and run time effi- 3 PROBABILISTIC INTERPRETATION
ciency, the sensitivity of results to the number of
Monte Carlo simulations is examined. The sensi- The estimation of the probability that a determin-
tivity analysis indicated that 2000 realizations are istic degree of consolidation overestimates the true
sufficient to give reasonably stable output statistics consolidation value is one of the main objectives of
for each analysis of interest. Based on this observa- the stochastic consolidation analyses. Such prob-
tion, the process of generating a random field of ability can be represented either by the probability
both k and mv and the subsequent finite element of achieving a target degree of consolidation, Us,
analysis is repeated 2000 times. Huang et al. (2010) (i.e., P[U(ts) ≥ Us(ts)]) at any specified consolida-
also performed successful probabilistic analy- tion time, ts, or the probability of required time t
sis on soil consolidation using 2000 simulations. to achieve Us that is less than or equal to ts (i.e.,
Although each k and mv field realizations out of P[t(Us) ≤ ts(Us)]). In this study, the later process is

179

ISGSR2013.indb 179 10/18/2013 9:38:18 AM

 employed, i.e., P[t(Us) ≤ ts(Us)] is estimated. This nsim
1 i
is because determining probability from a set of σ t (U ) ∑ [ti (U s ) μt (U ) ]2 (9)
i − 1 i =1
nsim
s s
data requires establishment of a reasonable prob-
ability distribution for the data set. However, the
obtained fit using the raw data of U(ts) was typi- where: ti(Us) is the t from the i’th realization (i = 1,
cally poor while the distribution of t(Us) obtained 2, 3, …, nsim) at given Us and nsim = total number of
from the suite of the 2000 realizations is reason- realizations = 2000. As 90% consolidation is usu-
ably fitted with lognormal distribution and gives ally acceptable for the purpose of design of any
sufficiently reasonable approximation to the soil improvement project (Bo et al. 2003), in this
P[t(Us) ≤ ts(Us)]. The legitimacy of the lognormal study, it is assumed that the target degree of con-
distribution hypothesis for t(Us) is examined by solidation is 90% and for convenience, it is simply
the well-known Chi-square test through the fre- denoted as U90. The probability that t is less than
quency density plot of t(Us) data obtained from or equal to ts that achieves U90 can then be obtained
the 2000 realizations and a fitted lognormal distri- from the following lognormal probability distribu-
bution is superimposed. This process is performed tion transformation:
for many combinations of υ and θ at several differ-
ent Us. For each of the cases considered, the good- ⎛ ln ts − μ lln t (U ) ⎞
ness-of-fit p-value is found to be high enough to P ⎡⎣t (U 90 ) ts (U 90 )⎤⎦ = Φ ⎜ 90
⎟ (10)
approve the rationality of the lognormal distribu- ⎜⎝ σ lln t (U ) ⎟⎠
90
tion hypothesis of simulated t(Us) data. Figure 2
illustrates a typical example of the histogram
where: P [.] is the probability of its argument; Φ(.)
of t(Us) for the case of υ ku = 50%, υ ks = 200%,
υ mvu = 10% , υ mvs = 30%, θ ku θ ks = θ mv = θ mv = 0.5 is the standard normal cumulative distribution
function; μ ln t (U ) and σ ln t (U ) are, respectively, the
at Us = 90%, along with their fitted lognormal dis-
u u
s s
mean and standard deviation of the underlying
tributions. The goodness-of-fit test yielded p-value
normally distributed ln t(Us) and can be estimated
of 0.4, indicating strong agreement between the
from μt (U ) and σ t (U ) with reference to Equations
histogram and the fitted distribution. Therefore, s s
3 and 4, as follows:
the lognormal distribution is certainly an appro-
priate assumption to the distribution of the simu-
1 2
lated t(Us) data. μ ln t (U ) l μt (U ) − σ ln (11)
By accepting the lognormal distribution for t s s
2 t (U s )
at any given Us, the statistical moments μt (U ) and
σ t (U s ) that are representing the mean and stand- ⎛ σ t2U ⎞
s

( )
ard deviation of the lognormally distributed t that σ ln t (U ) = lln ⎜1 + 2 s ⎟ (12)
achieves Us are calculated from the suite of 2000
s
⎜⎝ μt (U ) ⎟⎠
s
realizations using the following transformation
functions: Following the procedure set out above, prob-
nsim abilities of required time t to achieve Us that is less
1 i

μt (U ) =
s
nsim
∑ ti (U s ) (8) than or equal to ts can be estimated for any com-
bination of υ and θ, and the stochastic behavior
i i =1
of soil consolidation by PVDs can be investigated.

4 RESULTS AND DISCUSSION

In order to investigate the relative significance of

the spatially variability of the smear zone over the
undisturbed zone, a series of 3D consolidation anal-
yses are performed. The sensitivity of the statistics
of the degree of consolidation and the probability
of required time t to achieve Us that is less than or
equal to ts to the statistically defined input data (i.e.,
υ and θ) in relation to both k and mv is examined.
For each selected set of υ and θ, 2000 Monte Carlo
Figure 2. Typical example of frequency density histo- simulations are performed. The obtained consoli-
gram of simulated t(U90) with fitted lognormal distribu- dation responses are then statistically analyzed to
tion for υ ku = 50%, υ ks = 200%, υ m = 10%, υ mv = 30%,
estimate μU, σU and P[t(U90) ≤ ts(U90)] using the
θ ku θ ks = θ mv θ mv = 0.5 .
vu s

u s excess pore water pressure. Since the general trends

180

ISGSR2013.indb 180 10/18/2013 9:38:19 AM

 of μU, σU and P[t(U90) ≤ ts(U90)] remain unaltered
over the specified range of υ and θ, only the results
of a few of the tests conducted are presented in
Figures 3–8, which are believed to be sufficient to
demonstrate the main features of the influence of
spatial variability of k and mv on soil consolida-

Figure 5. Effect of υu and υs on σU for θu = θs = 0.5.

Figure 3. Effect of υu and υs on μU for θu = θs = 0.5.

Figure 6. Effect of θu and θs on σU for υ ku υ ks = 50%

and υ mv υ mv = 10%.
u s

tion by PVDs. In Figures 3–8, μU, σU and P[t(U90) ≤

ts(U90)] are expressed as a function of time t. Prior
to placing the stochastic analyses into context, an
initial deterministic solution has been performed
Figure 4. Effect of θu and θs on μU for υ ku υ ks = 50%
assuming a homogeneous soil. It should be noted
and υ mv υ mv = 10%.
u s that the deterministic solution of this case yields

181

ISGSR2013.indb 181 10/18/2013 9:38:25 AM

 also includes the deterministic solution of no soil
variability. It can be seen from Figure 3(a) that at
any consolidation time, there is a slight reduction
in μU for spatially varied soils compared to the
deterministic case. The nearly identical curves for
all cases of υu(υ ks and υ mvs are fixed at 50% and 10%
respectively) plotted in the figure indicate that the
effect of increasing υu on μU remains marginal. The
effect of υs on μU at fixed values of υ ku = 50% and
υ mv = 10% is illustrated in Figure 3b, which shows
that any change in υs has a significant impact on
u

the estimated values of μU. At any certain consoli-

dation time, μU decreases with the increase of υs,
and the decreasing rate of μU consistently increases
with the increase of υs. The comparison between
Figures 3a and b reveals that the effect of υs on μU
is dominating.
Figure 4 highlights the effects of increasing θu
and θs on μU at fixed values of υ ku υ ks = 50% and
υ mv υ mv = 10%. Virtually, the identical curves
of μU in Figure 4a for all θu at a fixed value of
u s

Figure 7. Effect of υu and υs on P[t(U90) ≤ ts(U90)] for θ ks θ mvs = 0.25 , indicate that μU is more or less
θu = θs = 0.5. independent of θu. Similar to θu, the influence of
θs (θ ku and θ mv are fixed at 0.25) on μU is also mar-
u
ginal as can be seen in Figure 4b.

4.2 Effect of variation of υ and θ on the standard

deviation of U
The influence of υu and υs on σU at a fixed value
of θu = θs = 0.5 is depicted in Figure 5. For a fixed
value of υ s (υ ks and υ mv are, respectively, 50% and
10% in this case), increasing υu has a marginal effect
s

on σU, as shown in Figure 5a. Figure 5b shows the

effect of υs on σU at fixed values of υ ku = 50% and
υ mv = 10%, and from which it can be seen that at
any certain consolidation time, σU increases signifi-
u

cantly with the increase of υs, implying the domi-

nant effect of υs on the estimated values of σU.
Figure 6 illustrates the effect of varying θu and
θs on σU at fixed values of υ ku υ ks = 50% and
υ mvu υ mvs = 10%. In Figure 6a, it can be seen
that similar to the effect of θu on μU, σU remains
almost identical for varying θu with a fixed value of
θ ks = θ mvs = 0.25 . On the other hand, the estimated
σU for different values of θs is plotted in Figure 6b
Figure 8. Effect of θu and θs on P[t(U90) ≤ ts(U90)] for at a fixed value of θ ku θ mv = 0.25, which illus-
trates that unlike θu, θs has a considerable impact
u
υ ku υ ks = 50% and υ mv υ mv = 10%.
u s
on the estimated values of σU.

U90 at t = 0.73 year (i.e., tD90 = 0.73 year). The results

4.3 Effect of variation of υ and θ on the
obtained from this study are described below.
probability of required time t to achieve U90
that is less than or equal to ts
4.1 Effect of variation of υ and θ on the mean
The influence of the smear zone parameters
of U
over the undisturbed zone parameters in relation
The effects of increasing υu and υs on μU at fixed to the probability of required time t to achieve
value of θu = θs = 0.5 is examined in Figure 3, which U90 that is less than or equal to ts are investigated

182

ISGSR2013.indb 182 10/18/2013 9:38:29 AM

 in Figures 7 and 8. The deterministic time of consolidation were found to be highly sensitive to
achieving 90% consolidation, tD90, is also shown υ and θ of the soil properties at the smear zone.
in the figures by vertical solid lines that give This result indicates that the probabilistic behav-
P[t(U90) ≤ ts(U90)] at that time, for any combina- ior of soil consolidation is governed by the spatial
tion of υ and θ. variation of the soil properties of the smear zone.
The effects of υu and υs on P[t(U90) ≤ ts(U90)] at Since the spatial variability of the smear zone will
a fixed value of θu = θs = 0.5 is demonstrated in possibly be different from that of the undisturbed
Figure 7. It can be seen from Figure 7a that, in zone, this observation has important implications
general, the effect of increasing υu (υ ks and υ mv are in the sense that, modeling soil consolidation
fixed at 50% and 10%, respectively) on P[t(U90) ≤ with the same υ and θ for both zones (i.e., undis-
s

ts(U90)] remains marginal. The effect of υs at fixed turbed and smear) that are equal to the υ and θ
values of υ ku = 50% and υ mvu = 10% is shown in of the smear zone does not significantly affect the
Figure 7b, which shows that varying the values of final results. Overall, the results obtained from
υs has a considerable impact on the estimated val- this research highlight valuable insights into the
ues of P[t(U90) ≤ ts(U90)]. At any certain consolida- impact of soil spatial variability on soil improve-
tion time, P[t(U90) ≤ ts(U90)] decreases significantly ment by PVDs and clearly demonstrate the
with the increase of υs. The overall observation potential of stochastic analyses in routine design
that can be derived from comparing the results in practice.
Figure 7 is that the effect of υs on P[t(U90) ≤ ts(U90)]
is dominant.
Figure 8 investigates the effects of θ on REFERENCES
P[t(U90) ≤ ts(U90)] at fixed values of υ ku υ ks = 50%
and υ mv υ mv = 10%. In Figure 8a, the influence Abuel-Naga, H.M., Pender, M.J. & Bergado, D.T. 2012.
of increasing θu on P[t(U90) ≤ ts(U90)] is shown
u s
Design curves of prefabricated vertical drains includ-
at θ ks θ mv = 0.25, and the results yield almost ing smear and transition zones effects. Geotextiles and
s
identical curves indicating that varying the values Geomembranes 32: 1–9.
Badaoui, M., Nour, A., Slimani, A. & Berrah, M.K.
of θu has little or no impact on the probabilistic 2007. Consolidation statistics investigation via thin
behavior of degree of consolidation. On the other layer method analysis. Transport in Porous Media
hand, the estimated P[t(U90) ≤ ts(U90)] for different 67(1): 69–91.
values of θs is plotted in Figure 8b at a fixed value Bari, M.W., Shahin, M.A. & Nikraz, H.R. 2012a. Effects
of θ ku θ mv = 0.25. It can be seen that unlike θu, of soil spatial variability on axisymmetric versus plane
θs has a considerable impact on the estimated val-
u
strain analyses of ground improvement by prefabri-
ues of P[t(U90) ≤ ts(U90)]. The comparison between cated vertical drains. International Journal of Geotech-
Figures 8a and b reveals that, the effect of θs on nical Engineering 6(2): 139–147.
P[t(U90) ≤ ts(U90)] is more significant than θu. It is Bari, M.W., Shahin, M.A. & Nikraz, H.R. 2012b.
Probabilistic analysis of soil consolidation via prefab-
interesting to know that the deterministic solution ricated vertical drains. International Journal of Geome-
yields P[t(U90) ≤ ts(U90)] < 50% for all combinations chanics, ASCE: (in press).
of values of υu, υs, θu, and θs, as can be seen in Beacher, G.B. & Christian, J.T. 2003. Reliability and
Figures 7 and 8. Statistics in Geotechnical Engineering, John Wiley &
Sons: Chichester, England.
Bo, M.W., Chu, J., Low, B.K. & Choa, V. 2003. Soil
5 CONCLUSIONS Improvement: Prefabricated Vertical Drain Techniques,
Thomson Learning: Singapore.
This paper has used the random field theory and Chu, J. 2004. Practical considerations for using vertical
drains in soil improvement projects. Geotextiles and
finite element modeling to investigate the rela- Geomembranes 22(1): 101–117.
tive significance of soil spatially variability in the Dagan, G. 1989. Flow and transport in porous media,
smear and undisturbed zones in soil improve- Springer: New York.
ment by prefabricated vertical drains. Both the Elkateb, T., Chalaturnyk, R. & Robertson, P.K. 2003.
coefficient of permeability, k, and coefficient of An overview of soil heterogeneity: Quantification and
volume compressibility, mv, were treated as inde- implications on geotechnical field problems. Canadian
pendent random variables and uncoupled 3D Geotechnical Journal 40(1): 1–15.
finite element analysis was applied. The effect of Fenton, G.A. & Griffiths, D.V. 2008. Risk assessment in
coefficient of variation, υ, and spatial correlation geotechnical engineering, Wiley: New York.
Fenton, G.A. & Vanmarcke, E.H. 1990. Simulation of
or scale of fluctuation, θ, of the undisturbed zone random fields via local average subdivision. Journal of
on the estimated mean and standard deviation Engineering Mechanics 116(8): 1733–1749.
of the degree of consolidation was found to be Freeze, R.A. 1977. Probabilistic one-dimensional consol-
marginal. On the other hand, the estimated statis- idation. Journal of Geotechnical Engineering Division
tics and probability associated with the degree of 103(GT7): 725–742.

183

ISGSR2013.indb 183 10/18/2013 9:38:35 AM

 Hong, H.P. & Shang, J.Q. 1998. Probabilistic analysis Pyrah, I.C. 1996. One-dimensional consolidation of lay-
of consolidation with prefabricated vertical drains ered soils. Géotechnique 46(3): 555–560.
for soil improvement. Canadian Geotechnical Journal Rowe, P.W. 1972. The relevance of soil fabric to site inves-
35(4): 666–677. tigation practice. Géotechnique 22(2): 195–300.
Huang, J., Griffiths, D.V. & Fenton, G.A. 2010. Proba- Shahin, M.A. & Bari, M.W. 2012. Modeling of Ground
bilistic analysis of coupled soil consolidation. Journal Improvement by Prefabricated Vertical Drains in
of Geotechnical and Geoenvironmental Engineering Highly Variable Soils. International Conference on
136(3): 417–430. Ground Improvement and Ground Control (ICGI 2012),
Hwang, D. & Witczak, M.W. 1984. Multidimensional University of Wollongong, Australia: 321–335.
probabilistic consolidation. Journal of Geotechnical Smith, I.M. & Griffiths, D.V. 2004. Programming the
Engineering 110(8): 1059–1077. finite element method, John Wiley & Sons.
Jaksa, M.B., Brooker, P.I. & Kaggwa, W.S. 1997. Vanmarcke, E.H. 1977. Probabilistic modelling of soil
Inaccuracies associated with estimating random meas- profiles. Journal of Geotechnical Engineering Division
urement errors. Journal of Geotechnical and Geoenvi- 103(11): 1227–1246.
ronmental Engineering 123(5): 393–401. Walker, R.T. 2006. Analytical solutions for modeling soft
Kulhawy, F.H., Roth, M.J.S. & Grigoriu, M.D. 1991. soil consolidation by vertical drains, University of
Some statistical evaluations of geotechnical properties. Wollongong: Wollongong, Australia.
Proceedings of the 6th International Conference on Zhou, W., Hong, H.P. & Shang, J.Q. 1999. Probabilistic
Applied Statistical Problems in Civil Engineering design method of prefabricated vertical drains for soil
(ICASP6), Mexico City: 705–712. improvement. Journal of Geotechnical and Geoenviron-
Lee, I.K., White, W. & Ingles, O.G. 1983. Geotechnical mental Engineering 125(8): 659–664.
engineering, Pitman: London.

184

ISGSR2013.indb 184 10/18/2013 9:38:38 AM

 Geotechnical Safety and Risk IV – Zhang et al. (eds)
© 2014 Taylor & Francis Group, London, ISBN 978-1-138-00163-3

Regularity of the variance reduction function in Tianjin Port

Shu-Wang Yan & Lin-Ping Guo

State Key Laboratory of Hydraulic Engineering Simulation and Safety, Tianjin University, Tianjin, China

Yong-Hua Cao
CCCC-Tianjin Port Engineering Institute Ltd., Tianjin, China

ABSTRACT: In the process of reliability analysis in geotechnical engineering, the variance reduction
function is the key parameter to transit the “point property” into “spatial average property”. In the paper,
improvement on the method to calculate the no correlation distance put forward by Yan Shu-wang and
Zhu Hong-xia was made, which can greatly minimize the computation efforts. The procedure of calcula-
tion of reduction function was analyzed, and general regularity of determining reduction function through
calculation of no correlation distance was obtained, which was for the first time put forward by us.

1 INSTRUCTION where δu = the correlation distance, which is a

constant; Δz = sampling distance; h = depth to be
Recent years, the essence of the rapidly developed averaged; ρ(Δz) = correlation function; Γ2(h) = the
reliability analysis is considering various kinds of variance reduction function. The correlation dis-
uncertainty in engineering. In reliability analysis, tance can be approximately written as:
the self-correlation theory is generally employed
to simulate spatial variability and relativity of h 2
( h) ≈ δ u (2)
the soil. The soil profile random field model sug-
gested by Vanmarcke (1977, 1983) can connect And the variance reduction factor can be
the “spatial variance” with “point property”. In obtained as:
the theory, the reduction function of the variance
is determined by the correlation distance of soil,
⎧ 1 ( δu )
therefore, the correlation distance is the key to the ⎪
analysis (Zhu & Gao, 2003; Zhu, 2007). Γ 2( ) = ⎨ δu (3)
⎪⎩ h ( δu )
The correlation function method is employed to
calculate the correlation distance of typical soil lay-
ers in Tianjin port. Improvement on no correlation To obtain the Eq. (3), the correlation function
distance method put forward by Yan & Zhu (2007) ρ(Δz) in the stochastic field model should be taken
is made, on the basis of which, reduction functions as:
of variance of Tianjin Port are calculated. And the
general regularity of determining reduction func- δu )
⎧1 (
tion through calculation of no correlation distance ρ( )= ⎨ (4)
is obtained. ⎩0 ( δu )

It can be known from the theory mentioned

2 CONCEPTION AND CALCULATION
above that the correlation distance is not a random
METHOD OF CORRELATION
constant. It can be regarded as an essential dis-
DISTANCE
tance, named as “correlation distance” measuring
correlation degree of two random variables. When
2.1 Conception of correlation distance
the distance between two points is smaller than the
Vanmarcke (1977) put forward that, if: value δu, they are strongly related; otherwise, they
are considered as uncorrelated (Jia, 1985; Yan
h⎛ Δz ⎞ et al., 2006). However, the “correlation distance” is
lim h 2
( ) = 2 lim ∫ 1 − ρ( )d ( )
h→ h →∞ 0 ⎝ h⎠ just a conception used in practical project, which is
h
not strictly the same as perfect correlated (ρ = 1) or
= 2∫ ρ ( )d ( ) = δu (1) uncorrelated (ρ = 0) in mathematics.
0

185

ISGSR2013.indb 185 10/18/2013 9:38:38 AM

 It can be known from Eq. (3) that the variance
reduction factor can be calculated as long as the
correlation distance is obtained.

2.2 The improved correlation function method

There are several methods to determine correla-
tion distance since the stochastic field model put
forward by Vanmarcke (1977, 1983). Such as
the space average method, correlation function
method, statistical simulation method, average
zero span method and half a variation coefficient
method. The most specific and convenient method
to obtain correlation distance is the space aver-
age method and the correlation function method
(Li et al., 2003; Yan et al., 2013). In this paper, to Figure 1. Fitting result of the improved correlation
the application of the no correlation distance, the function method.
improved correlation function method is employed
to calculate the correlation distance.
The improved correlation method calculate the
correlation distance through fitting the correlation plot the calculated correlation function accord-
function ρ(Δz). According to the definition of cor- ing to Δz;
relation distance and the variance reduction func- determine the fitting correlation function
tion, it can be obtained: according to the plot ρ(Δz) ∼ Δz;
suppose a set of parameter for the correlation
2 h⎛ Δz ⎞ function and plot the corresponding curve
h ∫0 ⎝
δu 2
ρ ( z )d ( z ) ρ(Δz) ∼ Δz;
h→ h →∞ h⎠
check whether the theoretical line can fit calcu-
∞ lated points in the front part;
= 2∫ ρ ( z )d ( ) (5)
0 if not, adjust parameters, till they fit each other
well in the front part (as shown in Fig. 1). And
Several commonly used correlation function and these parameters are what wanted;
corresponding expressions of correlation distance the correlation distance can be obtained accord-
are listed in Table 1 (Peng, 1992). ing to Table 1.
The correlation distance can be obtained once
the correlation function and corresponding param-
eters are determined. Steps are as follows: 3 CALCULATION METHOD AND
IMPROVEMENT OF NO CORRELATION
Take different values of i for Δz = iΔz0, and
DISTANCE
calculate:

ρ( ) ρ( E [ ( z ) ( z + Δz )] 3.1 Calculation method of no correlation

0)
distance
1 n i
= ∑Y ( k )Y ( k + i )
n i k =1
(6) Coefficient of reduction of variance is affected
by the length to be averaged and correlation dis-
tance. Once the correlation distance is determined,
And a series of value for correlation function the larger the value of h, the more to be reduced.
are obtained. Therefore, it’s of vital importance to confirm the
value of h and the no correlation distance. In prac-
tice, the no correlation distance can be determined
Table 1. Correspondence between correlation function
and correlation distance.
by means of plotting. The correlation function
should be confirmed firstly, and the Γ2(h) ∼ h/
Correlation function ρ(τ) Correlation distance δu δu curve can be plotted according to Eq. (3) and
Table 1. The point of intersection can be obtained,
e−b|τ| 2/b which is marked as n*, and consider the L* = n*δu,
e−(bτ)2 π /b of which the L* is the no correlation distance. The
no correlation distance is half value of L*. Regu-
e−b|τ| ⋅ cos(bτ) 1/b
larity of no correlation distance will be discussed
e−b|τ| ⋅ cos(ωτ) (2b/b2) + ω2
corresponding to different situations.

186

ISGSR2013.indb 186 10/18/2013 9:38:39 AM

 1. When the correlation function is in exponen- Rewrite the correlation function and the
tial type, the first type in Table 1, the variance variance reduction function with correlation
reduction function is: distance, there can be:

2 τ
Γ2 ( ) = ( bh
1) (7) ρ (τ ) = e −τ δu
cos ,
b2 h2 δu
Replace b in Eq. (7) with δu, and it can be δ u ⎛ δ u ⎞ 2 −h /δu h
Γ2 ( ) = − e sin (10)
obtained that: h ⎝ h ⎠ δu

ρ (τ ) = e −2τ δ u , The curve Γ2(h) ∼ h/δu can be plotted. Accord-

δ 1 ⎛ δu ⎞
2 ing to Eqns. (3) and (10), which is shown in
2h /δ
Γ2 ( ) = u + ( ) (8) Figure 3.
h 2⎝ h ⎠ Figure 3 shows that, the intersection point
is n* =4, of which the reduction coefficient is
The curve Γ2(h) ∼ h/δu can be plotted accord- approximately 0.25 and the calculated value
ing to Eqns. (3) and (8), which is shown in from Eq. (10) is 0.251, which can be approxi-
Figure 2. mately considered as equal.
Figure 2 shows that, the intersection point According to Eq. (10), there can be
is n* =10, of which the reduction coefficient is L*Γ2(L*) = 1.004δu ≈ δu, meanwhile, the cor-
approximately 0.1 and the calculated value from relation function is ρ(L*) = −0.012 ≈ 0.
Eq. (8) is 0.095, which can be approximately Therefore, the calculated L* is the lower limit
considered as equal. of h for hΓ2(h) ≈ δu, half of which is the no
According to Eq. (8), there can be correlation distance. Therefore the no corre-
L*Γ2(L*) = 0.095 δu ≈ δu, meanwhile, the correla- lation distance can be confirmed as h* = 2δu,
tion function is ρ(L*) = 2.06 × 10−9 ≈ 0. There- and the reduction coefficient of variance is
fore, the calculated L* is the lower limit of h for Γ2(h*) = 0.469.
hΓ2(h) ≈ δu, half of which is the no correlation 3. When the correlation function is in exponential
distance. Therefore the no correlation distance type with cosine, the last type in Table 1. And
can be confirmed as h* = 5δu, and the reduction the variance reduction function will be:
coefficient of variance is Γ2(h*) = 0.18.
2. When the correlation function is in exponential
type with cosine, the third type in Table 1. The
variance reduction function will be:
Γ2 ( ) =
h2 ( 2
2
2 2
)
{bbh( 2 2
) ( 2 2
)

2 2
Γ ( )= 2 2 ( bh
1) (9)
−e bh ⎡2ω bsin(
⎣ b siin((ω hh)) (ω
2
c ω h )⎤⎦
b2 ) cos( }
b h (11)

Figure 2. Reduction process of variance for correlation Figure 3. Reduction process of variance for correlation
function in exponential type. function in exponential type with cosine.

187

ISGSR2013.indb 187 10/18/2013 9:38:40 AM

 Table 2. No correlation distances under given conditions.

ω/b δu Γ2(h) h* Γ2(h*)

⎛δ u⎞ ⎡−0.495 − e − h / 0.505δ ⎛ h ⎞⎤
2
δu h
⎜⎝0.1sin 5.05δ − 0.99 cos 5.05δ ⎟⎠⎥
2
0.1 1/0.505b Γ ( )= + 0.095
⎝ h ⎠ ⎢⎣
u
5δu
h u u ⎦

⎛δ u⎞ ⎡−0.455 − e − h / 0.545δ ⎛ h ⎞⎤
2
δu h
⎜⎝0.3 sin 1.82δ − 0.91cos 1.82δ ⎟⎠⎥
2
0.3 1/0.545b Γ ( )= + 4.5δu 0.106
⎝ h ⎠ ⎢⎣
u

h u u ⎦

⎛δ u⎞ ⎡−0.375 − e − h / 0.625δ ⎛ h ⎞⎤
2
δu h
⎜⎝0.5 sin 1.25δ − 0.75 cos 1.25δ ⎟⎠⎥
2
0.5 1/0.625b Γ ( )= + 0.119
⎝ h ⎠ ⎢⎣
u
4δu
h u u ⎦

δu ⎛δ u⎞
2
⎡ − h / 0.745δ ⎛ h h ⎞⎤
⎜⎝0.7 sin 1.064δ − 0.51cos 1.0643δ ⎟⎠⎥
2
0.7 1/0.745b Γ ( )= + ⎢0.255 − e
u
3δu 0.189
h ⎝ h⎠ ⎣ u u ⎦

2
2 δu ⎛ δ u ⎞ e − h/δ sin h
1 1/b Γ ( )= − u
2δu 0.251
h ⎝ h⎠ δu
δu ⎛δ u⎞
2
⎡ − h / 1.345δ ⎛ h h ⎞⎤
⎜⎝1.3 sin 1.035δ + 0.69 cos 1.035δ ⎟⎠⎥
2
1.3 1/1.345b Γ ( )= + ⎢0.345 − e
u
3δu 0.215
h ⎝ h⎠ ⎣ u u ⎦

δu ⎛δ u⎞
2
⎡ − h//1.625δ ⎛ h h ⎞⎤
⎜⎝1.5 sin 1.083δ + 1.25 cos 1.083δ ⎟⎠⎥
2
1.5 1/1.625b Γ ( )= + ⎢0.625 − e
u
5δu 0.096
h ⎝ h⎠ ⎣ u u ⎦

⎛ δ u ⎞ ⎡0.945 − e − h / 1.945δ ⎛ h ⎞⎤
2
δu h
⎜⎝ 1.7 sin 1.144δ + 1.89 cos 1.144δ ⎟⎠ ⎥
2
1.7 1/1.945b Γ ( )= + 6.5δu 0.076
⎝ h ⎠ ⎢⎣
u

h u u ⎦

δu ⎛ δu ⎞
2
⎡ − h / 2.5δ ⎛ h h ⎞⎤
⎜ 2.sin 1.25δ + 3 cos 1.25δ ⎠⎟ ⎥
2
2 1/2.5b Γ ( )= + ⎢1.5 − e
u
9.5δu 0.057
h ⎝ h⎠ ⎣ ⎝ u u ⎦

The two parameters b and ω should be firstly

determined if the reduction function rewritten
with δu. Regularity of no correlation distance will
be discussed by means of the followed several situ-
ations, as shown in Table 2.
Curve Γ2(h) ∼ h/δu corresponding to different
value of ω/b are plotted according to Eq. (3) and
Table 2 as shown in Figure 4(a) and 4(b).
It can be known from Figure 4(a) and 4(b) that,
with increasing of ω/b, the no correlation dis-
tance decreases when the ratio is smaller than 1,
and the reduction coefficient increases; conversely,
with increasing of ω/b, the no correlation distance
increases when the ratio is larger than 1, and the
reduction coefficient decreases. The no correlation
distance is the largest and the reduction coefficient
is the smallest when the ratio ω/b is 1. Regulari-
ties described above can be described as Figures 5
and 6 as scatter plots.

3.2 Improvement of no correlation distance

There will be a value of ω/b for each layer of
each boring. The variance reduction function
corresponding to each ω/b should be calculated Figure 4. Reduction processes of variance with differ-
according to Table 2, of which the process seems ent value of ω/b.

188

ISGSR2013.indb 188 10/18/2013 9:38:42 AM

 to be greatly complicated. Through analyzing the
process mentioned above, an equation appropriate
for any ratio of ω/b is obtained:

δu ⎛ δu ⎞ 2 ⎡
Γ2 ( ) = + α 1 − e − h / α 2δ u
h ⎝ h⎠ ⎣
⎛ h h ⎞⎤
× ⎜ α 3 sin + α 5 cos
α 4δ u ⎟⎠ ⎥⎦
(12)
⎝ α 4δ u

where, ω/b = α, α1 = (α2 − 1)/2, α1 = (α2 + 1)/2,

α3 = α, α4 = 2α/(α2 + 1), α5 = α2 − 1.
In practical project, parameters b and ω should
be determined first according to correlation func-
tion, and the correlation distance will be obtained.
Secondly, plot the curve Γ2(h) ∼ h/δu and determine
the no correlation distance h*. Finally, the variance
reduction function can be obtained.
Figure 5. Change of reduction coefficient with the Analysis mentioned above supposing that, the
ratio ω/b. effective influence depth is no less than the no cor-
relation distance. The corresponding equation for
variance reduction function should be employed
in case the effective influence depth is smaller than
the no correlation distance.

3.3 Principle of determining variance reduction

function
As mentioned above, reduction coefficient
of variance can be determined according to
Table 3.

4 CASE STUDY

Test data of the forty cone penetration tests of

the Wharf engineering in Tianjin are analyzed.
Correlation distances and reduction functions of
variance corresponding to typical layers in Tianjin
Figure 6. Change of no correlation distance with the Port are obtained, which can be a reference to
ratio ω/b. other projects in this area.

Table 3. Principle of determining variance reduction function.

Variance reduction function Γ2(h)

Correlation function ρ(Δz) δu ≤ L ≤ h* L ≥ h*

2 bh
Exponential type ρ(τ) = e−b|τ| 2 2
( bh e 1) (where, h = L ) 0.095 1
b h
1
1. Type of exponential with ⎡
bh − e
( bh )
si bh ⎤⎦ (where, h = L ) 0.251 1
b h ⎣
2 2
Cosine ρ(τ) = e−b|τ| ⋅ cos(bτ)
δu/h* 1
δu⎛δ ⎞
2
⎡α − e − h / α δ ⎛ h h ⎞⎤
2. Type of exponential with + u ⎢1 ⎜⎝α 3 sin α δ + α 5 cos α δ ⎟⎠⎥ (where, ω/b =α,
2 u

Cosine ρ(τ) = e−b|τ| ⋅ cos(ωτ) h ⎝ h⎠ ⎣ 4 u u ⎦

α1 = (α2 − 1)/2, α1 = (α2 + 1)/2, α3 = α, α4 = 2α/(α2 + 1), α5 = α2 − 1, h = L)

189

ISGSR2013.indb 189 10/18/2013 9:38:46 AM

 4.1 Calculation and statistics of vertical The cone tip resistance of CPT parameter is
correlation distance used to calculate the vertical correlation distance
(Xu et al., 1998; Li et al., 2003; Yan et al., 2007).
Stochastic model of the soil profile in this area
For preparation of the application of no correla-
possess the stationarity and ergodicity (Yan et al.,
tion distance method, correlation function method
1995). Therefore, it can be analyzed by the stochas-
is employed. Results are shown in Table 4 (Zhu,
tic theory put forward by Vanmarcke (1977).
2007).

Table 4. Vertical correlation distance of Tianjin Port. 4.2 Variance reduction function
Correlation distance (m) Figure 7 describes laws for the reduction function
Cone penetration with ω/b. Results of this area can be employed to
test number Silt Silt clay Silty clay prove whether the law is ubiquitous. Reduction
functions of variance corresponding to the ratio
1 0.139 0.533 0.400 ω/b are plotted in Figure 7.
2 0.330 1.730 0.358 It can be known from Figure 7 that the law men-
3 0.209 1.220 0.431 tioned above is ubiquitous. That is with increas-
4 0.182 1.067 0.400 ing of ω/b, the no correlation distance decreases
5 0.190 0.273 0.483 when the ratio is smaller than 1, and the reduc-
6 0.412 0.800 0.431 tion coefficient increases; conversely, with increas-
7 0.067 1.077 0.297 ing of ω/b, the no correlation distance increases
8 0.270 1.239 0.325 when the ratio is larger than 1, and the reduction
9 0.984 1.239 0.308 coefficient decreases. The no correlation distance
10 0.175 0.400 0.492
is the largest and the reduction coefficient is the
11 0.543 0.554 0.335
smallest when the ratio ω/b is 1. To conservatively
12 0.176 1.103 0.224
consider in practical project, the smallest ratio ω/b
13 0.649 0.690 0.163
will be taken to determine the reduction function
14 0.130 0.825 0.400
of the variance. In addition, the value of variance
15 0.341 0.226 0.400
16 0.113 0.588 0.431
17 0.701 0.865 0.400
18 0.181 0.550 0.400
19 0.452 1.000 0.457
20 0.511 1.000 0.400
21 0.923 0.650 0.398
22 0.195 0.550 0.500
23 0.320 0.667 0.276
24 0.088 0.738 0.328
25 0.202 1.077 0.400
26 0.400 0.500 0.363
27 0.667 0.320 0.363
28 0.690 0.550 0.328
29 0.546 0.054 0.384
30 0.267 0.780 0.328
31 0.495 0.832 0.303
32 0.100 0.780 0.328 Figure 7. The change law of variance reduction
33 0.100 1.429 0.260 function.
34 0.100 0.667 0.355
35 0.100 0.920 0.328
36 0.100 1.107 0.328 Table 5. Variance reduction function of typical layers
in Tianjin Port.
37 0.098 1.176 0.303
38 0.183 1.440 0.270 No Reduction
39 0.175 0.976 0.303 Name of ω/b correlation function of
40 0.180 0.148 0.270 the soil (min) distance the variance
Minimum value 0.040 0.054 0.163
Maximum value 0.984 1.730 0.500 Silt 0.17 4.75 δu 0.374
Average 0.317 0.808 0.356 Silt clay 0.21 4.75 δu 0.312
Standard deviation 0.237 0.368 0.072 Silty clay 0.38 4.50 δu 0.378

190

ISGSR2013.indb 190 10/18/2013 9:38:47 AM

 reduction function is determined only by the ratio Li, Xiao-Yong, Xie, Kang-He, Yu, Yan. 2003. Research
ω/b but not the value of ω or b, which can also be of the characteristics of correlation distance on soil
known from Eq. (12). properties indexes. China Civil Engineering Journal
36(8): 91–95.
Peng, Da-Peng. 1992. Application of homogeneous ran-
dom field to analyze indexes of soil properties. Trans-
5 CONCLUSIONS action of Tianjin University 2: 118–124. (In Chinese).
Vanmareke, E.H. 1977. Probability modeling of soil
1. An equation appropriate for any ratio of ω/b Profile. J. Geotec. Eng. Division 103: 1227–1245.
for the process of calculating no correlation dis- Vanmarcke, E.H. 1983. Random fields: analysis and syn-
tance is obtained, which can greatly minimize thesis. MIT Press, Cambridge, Mass.
the computation efforts. Xu, Bin, Wang, Da-Tong, Gao, Da-zhao. 1998. Discussion
2. By analyzing regularity of determining vari- on calculation of correlation distance with cone pen-
ance reduction function with the method of no etration curve by correlation function method. Rock
and Soil Mechanics 19(1): 55–58. (In Chinese).
correlation distance, the law is obtained: with Yan, Shuwang, Guo, Linping. 2013. Estimation of scale
increasing of ω/b, the no correlation distance of fluctuation of soil properties indexes. Applied
decreases when the ratio is smaller than 1, and Mechanics and Materials. Guilin, China. 256–259:
the reduction coefficient increases; conversely, 149–152.
with increasing of ω/b, the no correlation dis- Yan, Shu-Wang, Jia, Xiao-Li, Guo, Huai-zhi. 1995. Test
tance increases when the ratio is larger than 1, on stationarity and ergodicity of stochastic model
and the reduction coefficient decreases. The no of soil profiles. Chinese Journal of Geotechnical
correlation distance is the largest and the reduc- Engineering 17(3): 1–9. (In Chinese).
tion coefficient is the smallest when the ratio Yan, Shu-Wang, Zhu, Hong-Xia, Liu, Run, Sun, Wan-he.
2007. Study on methods for estimating correlation dis-
ω/b is 1. To conservatively consider in practi- tance of soil layers. Rock and Soil Mechanics 28(8):
cal project, the smallest ratio ω/b will be taken 1581–1586. (In Chinese).
to determine the reduction function of the Zhu, Deng-Feng, Gao, Da-Zhao. 2003. Space statistical
variance. characteristics analysis of smooth stochastic field of
soil characteristics. Rock and Soil Mechanics 24(3):
455–462. (In Chinese).
REFERENCES Zhu, Hong-xia. 2007. Study of using random field
theory in soil foundation reliability analysis. Tianjin
Jia, Xiao-li. 1985. Study on properties of stochastic field University. (In Chinese).
of soil profiles. Tianjin University. (In Chinese).
Li, Jing-Pei, Shu, Xiang, Ding, Shi-Jun. 2003. Discussion
on characteristic parameter of self-correlation of soil
index and its determining principle. Transaction of
Tongji University 31(3): 287–290. (In Chinese).

191

ISGSR2013.indb 191 10/18/2013 9:38:48 AM

This page intentionally left blank
 Geotechnical Safety and Risk IV – Zhang et al. (eds)
© 2014 Taylor & Francis Group, London, ISBN 978-1-138-00163-3

Seepage through anisotropic non-stationary random fields

H. Zhu & L.M. Zhang

The Hong Kong University of Science and Technology, Hong Kong, China

ABSTRACT: Natural slopes are highly heterogeneous. When inverse segregation takes place in a land-
slide soil deposit, the coefficient of permeability of the soil statistically decreases from the surface to the
bottom of the deposit. Such material spatial variability can be characterized using non-stationary random
fields. The objectives of this paper are to develop algorithms to generate non-stationary random fields for
a slope with the mean coefficient of saturated permeability decreasing with depth and to carry out analy-
sis of infiltration in an anisotropic non-stationary slope. Basic anisotropic stationary random fields with
orthogonal major and minor scales of fluctuation are considered first. Non-stationary random fields are
then formulated by adding a trend to the stationary random fields. The Cholesky decomposition method
is applied to produce correlated random numbers following an exponential correlation structure. After
generating the anisotropically heterogeneous non-stationary random fields, the steady-state pore-water
pressures in each random field are obtained by solving the Richards equation. The effects of coefficient
of variation of the saturated permeability and the ratio between the major and minor scales of fluctuation
on the pressure bound are investigated.

1 INTRODUCTION Zhu et al. 2012) and bearing capacity of shallow

and deep foundations (Fenton & Griffiths 2003).
Natural slopes are highly heterogeneous as a Stationary random fields were widely simulated
result of depositional and post-depositional in geotechnical applications. A linear trend for
processes (DeGroot & Baecher 1993; Lacasse & undrained shear strength was considered as in
Nadim 1997). Spatial variability can be character- Lumb (1966). Hicks & Samy (2002) investigated
ized using random field theory (e.g. Vanmarcke the influence of heterogeneity on undrained clay
1977). In a random field, the variables exhibit slope stability, in which the effective shear strength
autocorrelation, which is a tendency for soil parameter was treated as a non-stationary ran-
properties at one point to be correlated to soil dom field with the mean strength parameter
properties at nearby points. A classic paper that increasing linearly with depth. Zhang et al. (2011)
introduces the spatial correlation concept was and others observed inverse segregation of par-
published by Vanmarcke (1977). According to ticles in a giant landslide event. The observation
the theory, random fields are classified into two confirms that segregation mechanism works by
groups. A random field is called stationary if the inverse grading, with large particles moving to the
joint probability distribution that governs the top and small particles accumulating at the bot-
field is invariant when translated over the param- tom (Miyamoto et al. 2007). The upper zone of
eter space; a random field is noted non-stationary the landslide deposit is more permeable than that
when a deterministic trend is added to the space. of the lower zone. Therefore, it is of great concern
Treatment must first be given to transform a non- to treat the parameter space of saturated perme-
stationary dataset into a stationary dataset by ability as a non-stationary random field with the
removing the deterministic trend when analyzing mean value decreasing from the top to the bottom
spatial correlation (Dasaka & Zhang 2012). Two of a slope, which realistically represents a land-
principal directions are defined in this study: the slide deposit.
major principal direction along which the proper- The objectives of this paper are (1) to develop
ties show the smoothest variation and the minor algorithms to generate non-stationary random
principal direction along which the properties fields with mean values varying with depth and
show the rapidest variation. (2) to figure out the extent the coefficient of satu-
Random fields have been extensively applied rated permeability and the ratio between the prin-
to geotechnical problems such as slope stabil- cipal scales of fluctuation affects the pore-water
ity (e.g. Huang et al. 2010; Griffiths et al. 2011; pressures in a slope from infiltration analysis.

193

ISGSR2013.indb 193 10/18/2013 9:38:48 AM

 2 CHARACTERIZING NON-STATIONARY
RANDOM FIELDS

2.1 Random field theory

Figure 1 shows a schematic of a One-Dimensional
(1D) and a Two-Dimensional (2D) stationary
random field of a variable, z, with a mean, E(z),
a variance, σ2 and a cumulative probability dis-
tribution function, F(z). These three features can
be obtained by analyzing field data collected via
ground investigation. The scale of fluctuation, θ, is
a constant if an isotropic random field is assumed
as shown in Figure 1(a). While in a 2D anisotropic
random field in Figure 1(b), θ is a function of the
directional angle, φ, and prescribed principal scales
of fluctuation. Spatial correlation models are par-
ametric relationships used to fit the experimental
variograms or covariance functions from analysis
of field data. Phoon & Kulhawy (1999) and Guide
(2005) provided an excellent summary of common
variogram models used in practice. Evaluation
of correlation structures was also performed by
Liu & Chen (2010). Listed below are three com-
mon correlation structures in the isotropic form
(from Equations 1–3: exponential, Gaussian and
spherical):
Figure 1. A realization of (a) a 1D random field and
ρ ( 2h/θ ) (1) (b) a 2D random field with a mean E[z], variance σ2, and
cumulative probability function F(z).
ρ ( ( /θ ) 2 ) (2)
ρ 1 1 ( h / θ ) + 0.5( h / θ ) , h θ ; ρ = 0, h > 0
3
(3)
space. The coefficient of correlation in the basic
where ρ is the correlation coefficient of the ran- anisotropy case is calculated as an example,
dom variable at two locations at a separation dis- assuming that the scale of fluctuation follows an
tance of h and θ is the scale of fluctuation. This elliptical function and the correlation structure is
study adopts the exponential correlation structure exponential.
as it fits well with geotechnical properties. The ani- Figure 2 explains the evaluation of the correla-
sotropic form of exponential correlation structure tion matrix. The grid points which represent the
can be generally written as property values in each finite element are involved
in the computation of correlation coefficients.
⎡ Δx 2 Δy2 ⎤
For example, if the distance between points 1 and
ρ = exp ⎢ −2 + 2 ⎥ (4) 2 is hx, the correlation between these two points
⎢⎣ θ12 θ 2 ⎥⎦ is calculated by putting the values Δx = hx and
Δy = 0 in Equation 4. Similarly, the correlations
where Δx and Δy are the horizontal and vertical of point 1 with points 3, 4, 5 can be established
separation distances between two observations in by placing Δx = 2hx and Δy = 0, Δx = 3hx and
the space, θ1 and θ2 denote the major and minor Δy = 0, Δx = 4hx and Δy = 0, respectively. The
principal scales of fluctuation, respectively. In this correlations of point 1 with points 41, 42 and 43
case, the major direction is parallel to the horizon- can be calculated by placing Δx = 0 and Δy = hy,
tal axis and the minor direction is parallel to the Δx = hx and Δy = hy, and Δx = 2hx and Δy = hy,
vertical one. Let denote θh = θ1 and θv = θ1. respectively. Therefore, values in the first row of
the correlation matrix are the correlation coef-
ficients between grid point 1 and other points,
2.2 Creating covariance matrix
1600 values in a row when the number of zones is
The covariance matrix is created by computing 40 × 40. Hence, considering all the zones, a cor-
the correlation coefficient of two points spaced at relation matrix ρx with 1600 × 1600 elements can
any distance and in any direction in the parameter be established.

194

ISGSR2013.indb 194 10/18/2013 9:38:48 AM

 X-direction, the general form of f(y) can be writ-
ten as:

f y ) = a b ( y Ymin ) (Ymax Ymin ) (7)

where y is the vertical coordinate of a desired posi-

tion in an analysis domain; a and b are parameters
in the linear function; Ymax and Ymin are the maxi-
mum and minimum Y-coordinates, respectively.
Hence, the non-stationary random field dataset
Y( x, y ) can be expressed as:

Y( x, y ) x, y ) × f ( y )
X( x, (8)

In order to better understand the trend in the

non-stationary random fields simulated in this
study, an example of a rectangular random field
Figure 2. Discretization of a random field with 40 × with X-coordinates ranging from 0 m to 10 m and
40 grid points.
Y-coordinates ranging from 15 m to 19 m is taken
into account. Values of a and b are assumed as 1
and 10, respectively. Thereafter, Y( x, y ) is written
2.3 Cholesky decomposition as:
The correlation matrix is decomposed into the
product of a lower triangular matrix and its trans- Y( x, y ) x, y ) × (1 + 10 × ( y −
X( x, ) / 4) (9)
pose by Cholesky’s decomposition (e.g. Press et al.
2002): The mean of the random set in X( x, y ) is 2 × 10−5,
and Y( x, y ) reaches the maximum mean value of
L LT = ρk (5) 2 × 10−5 at y = 19 m and the minimum mean value
of 2.2 × 10−4 at y = 15 m.
Given the matrix, L, a correlated standard nor- The values for X( x, y ) and Y( x, y ) along an
mal random field is obtained as follows: arbitrary cross section taken from one realiza-
tion of random field are shown in Figure 3. The
i mean value of ks for the stationary random field is
Gi ∑ Lij Z j , i = 1, 2, 3, ... n (6) assumed as 2 × 10−5 m/s. According to Equation 9,
the mean values for the non-stationary random
j =1
field decreases from the top (y = 19 m) to the bot-
where Zj is a sequence of independent standard tom (y = 15 m). It should be noted that the coef-
normal random variables. A correlated lognor- ficients of variation of ks are implicitly assumed
mal random field is obtained by transforming the constant since both the mean and the standard
standard normal data. deviation are enlarged simultaneously.

2.4 Adding a trend to the lognormal stationary

random field
Stationarity of the data is achieved by removing
a low-order polynomial trend of order not higher
than quadratic (Brooker 1991), which is usu-
ally estimated by the ordinary least squares error
approach (Journel & Huijbregts 1978). In most of
the studies, the trend function is simply estimated
by regression analysis using either linear or poly-
nomial curve fitting (Kulhawy et al. 1992).
Let X( x, y ) denote the generated lognormal
stationary anisotropic random field dataset at the
desired location X x , Y y in the X-Y plane. Figure 3. Random values along one arbitrary cross-
A deterministic trend f(y) is added to the verti- section of a stationary random field and a non-stationary
cal direction of X( x, y ). For all positions in the random field with a trend.

195

ISGSR2013.indb 195 10/18/2013 9:38:50 AM

 where h is the total head; Ψ is the pressure head;
y is the elevation head; q is the flux; k is the unsatu-
rated permeability. Considering mass conservation,
the governing equation for steady-state seepage
can be written as:

∂q ∂q
+ =0 (12)
∂x ∂y

An exponential permeability equation by

Leong & Rahardjo (1997) is employed:
p
⎛ θ − θr ⎞
k ks ⎜ w (13)
⎝ θ s − θ r ⎟⎠

where θw is the volumetric water content; θs is the

saturated volumetric water content; θr is the resid-
ual water content at which a large suction change is
required to remove additional water from the soil;
(θw – θr)/(θs – θr) is the normalized water content;
p is a constant depending on the soil type. From
Equation 13, the uncertainty of permeability func-
tion is characterized by the uncertainty of ks if p
remains unchanged for points within the slope.
This paper makes use of FlexPDE (Guide 2005)
to solve the seepage problem. It allows users to
input variables in tabulated files by which stochas-
Figure 4. Realizations of non-stationary random fields: tic finite element analyses are performed repeat-
(a) stationary isotropic random field with μ = 2 × 10−5 m/s, edly. The random field grid is superimposed on
COV = 0.5, θh = θy = 1 m; (b) non-stationary random the finite element mesh in such a way that their
field with a trend, COV = 0.5, θh = 10 m, θv = 1 m;
(c) non-stationary random field with a trend, COV = 0.5,
corresponding bottom right, bottom left and top
θh = 10 m, θv = 0.1 m. left corners coincide with those of the slope. The
ks value at each grid point of the rectangular ran-
dom field grid is then transferred into the finite ele-
ment program as an input soil parameter. Bilinear
2.5 Numerical examples of non-stationary interpolation is performed to map from the rectan-
random fields gular random field grid to the finite element mesh.
Typical realizations of random fields are presented The region that is outside of the boundary of the
in Figure 4. The mean value of coefficient of satu- finite element mesh is truncated automatically in
rated permeability, ks, for the stationary random the finite element program.
field is 2 × 10−5 m/s. The trend form is taken flowing Numerical modelling of seepage in a hypo-
Equation 9. The Coefficient Of Variation (COV) thetic slope with the same geometry of the slope
of ks is selected as 0.5 (Duncan 2000). The scales in Figure 4(c) is conducted. The boundary con-
of fluctuation are selected as θh = 10 m, θv = 0.1 m, ditions are considered deterministic. There is no
and θh = 10 m, θv = 1 m, respectively (Phoon and change in hydraulic head on both sides (B–C,
Kulhawy 1999). D–E), and the bottom boundary conditions as
well as the side boundaries above the water table
(A–B, E–F) are impermeable. A groundwater table
3 PROBABILISTIC INFILTRATION is fixed at the side boundaries at points B and E.
ANALYSIS A constant flux of 2 × 10−7 m/s is applied to the sur-
face boundaries A–H, H–G and G–F. For the soil
In this paper, the steady-state flow in a heteroge- properties in concern, the p value in the permeabil-
neous unsaturated-saturated soil is described by ity function is assumed as 3 for fine sands, and the
Darcy’s law in two dimensions: saturated water content is taken as 0.4. The COV
of ks is suggested as 0.1, 0.5 and 1.0 according to
h y (10)
Duncan (2000). The horizontal scale of fluctuation
q k ∇h (11) is assumed as 10 m, i.e. five times the slope height.

196

ISGSR2013.indb 196 10/18/2013 9:38:55 AM

 Table 1. Statistics of ks in each study case.

Case Coefficient
no. of variation Scale of fluctuation

1 1.0 θh = 10 m, θv = 0.1 m
2 0.5 θh = 10 m, θv = 0.1 m
3 0.1 θh = 10 m, θv = 0.1 m
4 1.0 θh = 10 m, θv = 1.0 m
5 0.5 θh = 10 m, θv = 1.0 m
6 1.0 θh = 10 m, θv = 0.5 m

The vertical scale of fluctuation is assumed as 1 m,

0.5 m and 0.1 m according to Phoon & Kulhawy
(1999). Since the purpose of this study is to inves-
tigate how the coefficient of variation of ks and the
ratio between the horizontal and vertical scales of
fluctuation affect the pore-water pressures in the
slope, Table 1 summarizes the statistics of ks in
each case of non-stationary random field.

4 RESULTS AND DISCUSSIONS

OF PROBABILISTIC ANALYSIS

For each case in Table 1 together as well as an

isotropic case, 1000 realizations of ks random
fields are simulated with the number of realiza-
tions sufficient to reach required computational
accuracy. For each realization, the pore-water
pressures are calculated by implementing the
finite element method via the FlexPDE platform.
The 25%, 50% and 75% quantiles of pore-water
pressures above the groundwater table along the
middle of slope for cases in Figures 4(a)–(c) are
estimated at each elevation. Results are shown in
Figures 5(a)–(c). The uncertainties in pore-water
pressures are reflected by the width of pressure Figure 5. The 25%, 50% and 75% quantiles of pore-
bound, which is defined as the difference between water pressures along the middle of the slope for cases
the 75% and 25% quantiles. A wider pressure (a), (b) and (c) in Figure 4.
bound implies larger uncertainty. The results
regarding the bound are further interpretated as
shown in Figures 6 and 7.
The effects of the COV of ks and ratios between
θh and θv on the width of pressure bound are inves-
tigated. Figure 6 shows the variation of the width
of pressure bound with the COV of ks. It implies
that the pore-water pressures have a wider range
of values at a higher value of COV. Figure 7 shows
the changes in the width of pressure bound with
the ratio between the horizontal and vertical scales
of fluctuation. The horizontal scale of fluctuation
is constant for all the study cases. The change in
the vertical scale of fluctuation controls the rate of
variation of ks with depth. A ratio of 100 induces Figure 6. Variation of width of pressure bound with
the rapidest variation of ks in the study cases. coefficient of variation of saturated permeability.

197

ISGSR2013.indb 197 10/18/2013 9:38:57 AM

 Dasaka, S. & Zhang, L.M. 2012. Spatial variability of in
situ weathered soil. Géotechnique 62(5): 375–384.
DeGroot, D.J. & Baecher, G.B. 1993. Estimating autoco-
variance of in-situ soil properties. Journal of Geotech-
nical Engineering 119(1): 147–166.
Duncan, J.M. 2000. Factors of safety and reliability in
geotechnical engineering. Journal of Geotechnical and
Geoenvironmental Engineering 126(4): 307–316.
Fenton, G.A. & Griffiths, D.V. 2003. Bearing-capacity
prediction of spatially random c-ϕ soils. Canadian
Geotechnical Journal 40(1): 54–65.
Griffiths, D.V., Huang, J.S. & Fenton, G.A. 2011.
Probabilistic infinite slope analysis. Computers and
Geotechnics 38(4): 577–584.
Figure 7. Variation of width of pressure bound with Guide, F.U. 2005. Version 5.0, PDE Solutions. Inc.,
the ratio between horizontal and vertical scales of Antioch, CA, USA.
fluctuation. Hicks, M.A. & Samy, K. 2002. Influence of heterogeneity
on undrained clay slope stability. Quarterly Journal of
Engineering Geology and Hydrogeology 35(1): 41–49.
Huang, J.S., Griffiths, D.V. & Fenton, G.A. 2010. System
It can be noted from Figure 7 that a more ragged reliability of slopes by RFEM. Soils and Foundations
distribution of ks in the vertical direction leads to a 50(3): 343–353.
smaller uncertainty of pressures. Journel, A.G. & Huijbregts, C.J. 1978. Mining geostatistics,
Academic press London.
Kulhawy, F., Birgisson, B. & Grigoriu, M. 1992.
5 SUMMARY AND CONCLUSIONS Reliability-based foundation design for transmission
line structures. Electric Power Research Inst., Palo
Numerical algorithms were developed to model Alto, CA; Cornell Univ., Ithaca, NY, USA.
non-stationary random fields with mean values Lacasse, S. & Nadim, F. 1997. Uncertainties in character-
varying with depth. Steady–state analysis of seep- izing soil properties. Publikasjon-Norges Geotekniske
Institute 201: 49–75.
age in the generated non-stationary random fields Leong, E.C. & Rahardjo, H. 1997. Permeability func-
was carried out. tions for unsaturated soils. Journal of Geotechnical and
The generated profiles of pore-water pressures Geoenvironmental Engineering 123(12): 1118–1126.
from infiltration analysis were further interpre- Liu, C.N. & Chen, C.H. 2010. Estimating spatial cor-
tated. The 25%, 50% and 75% quantiles of pore- relation structures based on CPT data. Georisk 4(2):
water pressures were estimated along the depth 99–108.
of the slope. The width of pressure bound can Lumb, P. (1966). The variability of natural soils. Canadian
be used to represent the uncertainty of hydraulic Geotechnical Journal 3(2): 74–97.
responses. Miyamoto, H., Yano, H., Scheeres, D.J., Abe, S.,
Barnouin-Jha, O., Cheng, A.F., Demura, H.,
Results show that the COV of ks and ratios Gaskell, R.W., Hirata, N. & Ishiguro, M. 2007.
between the horizontal and vertical scales of Regolith migration and sorting on asteroid Itokawa.
fluctuation have impacts on the pressure bounds. Science 316(5827): 1011–1014.
A wider range of pore-water pressures is expected Phoon, K.K. & Kulhawy, F.H. 1999. Characterization
at a higher value of COV. A more rapid variation of geotechnical variability. Canadian Geotechnical
of ks in the vertical direction leads to a smaller Journal 36(4): 612–624.
uncertainty of pressures. Press, W.H., Teukolsky, S.A., Vetterling, W.T. &
Flannery, B.P. 2002. Numerical recipes, Cambridge
University Press Cambridge.
ACKNOWLEDGEMENTS Vanmarcke, E.H. 1977. Probabilistic modeling of soil
profiles. Journal of the Geotechnical Engineering
Division 103(11): 1227–1246.
The work in this paper was substantially sup- Zhang, L.M., Xu, Y., Huang, R.Q. & Chang, D.S. 2011.
ported by the National Basic Research Program of Particle flow and segregation in a giant landslide event
China (973 Program) (No. 2011CB013506) and the triggered by the 2008 Wenchuan earthquake, Sichuan,
Research Grants Council of the Hong Kong Special China. Natural Hazards and Earth System Sciences
Administrative Region (No. HKUST09/CRF/9). 11: 1153–1162.
Zhu, H., Zhang, L.M., Zhang, L.L. & Zhou, C.B. 2013.
Two-dimensional probabilistic infiltration analy-
REFERENCES sis with a spatially varying permeability function.
Computers and Geotechnics 48: 249–259.
Brooker, P.I. 1991. Geostatistical Primer, World Scientific,
Singapore.

198

ISGSR2013.indb 198 10/18/2013 9:38:58 AM

 5 Geohazards

ISGSR2013.indb 199 10/18/2013 9:38:58 AM

This page intentionally left blank
 Geotechnical Safety and Risk IV – Zhang et al. (eds)
© 2014 Taylor & Francis Group, London, ISBN 978-1-138-00163-3

Integrating seismic hazard analyses with geotechnical site

characterization for liquefaction potential assessment
in Kaohsiung area

D. Huang & J.P. Wang

Department of Civil and Environmental Engineering, Hong Kong University of Science and Technology,
Kowloon, Hong Kong, China

ABSTRACT: Probabilistic Seismic Hazard Analysis (PSHA) has been widely used to estimate the
ground motion intensity that should be accounted for in structure design. Here, an in-depth PSHA study
for Kaohsiung, the metropolitan area in South Taiwan, is carried out. On the other hand, soil liquefaction
and subsequent structural damage are usually reported in a catastrophic earthquake event, such as numer-
ous significant incidents of liquefaction in inland alluvial areas during the 1999 Chi-Chi earthquake.
A quantitative framework to evaluate earthquake-induced soil liquefaction potential is proposed herein,
including earthquake hazard assessment and geotechnical liquefaction evaluation. This framework is fur-
ther demonstrated with a case study for a thoroughly-investigated site in Kaohsiung. Results show that
the liquefaction potential evaluated with existing methods is to some extent dependent on the earthquake
hazard in terms of the occurrence probability of large-size earthquakes at the study site. Ground improve-
ment is suggested considering severe consequences of ground failure.

1 INTRODUCTION is needed at the site. The framework is presented

with a case study in Kaohsiung metropolitan area,
Soil liquefaction is one of the disasters during earth- where some important civil engineering projects
quakes. For example, soil liquefaction and subse- are under construction or development at the site.
quent structural damages were reported in many
locations in Taiwan during the Chi-Chi earthquake
in 1999 (Chu et al. 2004, Juang et al. 2002). Serious
2 LIQUEFATION ANALYSES
consequences include lateral spreading of sloping
ground and settlement caused by reconsolidation
2.1 Geotechnical site characteristics
of the liquefied soils (Seed & Idriss 1971). A few
empirical methods have been proposed to evalu- Figure 1 shows the location of the site within
ate the liquefaction potential by comparing “load” Kaohsiung City (the second most important city
induced by one or more earthquakes with “resist- in Taiwan), together with the seismicity since 1900
ance” based on the in-situ methods at the site. (i.e., more than 55,000 events). It is worth not-
(Seed & Idriss 1971, Iwasaki et al. 1982, Iwasaki ing that this earthquake catalog was used for the
1986, Tokimatsu & Yoshimi 1983). seismic hazard analyses, as it has been analyzed
However, most studies use deterministic meth- for quantifying earthquake potentials for Taiwan
ods to evaluate earthquake loading. In contrast, (Wang et al. 2011, 2012a, 2012b).
only a few studies consider both uncertainties The geotechnical condition at the site was inten-
from geotechnical properties and earthquake sively investigated during the construction of the
geology around the site (Lee et al. 2010, Wang & mass-rapid-transit system (i.e., the subway system)
Huang 2012). As a result, the underlying scope of of this city starting in the 1990s. The soil at the
this study is to evaluate soil liquefaction given dif- site is mainly silty sand, with SPT N-values around
ferent durations of time, or the time-variant soil 4∼18 at different depth (Lu 2006). Moreover,
liquefaction potentials, with equal importance and at this study site the ground water table is about
effort spent to geotechnical site characteristics and 3 m below the ground surface (Lu 2006). Figure 2
earthquake geology. Both analyses are independ- shows a typical boring log at the site. Given such
ently carried out with existing approaches, and geotechnical information, the factor of safety of
the results are then combined to attain to the ulti- soil liquefaction was found greater than 1.0 at most
mate goal: Whether or not ground improvement depths down to 20 m below the ground.

201

ISGSR2013.indb 201 10/18/2013 9:38:58 AM

 where z is the depth below the ground, and A
is a random variable related to the Factor Of
Safety (FOS) of soil liquefaction at a given
depth:

⎧ 1 − FOS FOS < 1

A= ⎨ (2)
⎩ 0 FOS ≥ 1

In addition to the algorithms, Iwasaki

also offered some suggestion to the messages
behind this index. When L is greater than 15,
the site is considered “Very High Potential”
in liquefaction, followed by “Relatively High
Potential” and “Relatively Low Potential” as L is
in the ranges of 5∼15 and 0∼5, respectively. The
last category is “Very Low Potential” when L is
equal to zero.
It is worth noting that a few methods have
been proposed to calculate the FOS for soil liq-
uefaction. In short, it is based on the proportion
of Cyclic Resistance Ratio (CSR) to Cyclic Stress
Ratio (CRR). As for CRR computation, we fol-
Figure 1. Spatial distribution of the seismicity around lowed the framework of Seed & Idriss (1971); as
Taiwan since 1900 and the location of study site (Cheng for CSR, we adopted three common methods as
et al. 2007, Wang et al. 2011, 2012a, 2012b). to account for the imperfect models or the so-
called epistemic uncertainty (Ang & Tang 2007),
namely the NCEER approach (National Center
for Earthquake Engineering Research), JRA
approach (Japan Road Association), and T & Y
approach (Tokimastu and Yoshimi). Since the
three approaches have been generally accepted
in the community of earthquake engineering,
and the underlying scope of this study is not
to modify them to develop a new method, the
three analyses are not elaborated in this paper,
but readers can refer to the original papers or
technical guideline for methodological details
(Tokimatsu & Yoshimi 1983, Japan Road Asso-
ciation 1996, Youd & Idriss 1997).

2.3 Liquefaction potential for the site

With the existing liquefaction analyses, the rela-
Figure 2. Geotechnical site characteristics and the fac- tionship between liquefaction potential index
tors of safety of soil liquefaction at different depths given
(L) and Peak Ground Acceleration (PGA) at
a deterministic PGA value equal to 0.2 g.
the study site is shown in Figure 3. Accordingly,
as PGA is greater than 0.31 g in an earthquake
event, the site is “Very High Potential” (i.e., L ≥
2.2 Empirical method for liquefaction triggering 15) to liquefy.
Given the geotechnical site condition, the liquefac- However, it should be noted again that this
tion potential index (here denoted as L) proposed information alone should not be enough to
by Iwasaki (1986) and Iwasaki et al. (1982) can be answer the question if the soil liquefaction will
calculated as follows: occur in a given time span. This question should
only be properly managed until the earthquake
20 potential or the probability of PGA > 0.31 g
L ∫ A( − .5 z )ddz (1) within a finite time interval of interest also
z= 0 becomes available.

202

ISGSR2013.indb 202 10/18/2013 9:38:58 AM

 of motion attenuation. Moreover, v is the annual rate
of earthquakes; NS is the number of seismic sources;
NM and ND are the number of data bins in magnitude
and distance probability functions, respectively.
After the annual rate of seismic hazard (i.e., Y >
y*) calculated with Equation (3), its occurrence prob-
ability in time can be computed by assuming seis-
mic hazard is a Poissonian random variable, which
makes its occurrence probability given a time span
follow the exponential distribution (Kramer 1996):

Pr( y* λ ,,t)
t) 1 e− λ t (4)

where λ is the mean annual rate of Y > y*, and t is

the time of interest.
Figure 3. The relationship between liquefaction poten-
tial index L and PGA at the study site with three exist- 3.2 The recent PSHA study for Taipei
ing approaches; this relationship is mainly a refection to
geotechnical site characteristics and to a deterministic Wang et al. (2013) recently conducted a PSHA case
scenario in PGA. study for Taipei. Moreover, that specific PSHA
study endeavored to follow the framework of the
so-called “robust” seismic hazard analysis (Klugel
3 PROBABILISITC SEISMIC HAZARD 2008), with the calculation being repeatable and
ANALYSIS the inputs being traceable and verifiable. The con-
cept of analytical “robustness” is somewhat a solu-
3.1 Overview of the method tion and response to the criticism of PSHA being
Probabilistic Seismic Hazard Analysis (PSHA) somehow “egocentric” to those who perform the
developed in the late 1960s (Cornell 1968, McGuire analysis, when the underlying inputs cannot be sci-
1995) has become a common approach in evalu- entifically justified (Krinitzsky 2003).
ating the annual rate of site-specific earthquake As a result, the following PSHA for the study
motions of exceedance, e.g., PGA > 0.31 g. site is basically to follow the key inputs given in the
A number of PSHA case studies, including those most recent PSHA for Taiwan (Wang et al. 2013).
for Taiwan, have been reported in the last few years It is worth noting that although the referred study
(Cheng et al. 2007, Wang et al. 2013). Moreover, focused on a site in North Taiwan, the key inputs,
a recent technical guideline (USNRC 2007) pre- such as seismic source models (shown in Fig. 4),
scribing the use of PSHA to develop site-specific
earthquake-resistant designs for critical structures
should also reflect to its general acceptance in the
community of earthquake engineering.
The essence of PSHA is to account for the
uncertainties of the earthquake’s size, location,
and motion attenuation. Its governing equation
can be expressed as follows (Kramer 1996):

NS NM N D
λ( y*)
*) ∑ i ∑ ∑ Pr ⎡⎣Y y* | m j , d k ⎤⎦
i =1 j =1 k =1

× Pr ⎡⎣M m j ⎤⎦ × Pr [ D dk ] (3)

where λ ( y*) is the annual rate of ground motion

Y > y* (Y can be PGA or a specific spectral accelera-
tion), and the three probability terms are the under-
lying components of this analysis, calculating the
respective probability in earthquake size (M), source-
to-site distance (D), and the probability of motion Figure 4. The up-to-date seismic source model for
of exceedance (i.e., Y > y*) owing to the uncertainty Taiwan (after Cheng et al. 2007, Wang et al. 2012, 2013).

203

ISGSR2013.indb 203 10/18/2013 9:38:59 AM

 model parameters (summarized in Table 1), should
be generic for Taiwan, as they were calibrated with
nationwide earthquake data around the region.

3.3 Seismic hazard recurrence probability

With the annual rate of seismic hazard
(see Fig. 5), Figure 6 shows the hazard’s occur-
rence probability in four time spans through Eq. 4.

Table 1. Summary of recurrence parameters and maxi-

mum magnitudes of each source zone (Cheng et al. 2007).

Maximum
Source zone a-value b-value magnitude
Figure 6. PGA exceedance probabilities within four
A 2.670 0.828 6.5 time spans at the study site, given the annual rate of seis-
B 3.790 1.074 6.5 mic hazard calculated from PSHA (Fig. 5).
C 3.579 1.014 7.1
D 4.802 1.334 7.3
E 4.420 1.087 7.3 Given PGA > 0.31 g with an annual rate of 0.0022
F 4.720 1.117 7.3 through PSHA, the occurrence probability for
G 3.580 0.876 6.5 such an event is about 20% as far as a time span
H 3.071 0.756 7.3 of 100 years is concerned. On the other hand, the
I 4.450 1.137 6.5 occurrence probability is reduced to 2% when the
J 3.210 0.979 6.5 time of interest is decreased to 10 years.
K 4.490 1.101 6.5
L 3.510 0.918 7.3
M 3.307 0.803 6.5 4 LIQUEFACTION POTENTIAL
N 5.345 1.134 8.0 ASSESSED OVER TIME
O 5.440 1.107 8.3
P 4.920 1.058 7.8 Accordingly, at this study site we found that L = 15
Q 4.170 0.898 7.8 was on the basis of PGA > 0.31 g, and we have
R 3.640 0.913 7.8 also characterized a 20% occurrence probability
S 5.104 1.036 8.0 for PGA > 0.31 g in 100 years, we can therefore
T 4.110 0.900 7.8 combine the two pieces of underlying information
to evaluate time-variant soil liquefaction with a
new index referred to as “soil liquefaction index
per
p~
unit time,” denoted as LT t* . In this instance,
LT 100 yrs is equal to 3.0 per 100 years given a
unique scenario PGA > 0.31 g. Because there are
infinite scenarios other than PGA > 0.31 g that can
cause different levels of soil liquefaction p
potential,
~
Figure 7 shows the relationships between LT t* and
PGA given four periods of time, basically a combi-
nation of Figures 3 and 6. As a result, the integra-
tion of the curve is the resulting soil liquefaction
index per unit time covering any possible scenario
with
~
a small PGA to a large PGA. For example,
LT 100 yrs is about 1.18 at the study site.
The relationship between LT t* and t* at the
site with repetitive calculations from t* = 1 year
to t* = 1,000 years is nearly a perfect polynomial
function obtained with regression analysis on the
Figure 5. PSHA hazard curves at the study site with the 1,000 points:
use of four PGA attenuation relationships to account for ~
such an epistemic uncertainty. T t 0.000003 t*2 + 0.0112 t* + 0.0887 (5)

204

ISGSR2013.indb 204 10/18/2013 9:39:01 AM

 assessment, to decision-making whether or not
ground improvement is needed at the site. This
framework was demonstrated with a case study for
a site in South Taiwan. The result shows that the
liquefaction potential evaluated with existing meth-
ods is rather high based on geotechnical character-
istics at the study site. However, the liquefaction
potential is significantly reduced given a normal
engineering time span (e.g., 100 years), because of
the low occurrence probability of large-size earth-
quakes within such a “short” geological time. This
relationship between highly liquefiable soil and
low liquefaction risk is comprehensible: No matter
how liquefiable the soil would be, soil liquefaction
at the site will not be induced due to very low prob-
ability of large magnitude earthquakes.
Figure 7. The relationship between liquefaction poten-
tial index per unit time and PGA at the study site.
REFERENCES

Ang, A. & Tang, W. 2007. Probability Concepts in

Engineering: Emphasis on Applications to Civil and
With this time-variant relationship based on Environmental Engineering. John Wiley & Sons, New
geotechnical site characteristics and regional earth- Jersey.
quake geology, the liquefaction potential within a Cheng, C.T., Chiou, S.J., Lee, C.T., Tsai, Y.B. 2007. Study
normal engineering time span is “Relatively Low on probabilistic seismic hazard maps of Taiwan after
Potential” at the site. As the time span of interest Chi-Chi earthquake. Journal of GeoEngineering 2:19–28.
is increased to around 500 years, the site becomes Chu, D.B., Stewart, J.P., Lee, T.S., Tsai, J.S., Lin, P.S.,
“Relatively High Potential” as L > 5, according to Chu R.B., Hsu, S.C., Yu, M.S., Wang, C.H. 2004.
the framework of Iwasaki. Documentation of soil conditions at liquefaction and
non-liquefactions site from 1999 Chi-Chi (Taiwan)
earthquake. Soil Dynamics and Earthquake Engineer-
ing 24:647–657.
5 DISCUSSIONS Cornell, C.A. 1968. Engineering Seismic Risk Analysis.
Bulletin of the Seismological Society of America 58:
5.1 Rare events and low liquefaction potential 1583–1606.
Iwasaki, T. 1986. Liquefaction and flow during earth-
As seeing the result, we were also somehow sur- quakes. Geotechnique 43:351–415.
prised learning that the study site is subject to low Iwasaki, T., Tokida, K., Tatsuoka, F., Watanabe, S.,
liquefaction potential given a normal engineering Yasuda, S., Sato, H. 1982. Microzonation for soil liq-
time span, because the soil at the site seems rather uefaction potential using simplified methods. Proc. 3rd
liquefiable. However, the result is comprehensible International Conference on Microzonation, Seattle,
given the fact that large-size earthquakes are the 3:1319–1330.
underlying driving force to soil liquefaction, which Japan Road Association, 1996. Guidelines and Comments
are by all means rare, say, once in hundreds of years to Design of Roads and Bridges. Volume V: Earth-
quake Resistant Design, (in Japanese).
or longer. In other words, the low occurrence prob-
Juang, C.H., Yuan, H., Lee, D.H., Ku, C.S. 2002. Assess-
ability of “liquefaction earthquakes” is the under- ing CPT-based methods for liquefaction evaluation
lying cause to low liquefaction potential within an with emphasis on the cases from the Chi-Chi, Taiwan
engineering time span. In other words, no matter earthquake. Soil Dynamics and Earthquake Engineer-
how liquefiable the soil would be, the liquefaction ing 22:241–258.
probability is still nominal within a “short” period Klugel, J.U. 2008. Seismic hazard analysis—Quo vadis?
of time (in geological scale), especially when the Earth-Science Reviews 88:1–32.
site is located in a geologically stable region, such Kramer, S.L. 1996. Geotechnical Earthquake Engineering.
as Hong Kong. Prentice Hall Inc., New Jersey, pp. 117–133.
Krinitzsky, E.L. 2003. How to combine deterministic
and probabilistic methods for assessing earthquake
hazards. Engineering Geology 70:157–163.
6 CONCLUSIONS Lee, Y.F., Chi, Y.Y., Juang, C.H., Lee, D.H. 2010. Annual
probability and return period of soil liquefaction in
This study presents a new integrated approach Yuanlin, Taiwan attributed to Chelungpu Fault and
from soil liquefaction evaluation, to seismic hazard Changhua Fault. Engineering Geology 114:343–353.

205

ISGSR2013.indb 205 10/18/2013 9:39:04 AM

 Lin, P.S., Lee, C.T., Cheng, C.T., Sung, C.H. 2011. Wang, J.P., Huang, D. 2012. Comment on “Annual proba-
Response spectral attenuation relations for shallow bility and return period of soil liquefaction in Yuanlin,
crustal earthquakes in Taiwan. Engineering Geology Taiwan attributed to Chelungpu Fault and Changhua
121:150–164. Fault” by Lee et al., in Engineering Geology, 114: 343–
Lu, C.W. 2006. Report for hazard prevention and mitiga- 353. Engineering Geology 149–150:93–96.
tion project for Kaohsiung city. Sub- project: Liquefac- Wang, J.P., Huang, D., Chang, S.C., Wu, Y.M. 2012b. New
tion hazard map. evidence and perspective to the Poisson process and
McGuire, R.K. 1995. Probabilistic seismic hazard analy- earthquake temporal distribution from 55,000 events
sis and design earthquakes: closing the loop. Bulletin of around Taiwan since 1900. Natural Hazards Review-
the Seismological Society of America 85:1275–1284. ASCE, doi: 10.1061/(ASCE)NH.1527-6996.0000110.
Seed, H.B., Idriss, I.M. 1971. Simplified procedure Wang, J.P., Huang, D., Cheng, C.T., Shao, K.S.,
for evaluating soil liquefaction potential. Jour- Wu, Y.C., Chang, C.W. 2013. Seismic hazard analysis
nal of the Soil Mechanics and Foundations Division for Taipei City including deaggregation, design
97:1249–1273. spectra, and time history with Excel applications.
Tokimatsu, K., Yoshimi, Y. 1983. Empirical correlation- Computers & Geosciences 52: 146–154.
ship of soil liquefaction based on NSPT value and Wang, J.P., Xu, Y. 2012. The site selection of the Long-
fines content. Soil Foundation 23:56–74. man nuclear power plant in Taiwan. International
U.S. NRC, 2007. A performance-based approach to define Journal of Architecture, Engineering and Construction
the site-specific earthquake ground motion. United 36:251–256.
States Nuclear Regulatory Commission, Washington. Youd, T.L., Idriss, I.M. 1997. NCEER Workshop on
Wang, J.P., Brant, L., Wu, Y.M., Taheri, H. 2012a. Evaluation of Liquefaction Resistance of Soils, Nat.
Probability-based PGA estimations using the double- Ctr. Earthq. Eng. Res., State Univ. of New York at
lognormal distribution: including site-specific seismic Buffalo.
hazard analysis for four sites in Taiwan. Soil Dynamics
and Earthquake Engineering 42:177–183.
Wang, J.P., Chan, C.H., Wu, Y.M. 2011. The distribution
of annual maximum earthquake magnitude around
Taiwan and its application in the estimation of cata-
strophic earthquake recurrence probability. Natural
Hazards 59:553–570.

206

ISGSR2013.indb 206 10/18/2013 9:39:04 AM

 Geotechnical Safety and Risk IV – Zhang et al. (eds)
© 2014 Taylor & Francis Group, London, ISBN 978-1-138-00163-3

Landslide considerations for low cost remedial works

for the Karakorum Highway, Pakistan

A.D. Mackay
Nishimatsu Construction Company Ltd., Hong Kong, China
Formerly SMEC Asia Ltd., Hong Kong, China

ABSTRACT: The Pakistan section of the Karokorum Highway runs from Attobad in the south, crossing
Pakistan’s north-western frontier mountainous region to the Khunjerab Mountain Pass, in the north.
This section of the highway is over 800 long and rises to a maximum elevation of about 4,700 m above
sea level, making it the world’s highest paved international highway. Along the majority its length the
alignment follows the route of the Indus River, situated between the Himalaya and Karakorum mountain
ranges. During 8 October 2005 a severe earthquake struck the region triggering landslides impacting the
highway and destabilizing the natural terrain and cut slopes. This paper provides an overview of the insta-
bility of the region with particular attention specific inspections along the highway, carried out prior to
2010 impacted by the 2005 earthquake event. Details of the landslide assessment approach, particularly
detailed site reconnaissance survey, are summarised for specific locations along the highway running
through the Jhelan Valley. The findings of the landslide characterization and stabilization measures based
on the landslide assessment are presented. The proposed stabilization measures included river erosion
protection, debris removal, slope cut back and slope support.

1 INTRODUCTION

The Pakistan Karakorum Highway provides a

vital connection between Pakistan’s main cities
and the People’s Republic of China (PRC). It
rises to about 4,700 m above sea level (asl) at its
highest level and once connected Pakistan with
the ancient Mongolian capital city of Karakorum
during the 13th century. Due to the tectonic activ-
ity in the region, the effects of undercutting by
the main river systems and the climate and steep
terrain, the area is prone to severe instability. In
particular a severe earthquake occurred on the
8 October 2005, with an epicenter located near
Muzaraffarabad, devastating large areas of the
Pakistan North West (NW) frontier, severely
impacting the highway. This paper provides an
overview of specific landslide locations along
the highway impacted by the earthquake events,
and outlines approaches to landslide evaluation,
including site reconnaissance, involving geologi-
cal mapping.
Figure 1. General location, Pakistan North West
Frontier (Google Earth, 2011).
2 SETTING
The Highway runs to a maximum altitude of
2.1 Geographical features impacting instability
4,700 m asl; commencing from Abbotobad to the
The Karakorum Highway runs through the NW south, running across the mountainous region to
Frontier of Pakistan. Refer to Figures 1 and 2 for the Khunjerab Mountain Pass in the north. Refer
the location. to Figure 3 for the alignment.

207

ISGSR2013.indb 207 10/18/2013 9:39:04 AM

 (Muslehuddin et al., 2005). Refer to Figure 3 for
the terrain relief in Pakistan.
The combination of increased annual rain-
fall, river erosion along the foot of the mountain
ranges, steep terrain and intense tectonic activ-
ity in the region exacerbates instability, particu-
larly sensitive locations along the Karakorum
highway.

2.2 Earthquake, 2005

Pakistan has regions of intense tectonic activ-
ity, generally coinciding with young mountain
formation. Tectonic activity is particularly
intense in the NW Frontier, as presented in
Figures 3 and 4.
On October 8, 2005, an earthquake of magni-
tude of 7.6 occurred with the epicenter located in
the NW frontier of Pakistan, about 17 km north
of Muzaffarbad and 100 km north-east of Islama-
bad. This earthquake coincided towards the end
of the Pakistan monsoon in the region. About
100,000 fatalities occurred and 3 million persons
were left homeless. The effects significantly altered
the character of the mountain slopes and drainage
basins. The earthquake intensity in the vicinity of
the Karakorum Highway and main zones of settle-
ment is shown in Figure 5.
Figure 2. Karakorum Highway alignment.
The tectonic intensity in this region is generated
from the incremental northern movement of the
Indian Plate, about 40 mm/year, into the Eurasian
Plate. This movement generates uplifts in excess
of 7 mm/year (Coward et al., 1985). As presented

Figure 3. Pakistan relief and 2005 earthquake epicentre.

Due to the elevated and abrupt relief in northern

Pakistan the annual rainfall exceeds 700 mm during
the monsoon, originating from the Arabian Sea.
During July to September this rainfall increases to
1000 mm annually towards the highest elevations Figure 4. Pakistan earthquake intensity.

208

ISGSR2013.indb 208 10/18/2013 9:39:05 AM

 Figure 5. Pakistan earthquake intensity and epicentre.

Figure 7. Plate collisions, NW Pakistan.

ram Block uplifts, has occurred north of this

zone and the faults, valleys, drainage catchments
and major rivers systems associated with this
uplift trend in similar alignment to the syntaxis.
Further north the Karakorum mountain range
bounds the Himalayas along the Khyber Pass,
Afghanistan. The “Hazara-Kashmir-Syntaxis”
is estimated to be the Miocene geological age
group, ranging from about 5 to 23 Ma (Coward
et al., 1985).
The uplifted sedimentary deposits in the
region date from the middle to late Tertiary
age groups, 2 to 50 Ma. This coincides with
the age of the Himalayan orogeny, which has
ultimately deformed all the rocks in the region.
Recent deposits include alluvium and col-
luvium formed from debris flows, talluvium,
debris fans, glacial deposits and river terrace
depositions.

3 GROUND CONDITIONS

3.1 Site investigation approaches

Figure 6. Continental plate locations. A Site Investigation (SI) was initiated to assess the
landslides impact to the highway. This included
a literature search and site inspection, involving
in Figure 6 complications in the tectonic activity satellite imagery, geomorphological and geologi-
occur due to the contortion in the collision zone cal mapping. A ground investigation, comprising
boundary, “syntaxis”, as the plates collide (Farah boreholes with sampling and in-situ and labora-
et al., 1984). tory testing was recommended based on the find-
The collision zone is referred to as the “Haz- ings. The stability affecting the roads was generally
ara-Kashmir-Syntaxis” and is associated with assessed based on the interpretation from Varnes,
intense folding. The Karakorum mountain range, 1978 and Hungr et al, 2001. Refer to Table 1 for a
comprising the Kohistan, Ladahk and Karoko- summary.

209

ISGSR2013.indb 209 10/18/2013 9:39:07 AM

 Table 1. Landslide types.

Type of material Type of material Type of material

Movement type (bedrock) (coarse soils) (fine soils)

Falls Rock fall Debris fall Earth fall

Topples Rock topple Debris topple Earth topple
Slides (rotational & Rock slide Debris slide Earth slide
translational)
Lateral spreads Debris spreads Earth spreads Debris spreads
Flows Rock flow Debris flow Earth flow
(deep creep) (soil creep) (soil creep)

Complex—Combination of two or more principal types of movement.

Figure 8. Inspection locations, Highway, S2, Muzaf-

Plate 2. Rock slide.
farabad—Chakoti.

Plate 1. Debris slide. Plate 3. Rock fall.

Examples of the approach to the site reconnais-

3.2 Findings
sance along the highway are provided from the
exposed geological formations and landslide types, The National Highway, S2, runs for about 50 km,
observed along portions of the National Highway, from Muzaffarabad, at Chainage (Ch.) 0, to
S2, Muzaffarabad to Chakoti. The examples are Chakoti, at Ch. 50, alongside the Jhelum River,
summarized in the Findings. Jhelum Valley (Fig. 8). Examples of the landslide

210

ISGSR2013.indb 210 10/18/2013 9:39:09 AM

 type and geological conditions are summarized steepened, overhanging portions of the slope with
below. increased instability. The original rock mass had
Debris Slide (Ch. 45.8)—Comprises recent experienced fracturing and disintegration from
deposition of boulders, cobbles, river terraces the earthquake effects increasing the opportuni-
derived from a shale rock mass. The underlying ties for rock block release. The extent of the rock
shale bedding dips vertically and is aligned paral- fall during the 2005 earthquake partially blocked
lel to the natural dip gradient at this location. The the river flow at the base of the valley below.
debris slide was triggered by the 2005 earthquake Refer to Plate 3.
in combination with river undercutting the slope Rock slide (Ch. 53.3)—The rock slides occurred
toe. Refer to Plate 1. along competent zones of inter-bedded shale and
Rock slide (Ch. 47.1)—Located near the Chinari sandstone. The bedding dipped sub-vertically and
village, comprised a mixture of debris flow and was generally orientated perpendicular to the high-
boulders released as wedges from the original rock way and river alignment. The rock mass was frac-
mass. The wedges were formed from within the S tured and disintegrated, possibly resulting from
shaped folds in the rock, dipping at a maximum the tectonic activity, which allowed release of the
angle of about 70 degrees with axes orientated rock blocks. The instability was exacerbated by
unfavourably out of the natural slope. The S folds river erosion at the slope toe and overlying glacial
were formed during the intense folding associated deposits and sub-rounded cobbles and boulders
with the uplift. Refer to Plate 2. were present. Refer to Plate 4.
Rock fall (Ch. 48.2)—Rock blocks were
released from persistent discontinuities, extend-
ing about 50 m vertically above the highway. The 4 INTERPRETATION
main zones of instability were derived from over-
To allow a rapid urgent assessment of the stabiliza-
tion measures required support was assessed from
the landslide type and scale, for both soil and rock
type failures, as summarized in Tables 2 and 3.

5 MITIGATION MEASURES

Due to the limited available options to place

the highway re-alignment was not considered.
Stabilization measures for the existing highway
alignment were therefore assessed based on a
combination of river erosion protection, suit-
able surface and sub-surface drainage relief and
slope stabilization measures, typically including
slope debris removal and cut back, rock bolting,
retaining structures, catchments or a combination
of available stabilization options. Generic solu-
tions are presented in Figures 9 to 11, including
Plate 4. Rock slide.

Table 2. Proposed rock slope cut design, exceeding 7–10 m vertical height.

Slope gradient Support (S./Crete; Benches

Rock type (H:V) wiremesh; rockbolts) Drainage (H/W m)

Medium jointed rock slightly 1:3 Locally—30 mm; Locally—50 mm dia. 12/2.6–3
weathered as required 6–9 m long slotted
PVC
Intensively jointed rock, 1:2 50 mm; N/A; 25 mm 4 m c/c, 50 mm dia., 12/3
unfavourable dia., 4–6 m length 6–9 m long slotted
discontinuities PVC
Soft rock, weak, highly 1:1 70 mm; N/A; 25 mm 4 m c/c, 50 mm dia., 7/3
fractured and weathered, dia., 6 m length 9–12 m long slotted
shale, partly disintegrated PVC

N/A—Not applicable.

211

ISGSR2013.indb 211 10/18/2013 9:39:11 AM

 Table 3. Proposed soil slope cut design, exceeding 5–7 m vertical height.

Slope gradient Support (S./Crete; Benches

Soil type (H:V) wiremesh; soil nails) Drainage (H/W m)

Conglomeratic debris, 2:1 30–50; locally—less Locally—50 mm dia. 12/3

river terrace, cohesion; 25 mm 6 m long slotted PVC
lodgement till dia., 4–6 m.
Cohesive soils, colluvium 1:1–1:1.5 50–70; N/A; Locally—50 mm dia. 7/3
and talluvium 25 mm dia., 6 m 6 m long slotted PVC
Clayey-silty soil, sensitive 2:1–1.5:1 70–100; N/A; Locally—50 mm dia. 5–7/3–5
to sliding, water 25 mm dia., 6 m 6 m long slotted PVC
saturated, landslide

N/A—Not applicable.

Figure 11. Slope cut back.

Figure 9. Slope stabilization measures.

Plate 5. River erosion protection.

6 CONCLUSIONS
Figure 10. Slope stabilization measures with slope cut
back. The detailed inspections assessing the ground
conditions along specific landslide locations along
the alignment, considering influences affecting the
rock dowel and/or soil nail and retaining walls stability, such as the tectonic effects triggering
with sub-surface drainage measures, cut back in landslides, rainfall, river erosion and the exisitngf
combination with slope support measures. River g steep natural gradients, were considered applica-
erosion protection measures are presented in ble in assessing preliminary landslide mitigation
Plate 5. proposals.

212

ISGSR2013.indb 212 10/18/2013 9:39:11 AM

 ACKNOWLEDGEMENTS Google Earth (2011). Searchable map/satellite view of
Pakistan.
The author wishes to thank Ronald Chan in Hungr, O., Evans, S.G., Bovis, M.J. & Hutchinson, J.N.
assistance in preparing this paper. The opinions (2001). Review of the classification of landslides of the
flow type, Environmental & Engineering. Geoscience
expressed in this paper are those solely of the 3(3): 221–238.
author and not of any other party. Muslehuddin M., Mir H. & Faisal N. (2005). Sindh sum-
mer (June–September) Monsoon Rainfall Prediction.
Pakistan Journal of Meteorology 2(4): 91.
REFERENCES Varnes, D.J. 1978. Slope movement types and processes.
In: Special Report 176: Landslides: Analysis and
Coward, M.P. & Butler, R.W.H. (1985). Thrust tecton- Control (Eds: Schuster, R.L. & Krizek, R.J.).
ics and the deep structure of the Pakistan Himalayas: Transportation and Road Research Board, National
Geology, v. 13, p. 417–420. Academy of Science, Washington D.C., 11–33.
Farah, A., Lawrence, R.D., & DeJong, K.A. (1984). An
overview of the tectonics of Pakistan: Marine Geol-
ogy and Oceanography of Arabian Sea and Coast,
p. 161–176.

213

ISGSR2013.indb 213 10/18/2013 9:39:13 AM

This page intentionally left blank
 Geotechnical Safety and Risk IV – Zhang et al. (eds)
© 2014 Taylor & Francis Group, London, ISBN 978-1-138-00163-3

Landslide mitigation considerations for low cost highway remedial

works, East Timor

A.D. Mackay
Nishimatsu Construction Company Ltd., Hong Kong, China
Formerly SMEC Asia Ltd., Hong Kong, China

ABSTRACT: East Timor has an extensive road network often prone to land-sliding; typically this
causes disruption to its transportation system which is the island’s life line. The instability impacting
the roads is mainly a combination of steep terrain; high annual precipitation, typical of a sub-tropical
climates, and complex ground conditions affected by the tectonic uplift and deep weathering zones com-
mon throughout the island. Other influences affecting the susceptibility include erosion sensitivity and
roads being placed on poorly prepared and un-compacted ground. Remedial measures against landslides
have considered optimization of road alignment corridors; the avoidance of areas of concern, were pos-
sible, and a reduction in the reliance of traditional slope stabilization measures. Traditional stabiliza-
tion measures have included stone pitching and gabion boxes, which have proven insufficient to support
modern transportation frequency and load impacts. Up to date landslide stability evaluation has allowed
assessment of landslide susceptibility for different ground conditions. Recommendations for suitable,
regular maintenance with particular attention to effective road drainage networks have been based on
this. Notwithstanding effective resource, with suitable training and allocation, is required to carry out
this maintenance, and allocation using suitable construction materials, sourced from suitable locations as
needed. This paper provides an overview of the road network throughout East Timor and factors affect-
ing the landslide susceptibility, such as precipitation, terrain and the geological conditions. Details of
some of the landslides, representing different types of susceptibility, impacting the roads are presented,
with suggested slope stability mitigation measures for each representative landslide.

1 INTRODUCTION 2 SETTING

East Timor’s road network has an extensive and 2.1 Geographical features impacting instability
large proportion of paved, surfaced roads com-
Timor Island is located south of the Banda Arc
pared with many other countries with similar capi-
Archipelago, which extends eastwards from Bali
tal expenditure. Notwithstanding there has been
Island. East Timor occupies the north eastern
a recent increase in the transportation frequency
portion of Timor Island. Refer to Figure 1 for the
and loading, which has impacted and/or been
location.
effected by the instability occurring along the road
network. The instability is typically a result of a
combination of poor road base preparation, com-
plex geology, including tectonic uplifting and deep
weathering zones, and high precipitation. To miti-
gate this instability initiatives have been taken to
carry out remedial measures and implement more
efficient road maintenance. These measures have
included more robust cut and fill support along
the road alignments, with particular attention to
effective surface and sub-surface drainage systems,
optimizing road corridors and avoiding zones sus-
ceptible to landslides were possible. This paper
provides an overview of the road network, geog-
raphy, climate, terrain and geology impacting the
landslide susceptibility. Details of proposed stabi-
lization initiatives for different and representative
landslides encountered are presented. Figure 1. East Timor location (Google Earth, 2011).

215

ISGSR2013.indb 215 10/18/2013 9:39:13 AM

 East Timor has mountainous terrain, rising to
an altitude of 2,963 metres above sea level (asl)
at Mount Tatamaiau, see Figure 2 for the general
locations and Figure 3 for the relief.
The precipitation has been measured from rain
gauges located throughout East Timor, with the
majority located on the higher altitudes towards
the west. As shown in Figure 4 the most intense
annual precipitation rates exceed 2500 mm per
annum, corresponding to the highest elevations.
The effect of the intense rainfall on roads with
uncontrolled drainage systems is presented in Plate 1. Uncontrolled surface drainage impacting roads.
Plate 1.

Figure 2. East Timor (Google Earth, 2011).

Figure 5. Major road network, East Timor.

The high annual precipitation, changeable relief

and altitude of many of the island locations con-
tribute to the landslide susceptibility.

2.2 Road network

East Timor has an extensive sealed road network,
which often traverses the island’s steep, unstable
terrain exposed to heavy rainfall. In the past the
road network catered for light transport; more
recently the transport loading and frequency has
Figure 3. East Timor relief. increased significantly, further impacting stability.
Refer to Figure 5 for the main road network.
The road maintenance typically relied on labour
intensive remedial measures constructed using
traditional slope stabilization measures, such as
masonry and gabion retaining walls. With the
increased demand on the road network efficient
construction of robust slope stabilization meas-
ures are necessary. Problems with slope instability
impacting the roads in east Timor have typically
been associated with:
Natural conditions—terrain steepness, requiring
side long cut and fill placement for road forma-
tion; weak, unstable ground conditions; intense
Figure 4. East Timor precipitation. rainfall; surface water erosion and scour.

216

ISGSR2013.indb 216 10/18/2013 9:39:14 AM

 Construction—difficulties sourcing, handling and
processing high quality construction material; over
reliance on masonry/gabion retaining walls often
founded on unstable ground; insufficient resource
for emergency landslide mitigation; lack of expe-
rienced staff (designers, supervisors and contrac-
tors) and lack of suitable construction plant.

2.3 Geological overview

Timor Island is located between two subduction Figure 8. Sectional geology of East Timor (Norvick,
zones, along its north-west edge within the Banda 1979).
Sea forming the boundary between the Australasian
Continental and the Eurasian Oceanic Plates, Table 1. Summary of the solid geology of East Timor
and along to the south east along the East Timor (Geological research, 1994 to 1995).
Trench. East Timor is located at the north west-
ern boundary of the Australian Plate. The island’s Geological
centre has experienced pronounced geological up- Formation age Facies/lithology (remarks)
thrust, with the oldest rock (Lolotoi complex of
pre-Permian age) located at its core (Thompson, Barique Oligocene Epineritic/basaltic tuffs, lavas
2011). Refer to Figures 6 to 8 for the geological and limestone fragments
plans and section. Dartollu Eocene Shallow marine/limestone,
calcarenite (brown
The lithologies and environment of for- colouration)
mation (facies), categorized according to Borolalo Cretaceous Deep marine/massive
geological age; Oligocene-23-34Ma; Eocene-34-56 limestone, chert and
Ma; Cretaceous-66-145 Ma; Jurassic-145-200 Ma; calcareous shale
Waibua Cretaceous Deep marine/shale, chert,
calcarinite and calcirudite
Wailuli Jurrasic Deep marine/shale, silt-
stone and limestone and
sandstone—lowest levels
(tubidites)
Aitutu Triassic Epineritic/calcilutite, marl,
calcareous shale and
calcarenites
Cribas Permian Shallow marine/shale,
claystone, siltstone,
sandstone and calcarenites
Maubissa Permian Shallow marine/reefal
limestone (reddish
colouration)
Figure 6. Banda arc plate boundaries (Thompson, Aileu Permian Shallow marine/phylite,
2011). schist, amphibolite, slate,
occasional volcanics
(effected by metamorphosis
and deformation)
Lolitoi Pre Phylite, schist and gneisses
permian (weak to medium effects
of metamorphosis). Has
experienced the relatively
intense metamorphism

Triassic-200-245 Ma and Permian-245-300 Ma;

are summarized in Table 1.
Due to the compressive effects and island uplift
between the subduction zones the island has expe-
rienced intense folding and shearing. This has
influenced the competency and susceptibility to
Figure 7. Geology of East Timor (Charlton, 2002). erosion and weathering of each rock formation.

217

ISGSR2013.indb 217 10/18/2013 9:39:16 AM

 3 GROUND CONDITIONS The Wailuli formation tended to weather to a
colluvium in steeper gradients. Due its impermeable
3.1 Site investigation approaches characteristics, comprising a large proportion of
fine grained weathered constituents uncontrolled
A Site Investigation (SI) was initiated to assess the
surface drainage was observed (see Plate 2).
landslide susceptibility impacting the road network
Due to the recent, un-consolidated deposition
generally following the principles set out in TRL,
of the Bobonaro complex the formation tended to
2005. Key considerations included the engineering
be susceptible to instability, as shown by the effects
properties of the classified rocks and soils, natural
of uncontrolled drainage in Plate 3.
instability, rainfall, hydrology, topography and seis-
The Viqueque formation has been deposited
mic activity. Engineering considerations included
recently. As a result it is highly susceptible to ero-
the presence and condition of existing engineer-
sion and instability as shown by the undercutting
ing structures, such as the retaining walls; natural
effects of river erosion in Plate 4.
and man-made hazards associated with the ground
The Ailue formation is a relatively old and
conditions; the road sub-grade conditions and con-
therefore typically stable. However Plate 5 shows a
struction materials. Particular attention was given
location with a reduced strength from weathering
to the behaviour of tropical weathering effects. The
effects with colluvium covering the uppermost sur-
SI relied on detailed inspections of landslide and
face. The failure was triggered by discharge from
rock exposure geological formations.
a culvert.
Plate 6 presents a mudflow within the Cablaci
3.2 Findings
limestone. The failure is located at the edge of a
The more recent geological formations, deposited limestone Plateau exacerbated by erosion from a
during the Pliocence (2.6 to 5.3 Ma) and Miocene stream course, transporting the failed material
(5.3 to 23 Ma) geological age series were inspected away from the failure toe. As the failure is located
during the SI as summarized in Table 2. The inspec- away from roads it does not impose a hazard.
tion locations referenced are presented in Figure 9.

Table 2. Summary of the recent geology of East Timor

(Geological research, 1994 to 1995).

Geological Facies/lithology
Formation age (remarks)

Viqueque Pliocene Littoral/claystones,

interbedded with
limestone, tuffs,
silts and sands
Bobonaro Miocene Deep marine/chaotic
complex boulders in a clay
matrix Plate 2. Wailuli formation (river flooding across road).
Cablaci Miocene Shallow marine/clastic
limestone Limestone (crystalline
matrix)

Plate 3. Borbanaro complex: Instability above and below

Figure 9. Site inspection locations. road.

218

ISGSR2013.indb 218 10/18/2013 9:39:17 AM

 Table 3. Ground conditions encountered in the SI.

Zone Description

1 Coastal and alluvial plains. Stability generally

not problematic unless embankments or
structures are required (See Plate 7)
2 Older geological formations exhibiting competent
rock standing in steep slopes. These include the
Alieu and This is particularly prone to
instability in more locations with more intense
rainfall and deeper weathering
3 Less competent rock with minor instability
4 Moderate instability
Plate 4. Viqueque formation: gabion wall failure above 5 Weak and unstable geological formations with
river. potentially severe instability. In these locations
retaining wall support is insufficient

Table 4. Ground conditions encountered in the SI.

Zone Description

2 Cribas, Mabisse, Alieu

3 Lolotoi, Wailuli, Waibua, Borolalo, Dartolla
limestone, Ainaro, Bancau limestone,
Surebaco and Aliambata
4 Atitu, Barique, Cablaco, Lariquiti, Dilor
5 Viqueque, Barbanoro formation and complex

Plate 5. Instability beneath road, Aileu Formation.

Plate 7. Zone 1, flat coastal and alluvial plains.

Plate 6. Mudflow, away from road, Cablaci Limestone. The zoning generally corresponded with the
source geological conditions, the weathering inten-
sity, natural slope gradient and precipitation. An
4 INTERPRETATION approximate zoning corresponding with the source
geological formation encountered is summarized
The stability affecting the roads could generally in Table 4, as Zone 1 comprises flat recently depos-
be zoned based on the general scale of ited material, such as alluvium, this has not been
instability associated with ground conditions, included.
were information is available. A generic zoning Generally instability increases with younger,
presenting basic landslide susceptibility, is uplifted geological formations; hence the inclu-
presented in Table 3. sion of the Viqueque, and Barbanoro Formations

219

ISGSR2013.indb 219 10/18/2013 9:39:18 AM

 in Zone 5. More recent deposition in the Miocene
age group also includes the Aliambata Formation
in Zone 3 and the Lariquiti and Dilor formations
in Zone 4.

5 MITIGATION MEASURES

Based on the findings of the instability impacting

the roads more robust stabilization solutions are
required. These can include:
Major realignment—This is required were severe Figure 11. Skin wall.
instability impacts the road. Difficulties often
occur with the land ownership and impact to exist-
ing settlements when realignments are considered Drainage is a requirement for all solutions
Slope stabilization: These can include combina- ensuring surface drainage is taken away from
tions of shotcrete, mesh, rock dowels and bolts, areas susceptible to landsliding and that culverts
catch fences, reinforced back-fill and retaining are functioning effectively. Resource also requires
structures. Considerations should be given for free allocation to ensure all roads are maintained,
drainage, such as suitable permeable backfill and particularly the drainage networks.
weepholes, in the shotcrete and for use of backfill
behind retaining structures. Generic solutions are
presented in Figures 9 to 11. 6 CONCLUSIONS
Do nothing: Due to the prohibitive expense and
technical problems that may occur with some of The detailed inspections assessing the ground
the solutions it may be advantageous to allocate conditions in combination with an assessment of
resources to for emergency maintenance and/or factors influencing the stability, such as construc-
repair when failures occur. tions methods, natural slope gradient and annual
rainfall, were considered successful in assessing
preliminary landslide susceptibility. The use of
these findings to categorize slope stability can be
considered useful in efficient allocating resources
to areas were the need is greatest and stabilization
requires greatest attention.

ACKNOWLEDGEMENTS

The author wishes to thank Ronald Chan in

assistance in preparing this paper. The opinions
expressed in this paper are those solely of the
author and not of any other party.
Figure 9. Retaining wall.

REFERENCES

Charlton, T.R. (2002). The Petroleum Potential of East

Timor. APPEA Proceedings 32, pp 20–38.
Geological Research and Development Centre, Indonesia
(1994–1995). Geological Maps and memoirs of Bacau
Quadrangle, Dili Street, Atambua Sheet, 1:250,000
scale.
Norvick (1979). The tectonic history of the NBanda
Arcs, Eastern Indonesia; a Review. Journal of the
Geological Society, version 136, No. 5, pp 519–526.
Thompson S.J. (2011). The Geology and Soils in Timor
Leste (internet).
Transport Research Laboratory (TRL, 2005). Overseas
Figure 10. Skin wall with soil nails. Road note, A guide to road Appraisal.

220

ISGSR2013.indb 220 10/18/2013 9:39:20 AM

 Geotechnical Safety and Risk IV – Zhang et al. (eds)
© 2014 Taylor & Francis Group, London, ISBN 978-1-138-00163-3

The landslide of Kirf: A chain of governance failures

S. Van Baars & M. Sosson

University of Luxembourg, Luxembourg

S. Jung
Dr. Jung and Lang Engineering, Germany

R. Becker
Holmalux, Luxembourg

ABSTRACT: In January 2011 a landslide occurred in the German village of Kirf, which is between
the town Saarburg and the border of Luxembourg. This slope failure caused a destruction of a federal
main road over 50 meters. Stability calculations were made by a geotechnical consulting company,
commissioned from the Federal Road Administration, in order to design the repair works and also to find
out what really had happened. These calculations showed remarkably high safety factors. Nevertheless
the consulting company blamed the families of the four houses down the slope for the failure, because
they had made gabion walls in the toe of the slope. A research team of the University of Luxembourg
discovered that this hilly area originally was called “On the slide” and has never been stabile according to
their own calculations. So, the real question is why so many apparently uncorrelated mistakes were made
in a single case.

1 LANDSLIDE OF KIRF about a decimeter had been recorded. On January

7 and 8, after a time of snow melt, the slope slid
1.1 Slope failure and the road on this slope collapsed and had to be
closed off. The families of the four houses down
In January 2011 a landslide occurred in the German
the slope were temporarily evacuated.
village of Kirf, which is between the town Saarburg
and the border of Luxembourg. This slope failure
1.2 History
caused a destruction of the federal main road B407
over a length of 50 meters, see Figure 1. The Federal main road B407 was first build as
Cracks had been noticed in this road since “Reichstrasse 327” in 1938–1939 as a strate-
August 2010 and in October 2010 settlements of gic road connection. First in the early sixties the
nearby southern part of the highway A1 was con-
structed and the “Reichstrasse 327” was trans-
formed into the federal main road “Bundesstrasse
407”. It is important to know that during all these
years never a single stability analyses was made of
the road-slope construction of the B407 near Kirf,
but also never a large landslide had occurred until
January 2011.
The village municipality has granted building
permits for four houses at the toe of this slope. The
first house (most North, with house number 16)
was built in 1996. The other three houses were con-
structed between 2006 and 2009. All four houses
have basements which were constructed in tem-
porarily excavations at the toe of the slope. The
families of the four houses ordered a constructing
company in December 2009 to excavate the toe of
Figure 1. Landslide at the federal road B407 near Kirf. the slope and to construct gabion walls, in order to

221

ISGSR2013.indb 221 10/18/2013 9:39:21 AM

 increase the lower part of their gardens, see also the calculations, which are the safety factors (the
Figure 3. The constructing company did neither slope strength divided by the load), are found in
demand a stability calculation of the gabion wall, Table 1.
nor a stability calculation of the total slope. The stability calculations of the consult-
ing company resulted in safety factors between
SF = 1.55 and SF = 1.94 before the construction
2 RESULTS OF THE CONSULTING of the gabion walls and a strong reduction down
COMPANY to safety factors between SF = 1.27 and SF = 1.52
after the construction of the gabion walls. Since
2.1 Introduction the safety factors were still far above SF = 1.00
(point of failure), the soil strength parameters
For the repair works of the slope and road after
were reduced and based on residual strength (after
the landslide, a type of crib wall were used, which
shearing) for additional calculations for a safe
is a combination of concrete beams and soil. In
design of the crib walls.
order to design these works and also in order to
The consulting company blamed the families
find out what really had happened, the Federal
of the four houses down the slope for the failure,
Road Administration (LBM Trier) defined one
because they had made gabion walls in the toe
cross sections at each of the four houses, and com-
of the slope, see Figure 3. Especially the family
missioned a geo-technical consulting company to
of the first house (No. 16) was found responsi-
make a map of the cracks in the slope and to make
ble because they had reduced the safety factor
slope stability calculations for the four cross sec-
relatively the most.
tions, see Figure 2.
Based on these results, the Federal Road
Administration charged in November 2011 the four
2.2 Block slide calculations families to pay the repair works of € 900.000—(LBM
Trier, 2011). The four families feel betrayed because
For these four cross-sections the soil layering was
determined and for each soil layer the soil param-
eters were determined. It was said that the soil
Table 1. Safety Factors (SF) according to consulting
contained too many large stones for a normal soil company.
sampling and testing. The geotechnical consulting
company chose to make block slide failure calcu- Before gabions After gabions
lations, since they were convinced that the slope
slid as a single block over a strong and deeper soil Safety factor SF SF
layer. For each of the four cross sections, calcula-
tions were made. These calculations were based on House No. 16* 1.94 1.39
a diagonal translation of a soil body. The results of House No. 18 1.89 1.52
House No. 20 1.64 1.44
House No. 22 1.55 1.27

*First house from the North.

Figure 2. The road, four houses, four parallel cross

sections and cracks in the ground (under laying photo: Figure 3. Pushed-over gabion wall at toe of slope near
Google Maps). House No. 16.

222

ISGSR2013.indb 222 10/18/2013 9:39:22 AM

 they do feel they did neither anything wrong, nor the biggest reduction is important for failure, but
did they know anything about soil mechanics. the lowest safety factor.
The consulting company assumed also that the
slope slid partly horizontally (parallel to the road
3 RESULTS OF THE UNIVERSITY axis) towards the first house, indicating that the
family of this house is most responsible. There is
3.1 A chain of governance failures however no evidence (measurements or photos)
At the end of 2012, one of the families asked the indicating that such a parallel movement took
University of Luxembourg to have a look at this place, nor is this possible, since there is no energy
case. A research team was formed consisting of source to compensate for the energy loss due to
the four authors of this article. They discovered a the frictional parallel movement, because the only
chain of mistakes in governance in this case. First source here; gravity, acts purely vertical.
of all, the geology, topography and also the veg-
etation show irregularities on the side, which could 3.3 Slip circle calculations
be a first indication of an instable slope. Besides, Remarkably, the unreduced stability calculations of
according to one of the maps this area is called “On the consulting company, mentioned before, proved
the slide” (Consultancy Spoo & Pittner GmbH, the fitness of the slope, in spite of the occurred
2002). The road was built already for many decades landslide. However, these calculations were by
on this slope and the traffic load had risen over the mistake based on “block failure”. Therefore the
years but never a stability calculation had been University of Luxembourg has made new “slip cir-
commissioned. The Federal Road Administration cle” calculations for the four cross-sections, based
also never acquired the slope in order to protect it, on Bishop’s method, both before and after the
nor did they warn or regulate the (new) owners of construction of the gabion walls, see Figure 4.
the slope. Neither did the village municipality when For the selected soil parameters see Table 2. The
they granted a building permit for these houses. consulting company used an unrealistically high
Also none of the architects and builders of the friction angle for the Top Loam of φ = 35 degrees,
four houses required a stability calculation for the which has been reduced by the authors to φ = 25
construction of the building pits for the basements degrees.
of these houses, right next to the toe of the slope
(only in one report we find a recommendation, no
requirement). Also when the four families installed
the gabion walls, nobody required a stability calcu-
lation. This is a clear chain of governance failure.

3.2 Review of consulting company

The review of the report of the consulting company
revealed several major mistakes. Their assumption
that the soil body of the slope slid diagonally as a
block was incorrect. Detailed photos show clearly
that the road and the guardrail moved vertically
and the slope toe horizontally (perpendicular to
the road axis), indicating that there was a rota- Figure 4. Slip circle calculation at cross-section House
tional movement. That is why “slip circle” calcula- No. 20.
tions had to be made.
The fact that their safety factors were still far
above SF = 1.00, should not have immediately lead Table 2. Soil parameters (unit weight, friction angle
to the use of lower soil strength parameters, but and cohesion).
to the search of the real error in their modeling.
Besides, their reduction of the soil strength param- γ φ c
eters certainly has led to a strong over dimension-
Parameters kN/m3 degrees kPa
ing of the repair works, which means that a large
part of the costs of the repair works is on the Fill 19 35 2
responsibility of the consulting company itself. Top loam 19 25* 2
The consulting company concluded that the Grey clay 20 27.5 10
family of the first house was most accountable Red clay 20 22.5 4
because here the largest reduction in safety factor
was found. This reasoning is incorrect, because not *The consulting company used here φ = 35 degrees.

223

ISGSR2013.indb 223 10/18/2013 9:39:23 AM

 Table 3. Safety factors according to the authors. 9. The owners of the four houses all decide to install
large gabions walls in the toe of the slope.
Before gabions After gabions 10. The constructor(s) of the gabion walls of all
four houses does not require stability calcula-
Safety factor SF SF
tions to be made.
House No. 16* 1.55 0.99 These 10 events seem rather uncorrelated. The
House No. 18 1.04 0.93 real question which remains is; why such a chain of
House No. 20 1.05 0.99 apparently uncorrelated mistakes could be made in
House No. 22 0.96 0.89 a single case. The fact therefore that all 10 events
*First house from the North.
were required to take place and also did take place,
means that the chance of occurrence of each of
the events is unfortunately too high. Therefore no
other conclusion is possible than that in this case,
The calculations of the University of and therefore probably also in other cases, the cur-
Luxembourg were made with a high groundwater rent requirements for good governance and risk
table in order to simulate the effect of the melting management of earthworks were not followed by
snow but without traffic load. any of the stakeholders.
The results of the calculations indicate that There are also other problems. At the first stability
the stability factors before the construction of the analysis which was performed by the consulting
gabion walls were, for the profiles of three of the company after the landslide, an unacceptable high
four houses, between SF = 0.96 and SF = 1.05 number of mistakes were made. And what is worse,
and after its construction between SF = 0.89 and they were not discovered by the normal governance
SF = 0.99, see Table 3. system. The fact that soil strength parameters were
strongly reduced in order to find failure in their
calculations was unacceptable.
4 CONCLUSIONS Finally the most unacceptable aspect of this
case is the fact that the four families living at the
The occurrence of the landslide is not a surprise toe of the slope were held fully accountable for the
since the results from the slip circle calculations failure, and this conclusion was reached based on
show safety factors SF < 1. calculations with many errors and the lack of gov-
However, in order to get to the condition of ernance described.
such a failure, the following chain of events and
boundary conditions is needed:
1. There is an unstable slope. REFERENCES
2. A road is built on the slope.
Clayton Environmental Consultancy GmbH, 2006.
3. No stability calculation is made before the con- Geotechnische bericht zum Bauvorhaben Dostert “zum
struction of the road. Altenberg” in Kirf, Ludwigshafen, Germany.
4. Never a stability calculation is commissioned Consultancy Spoo & Pittner GmbH, 2002.
during all the years of maintenance. Bodengutachten “Im Hübel auf der Maaswies” in Kirf,
5. The Federal Road Administration never Mertesdorf, Germany.
acquires the slope in order to protect it. Landesbetrieb Mobilität Rheinland-Pfalz LBM Trier, 2011,
6. Nor do they prescribe anything to the slope Letter from G. Wiesel to lawyer Martini-Mogg-Vogt in
owners, in order to protect the slope and the Koblenz of the families houses no. 16 and 18.
Topographical Maps Page Koblenz 1925 and 1939 and
owners.
Page Trier 1925 and 1940.
7. The village municipality grants building permits Trierischer Volksfreund, Newspaper, http://www.
without any prescription for the slope. volksfreund.de—Archive: 10.01.2011—Hangrutsch
8. The architect and builders of the houses and bei Kirf: Bewohner verlassen Haus.
the excavations of the basements, never require WPW Geoconsult GmbH, 2011, Geotechnischer Bericht
stability calculations to be made. “Schadensanalyse”, Trier, Germany.

224

ISGSR2013.indb 224 10/18/2013 9:39:24 AM

 Geotechnical Safety and Risk IV – Zhang et al. (eds)
© 2014 Taylor & Francis Group, London, ISBN 978-1-138-00163-3

A risk-based active fault classification

J.P. Wang
Department of Civil and Environmental Engineering, Hong Kong University of Science and Technology,
Kowloon, Hong Kong, China

ABSTRACT: Engineering risk is referred to as occurrence probability multiplied with the consequence
when the event occurs. Few engineering decisions are now risk-based, using an alternative with the lowest
risk as the solution to a problem. Under such a risk framework, this study develops a classification system
for active faults. In addition to population at risk, this classification also considers the earthquake mag-
nitudes and earthquake energy release. Applying such a classification to six active faults in Taiwan, we
suggested that the Sanchiao Fault and Changhua Fault should be Category III active faults, the highest
level in this classification system.

1 INTRODUCTION 2 RISK SCORE

Engineering risk is referred to occurrence The risk score of an active fault in this study is
probability multiplied with the consequence when referred to the relative level of risk among the
the event occurs (Wang et al., 2012a). As for risk- faults analyzed. Note that the original score is
based decision making, the decision with the lowest normalized, so that the resulting risk score is
risk will be adopted. Few engineering decisions are dimensionless. In other words, after normalizing
now risk-based, including site selections for criti- the units of the underlying parameters should not
cal structures (Wang et al., 2012b), and earthquake affect the classification.
early warning (Wang et al., 2012a). With the concept of risk, the risk score (RS) of
Understandably, probability is used to quan- an active fault in this study can be first expressed
tify the level of uncertainty or randomness. For as follows:
example, the probability in a random process that
“Point One” occurs while tossing a fair dice is 1/6, RS P (M > m*) CS (1)
which reflects the randomness of this stochastic
process. Since the earthquake is hardly predictable where Pr( *) is the probability of magnitude
given our limited, imperfect understanding (Geller exceedance for M > m* when an earthquake occurs
et al., 1997), probabilistic analyses are gradu- at an active fault; CS denotes the consequence
ally applied to evaluating earthquake potentials when such an event occurs. This expression is the
in terms of probability. For example, a couple underlying governing equation of this study.
of probabilistic studies (Wang et al., 2013a) were
performed and suggested that indeed the earth-
2.1 Earthquake energy release
quake’s temporal occurrence follows the statistical
Poisson model that has been commonly used in Earthquake energy release exponentially increases
earthquake analyses (Wang et al., 2012c, 2013b). with magnitude. For example, an M 7.0 earthquake
Another instance is the analysis evaluating large- can release energy 32 times as large as an M 6.0
earthquake probabilities from studying the pool event (Keller, 1996). Figure 1 shows the energy
of annual maximum earthquakes in Taiwan since release ratio on the basis of M 6.0 earthquakes. It
1900 (Wang et al., 2011). is around a 1,000-time difference in energy release
This paper presents a risk-based classicification between M 6.0 and M 8.0 earthquakes.
system for active faults. In addition to Population
At Risk (PAR), the best-estimate earthquake
2.2 Earthquake energy release and consequence
magnitude and earthquake energy release are
accounted for in the classicification developed. Relatively speaking, the relationship between earth-
The method was then applied to six active faults in quake energy release and the consequence caused by
Taiwan, which is considered one of the earthquake earthquake is more linear than that between earth-
“hot zones” in the world. quake magnitude and the consequence. Therefore,

225

ISGSR2013.indb 225 10/18/2013 9:39:24 AM

 computed by the summation of Pr( ) R6.0 ( ),
with M from m* to a reasonable upper-bound
value (e.g., 9.0). As mentioned, the upper-bound
magnitude has little influence on the resulting
risk score because normalizing will be performed
eventually.

3 RISK SCORE COMPUTATIONS

This study applied such a risk-based classification

to six active faults in Taiwan. Figure 2 shows their
locations. According to Cheng et al. (2007), the
best-estimate maximum earthquake magnitudes
Figure 1. Earthquake energy release ratios on the basis
of M 6.0 events.
are summarized in Table 1, with the largest one
equal to M 7.6 associated with the Changhua
Fault, and the smallest one of M 6.4 associ-
we used energy release ratios as one of the factors ated with the Hsinhua Fault. Also based on that
to estimate the earthquake consequence in this
study. In other words, the consequence caused by
an M 7.0 is considered 32 times as severe as that
caused by an M 6.0 event. Although this estimate
can not be verified on a more scientific basis, it is
more reasonable than considering a linear relation-
ship between magnitude and consequence. That is,
if the consequence increment is equal to 1 from
M 6.0 to M 7.0 earthquakes, the increment is then
equal to 2 considering the magnitude increment
from 6.0 to 8.0.

2.3 Fault location and Population At Risk (PAR)

Understandably, the consequence caused by a large
earthquake is more severe when it occurs in major
cities than in remote areas. In other words, the fault’s
location is related to earthquake consequences.
Considering human life is the most important asset
to be protected, this study employed population at
risk as another factor to characterize the conse-
quence in terms of loss of life.
Therefore, with the earthquake consequence
characterized by Earthquake energy Release Figure 2. Locations of six active faults in Taiwan in this
study.
(ENR) and Population At Risk (PAR), the govern-
ing equation (i.e., Eq. 1) to estimate the risk score
of active faults can be rewritten as follows: Table 1. Summary of maximum probable earthquake
magnitudes induced by the six fault in Taiwan.
RS P (M > m*) ( ENR
E × PAR ) (2)
Faults Magnitude (Mw*) COV** (%)
Using the ratio R6.0 to quantify energy releases
on the basis of M 6.0 events, Eq. 2 becomes: Sanchiao 7.0 2.9
Hsinchu 6.8 2.9
RS P (M > m*) R6.0 (M m*) × PAR (3) Changhua 7.6 2.6
Meishan 6.5 3.1
Hsinhua 6.4 3.1
Therefore, risk score RS is a function of mag- Fenshan 6.7 3.0
nitude m* that appears in the first two terms, also
a function of the fault’s location determining the *Mw: moment magnitude; **COV: coefficient of varia-
level of PAR. Note that R6.0 (M m*) can by tion = standard deviation/mean.

226

ISGSR2013.indb 226 10/18/2013 9:39:25 AM

 study, the best-estimate standard deviation was
reported at 0.2.

3.1 Magnitude exceedance probability

The first step to calculate the risk score with Eq. 3
is to compute the magnitude exceedance probabil-
ity Pr( *). However, one more piece of infor-
mation is needed for this calculation, in addition to
the mean values and standard deviations of maxi-
mum earthquake magnitudes available. That is, the
probability distribution that can properly model
this random variable.
Understandably, not too many data are avail-
able to perform statistical goodness-of-fit tests on Figure 4. Population at risk for each fault considering a
earthquake magnitudes induced by the same fault 50-km radius from the fault.
owing to the long return periods. Alternatively,
Wang et al. (2011) examined 110 annual maximum
earthquakes around Taiwan since 1900, finding
that the lognormal distribution can properly model
the random magnitude.
Using this best-estimate information to
model the distribution of maximum earthquake
magnitudes induced by an active fault, Figure 3
shows the magnitude exceedance probability of the
six faults. With this probabilistic analysis, the maxi-
mum earthquake magnitude exceeding 7.0 induced
by the Hsinhua Fault is in probability of 0.2% (still
possible), given the best-estimate mean value equal
to 6.4. In contrast, the probability is near 100%
for an earthquake induced by the Changhua Fault
exceeding 7.0, given the best-estimate mean = 7.6.
Figure 5. Risk scores and the classification for the six
active faults in Taiwan.
3.2 Estimates of Population At Risk
This study considers the residents within a distance
of 50 km from the fault are at risk. Based on the Taipei, the most populated city in Taiwan, results
demographic data of Taiwan, Figure 4 shows the in a PAR in 7.5 million. In contrast, the Meishan
estimates of PAR on such a condition for the six Fault located in Chaiyi leads to a relatively low
active faults. The Sanchiao Fault located close to PAR less than one million.

3.3 Risk score and fault classicification

With the earthquake energy release ratios (i.e.,
Fig. 1), the best-estimate magnitude exceedance
probability (i.e., Fig. 3), and the best-estimate
population at risk (i.e., Fig. 4), we calculated the
(normalized) risk scores for the six faults in Taiwan,
as shown in Figure 5.
The result shows that the six risk scores are very
distinct, ranging from 1.0 up to 200. The main
causes to such a wide distribution of risk scores are
that earthquake energy release that should be lin-
early related to earthquake consequence increases
exponentially with magnitude, and that the mag-
nitude exceedance probability at the same m* can
Figure 3. Earthquake magnitude exceedance probabil- be distinctly different between two faults, say, 0.2%
ity for the six faults analyzed in this study. and 100% for earthquake magnitudes exceeding 7.0

227

ISGSR2013.indb 227 10/18/2013 9:39:27 AM

 induced by the Hsinhua Fault and Changhua Fault Changhwa Fault in Central Taiwan are Category III
as previously mentioned. active faults, the highest level in this study.
Based on the score distribution, three categories
were employed in this fault classicification:
Category III Extremely High Risk; Category II REFERENCES
High Risk; Category I Moderate Risk. The risk
score ranges for each category are 1∼3, 3∼30, and Cheng, C.T., Chiou, S.J., Lee, C.T. & Tsai, Y.B. 2007.
above 30 from Category I to Category III. As a Study on probabilistic seismic hazard maps of Taiwan
result, we considered the Sanchiao Fault in North after Chi-chi earthquake. Journal of GeoEngineering
Taiwan and Changhua Fault in Central Taiwan are 2(1): 19–28.
Geller, R.J., Jackson, D.D., Kagan, Y.Y. & Mulargia, F.
Category III active faults based on this risk-based 1997. Earthquake cannot be predicted. Science 275:
classification. In other words, a more conservative 1616.
earthquake-resistant design should be employed to Keller, E.A. 1996. Environmental Geology 7th ed.
ensure the seismic safety of cities and townships Prentice Hall, New Jersey, USA.
near the two faults. Wang, J.P., Chan, C.H. & Wu, Y.M. 2011. The distribution
of annual maximum earthquake magnitude around
Taiwan and its application in the estimation of
catastrophic earthquake recurrence probability.
4 DISCUSSIONS Natural Hazards 59(1): 553–570.
Wang, J.P., Wu, Y.M., Lin, T.L. & Brant, L. 2012a.
Understandably, other properties of the fault can The uncertainty of a Pd3-PGV onsite earthquake
be further integrated to this risk-based classifica- early warning system. Soil Dynamics and Earthquake
tion system, such as the return period of the fault Engineering 36: 32–37.
activity or the recurred earthquake, mainly because Wang, J.P., & Xu, Y. 2012b. The site selection of the
the earthquake occurrence probability within a Lungmen nuclear power plant in Taiwan. International
finite interval (e.g., 50 years) should be different Journal of Architecture, Engineering and Constructions
1: 83–89.
if the return periods of two faults are 100 years Wang, J.P., Lin, C.W., Taheri, H. & Chen, W.S. 2012c.
and 1,000 years. It should not be too difficult to Impact of fault parameter uncertainties on earthquake
re-calculate the risk score with such a fault prop- recurrence probability by Monte Carlo simulation—an
erty taken into account. The issue could be the example in central Taiwan. Engineering Geology 126:
best-estimate return periods, which are better to 67–74.
be somewhat verifiable, otherwise the analysis is Wang, J.P., Huang, D. & Yang, Z. 2012d. The deterministic
not on a scientific ground although the governing seismic hazard map for Taiwan developed using
equation is more logical covering more aspects. an in-house Excel-based program. Computers and
Geoscience 48: 111–116.
Wang, J.P., Huang, D. Chang, S.C. & Wu, Y.M. 2013a.
New evidence and perspective to the Poisson process
5 CONCLUSION and earthquake temporal distribution from 55,000
events around Taiwan since 1900. Natural Hazards
This paper presents a risk-based classification Review ASCE doi: 10.1061/(ASCE)NH.1527-
framework for active faults in a region, considering 6996.0000110.
the fault’s location and population at risk, the best- Wang, J.P., Huang, D. & Chang, S.C. 2013b. Assessment
of seismic hazard associated with the Meishan fault in
estimate earthquake magnitude, magnitude exceed- Central Taiwan. Bulletin of Engineering Geology and
ance probability, and earthquake energy release. the Environment doi: 10.1007/s10064-0.13-0471-x.
This new method was then applied to six faults Wang, J.P., Huang, D. Chang, S.C. & Wu, Y.M. 2013c.
in Taiwan. Accordingly, we considered that the A seismic hazard analysis for the region of Tehran.
Sanchiao Fault located in North Taiwan and the Natural Hazards Review ASCE (accepted).

228

ISGSR2013.indb 228 10/18/2013 9:39:28 AM

 Geotechnical Safety and Risk IV – Zhang et al. (eds)
© 2014 Taylor & Francis Group, London, ISBN 978-1-138-00163-3

Introducing non-stationary earthquake process concept: Including

an analytical model and a case study in Central Taiwan

Y. Xu & J.P. Wang

Department of Civil and Environmental Engineering, Hong Kong University of Science and Technology,
Kowloon, Hong Kong, China

ABSTRACT: From observation and experience, earthquake recurrence induced by the same fault should
be influenced by “memory”, or such a stochastic process should be non-stationary rather than being
stationary. For instance, the same fault triggering the recent Tohoku earthquake in Japan is less likely to
trigger another one in the coming few years, but the probability will be increasing with time. However, the
commonly used Poisson process considers the earthquake as a stationary stochastic process, resulting in the
identical probability regardless whether the next earthquake occurs in 2020∼2030 or 2220∼2230. As a result,
based on the mechanism of faulting, this paper introduces a new analytical model to properly take the earth-
quake memory effect into account. A case study in Central Taiwan is given, which is considered to provide a
more realistic result compared to that from the stationary stochastic process suggested by Poisson model.

1 INTRODUCTION characterization, to earthquake hazard analysis, to

earthquake early warning system (e.g. Cheng et al.,
The Poisson process is commonly used to model the 2007; Wang et al., 2011; Wang et al., 2012a, 2012b,
earthquake recurrence within a given time (Ang and 2012c; Wu and Kanamori, 2005). A fault inves-
Tang, 2007; Wang et al., 2013). Owing to the proper- tigating program was carried out by the Central
ties of the statistical model, the event of the Poisson Geological Survey Taiwan (CGST), and the best-
process occurs randomly, independent of time, which estimated return period and earthquake magnitude
is so-call “memory-less” effect or a non-stationary of 35 active faults on the island were published
stochastic process (Kramer, 1996). However, a series (Lin et al., 2008, 2009).
of equivalent earthquakes induced by the same fault The Meishan Fault, located in the Central
seem not memory-less in reality. Based on the elastic Taiwan (Fig. 1), was reportedly capable of
rebound theory, earthquakes occur when the shear
stress reaches the shear strength of the rock along
the fault, releasing the strain energy accumulated on
the fault (Reid, 1911). As a result, the chances of the
earthquake recurrence must relate to the time that
has elapsed since the last earthquake.
Therefore, this paper introduces a new analyti-
cal framework to assess the non-stationary earth-
quake probability from the underlying mechanism
of faulting, of which the changing probability with
time is a result of time-dependent external stress
accumulation and time-independent ultimate resist-
ance on the fault plane. The method is then applied
to the Meishan Fault in Central Taiwan, whose
return period was best estimated at 162 years with
the very last event occurring in 1906. Altogether,
the overview of this fault, the development of this
approach, and a case study are given in this paper.

2 OVERVIEW OF THE MEISHAN FAULT

Given the highly active seismic region around

Taiwan, efforts have been spent from seismogenic Figure 1. Meishan fault in Taiwan.

229

ISGSR2013.indb 229 10/18/2013 9:39:28 AM

 inducing a M 7.1 event every 162 years. The last
occurrence in 1906 resulted in hundreds of casu-
alties and thousands of building collapses, most
of which were around the Chaiyi City (Yen et al.,
2008). With the best-estimated data, the next strike
caused by the Meishan Fault will be on 2068, and
the Chaiyi City, lying very close to the fault, is
under high earthquake risk.

3 EARTHQUAKE PROBABILITY
INDUCED BY THE MEISHAN FAULT
WITH THE POISSON PROCESS

To characterize the temporal distribution of the

earthquake recurrence, the Poisson probability is
expressed as follows (Kramer, 1996): Figure 2. Systematic diagram showing the mechanism
of faulting: when shear stress is greater than shear
strength, the faulting should be activated. Note σ 1 and
Pr( ; t1 ≤ T t1 + Δt; λ ) σ 3 on the figure denotes the major and minor principal

=
( × t ) e − λ × Δt
n
(1)
stresses in horizontal and vertical directions, respectively.

n!
stress accumulation (Δτ ) and stress accumulation
where n = the number of the event; t1 = the starting (τ t) at time t on the fault plane can be formulated
date; Δt = time interval; and λ = the mean annual as follows:
rate, respectively.
Given the occurrence probability of at least one T t
exceedance in a specific time window can be esti- τt ∫ T = 0 Δτ (3)
mated as follows:
Applying the boundary condition derived
Pr( ; t1 ≤ T t1 + Δt; λ ) beforehand, the relationship between the shear
stress and the strength at return period R can be
= 1 − e − λ × Δt (2) expressed as follows:

Therefore, if the model is considered suitable for T R

simulating the earthquake probability induced by
τR ∫T = 0 Δτ = S (4)
the Meishan Fault with the best-estimated annual
rate, λ = 1 / 162, the probability that the earthquake where T = 0 represents the very moment right after
could occur in 2020∼2040 is 12%, identical to the the latest occurrence; S = the strength of the rock.
likelihood in 2040∼2060 given the same interval. Moreover, given the faulting governed by the
simple law of physics, the probability of faulting
can be expressed as follows:
4 EARTHQUAKE PROBABILITY
ASSESSMENT WITH A PHYSICS-BASED Pr( f ) = Pr(
P( t ) (5)
MODEL

4.1 Mechanism of faulting and faulting 4.2 Model verification on a deterministic basis
probability
On a deterministic basis, the earthquake probabil-
Figure 2 shows a systematic diagram of the under- ity is zero before the return period is due. Since the
lying mechanism of faulting. The model is dictated stress accumulation in every 162 years in this case is
by the law that when the shear stress exceeds the equal to the strength S, the unit stress accumulation
strength of the rock along the fault, the faulting is per year is S /162. Therefore, the stress accumula-
initiated and leads to an earthquake (Reid, 1911). tion from 1906 to 2020 is equal to ( ) /162,
Therefore, the first boundary condition in this and ( ) /162 till 2040. As a result, the earth-
model is that the accumulated shear stress equals the quake probability is zero in 2020∼2040 because
strength of the rock when the return period is due. at any moment during this time period the stress
Given the ground subject to a stress accumula- is less than the strength. Applying the calcula-
tion with time, the relationship between transient tion to evaluating the earthquake probability in

230

ISGSR2013.indb 230 10/18/2013 9:39:29 AM

 2040∼2060, the same outcome (probability is zero) With the mean and standard deviation of stress
can be attained. As a result, it is a zero probability accumulation at time t, we can go on calculating
that the earthquake will occur before 2068 from a the probability of τ t greater than S, which is the
deterministic analysis, either it is from 2020∼2040 likelihood of earthquakes based on the faulting
or 2040∼2060 without accounting for the uncer- model. It is worth noting that during such calcu-
tainty in stress accumulation. lation the probability distribution for this variable
With the two verifications, the results are robust is needed as well, no matter it is based on theory,
compared to the deterministic analysis, so that this observation, or assumption.
analytical model from the mechanism of faulting
is reliable. 5.3 Earthquake probability induced by the
Meishan Fault with stress accumulation
uncertainty and sensitivity analysis
5 PROBABILISTIC ANALYSIS:
THE UNCERTAINTY OF STRESS Since the variability and probability distribution
ACCUMULATION IN TIME of stress accumulation in Central Taiwan are not
available, we carried out sensitivity analyses in
5.1 The mean value of stress accumulation the calculation of earthquake probability induced
at time t by the Meishan Fault. The levels of variability in
terms of the coefficient of variation (= standard
Logically speaking, it is very unlikely that the deviation/mean) are 0.5, 1.0 and 2.0, and we
earthquake really returns at the return period, considered that stress accumulation is a variable
which makes the deterministic analysis less real- following either the normal distribution or the log-
istic. One source of uncertainties in this problem normal distribution.
is in the stress accumulation. In other words, the Figure 3 shows the earthquake probability
stress accumulation at time t is a random vari- on a variety of conditions. First of all, with the
able. For example, at the return period the stress
accumulation should be a distribution with the
mean value equal to the strength S given the return
period, which can be expressed as follows:

E ⎡∫ Δτ ⎤
T R
E [τ R ] = S (6)
⎣ TT= 0 ⎦

where E denotes the expectation.

As a result, the mean value of stress accumula-
tion at time t can be derived as follows (apparently
t and R are constants):

⎡t T ⎤
E ⎡ ∫ Δτ ⎤
T t R
E [τ t ] = E ⎢ ∫ Δτ ⎥
⎣T o ⎦ ⎣R T o ⎦
t
= ×S (7)
R

5.2 The variance of stress accumulation at time t

Similarly, given the best-estimated standard devia-
tion of stress accumulation at return period is
equal to σ * (i.e. constant), the variance of stress
accumulation at time t can be derived as follows
based on probability and statistics.:

VAR ⎡ ∫ Δτ ⎤⎥ = VAR[τ R ] σ
T R *
⎣⎢ T = 0 ⎦
⎡t T R ⎤
⇒ VAR ⎡ ∫ Δτ ⎤ = VAR ⎢ ∫
T t
Δτ ⎥
⎣⎢ T = 0 ⎦⎥ ⎣ R T =0 ⎦
2
⎛t⎞
= ×σ* (8) Figure 3. Earthquake probability with respect to differ-
⎝ R⎠ ent coefficient of variation.

231

ISGSR2013.indb 231 10/18/2013 9:39:34 AM

 new analytical model proposed, the earthquake used for modeling stress accumulation, the earth-
probability induced by a specific fault is non- quake probability, Pr( t ),is increased with the
stationary, and it increases with the time that has variability in τ t . On the other hand, Pr( t ) is
elapsed in reflecting the continuing stress accumu- slightly smaller with COV = 0.5 in τ t than with
lation. Secondly, given the same period of time, COV = 2.0 when this variable is considered fol-
the probability is substantially affected by the vari- lowing the asymmetrical lognormal distribution.
ability and probability distribution assigned for Therefore, as most probabilistic analyses, the
stress accumulation. For example, in 2020∼2040, underlying statistical attributes are the key factor
the probability can vary in a range from 23% to to estimation. However, it must be noted that the
44%, depending on the combination of levels of characterization in the statistical attribute of this
variability and probability distribution associated variable is not the scope of this study. An inde-
with the stress accumulation. In this case study, the pendent, comprehensive study with such a focus
increase in variability is not necessarily causing the should be suggested to address the problem.
increase in earthquake probability, as the calcula-
tions show an opposite trend with the normal and
lognormal distributions employed to simulate this 6 DISCUSSIONS
statistical attribute of stress accumulation.
From statistics and probability, such an influence Contradictory to the stationary results suggested
in probability distribution can be fully expected by the Poisson process, the suggested analytical
and explained. Figure 4 shows the cases calcu- model provides a non-stationary earthquake proc-
lating the earthquake probability in 2040∼2060. ess, which is more realistic to both the observation
When the symmetrical normal distribution was and the experience. However, the Poisson model is
still valid in calculating the earthquake probability.
The confusion comes from the underlying differ-
ence of the two problems.
For the new proposed analytical model, a specific
fault with its best-estimated return period and last
occurrence is required. For a specific fault, once the
earthquake happens, the train energy is released.
The shear stress needs time to accumulate to trig-
ger another earthquake, so the earthquake prob-
ability drops at once when the earthquake happens
and increases as the time passes by. Apparently,
for a specific fault, it is a non-stationary process.

Figure 5. Systematic diagram showing the fundamental

condition resulting in the memory-less effect in the
Poisson model; given the region containing eight faults
and the last event being triggered by Fault A, in the next
few years, although the recurrence probability by Fault A
is believed to be low, the increasing probability associated
Figure 4. Earthquake probability in 2040∼2060s with with other faults in the system makes up Fault A’s low
respect to different coefficient of variation. probability.

232

ISGSR2013.indb 232 10/18/2013 9:39:37 AM

 On the other hand, when a broad region with more robust stationary process because the respective
active faults, say 20 faults, is taken into consid- variation in faulting probability with time is can-
eration, the individually non-stationary process celed out owing to the nature of statistics, resulting
becomes a robust stationary process, as suggested in an overall, memory-less probability estimation
by the Poisson model. with a stationary statistical model.
To be more specific, Figure 5 shows a systematic
diagram illustrating the relationship between many
non-stationary processes and a stationary process REFERENCES
after combining. Given Fault A occurring at year
“zero,” the overall earthquake probability in the Ang, A.H. & Tung, W.H. 2007. Probability concepts in
very next moment is unchanged, or memory-less, engineering: emphasis on applications to civil and
because the earthquake probabilities induced by environmental engineering. 2nd edn., Wiley, N.J.
the other four faults keep increasing. Those prob- Cheng, C.T., Chiou, S.J., Lee, C.T. & Tsai, Y.B. 2007.
Study on probabilistic seismic hazard maps of Taiwan
ability increases compensate the “probability reset- after Chi-chi earthquake. Journal of GeoEngineering
ting” in Fault A, which is more believed to be a 2(1): 19–28.
non-stationary process. Kramer, S.L. 1996. Geotechnical Earthquake Engineering.
Prentice Hall Inc., N.J.
Lin, C.W., Chen, W.S., Liu, Y.C. & Chen, P.T. 2009.
7 CONCLUSIONS Active faults of Eastern and Southern Taiwan. Special
Publication of Central Geological Survey 23: 178 (In
This paper introduces a new analytical model Chinese with English abstract).
to tackle the same problem from the underly- Lin, C.W., Lu, S.T., Shih, T.S., Lin, W.H., Liu, Y.C. &
ing mechanism of faulting. That is, the faulting Chen, P.T. 2008. Active faults of Central Taiwan.
or earthquake is induced because the strength is Special Publication of Central Geological Survey 21:
148 (In Chinese with English abstract).
exceeded by the external stress on the fault plane. Reid, H.F. 1911. The elastic rebound theory of
Using this physics-based model with some proba- earthquakes. Bulletin of the Department of Geology,
bilistic consideration in stress accumulation, the University of Berkeley 6: 413–444.
sensitivity study shows that the earthquake prob- Wang, J.P., Brant, L., Wu, Y.M. & Taheri, H. 2012a.
ability in 2020∼2040 is indeed lower than that in Probability-based PGA estimations using the
2040∼2060 given the same hypothetical informa- double-lognormal distribution: including site-specific
tion in the variability and probability distribution seismic hazard analysis for four sites in Taiwan. Soil
in stress accumulation. In short, the plausible non- Dynamics and Earthquake Engineering 42: 177–183.
stationary earthquake process induced by a spe- Wang, J.P., Chan, C.H. & Wu, Y.M. 2011. The distri-
bution of annual maximum earthquake magnitude
cific fault is properly modeled by this new model around Taiwan and its application in the estimation
proposed. of catastrophic earthquake recurrence probability.
Based on the stationary Poisson process, the Natural Hazards 59(1): 553–570.
earthquake probability in 2020∼2040 and in Wang, J.P., Huang, D. & Yang, Z. 2012c. The determinis-
2040∼2060 is of no difference (=12%) given the tic seismic hazard map for Taiwan developed using an
best-estimated return period (i.e. 162 years) of in-house Excel-based program. Computers and Geo-
the Meishan Fault and its last event in year 1906, science 48: 111–116.
owing to the nature of this memory-less statistical Wang, J.P., Huang, D.R., Cheng, C.T., Shao, K.S.,
model. This stationary estimation is considered less Wu, Y.C. & Chang, C.W., 2013. Seismic hazard analysis
for Taipei city including deaggregation, design spectra,
realistic. Take the Tohoku earthquake in 2011 for and time history with excel applications. Computers
example, an equivalent event induced by the same and Geosciences 52: 146–154.
fault is less likely to recur in the coming few years, Wang, J.P., Lin, C.W., Taheri, H. & Chen, W.S. 2012b.
but much more probable in few hundred years. In Impact of fault parameter uncertainties on earthquake
short, such a problem is a non-stationary process, recurrence probability by Monte Carlo simulation—an
implying the earthquake memory effect. example in central Taiwan. Engineering Geology 126:
It is worth noting that as many suggested, the 67–74.
customary use of the stationary Poisson model in Wu, Y.M. & Kanamori, H. 2005. Experiment on an
earthquake prediction is robust from a statistical onsite early warning method for the Taiwan early
warning system. Bulletin of the Seismological Society
point of view. The problem is that the two situa- of America 95(1): 347–353.
tions should not be considered the same. Given a Yen, J.Y., Chen, K.S., Chang, C.P. & Boerner, W.M. 2008.
broad region where there are, say, 20 active faults, Evaluation of earthquake potential and surface defor-
the earthquake probability in the region is a result mation by Differential Interferometry. Remote Sensing
of 20 non-stationary processes, which becomes a of Environment 112(3): 782–795.

233

ISGSR2013.indb 233 10/18/2013 9:39:39 AM

This page intentionally left blank
 Geotechnical Safety and Risk IV – Zhang et al. (eds)
© 2014 Taylor & Francis Group, London, ISBN 978-1-138-00163-3

Modelling techniques of submarine landslide in centrifuge

C. Zhao, W.Z. Zhang, J.H. Zhang, K.Z. Wang & Z.S. Xiong
State Key Laboratory of Hydroscience and Engineering, Tsinghua University, Beijing, China

ABSTRACT: Gas hydrate is ice-like compound consisting of mostly methane and water. It is a
promising energy source but very unstable and easy to dissociate when it is disturbed. If gas hydrate in
the sediments under the sea dissociates, it may trigger a submarine slide and render high risk to human
beings and facilities in the sea. In this paper, modeling techniques were developed to simulate changes
of strength and pore pressure in sediments caused by the dissociation of gas hydrates in a geotechnical
centrifuge in order to assess the risk potential associated with submarine slides. The soft toothpaste,
Kaolin paste and Kaolin paste underlying sand were used to simulate sediments containing gas hydrate.
Centrifuge modeling tests were carried out on submarine landslides using the materials. The effectiveness
of the methods is discussed.

1 INTRODUCTION geotechnical centrifuge. Different methods were

used to simulate the changes of strength and
1.1 Type area pore pressure in sediments caused by the disso-
ciation of gas hydrates in order to trigger sub-
Gas hydrate is ice-like compound consisting of
marine landslides. The purpose of the study was
mostly methane and water. It is a promising energy
to develop an effective method so as to initiate
source because of its great storage under the sea.
a submarine landslide triggered by gas hydrate
However, gas hydrate is very unstable and easy to
dissociation.
dissociate when it is disturbed. If gas hydrate in the
sediments under the sea dissociates, it may trigger
a submarine slide render high risk to human beings
and facilities in the sea (Field 1990, McIver 1982, 2 METHODOLOGY
Paull et al. 1996, Mienert et al. 2005).
Theoretical and experimental investigations The mechanism of the dissociation of gas hydrate
have been made to identify the relationship affecting the stability of the submarine slope is
between the strength variation of sediment con- generally considered in two ways: 1) the disso-
taining gas hydrate and the dissociation so as ciation of gas hydrates reduces the strength of
to assess the risk potential imposed on natural sediments and thus forming a weak layer which
resource development, sea-floor transport and leads to instability of the sediments; 2) excess pore
communication, coastal development and marine pressure generates from gas hydrate dissociation
environment protection (e.g. Sultan et al. 2004, and reduces the effective stress in the sedi-
Xu 2006, Nixon & Grozic 2007, Winters et al. ments. The methodology of simulating sediment
2007). Great efforts also have been dedicated containing gas hydrate falls in three categories:
on the mechanism and consequences of subma- 1) using soft materials, such as toothpaste or
rine landslides (e.g. Locat & Lee 2002, Masson soft clay, to form the weak layer in sediments;
et al. 2006, Chau 2009, Brune & Ladage 2010, 2) introducing excess pore pressure in sedi-
Sassa & Sekiguchi 2011, Xiong & Zhang 2012, ments to simulate high pressures resulted from
Boukpeti et al. 2012). Physical modeling is one gas hydrate dissociation; 3) producing similar
of the important approaches used to improve the environment analog to the ocean (Xiong 2010,
understanding of the submarine landside and its Zhang 2012, Wang 2012). Modeling techniques
influence over offshore structures such as pipe- in a geotechnical of centrifuge were developed to
lines (e.g. Gaudin et al. 2009, Boylan et al. 2009, simulate submarine landslides with methodology
Truong et al. 2010, Chi et al. 2011). mentioned. The centrifuge tests were carried out
This paper discusses modeling techniques on the 50 gt centrifuge at Tsinghua University, as
developed to initiate a submarine landslide in a shown in Figure 1.

235

ISGSR2013.indb 235 10/18/2013 9:39:40 AM

 Figure 1. The 50 gt centrifuge at Tsinghua University. Figure 2. Centrifuge model with a layer of toothpaste.

3 TESTS AND RESULTS

3.1 Simulation of weak layer

A medium sand was used to simulate the sediment
containing gas hydrates. A 15 mm thick layer of
toothpaste forms a weak layer interlayering between
two layers of medium sand. The toothpaste made in
China can be bought in any supermarket. The und-
rained strength of the toothpaste was determined
by vane shear test to be 0.5 kPa. The medium sand
above the toothpaste, having a thickness of 50 mm,
was prepared by sand raining method. The dry den-
sity of the sand is 1600 kg/m3. Figure 2 shows the
preparation of a centrifuge model of a slope. The
strong box has an inner dimension of 600 mm in
length, 500 mm in height and 200 mm in width. The
slope angle was 10 degree. The toothpaste was so
soft that it was manually squeezed onto the lower
layer of medium sand to form the weak layer. The
model was filled with tap water before being put
onto the centrifuge platform.
The slope failed at a centrifugal acceleration of Figure 3. Centrifuge model failed at an acceleration of
10 g. Figure 3 shows the model after failure. The 10 g.
top sand layer moved downward along the layer
of toothpaste. However, as the toothpaste was so
soft that it was squeezed out and smeared on the first to liquid and then to gas phase, therefore high
plexiglass. The main drawback of the method was pressure water then air was introduced within the
that the toothpaste was not stable as it was par- sediment to trigger submarine slides in the centri-
tially dissolved in the water and lost most of its fuge model. Figure 4 illustrates the layout of model
strength, leading to conspicuous deformation in set-up for a slope. A plexiglass plate was used to
the layer in 1 g condition. If the model was soaked form an inclined base. Rotating the plexiglass plate
for a long time, the toothpaste will be squeezed out around the lower end will change the slope angle.
rendering sand-sand contact. Later the Kaolin clay The slope consisted of Kaolin clay of 20 mm thick
mixed with tap water was also used, but its strength underlying the medium sand. Before placing the
was too high after consolidation in the centrifuge. sand and clay, a set of tiny pipes were installed on
It cannot form the weak layer in sand. the plate. High pressure water or air will flow in
through the pipes to apply excess pore pressure
in the sand. Three pore water pressure transduc-
3.2 Simulation of high pore pressure in sediment
ers were placed in the sand near the outlets of the
Gas releasing from hydrate dissociation will pipes to monitor the variation of the pore pressure.
increase excess pore pressures in the sediment. Soil pressure cells were installed at the toe of the
As the gas hydrate may change from solid phase slope to measure the impact of the landslides on

236

ISGSR2013.indb 236 10/18/2013 9:39:40 AM

 Figure 6. Centrifuge model set-up using salty water.

Figure 4. Layout of centrifuge model set-up.

Figure 5. A model with 15° slope angle after failure.

structures, such as pipelines. The Kaolin clay was

prepared as paste using tap water. After the prepa-
ration, the model was filled with tap water.
The consolidated undrained shear strength of Figure 7. A model with 15° slope angle after failure.
the Kaolin clay was determined by means of triax-
ial tests. The clay slope will not fail at slope angles
below 25° at a centrifugal acceleration of 50 g. It with the concentration of 3.5% analog to the
was the excess pore pressure rendering the failure South China Sea. Figure 6 shows the model set-up.
of the gentle slope. Figure 5 shows a model with The layout of the transducers and water pipes was
15° slope after failure. It was found from the tests identical to that in Figure 4.
that: 1) the water or air pressure was effective to Due to the presence of the salt, the Kaolin was
increase excess pore pressure in the sand and thus greatly dispersed and thus less sticky, featuring low
trigger a submarine landslide, 2) However, the cohesion strength. Tiny visible pores uniformly
state and strength of the clay was very sensitive to distributed in the clay, which rendering very differ-
the property of water. As tap-water did not reflect ent properties from that in fresh tap water. With
the ocean environment, the Kaolin clay paste was the same preparing procedure, the dry density of
very sticky and its dry density was greater than the clay in the model was measured to be only
1600 kg/m3 after centrifuge test. 1140 kg/m3, significantly lower than that in fresh
water. Figure 7 shows a model with 15° slope failed
3.3 Simulation of ocean water at 50 g, however it was quite stable in fresh water
owing to the cohesion strength of the Kaolin clay.
The ocean is composed of water with salt
concentration of 3.5%. It is the basic feature of the
ocean environment and highly affects the behavior 4 DISCUSSIONS AND CONCLUSIONS
of ocean sediment soil. During the preparation of
the model, edible salt of sodium chlorde bought The key factors influencing landslide occurrence
in supermarket was used to prepare the sea water are elevated pore pressures (leading to decreased

237

ISGSR2013.indb 237 10/18/2013 9:39:41 AM

 frictional resistance to sliding) and specific weak Field, M.E. 1990. Submarine landslides associated with
layers within stratified sequences. The deposit shallow seafloor gas and gas hydrates off northern
condition in ocean environment also contributes California, AAPG Bull. 74(6): 971–972.
greatly to the instability of submarine sediments. Gaudin, C., White, D.J. & Boylan, N. 2009. A wireless
high-speed data acquisition system for geotechni-
This paper discuss the modeling technique in cal centrifuge model testing. Measurement Science &
centrifuge with the main purpose to develop an Technology 20(9): 095709.
effective method so as to initiate a submarine land- Locat, J. & Lee, H.J. 2002. Submarine landslides: advances
slide triggered by gas hydrate dissociation. In the and challenges. Can. Geotech. J. 39: 193–212.
early stage of the development of modeling tech- Masson, D.G., Harbitz, C.B., Wynn, R.B., Pedersen, G. &
nique, tap water was used to prepare Kaolin clay Løvholt, F. 2006. Submarine landslides: processes,
resulting in very strong material with high shear triggers and hazard prediction. Philosophical Transac-
strength. tion Royal Society A 364: 2009–2039.
When using salty water was used to simulate sea Mienert, J., M. Vanneste, S. Bünz, K. Andreassen, H.
Haflidason & H.P. Sejrup 2005. Ocean warming and
water, the Kaolin clay featured low strength and gas hydrate stability on the mid-Norwegian margin at
low density, which was more analog to the ocean the Storegga Slide, Mar. Pet. Geol. 22: 233–244.
environment. High pressure water was introduced Nixon, M.F. & Grozic, J.L.H. 2007. Submarine slope
followed with high pressure air in the sediment can failure due to gas hydrate dissociation: a preliminary
well simulate the sediment containing gas hydrate quantification, Canadian Geotechnical Journal 44(3):
and its dissociation process. 314–325.
The method described in the paper was effective Paull, C., W.J. Buelow, W. Ussler, & W.S. Borowski 1996.
to trigger underground landslides in the centri- Increased continental-margin slumping frequency
fuge. It provides a sound basis for assessing hazard during sea-level lowstands above gas hydrate-bearing
sediments, Geology 24: 143–146.
potential related to submarine landslides. Sassa, S. & Sekiguchi, H. 2011. Dynamics of Submarine
Liquefied Sediment Flows: Theory, Experiments and
Analysis of Field Behavior. 5th International Sym-
ACKNOWLEDGEMENTS posium on Submarine Mass Movements and Their
Consequences, eds. Yamada, Y, Kawamura, K, Ike-
The research work was funded by the National hara, K., Kyoto, Japan, Oct 24–26, 2011. Submarine
Natural Science Foundation of China (51038007) Mass Movements and Their Consequences, Advances
the National Basic Research Program of China in Natural and Technological Hazards Research 31:
(2013CB035902), and State Key Labora- 405–416.
Sultan, N., Cochonat, P., Foucher, J.P., Mienert, J. 2004.
tory of Hydroscience and Engineering Project Effect of gas hydrates melting on seafloor slope insta-
(2012-KY-04, 2013-KY-04). bility, Mar. Geol. 213: 379–401.
Truong, Q.H., Lee, C., Cho, G.C. & Lee, J.S. 2010.
Geophysical Monitoring Techniques for Underwater
REFERENCES Landslide in 1 g Model. Journal of Environmental and
Engineering Geophysics 15(1): 1–19.
Boukpeti, N., White, D.J. & Randolph, M.F. 2012. Wang, K.Z. 2012. Mechanism of submarine landslide
Analytical modelling of the steady flow of a considering high excess pore pressure from gas hydrate
submarine slide and consequent loading on a pipeline. dissociation. Thesis for Master degree, Tsinghua
Geotechnique 62(2): 137–146. University. (in Chinese).
Boylan, N. Gaudin, C., White, D.J., Randolph, M.F. & Winters, W.J., Waite, W.F., Mason, D.H., Gilbert, L.Y. &
Schneider, J.A. 2009. Geotechnical centrifuge Pecher, I.A. 2007. Methane gas hydrate effect on
modelling techniques for submarine slides. Offshore sediment acoustic and strength properties, Journal of
Geotechnics—Petroleum Technology 7: 65–72. Petroleum Science & Engineering 56(1–3): 127–135.
Brune, S. & Ladage, S. 2010. Submarine landslides at Xiong, Z.S. 2010. Mechanism of submarine landslide
the eastern Sunda margin: observations and tsunami initiated by gas hydrate dissociation. Thesis for
impact assessment. Natural Hazards 54(2): 547–562. Bachelor degree, Tsinghua University. (in Chinese).
Chau, K.T. 2009. Numerical simulations of submarine- Xiong, Z.S. & Zhang, J.H. 2012. Effect of Dissociation
landslide-induced and sea-floor-collapse-induced tsu- of Gas Hydrate on the Stability of Submarine Slope.
nami along coastline of South China. Prediction and 31st International Conference on Ocean, Offshore
Simulation Methods for Geohazard Mitigation, eds. and Arctic Engineering, Rio de Janerio, Brazile,
Oka, F; Murakami, A; Kimoto, S, pp. 247–252. Jun 10–15, 2012, OMAE2012-83190.
Chi, K., Zakeri, A. & Hawlader, B. 2011. Impact drag Xu, W. & Germanovich, L.N. 2006. Excess pore pressure
forces on pipelines caused by submarine glide blocks or resulting from methane hydrate dissociation in marine
out-runner blocks. 5th Int. Symp. on Submarine Mass sediments: A theoretical approach, J. Geophys. Res.,
Movements and Their Consequences, eds. Yamada, Y, 111, B01104, doi: 10.1029/2004JB00360.
Kawamura, K, Ikehara, K., Kyoto, Japan, Oct 24–26, Zhang, W.Z. 2012. Initiation of submarine landslide
2011. Advances in Natural and Technological Hazards in centrifuge. Thesis for Bachelor degree, Tsinghua
Research 31: 429–439. University. (in Chinese).

238

ISGSR2013.indb 238 10/18/2013 9:39:43 AM

 6 Reliability and risk analysis

ISGSR2013.indb 239 10/18/2013 9:39:43 AM

This page intentionally left blank
 Geotechnical Safety and Risk IV – Zhang et al. (eds)
© 2014 Taylor & Francis Group, London, ISBN 978-1-138-00163-3

The schedule optimization for subway station construction based

on critical chain

G.X. Chen
School of Civil Engineering, Tongji University, Shanghai, China
School of Civil Engineering and Architecture, Henan University of Technology, Zhengzhou, China

J.L. Song, Z.R. Xiao, J. Wang & L.F. Yue

School of Civil Engineering and Architecture, Henan University of Technology, Zhengzhou, China

ABSTRACT: Subway construction schedule management is one of the major tasks for the government
departments, owner and construction units. As a new schedule management method, critical chain method
is effective at shortening the project duration, optimizing the resources scheduling and economizing the
project investment. Based on the analysis of basic theory and assumption of the critical chain, and com-
bined with the characters of subway station construction, steps applying critical chain to subway project
were illustrated in this paper, and methods to calculate the buffer size and activity duration were also
introduced. Finally, the critical chain method is systematically used in one subway station of Zhengzhou
Metro 1 route and the performance of critical chain method is satisfactory.

1 INTRODUCTION management theory (Kevin & Watson 2007).

Because of its good performance in schedule man-
With the positive characteristics of large capacity, agement and investment control, critical chain
fast speed, safety, punctuality, energy and land method has been widely used in the field of manu-
saving and so on, subway rail transit is develop- facturing (Han & Xu 2005), software research and
ing with such a high speed in recent years. For development (Dai & Ren 2009) and other aspects.
example, Zhengzhou City has planned 6 subway Starting from solving the main limiting factors
lines with a total length of 202.53 km, including of the project, CCPM overcomes the shortcomings
three horizontal lines, two vertical lines and one of traditional scheduling methods in just consider-
ring line, and the total investment is expected to be ing the logical relationship between processes and
100 billion Yuan. Accompanied with the construc- the duration redundancy of activities, and it is
tion of subway rail transit engineering, schedule especially good at schedule optimization. However,
management has become one of the key tasks for few researches are conducted applying CCPM to
government departments, owner and construction urban rail transit project. Considering the complex-
units to figure out. ity of rail transit projects, this paper applies CCPM
Along with the rapid advance of computer to the schedule management of urban rail transit
technology and mutual penetration of related dis- project. By taking the construction of one station
ciplines, and combined with the decision theory in Zhengzhou rail transit No.1 line for empirical
and cybernetics, certain kinds of schedule man- research, the basic procedure of applying CCPM
agement technologies such as Decision Network to rail transit project schedule management is intro-
(DN), Graphic Evaluation and Review Technique duced, and the critical chain method optimizes the
(GERT), Venture Evaluation and Review Tech- project schedule to a certain extent. This paper pro-
nique (VERT) are invented based on the network vides a more comprehensive description of applying
planning technology. However, as the character- critical chain method to urban rail transit project.
istics of schedule management are changing, all
the schedule management methods above are not
being able to meet the needs of modern project 2 SUBWAY PROGRESS
schedule management gradually, especially for the CHARACTERISTICS AND
subway project. Critical Chain Project Manage- INFLUENCE FACTORS
ment, known as CCPM—presented by Goldratt
in 1997 in his book named The Critical Chain—is The construction cycle of subway project is long-
considered to be a new era and stage for the project period. Usually, the average construction cycle of a

241

ISGSR2013.indb 241 10/18/2013 9:39:43 AM

 single metro line in China ranges from 4 to 5 years, certain extent. Furthermore, as the standards
and it takes 10 to 20 years to build up a general for the quality and technology of subway project
line network. In the course of general line network are very high, and the design units must spend some
building, difficulties such as layout of the subway more time and energy to ensure the quality of the
line network, construction sequence and construc- drawings, which will definitely result in delay of
tion resource sharing are with different characteris- the speed of drawings. Therefore, the contractors
tics from traditional project schedule management. frequently wait for the blueprint to work, resulting
The following are the main characteristics. in the three simultaneities project, namely designing
while constructing and modifying simultaneously.
In addition, because there are not complete and
2.1 Limited time
systematic drawings, it is quite easy to cause a lot
Subway project is a large city infrastructure project of design changes later, or even lead to rework,
with big social impact, which is generally regarded which will also contribute to the procrastination of
as the landmark project of a city, representing the construction drawings delivery.
the city’s image and receiving widespread con-
cern from the government and the public. There
3 THEORY AND BASIC ASSUMPTIONS
is almost no possibility of delaying the progress
OF CRITICAL CHAIN
because the completion date is always announced
and published to society after determined, and the
3.1 The basic theory of critical chain
rigid requirement of progress is in great demand.
In addition, as an indicator of their performance, Known as TOC, theory of constraint is also called
managers pursue the construction speed of sub- bottleneck theory, which originates from the Opti-
way project too much, which brings many new mized Production Technology (OPT), and it is first
problems for setting schedule goals and milepost generated and applied in manufacturing industry
program. (Goldratt 1997). Goldratt, the author of CCPM
held the opinion that simply pursuing the maxi-
mization of local production efficiency of each
2.2 Heavy preliminary work
department does not necessarily lead to global
The subway station is usually located at the pros- optimum, and proposed the production manage-
perous urban area and the main road of the city. ment theory focusing on the settlement of con-
There are various kinds of preliminary difficul- straint factors and bottleneck processes, namely
ties including land acquisition and resettlement, the theory of constraint.
afforest transplantation, pipeline move, bus line TOC is a systematic cyclical process including
change and traffic reconciliation, which demand discovering, improving and eventually utilizing the
much work of coordination. What’s more, the land constraints at the maximum. The core is to utilize
acquisition and resettlement are almost the most the bottleneck activities and resources restricting
difficult issues because there are many stakehold- system function, and the basic assumption is “any
ers involved and the procedures are cumbersome system exists one bottleneck at least, namely the
during implementing. Organizing reasonable restraining factor, which restricts the project’s max-
organization structure then establishing an efficient imum output”. What’s more, Goldratt also showed
construction team, preparing various documents clearly that in the process of project implementa-
submitted for approval and the implementation tion, utilization rate of the resources is not bal-
of specific tasks, all these include a large amount anced, and the output of a certain project is only
of complicated work. Therefore, the preliminary affected by a certain kind of bottleneck resources
workload of rail transit project is very heavy and instead of all the project resources. Therefore,
intricate, and we can say that the successful settle- we must improve the utilization of bottleneck
ment of preliminary problems is equivalent to half resources in order to speed up the progress of the
the success of subway construction. project (Ma & You 2007).
By applying the constraint theory to project
management, Goldratt presented the critical chain
2.3 Slow delivery of blueprint
project management. CCPM puts the importance
The schedule goals for rail transit project are of resource constraints to the same importance
tight and the design task is heavy. The time for level with logical relationship of the processes. It
design unit to prepare is short, and the influence introduces that the total construction period is
of non-technological factors like route rechan- determined by the duration of each activities, their
neling should be taken into consideration in the logical relationship, resource constraints and other
design process, which can lead to the decline in factors. By cutting down and conducting effective
drawing speed and quality of the drawings in a integration of the safety time to be the buffers, the

242

ISGSR2013.indb 242 10/18/2013 9:39:43 AM

 critical chain eliminates various possible uncertain end of critical chain or the junction of non-critical
factors that may delay the progress of project chain and critical chain in the form of time buffer.
implementation. The critical chain technique presents three kinds of
buffers: Project Buffer (PB), Feeding Buffer (FB)
and Resource Buffer (RB), in which RB exists as
3.2 Basic assumptions
an early warning indicator (Cheng et al. 2007).
As a new planning technique for project schedule At present, the frequent method to calculate the
management, the theory of critical chain project buffer size is mainly the cut-and-paste method and
management implies some basic assumptions and the square root of variance method. However, in
principles, and to be specific, the main points are general, there are many factors affecting the size of
below. the buffers, such as safety time of the process, the
characteristics of the project, the project manager’s
3.2.1 Progress impact factors risk preference and complexity of the network
There are many factors affecting the project graph, etc. Combined with the characteristics of
progress, such as logical relationship of the proc- different projects, there are some new methods sug-
esses, supply conditions of the resources, require- gested to calculate the buffer size by some scholars
ments of the construction period and other from the different points of view, which are still in
characteristics of the project. Traditional project the stage of theoretical research without practical
scheduling techniques considered only the influ- application (Cao & Liu 2010).
ence of logical relationship between the processes.
The technique of critical chain project manage- 3.2.4 Critical chain identification
ment based on Theory of Constraint presents Goldratt defined critical chain as “the longest
that, not only the logical relationship of proc- chain constituted by activities determining the ear-
esses affects project schedule, but also the supply liest completion time of the project”. The critical
conditions of resources have a great impact on chain is generally obtained through the following
schedule. Meanwhile, critical chain method also ways: Get the original project schedule through the
takes the factor of human resource into considera- traditional scheduling technology such as the criti-
tion, such as Students Syndrome, Parkinson’s Law cal path method; adjust the initial schedule based
(Zhao et al. 2009), etc, and the impact of various on resource constraints and time requirements
possible factors on project schedule is analyzed then to get the adjusted schedule; circulate this
comprehensively. process constantly to obtain the longest path, and
the longest path is the critical chain.
3.2.2 Estimation of the construction period
Construction period of the process restricts the
whole duration of the project. It is especially impor- 4 PRACTICAL APPLICATIONS
tant to take an appropriate method to calculate and
estimate the duration of the activities in subway This paper takes the construction of one station
project. General project schedule software obtains in Zhengzhou rail transit No.1 line for empirical
the estimated construction period through industrial research, combined with the critical chain project
standards and rules, or even the empirical data accu- management technique, and studies the subway sta-
mulated by the enterprise, which has a higher com- tion schedule. Relevant activities, duration and log-
pletion guarantee rate. However, construction period ical relations of the project are shown in Table 1.
obtained through the above ways has been proved to
be too long by many scholars (Chen & Guo 2010). 4.1 Determination of the initial schedule
The critical chain technique suggests that, construc-
tion period of the activities contains two parts of Use network planning technology to get activity-
estimation: the necessary time and the safety time. on-node network of this project, which is shown
The necessary time is considered to be the necessary in Figure 1.
estimation to complete the process, and safety time There are a total of 8 task nodes in the civil engi-
contributes less to the process completion, and the neering of the station, constituting 2 construction
safety time is thought to be unnecessary. paths, which are:
Path 1: A1000-A1010-A1020-A1030-A1070;
3.2.3 Buffer determining and affected factors Path 2: A1000-A1010-A1020-A1040-A1050-
There are various kinds of risks resulting in A1060-A1070.
progress delays in the process of project implemen-
tation. In order to avoid the risks, the critical chain Path 2 is the longest. Therefore, the critical path
technique eliminates the safety time out of the is Path 2. Initial scheduling of the project is shown
construction period and inserts it into the distal in Figure 2.

243

ISGSR2013.indb 243 10/18/2013 9:39:43 AM

 Table 1. The activity and coefficient.

Code No. Name Project duration/d Foreground task Logical relations

A1000 A Building demolition 45

A1010 B Pipeline moving 60 A SS+15
A1020 C Maintenance construction 350 B FS-15
A1030 D Dewatering construction 20 C FS-13
A1040 E Foundation pit construction 50 C FS-6
A1050 F Main construction 280 E FS-28
A1060 G Gateway construction 140 F SS+130
A1070 H Affiliated construction 190 G SS

Table 2. Critical chain activity duration and buffer size.

Code No. Name d′ Δti

A1000 A Building 23 22
demolition
A1010 B Pipeline 30 30
moving
Figure 1. Activity-on-node network of this project. A1020 C Maintenance 175 175
construction
A1030 D Dewatering 10 10
construction
A1040 E Foundation pit 25 25
construction
A1050 F Main 140 140
construction
A1060 G Gateway 70 70
construction
A1070 H Affiliated 95 95
construction
Figure 2. Initial scheduling.

4.2 Activity duration and buffer size calculation relocation construction can be carried out together
with building demolition. Then, adjust their logical
According to the critical chain by Goldratt, the relations as SS; adjust logical relations of pipeline
critical chain activity duration is half the amount relocation and maintenance construction as FS;
of planning duration, and the buffer size is half the adjust logical relations of dewatering construction
amount of safety time related. and foundation pit construction as FS-6; adjust
logical relations of main construction and founda-
1 tion pit construction as SS+24, and other logical
d′ d (1)
2 relations remain invariant. The logical relationship
of the activities and the critical chain activity dura-
n
⎛1 ⎞ tion after adjustment are shown in Table 3.
Buffer size = ∑ Δti (2)
⎝
i =1 2
⎠
4.4 Determining the critical chain
where: d ′ is critical chain activity duration, Δti and the buffer
stands for safety time, n is the number of the activi-
Take the necessary time of the activities as the
ties. Then the critical chain activity duration and
critical chain duration, then get PB and FB of the
buffer size are shown in Table 2.
project using the cut & past method and insert it
into the corresponding chain, which makes rea-
4.3 Adjusting the initial scheduling
sonable use of the safety time and ensures that the
After analyzing the actual situation of the project project can be implemented in accordance with
and optimizing the collective resources and the logi- the schedule, and the project duration is shortened
cal relationship of the activities, we find that pipeline as well. According to the calculation result, the

244

ISGSR2013.indb 244 10/18/2013 9:39:43 AM

 Table 3. The new coefficient of the activities.

Code No. Name Foreground task Logical relation Buffer size

A1000 A Building demolition 11

A1010 B Pipeline moving A SS 15
A1020 C Maintenance construction B FS 87
A1030 D Dewatering construction C FS-13 5
A1040 E Foundation pit construction C FS-6 12
A1050 F Main construction E SS+24 70
A1060 G Gateway construction F SS+130 35
A1070 H Affiliated construction G SS 47

limited at a small range of sub-projects or small

projects. Instead, the method of critical chain
project schedule management should be used sys-
tematically to achieve the optimization of project
resources, scheduling and other elements.

ACKNOWLEDGMENT

The authors wish to acknowledge the support from

Figure 3. Critical chain scheduling. Capital Foundation—Diaphysis Teacher’s subsidy
of Henan Province 2010GGJS-083.

subway station project scheduling is obtained based

on critical chain, which is shown in Figure 3. REFERENCES
According to Figure 2, the original planned
duration is 746 days. The critical chain project Cao, X.L. & Liu, R.H. 2010. Research on the calcula-
duration of subway station is 445 days and the tion method of the buffer of critical chain project
buffer size is 189 days. The total duration is management. Statistics and Decision (303): 69–71.
634 days, less than the original planned duration Chen, Y. & Guo, Y.L. 2010. Flexible Management of
Expressway Construction Scheduling Based on Key
112 days. That is to say the critical chain method
Chain. Journal of Changsha University of Science &
could effectively optimize the project schedule. Technology (Social Science) 25(3): 26–30.
Cheng, T.T., Li, W.B., Cheng, X. 2007. Research and
Application for Critical Chain in Project’s Process
5 CONCLUSIONS Management. Microcomputer Applications 28(5):
539–543.
As a new method of project schedule manage- Dai, Y.B. & Ren, X.L. 2009. Research on project critical
ment, CCPM is widely used in schedule planning chain based on Bayesian network. Computer Applica-
and time optimization. However, in the present tions and Software 26(10): 127–129.
Goldratt, E.M. 1997. Critical Chain. New York: North
study, the critical chain method is mostly applied
River Press.
in part of the project, and has not been applied to Han, W.M & Xu, X.Q. 2005. The critical chain technique
the whole project schedule management systemati- and the application in production plan. Market mod-
cally, which is contradictory to the theory of criti- ernization 453(12): 19–20.
cal chain that local optimum does not necessarily Kevin, J. & Watson, J.H. Blackstone, Stanley C. Gardiner.
lead to global optimum at a certain extent. 2007. The evolution of a management philosophy: The
Meanwhile, critical chain emphasizes the impor- theory of Constraints. Journal of Operations Manage-
tance of resources and supply chain management. ment 25(2): 387–402.
Applying the critical chain method partly and set- Ma, G.F. & You, J.X. 2007. Quantitative Analysis
of Critical Chain Multiple Project Scheduling
ting various kinds of buffers are likely to have a
Management. Systems Engineering-Theory & Practice,
great impact on the whole project resource sched- 27(9): 54–60.
uling, and then adversely affect the overall progress Zhao, Z.Y., You, W.Y., Lv, Q.L. 2009. Research on the
of the project, which does not achieve the purpose Making Construction Project Schedule with Critical
of optimizing the project schedule. Therefore, the Chain Method. Journal of Railway Engineering
utilization of critical chain method should not be Society, (5): 98–102.

245

ISGSR2013.indb 245 10/18/2013 9:39:45 AM

This page intentionally left blank
 Geotechnical Safety and Risk IV – Zhang et al. (eds)
© 2014 Taylor & Francis Group, London, ISBN 978-1-138-00163-3

A cell-based model for predicting runout distances of detached

materials in rainfall-induced slope failures

H.X. Chen & L.M. Zhang

Department of Civil and Environmental Engineering, The Hong Kong University of Science and Technology,
Clear Water Bay, Hong Kong, China

ABSTRACT: Rainfall-induced slope failures are one of the most frequent hazards in hilly terrains.
This paper presents a method that adopts a distributed cell model for predicting rainfall-induced slope
failures in a large area and the runout distances of detached materials. The model consists of five compo-
nents; namely, a digital terrain model, a spatial rainfall distribution model, a rainfall infiltration analysis
model, a slope reliability analysis model and a runout distance prediction model. A cluster of unstable
cells located together are viewed as an entity and the size effect of the combined detached materials is
considered. The materials move down the slopes along the steepest paths on the digital elevation map. The
method predicts the locations of rainfall-induced slope failures and their depositional zones reasonably
well. The material movement traces assessed by this model also offer information for the assessment of
landslide and rockfall risks. The presented method shows promise for use as a module in a real-time warn-
ing system for rainfall-induced slope failures.

1 INTRODUCTION Wenchuan earthquake (Fig. 1). Provincial Road

303 (PR303) goes through the study area. It has
Rainfall-induced slope failures are one of the most an area of 16.8 km2, an elevation range between
frequent hazards in hilly terrains, especially for those 1,000 m and 3,540 m, and a local relief of 2,540 m
with abundant supply of loose soil materials. The (Chen & Zhang 2013). The exposed lithology in
2008 Wenchuan earthquake triggered numerous the study area is mainly composed of four kinds
landslides, leaving a large amount of loose materi- of Proterozoic magmatic rocks; namely, diorite,
als in steep hill slopes and gullies. Much of the loose biotitic granite, granodiorite and hornblende
materials lost stability in the past few years when diorite. The maximum and mean annual pre-
triggered by heavy rains. The detached materials cipitations within the study area are 1,225 mm
ran down to lower locations, posing great danger and 828 mm, respectively. Approximately 68%
to roads and people, or becoming the source mate- of the total precipitation falls between June and
rials for debris flows. A cell-based model can be a September. Based on the surface geological
viable tool for predicting potential rainfall-induced characteristics, the soils and rocks in the study area
slope failures (e.g. Crosta & Frattini 2003, Baum are classified into four types: loose soil deposit,
et al. 2008, Takara et al. 2010, Arnone et al. 2011, vegetated soil, bedrock and riverbed. Among them,
Chen & Zhang 2013). However, there are still some the bedrock and riverbed are set to be stable.
pending issues for developing a cell-based model
in the Wenchuan earthquake zone. One of these
issues is to assess the runout paths and distances of
detached materials in a large area efficiently.
The objective of this paper is to develop a dis-
tributed cell-based stability analysis model for pre-
dicting rainfall-induced slope failures and runout
distances of detached materials in regional scale.

2 STUDY AREA

The study area, including the Xiaojiagou Ravine,

Pubugou Ravine and their vicinity, is only Figure 1. Locations of the study area (i.e. the box) and
5 km from the epicenter, Yingxiu of the 2008 rain gauges (i.e. the six dotted locations).

247

ISGSR2013.indb 247 10/18/2013 9:39:46 AM

 3 MODEL FRAMEWORK The values of φ′ and c′ for the loose soil deposits
are determined as 37° and 4 kPa, respectively and the
The distributed cell-based model consists of five value of φ′ for the vegetated land is also 37° (Chen &
components; namely, a digital terrain model, a spa- Zhang 2013). According to Coppin and Richards
tial rainfall distribution model, a rainfall infiltration (1990), c′ can typically increase by 1–12 kPa by
analysis model, a slope stability analysis model and vegetation roots. An average value, 6.5 kPa, is
a runout distance prediction model. The study area adopted here for the vegetated land. Other param-
is discretized into a grid first with information for eters are reported by Chen & Zhang (2013).
each cell assigned (e.g. the elevation, slope angle, soil The thicknesses of the two-layer soils are measured
depth, ground water level, soil properties, and rain- along the vertical direction in this study. The fissured
fall information). The cell size is 20 m × 20 m. Based rock layers in vegetated land and deposit zones are
on the universal Kriging method, a spatial rainfall estimated to be 4 m based on field investigations.
distribution model is used to interpolate the cumula- As the fissure rock layers have much higher shear
tive rainfall at each cell in the study area, which can strength, the potential slip surface is within the top
be represented by an optimal linear combination of soil layer. The infiltration parameters for the fissured
the rainfall records of the nearby rain gauges. The rock layers are similar with those of the vegetated
real-time hourly rainfall intensity at each cell can be land. Field investigations also indicate that the aver-
then obtained. The reason why universal Kriging is age soil thickness in flat areas is about 10 m. The
used is that it does not require that the mean rainfall interpretation of the landslide data triggered by the
in the study area is constant and it reflects the topo- Wenchuan earthquake near the study area shows
graphic effect. The algorithm has been explained in that the average thickness of the deposits retained
detail by Olea (1999). An infiltration model for two- on slopes with angles larger than 35° is less than 2 m.
layer soils is then applied to compute the infiltration Back analysis using the infinite slope model consid-
and runoff processes. The pore water pressure pro- ering relatively dry initial soil condition indicates that
file can be obtained. A slope stability analysis model a layer of 2 m thick soil can be retained on slopes as
is then used to compute the Factor of Safety (FS) steep as 75°. Hence, it is estimated that the soil thick-
of each cell at different times. If the factor of safety ness of the top layer in the vegetated land and loose
is smaller than 1, the cell is viewed as unstable. The soil deposit linearly decreases from 10 m to 2 m as
probability of slope failure can be calculated using the slope angle increases from 0° to 35°, and keeps at
Monte Carlo simulation. The runout paths and dis- 2 m when the slope angle is larger than 35°.
tances of the detached materials are then assessed From 12 August 2010 to 14 August 2010, a storm
using a runout distance prediction model. swept Yingxiu and its vicinity. The locations of six
rain gauges are shown in Figure 1. The rainfall
interpolation results and the material properties
4 SLOPE STABILITY ANALYSIS of the four surface material types are reported by
Chen & Zhang (2013). Figure 2 shows the results
Infiltration analysis is first conducted to obtain the of slope stability analysis. As shown in Figure 2a,
pore water pressure profiles in each cell. The rain- almost all the cells are stable at the beginning of
fall information of each cell is obtained through the rainfall. The number of unstable cells increases
universal Kriging interpolation. In the study area, sharply at about 15 hours after the rain starts (see
soils are layered in some locations. For the vege- Fig. 2b). The unstable cells increase continuously
tated land and loose soil deposit, the underlying until the peak rainfall intensity occurs (see Fig. 2c).
layer is fissured rocks. The infiltration param- The distribution of the unstable cells when the peak
eters for the top layer and the bottom layer may rainfall intensity occurs is almost the same with
be significantly different. Hence, an infiltration that at the end of the storm event (see Fig. 2d).
model for two-layer soils is adopted to consider A comparison between the rainfall-induced
the heterogeneity of the soils. Analytical solutions slope failures after the 2008 Wenchuan earthquake
to transient infiltration in layered soils under con- and the computed unstable cells is shown in
stant rainfall condition proposed by Srivastava & Figure 3. Most slope failures in the inventory have
Yeh (1991) and Zhan et al. (2012) are improved been predicted successfully, which indicates that
using Heaviside functions based on the principle the cell-based model can predict the locations of
of superposition. The improved model can con- rainfall-induced slope failures reasonably well.
sider arbitrary rainfall processes. An infinite slope A small portion of slope failures are not predicted
model is then adopted to compute the factor of successfully and some unstable cells are located at
safety for each of the large number of cells. The places where there is no slope failure, which maybe
adopted analysis method is described in detail by due to the resolution of the digital elevation data,
Chen & Zhang (2013). The slip surface is assumed the cell size, and the limitation of the infinite slope
to be at the wetting front in this study. stability model.

248

ISGSR2013.indb 248 10/18/2013 9:39:46 AM

 A slope failure of a larger size will move farther if
other conditions are the same. A method consider-
ing the size effect of the detached materials and the
elevation difference between the original location
and the final location of the detached materials is
developed to predict the runout path and distances
of the detached materials.
An empirical relationship is firstly developed
to predict the runout distance as follows based
on data of 31 soil slope failures in the Wenchuan
earthquake zone:

L 3 85 V 0.19 H 0.38 (1)

where L (m) is the runout distance; V (m3) is the

volume of slope failure; H (m) is the elevation dif-
ference. The runout distance here is defined as the
length of the horizontal path between the highest
Figure 2. Computed unstable cells after the rainfall
point of the slope failure and the lowest point of the
starts: a) 1 hour; b) 15 hours; c) 36 hours; d) 46 hours. deposit. The R2 coefficient for equation (1) is 0.74.
The steps for cell-based runout distance prediction
considering the effect of V and H are as follows:
1. Group the unstable cells that are bounded at
one or more sides, each cell belonging to a slope
group and each group is viewed as an individual
slope failure.
2. Compute the volume of the detached materi-
als of each group by summing the volumes of
the detached materials in the cells in that group.
The volume of the detached material of each
unstable cell is determined by the cell area and
the depth of the slip surface.
3. Conduct analysis cell by cell, allowing the
detached materials to move to lower cells along
the steepest path. The steepest path is the one
with the largest gradient between the cell where
the detached material is located and each of the
eight adjacent cells (i.e. the east, south, west,
north, northeast, southeast, southwest and
northwest cells). The detached material moves
along the steepest path on the hill slope first.
Figure 3. Comparison between rainfall-induced slope
After it reaches a channel, it will move along the
failure inventory after 14 Aug. 2010 storm and the com-
puted unstable cells. steepest path in the channel.
4. Compute H and L when the detached material
reaches a new cell.
5. The material stops movement when L is larger
5 RUNOUT DISTANCE PREDICTION
than the predicted value by Equation 1.
Although the runout distance of detached materials The runout traces and deposited locations of
has been widely researched (e.g. Corominas 1996, the detached materials are shown in Figure 4. The
Dai et al. 2003, Kuo et al. 2009), a cell-based runout movements of the detached materials reveal the
distance computation method is still to be developed. evolution of the loose materials. In Pubugou Ravine
Simple empirical approaches can be adopted to com- and Xiaojiagou Ravine, a small part of the detached
pute the runout distances of the detached materials materials from the rainfall-induced failures move for
from failure of one cell efficiently (e.g. Arnone et al. a short distance and stay on the hillslopes; the major-
2011), but the size effect of the detached materials ity of the detached materials run down for a long
on the runout distance is not considered previously, distance and are retained in the channels. They can
which tends to underestimate the runout distance. block the channels, forming barrier ponds during

249

ISGSR2013.indb 249 10/18/2013 9:39:46 AM

 ACKNOWLEDGEMENT

The authors acknowledge the support from

Sichuan Provincial Department of Transportation
and Communications, the National Basic Research
Program of China (No. 2011CB013506), and the
Research Grants Council of the Hong Kong SAR
(Grant No. 622210). The technical help from
Mr. Jozsef Danka is also acknowledged.

REFERENCES

Arnone, E., Noto, L.V., Lepore, C. & Bras, R.L. 2011.

Physically-based and distributed approach to analyze
rainfall-triggered landslides at watershed scale.
Geomorphology 133: 121–131.
Baum, R.L., Savage, W.Z., & Godt, J.W. 2008. TRIGRS-a
Fortran program for transient rainfall infiltration
Figure 4. Runout traces and deposited locations of the and grid-based regional slope-stability analysis,
detached materials at the end of the 14 Aug. 2010 storm. Version 2.0. U.S. Geological Survey Open-File Report
2008 75 pp.
Chen, H.X. & Zhang, L.M. 2013. A cell-based reliability
analysis model for predicting regional rainfall-induced
the storm or simply become the source materials for slope failures. In Proceedings of the 2013
channelized debris flows. Some hills lopes along the Geo-Congress, Entitled Stability and Performance of
PR303 fail; part of the detached materials moves Slopes and Embankments III, San Diego, 3–7 March,
down, goes across the road or deposits on the road, 2013, 982–990.
which poses hazard to the people and properties. Coppin, N.J. & Richards, I.G. 1990. Use of vegetation in
Hence, the prediction of runout distance not only civil engineering. Butterworths, London.
Corominas, J. 1996. The angle of reach as a mobility
provides information for assessing debris flow risks,
index for small and large landslides. Can. Geotech. J.
but also directly helps assess the risk posed by the 33: 260–271.
rainfall-induced slope failures along PR303. Crosta, G.B. & Frattini, P. 2003. Distributed mod-
elling of shallow landslides triggered by intense
rainfall. Natural Hazards and Earth System Science
6 CONCLUSIONS
3: 81–93.
Dai, F.C., Lee, C.F. & Wang, S.J. 2003. Characterization
A distributed cell-based model for predicting the of rainfall-induced landslides. International Journal of
runout distances of detached materials from rain- Remote Sensing 24: 4817–4834.
fall-induced slope failures is developed in this study. Kuo, C.Y., Tai, Y.C., Bouchut, F., Mangeney, A., Pelanti, M.,
The model consists of five components; namely, Chen, R.F. & Chang, K.J. 2009. Simulation of
a digital terrain model, a spatial rainfall distribu- Tsaoling landslide, Taiwan, based on Saint Venant
tion model, a rainfall infiltration analysis model, equations over general topography. Engineering
a slope stability analysis model and a runout dis- Geology 104: 181–189.
Olea, R.A. 1999. Geostatistics for engineers and earth
tance prediction model.
scientists. Kluwer Academic Publishers, Norwell.
The slope stability and runout distances of the Srivastava, R. & Yeh, T.J. 1991. Analytical solutions for
detached materials can be computed efficiently. one-dimensional, transient infiltration toward the
The runout distance prediction model takes into water table in homogeneous and layered soils. Water
account the effect of V and H. Resources Research 27: 753–762.
A comparison between the slope failure inven- Takara, K., Yamashiki, Y., Sassa, K., Ibrahim, A.B. &
tory and the simulation results shows that the Fukuoka, H. 2010. A distributed hydrological–
model can predict the locations of the slope failures geotechnical model using satellite-derived rainfall
reasonably well. The runout paths and distances of estimates for shallow landslide prediction system at a
catchment scale. Landslides 7: 237–258.
the detached materials can also be predicted suc-
Zhan, T.L.T., Jia, G.W., Chen, Y.M., Fredlund, D.G. &
cessfully. The material movements assessed by this Li, H. 2012. An analytical solution for rainfall infiltra-
model offer information for the assessment of land- tion into an unsaturated infinite slope and its applica-
slide and debris flow risks. The presented method tion to slope stability analysis. Int. J. for Numerical and
shows promise for use as a module in a real-time Analytical Methods in Geomechanics DOI: 10.1002/
warning system for rainfall-induced slope failures. nag.2106.

250

ISGSR2013.indb 250 10/18/2013 9:39:49 AM

 Geotechnical Safety and Risk IV – Zhang et al. (eds)
© 2014 Taylor & Francis Group, London, ISBN 978-1-138-00163-3

Dike failure mechanisms from the perspective of risk assessment

J. Danka & L.M. Zhang

The Hong Kong University of Science and Technology, Clear Water Bay, Hong Kong SAR, China

ABSTRACT: The failure of a dike system, as a characteristic of long linear systems, is ruled by its
weakest link. Meanwhile the most serious hazard in a dike section is its breaching, which can develop
due to different failure mechanisms. The objective of this paper is to identify the most common fail-
ure mechanisms and evaluate them from the point of view of risk assessment. As illustrated via four
case studies, the most frequent problem in the design and risk assessment is the inappropriate address-
ing of potential threats. The results of risk analysis are not reliable if relevant failure mechanisms are
neglected, the structural behaviour of the dike-section is not understood or inadequate maintenance is
provided.

1 INTRODUCTION affect the risk management. Four case studies are

reported in which breaching developed as a result
From the dike risk management point of view, the of unexpected failure mechanisms, misunder-
most serious threat is related to the breaching of standing of the structural behaviour, and failure of
the dike, which may cause the inundation of the maintenance.
flood zone. Meanwhile from the geotechnical engi- Figure 1 illustrates a possible event tree for
neering aspect the breaching process is governed risk management purposes. The importance of
by several failure mechanisms. the topic can be clearly identified if one of the
The first objective of this paper is to identify the potential failure mechanisms is considered as an
most common failure mechanisms which may lead “unlikely” event. In that case the risk assessment
to breaching. This study intends to recognize seri- would fail to identify a threat, therefore may cause
ous threats from the risk assessment aspect. Even a a false sense of security and potential problems
failure mechanism occurs frequently, if it is easily during the flood event.
noticeable and/or predictable, the risk could be low-
ered since in such cases warning, evacuation and
countermeasures could be implemented. If the pro- 2 FAILURE MECHANISMS
gression of a failure mechanism is hardly noticeable
or the failure will develop in a less ductile way, then 2.1 Historical overview
the time for countermeasures or warning will be
During qualitative or quantitative risk analysis, the
limited and the corresponding risk might be high.
first step is the identification of potential hazards.
The second objective is to evaluate breaching
Comprehensive statistical analysis on past dike
cases in which different failure drivers significantly
breaching cases was carried out by a few research-
ers (Nagy 2012, Fukunari 2008, Baars & Kempen
2009). Based on their results, the most frequent
failure mechanisms were determined, which may
aid a rough estimation of likelihood of potential
hazards. Here the results of Nagy (2012) and Baars
& Kempen (2009) will be compared and evaluated,
which are presented in Table 1.
The analyses of Nagy and Baars & Kempen focus
on the Carpathian Basin and on the Netherlands,
respectively. Their studies cover different types of
dikes from a few aspects (construction material—
fine and/or coarse graded materials, peat; different
Figure 1. Illustration of an event tree related to a dike structural arrangements of the dike section—core,
section. drain, puppet wall etc.; sea and river defense

251

ISGSR2013.indb 251 10/18/2013 9:39:50 AM

 Table 1. Comparison of statistics related to dike failure mechanisms from Nagy (2012) and
Baars & Kempen (2009).

Nagy (2012) Baars & Kempen (2009)

Investigation period

(1564–2010) (1134–2006)

Failure mechanism Rel. freq. Failure mechanisms Rel. freq.

Not known 58.1% Erosion of inner slope prot. & crest 67%
Overtopping 32.4% Ice drift 11%
Human activity 2.3% Erosion or instability of 6%
Slope failure 2.0% outer slope prot.
Subsoil failure 1.8% Sliding in. slope 5%
Other known 1.2% External (human 4%
Structural failure 1.3% and animal)
Wave erosion 0.6% Sliding out. slope 3%
Liq. of shore line 2%
Piping 1%
Micro instability 0.5%
Horizontal shear 0.5%
Total 2858 cases 1735 cases

structures). Based on their results, the most initiations are hardly observable without a moni-
frequently observed cause of dike breaching was toring system. More likely they are only noticed
the external erosion due to overtopping. Erosion in the developed phase. By then the opportunity
of the outer slope, distresses caused by human of implementing rescue measures is limited. In
or animal activities and sliding of the inner slope Figure 2 the most common failure modes related
during flood event are the most serious threats in to external and internal erosion are summarized.
addition to overtopping. Moreover potential factors that may accelerate
erosion are highlighted.
Additionally non-erosion related failure modes
2.2 Most important failure modes
like slope failures must be evaluated both under
In the historical overview it was revealed that the flood conditions (the down-stream side) and under
most probable failure scenario is the overtopping rapid draw-down conditions (the upstream side).
of dike section. However the external erosion Local slope failures caused by seepage in the inner-
might not be governed by the elevation difference slope may be referred to as micro-instabilities in
between the flood water level and the dike crest, but the literature.
by waves. When the erosion starts in the upstream Other failures might be governed by settlement,
side (e.g. the Upper Jones Tract levee breach), the horizontal shearing (sliding), earthquakes, drift-
failure of the riverside slope will become the most ing ice or collision with floating objects. Non-
likely cause of breaching. erosion related failure modes are summarised in
Although the most probable failure mechanism Figure 3.
is the external erosion of the section it must be
stated that, from risk management point of view,
2.3 Failure modes vs. risk
their assessment may not be the most challenging
one. Nowadays flood water levels can be estimated The first issue in the risk management is whether a
with acceptable accuracy; therefore the overtop- threat is identified or not. One of the most common
ping of the section is quite predictable. Moreover, reasons for dike breaching is that a failure mode is
the observation of the external erosion type of fail- considered as “unlikely”, hence not included in the
ures is quite straightforward. Thus the warning of design or risk analysis. Once the failure unexpect-
the population at risk and their evacuation from edly develops, adequate countermeasures may not
the flood zone seem to be more manageable than be well prepared.
in the case of a less ductile failure. Other possible failure scenarios occur when the
The failure modes of dike sections originated potential failure modes are identified but the dike
from internal erosion are more deceitful since their construction does not meet the design criteria.

252

ISGSR2013.indb 252 10/18/2013 9:39:50 AM

 Figure 2. Most common failure modes related to external and internal erosions. (Based on Vrijling et al. 2011).

Such events may happen due to the lack of appro-

priate construction or inadequate maintenance.
Finally it has to be underscored that progression
time of failure is one of the most crucial factors
of a successful defence. In this regards the evalu-
ation of different structural materials and failure
modes are essential: soils may behave differently
(erodability, loss of strength due to water content
changes etc.), hence the progression time varies in
a wide range.

2.4 Progression of failure

The breaching probability is usually approximated
as the probability of initiation of failure because
of the complex mechanisms of dike breaching and
the uncertainties in the development time. This
approximation leads to a conservative estimate.
For instance in Figure 4 a piping-caused breach-
ing scenario is illustrated. If the piping failure is
approximated by the initiation probability, then
only one criterion has to be met: the hydraulic gra-
dient during the flood event (i) must be higher than
the critical gradient (ic).
However the time factor must be considered
additionally. If the duration of the flood event is
Figure 3. Non-erosion related failure modes (based on short, then the time for the development of piping
Vrijling et al. 2011). will be limited, hence no breaching may take place.

253

ISGSR2013.indb 253 10/18/2013 9:39:50 AM

 On 15 January 1989 at 3.30 pm, a large 350 m
wide breach opened in the north bank of the Grand
Canal as its dike slid horizontally. The distance of
sliding was 60 m at the eastern end and decreased
gradually to zero in the western end of the breach.
The breaching caused the water level in the canal
to decrease rapidly. As a result of the rapid drawn
down, 200 m of the south bank was damaged.
Approximately 200,000 m3 of materials were dis-
placed during the failure. Based on the report of
an eyewitness, a loud “tearing noise” accompanied
the lateral displacement of the north bank.
Investigation was carried out including labora-
tory measurements, in-situ vane tests and other
probes before the remedial works. Based on the
Figure 4. Illustration of differences between breach- analysis, the groundwater level was close to or at
ing probability and initiation probability in the case of the ground surface, the frictional resistance of the
piping. peat was low. The most probable scenario is sliding
failure as in the case of the comprehensively stud-
ied Wilnis dike breaching (Baars 2005).
Although the use of the initiation probability is It might be assumed that the sliding failure was
common in quantitative risk analysis, there have not expected during the design of the dike. Baars &
been significant developments in the research of Kempen (2009) reveals that even in the Netherlands
breaching processes. Relevant papers (Zhu et al. the sliding failure is considered as an “impossible
2005, Visser et al. 2006) are published related to failure mode” though former Dutch and New
the overtopping erosion failure of sand and clay Orleans case studies highlighted the possibility.
dikes. In their works 5 stages of breaching process
are distinguished. The analysis assumes a homo-
3.3 Foenna stream dike, Sinalunga, Italy
geneous and trapezoidal breaching section. In the
case of internal erosion, ICOLD bulletin (ICOLD A dike breach occurred in Sinalunga on the bank
2013) and Chang & Zhang (2013) derive relevant of the Foenna stream due to animal burrows and
information to access the time factor of breaching piping during an ordinary flood event on 1 January
and the critical hydraulic gradients in each phase 2006 (Bayoumi & Meguid 2011, Camici et al.
of internal erosion, respectively. 2010). The sings of animal presence and activi-
ties were observed earlier, which led to extensive
maintenance along the dike. However during the
3 CASE STUDIES flood event an outflow was located on the down-
stream side of the dike, 2 m below the dike crest.
3.1 Objectives On the upstream slope the inlet a flow channel
The investigation of potential failure modes plays was observed approximately at the same elevation.
an important role in both risk assessment and Counter-measures were applied to stop the piping
ordinary design process. Four case studies are process: blankets were installed on the upstream
introduced in this section to illustrate how unfore- hole. Unfortunately there was not enough time
seen failure mechanisms (not expected during the to stop the internal erosion and the piping chan-
design) and unexpected conditions (animal bur- nel collapsed. Consequently the crest settled hence
rows, or gas development in the subsoil) led to the section was overtopped. The formation of the
breaching or failure of dike sections. trapezoidal breaching was fast; the town next to
the stream was inundated in an hour.
After investigations, it was verified that por-
3.2 Grand canal, Edenderry, Ireland
cupine burrows were responsible for the develop-
In the 18th century, a canal dike was constructed ment of the internal erosion channel, even though
close to Edenderry. The canal was formed into maintenance works were finalized before the flood
the crest of a peat embankment and its bed was event.
covered with a clay layer (Pigott et al. 1992). The The required maintenance and remedial works
height of the embankment was 10 m. The height must be performed to satisfy the design conditions.
of the two dikes formed by cutting was 2 m and However the Sinalunga case study reveals that
their crest width was 6 m. The inclination of the even the potential problems (animal burrows) were
slopes was about 1:4. tackled, neither investigation of the burrows nor

254

ISGSR2013.indb 254 10/18/2013 9:39:53 AM

 Figure 6. Development of failure in 17th Street Canal:
(1) Rotation of the I-wall and development of gap; (2)
Increase of water level and mobilization of the passive
earth pressure; (3) Development of critical slip surface,
backward progression of the slip surface and sliding/
rotation of the section.

Figure 5. Development of an unexpected branch

of the event tree due to animal burrows and lack of accompanied by horizontal displacements,
maintenance. to withstand the horizontal hydrostatic load.
Mean-while on the top of the clay surface the criti-
cal slip surface was formed and extended landwards
the maintenance was performed adequately, which to the ground surface through the marsh layer.
led to the change in piping resistance. The burrows Although the case study only focused on the
formed an unexpected branch of the event tree as 17th Street Canal dike, the failure mechanism
shown in Figure 5. was common in New Orleans. A gap between the
I-wall and the soil was formed both on the London
3.4 17th Street Canal, New Orleans, USA Avenue (North and South) and on the Orleans
Avenue (North and South). Therefore an impor-
The inundation of the major part of New Orleans tant threat was revealed related to I-wall type of
and the large number of breaches triggered a com- flood defence structures.
prehensive investigation after hurricane Katrina
in 2005. The final reports underscored relevant
3.5 Bleiswijk, The Netherlands
threats in different fields: potential failure mecha-
nisms, structural behaviour, erodibility of materials During August 1990 deformations were observed
etc. The case of the 17th Street Canal is presented on the downstream slope of a polder dike in
here to improve the understanding of the structural Bleiswijk, The Netherlands, which consisted of
behaviour and failure mechanism of the I-wall peat and clay layers. Due to the settlement of the
structures. In this case the breach was caused by section, the water level in the belt canal was close to
the horizontal sliding of the I-wall (flood protec- the crest. Moreover, intensive piping was detected
tion wall) and the embankment. and cracks were observed. These phenomena acti-
Around 6.30 am on 29 August 2005 a portion vated immediate counteractions and investigations
of the I-wall and embankment moved about 15 m of the site. Fortunately breaching did not occur.
horizontally and opened a gap of approximately The deformation mechanism was examined and
135 m. The breach released storm surge floodwa- the final results showed that the effect of draught,
ter, which flooded homes and destroyed buildings shrinking and oxidation of the peat caused higher
in the Lakeview area. As a part of the governmen- permeability and changes in the structure of the
tal investigation, field surveys were carried out peat material.
including auger borings and CPT tests. The site At another site during the survey of Bleiswijk
investigation was further supported by centrifuge dikes, water spurted out of a hole after the removal
tests (Steedman & Sharp 2011), the observed fail- of a penetration cone. To stop the fountain, a stand-
ure mechanism is presented in Figure 6. pipe was installed and after 1.5 hours the water level
Based on the centrifuge tests, when the flood- in the pipe started to drop. Investigation showed
water reached the bottom of the I-wall, the fail- that expansion of gas bubble inside the subsoil
ure initiated with a slight landward rotation of the caused the water spring. Gas bubbles formed dur-
I-wall. Hence a gap was formed between the foun- ing the oxidation of the peat. The origin and age
dation of the I-wall and the soil on the canal site. of the gas could not be ascertained therefore the
The gap reached the underlying clay layer in which bubble could have developed thousands of years
the foundation was located. Upon formation, the ago. Irrespective of the origin of the gas, its pres-
gap was immediately filled with water. ence influenced unfavourable effective stress condi-
When the water level rose, the passive earth tions and the pressurized gas could be responsible
resistance on the defended site had to be mobilized, for significant uplift forces (Vonk 1994).

255

ISGSR2013.indb 255 10/18/2013 9:39:53 AM

 Some threats are significant not only because ACKNOWLEDGEMENTS
they have influence on one distinguished failure
mechanism, but also because they may trigger The research work was supported by the National
other failure mechanisms. For instance the uplift Science Foundation of China (No. 51129902).
force decreases the sliding resistance due to the
decrease of the effective stress and hence the fric-
tional resistance. In addition the case reveals that REFERENCES
anomalies (gas bubble, and highly permeable soil
deposits in the river) may influence or govern the Bayoumi, A., Meguid, M.A. 2011. Wildlife and safety
failure modes. of earthen structures: a review. Journal of Failure
Analysis and Prevention, 11(4): 295–319.
Chang, D.S., Zhang, L.M. 2012. Critical hydraulic
4 SUMMARY gradients of internal erosion under complex stress
states. Journal of Geotechnical and Geoenvironmental
In the first phase of dike risk assessment, the Engineering, online, doi:10.1061/(ASCE)GT.
1943-5606.0000871.
potential threats must be identified and studied. Fukunari, K. 2008. The Tone river case and others.
Based on historical dike breaching cases (Nagy Safety of River Levee in Japan Conference, PowerPoint
2012, Baars & Kempen 2009), overtopping presentation, Tokyo, Japan.
(external erosion) is the most common cause ICOLD 2013. ICOLD Bulletin on Internal Erosion of
of breaching. The erosion of the outer slope, Existing Dams, Levees and Dikes, and their Founda-
breaching caused by human or animal activities tions, Volume 1: Internal Erosion Processes and Engi-
and sliding of the inner slope during flood event neering Assessment, ICOLD, Paris.
are the most serious threats in addition to the Nagy, L. 2012. Statistical evaluation of historical dike
overtopping. failure mechanism. Riscuri si Catastrofe, Ed.: Sorocov-
schi V., Editura Casa Cartii de Stiinta, Cluj-Napoca,
Different failure mechanisms have different X(2): 7–20.
development times, therefore the duration of the NSF 2006. Investigation of the Performance of New
flood event cannot be neglected. Since the warning Orleans Flood Protection Systems in Hurricane
and evacuation time is an important factor in Katrina on August 29, 2005, online, Berkeley
risk management, failure mechanisms must be University, USA.
differentiated based on the development time and Pigott, P.T., Hanrahan, E.T., Somers, N. 1992. Major
on the perceptibility of the process. In this respect, canal reconstruction in peat areas. Proceedings of
external erosion type of failures is less threatening the ICE—Water Maritime and Engineering, 96(3):
but internal erosion is more deceitful. 141–152.
Steedman, R.S. & Sharp, M.K. 2011. Physical modelling
Case studies revealed that breaching will most analysis of the New Orleans levee breaches. Proceed-
likely occur due to a well-known failure mechanism. ings of the ICE—Geotechnical Engineering, 164(6):
However, some of those failure mechanisms were 353–372.
neglected during design or risk analysis. Possible Van Baars, S., Van Kempen, I.M. 2009. The causes and
scenarios for such negligence include bad under- mechanisms of historical dike failures in the Netherlands.
standing of failure mechanisms/structural behav- E-Water, Official Publication of the European Water
iour, and anomalies along the dike. Even though Association, online.
the presence of gas bubbles is unlikely, crossings Van Baars, S. 2005. The horizontal failure mechanism of
with river bed with highly permeable soil depos- Wilnis peat dyke, Géotechnique, 55(4): 319–323.
Visser, P.J., Thu, Y., Vrijling, J.K. 2006. Breaching of
its may be faced. The highly permeable formation dikes. Proceeding of the 30th International Conference
may result in unexpected and extensive piping. on Coastal Engineering, Sand Diego, USA.
Another relevant failure driver might be failure of 2893–2905.
maintenance, which may lead to the development Vonk, B.F. 1994. Some aspects of the engineering prac-
of an unforeseen branch of the event tree. tice regarding peat in small polder dikes, Advances
Due to the rising sea level and climate changes in Understanding and Modelling the Mechanical
which aggravate weather extremes, the demand on Behaviour of Peat, den Haan, Termaat & Edil (eds).
detailed risk analysis and management of dike sys- Balkema, Rotterdam, 389–399.
tems is growing. Such demand is magnified by the Vrijling, J.K., Schweckendiek, T., Kanning, W. 2011.
Safety standards of flood defenses. Geotechnical
rapid growth of elements at risk in the flood zones. Safety and Risk, Proceedings of ISGSR 2011, Bunde-
Hence dike failure mechanisms must be studied sanstalt für Wasserbau, Münich, 67–84.
more carefully to support successful coastal and Zhu, Y., Visser, P.J., Vrijling, J.K. 2005. A model for
flood defence and ensure the security of the inhab- breach erosion in clay-dikes. Proceedings of Coastal
itants in the flood zones. Dynamics 2005: State of Practice: 1–10.

256

ISGSR2013.indb 256 10/18/2013 9:39:54 AM

 Geotechnical Safety and Risk IV – Zhang et al. (eds)
© 2014 Taylor & Francis Group, London, ISBN 978-1-138-00163-3

Reliability assessment of earth pressure on rigid non-yielding retaining

walls

S.M. Dasaka, T.N. Dave & V.K. Gade

Indian Institute of Technology Bombay, Mumbai, India

ABSTRACT: Pressure transducers are increasingly being used within soil mass or at soil-structure inter-
face to monitor stresses at the point of installation. Calibration of these transducers provides unique rela-
tionship between applied pressure and measured stress. Routinely, calibration is performed by applying
fluid pressure on diaphragm of transducer, however interpretation of data using fluid calibration results
lead to measurement errors when transducer is used in soil for stress measurement. This paper presents
comparison of calibration results of pressure transducer using fluid and soil calibration techniques. The
bias factor in the earth pressure measurement using fluid calibration results are evaluated, which can be
used with fluid calibration results to precisely arrive at the values of earth pressures on the retaining wall,
in the absence of in-soil calibration test results.

1 INTRODUCTION interface for measurement of in-situ stresses.

Hence, it is suggested that the pressure transducers
Retaining structures are integral part of various should be calibrated under the conditions identical
infrastructural projects to support deep excava- to its intended use.
tions, or steep and deep basements. These walls
withstand pressures from retained materials, sur-
charge pressures due to movement of vehicular 2 LITERATURE REVIEW
traffic or loads from foundations of the adja-
cent buildings on their backfills and loads due Various approaches have been adopted in the
to natural calamities like earthquake. A properly past for laboratory fluid calibration of pres-
instrumented structure gives information about sure transducers. Ramirez et al. (2010) cali-
variation of stresses with time and space. Pressure brated pressure transducers by subjecting them
transducers of different varieties are used to meas- to external oil pressure by means of dead weight
ure stresses within soil mass or at the interface of calibrator. The other approaches are (I) use of
structure and soil. centrifugal technique to calibrate pressure trans-
Calibration of transducers is very important as ducers (Take 1997, Chen & Randolph 2006),
the calibration factors obtained from calibration (II) use of application of fluid pressure to cali-
process would give an idea about actual stresses brate pressure transducer (Clayton & Bica 1993,
at the point of measurement. The calibration of Labuz & Theroux 2005, Ramirez et al. 2010). The
pressure transducer involves the investigation of fluid calibration helps in assessing instrument’s
the unique relationship between the applied pres- physical condition; however data interpretation
sure and pressure cell output (Take 1997). Through using standard calibration curve obtained using
calibration, the output from pressure transducers fluid calibration leads to measurement errors
is related to normal stress (multiplying the output when EPC is used in soil for stress measurement.
voltage (or strain) with the calibration factor will (Clayton & Bica 1993). The alternative but more
give rise to actual in-situ pressure, that converts complex procedure of in-soil calibration of EPC
cell’s electrical output to the pressure). To obtain may reduce this error considerably as a more
calibration factors, the standard procedure is to realistic calibration curve can be obtained (Selig
calibrate the transducer in a fluid (air, water or oil) 1980, Weiler & Kulhawy 1982, Selig 1989). The
and to analyze the unique relationship between reliable measurement of stresses in soil is still
the input and output. However, recent studies difficult to achieve due to strong dependency
highlighted that the results of fluid calibration of of measurement on relationship between EPC
Earth Pressure Cell (EPC) will be highly mislead- and soil stiffness (Hadala 1967, Hvorslev 1976,
ing, if the EPC are used in soil or at soil-structure Dunnicliff 1988).

257

ISGSR2013.indb 257 10/18/2013 9:39:54 AM

 Various factors affecting EPC output during in-
soil calibration were broadly classified into inclu-
sion effects, EPC-soil interaction, placement effects,
environmental influence and dynamic response
(Weiler & Kulhawy 1982, Dunnicliff 1988). The
literature is full of examples illustrating the effect
of placement method, soil density/stiffness, geom-
etry, grain size, loading history and soil type on the
calibration of EPC. Askegaard (1994) suggested to
test EPC under as varied conditions as possible to
get an estimate of the accuracy obtainable when
the EPC are used in practice in unknown material
and loading histories. In-soil calibration of EPC
involves application of uniform vertical stress to
the upper surface of a large soil specimen placed
Figure 1. Details of in-house calibration device devel-
inside a stiff-walled chamber.
oped by modification in triaxial set up.
Review of literature suggested that studies
related to effect of using fluid and in-soil calibra-
tion results on the measured earth pressure and
statistical studies in relation to evaluation of bias
factors is missing. Hence, the present study is
aimed at evaluation of effect of using fluid and in-
soil calibration results on the pressure cell output,
and its influence on the measured earth pressures.
Based on the above, bias factors are derived for
earth pressure, which can be used with fluid cali-
bration results to estimate the earth pressures on
the retaining wall, in the absence of in-soil calibra-
tion test results. Bias factor is defined as the ratio
of average pressure cell output using fluid calibra-
tion and that corresponds to in-soil calibration.
In the present study in-soil calibration of EPC
are carried out using sand.

3 EVALUATION OF CALIBRATION
FACTORS

3.1 Fluid calibration

In order to obtain calibration factors, a calibration
device was developed by modifying a conventional
triaxial apparatus suitable for testing 100 mm diam-
eter soil specimen, as shown in Figure 1. Further
details about the developed device can be seen in
Dave and Dasaka (2012a). The calibration device Figure 2. Pressure transducers used in present study
was fully filled with de-aired water, and pressure (a) type 1 (b) type 2.
was applied by an air-water bladder cylinder, using
compressed air with an accuracy of 0.1 kPa.
The fluid pressure in the triaxial cell was (Type 1) and miniature transducer (TML, Japan
increased so as to apply pressure on pressure make—Model No. PDA PA) of diameter 6.5 mm
transducer for calibration. Pressure in increments (Type 2) presented in Figure 2, was carried out
of 20 kPa was applied on transducer up to a using the developed calibration device. A detailed
maximum value of 180 kPa. From the maximum specification of both the transducers is presented
value, the pressure in the chamber was reduced by in Table 1. The fluid pressure in the triaxial cell was
decreasing the fluid pressure in 20 kPa steps back increased to a maximum of 180 kPa, in increments
to zero gage pressure. Calibration of two different of 20 kPa. The transducer data was recorded in
transducers, medium size transducer (Haris Sensor terms of the output strain from the transducer at
Technologies make, India) of diameter 40 mm each step of loading. Readings corresponding to

258

ISGSR2013.indb 258 10/18/2013 9:39:54 AM

 Table 1. Details of transducer used in the present
study.

Terminology Type 1 Type 2

Pressure range 0–2 kg/cm2 0–2 kg/cm2

Sensitivity 1.420 mV/V at FS +946 μV/V
Non-linearity & 0.5% of FS 0.5% of FS
hysteresis
Thermal sensitivity 0.1% of FS/°C 1%/°C
shift
Dimensions 40 mm Φ × 6.5 mm Φ ×
10 mm 1 mm

each loading were obtained once the fluid pressure Figure 3. Particle size distribution of sand.
had stabilized, typically in 30 seconds.
The relationship between applied pressure and
acquired data (measured in terms of micro strain) Table 2. Physical and mechanical properties of sand.
and measurement non-linearity for both the trans-
Property Value
ducers was obtained and presented in Table 1.
Gs 2.65
3.2 In-soil calibration Cu 1.42
Cc 0.93
Effect of sand layer thickness on the performance γdmin (kN/m3) 14.58 (ASTM D4254-00)
of Type 2 transducer is observed. Same modi- γdmax(kN/m3) 17.10 (Pluviator)
fied triaxial setup with additional plastic tube of Cohesion, kPa (c′)* 0
2.5 mm thickness and 100 mm Internal Diameter Friction angle (φ′)* 39°
(ID) placed tightly on the brass pedestal was
used with the transducer fixed flush on it. In the *Obtained from direct shear test at 68% relative density.
present study silica sand (Indian Standard sand
of Grade III) was used. The typical particle size
distribution curve of the sand used in the study (Type 1 and Type 2) were compared, discussed in
is shown in Figure 3. Some of the physical and the following section.
mechanical properties of the sand used in the
present study are presented in Table 2.
Considering diameter of EPC (DEPC) as refer- 4 RESULTS AND DISCUSSION
ence, sand layer thicknesses of 10 mm (0.25DEPC),
20 mm (0.5DEPC), 40 mm (DEPC), 60 mm (1.5DEPC) Calibration performances of Type 1 and Type
and 100 mm (2.5DEPC) were placed above the EPC 2 transducers in the range of 0–180 kPa for first
to obtain relation between applied pressure and loading-unloading cycle were obtained. Typical
measured strain. Greased polyethylene sheets of fluid calibration results are presented in Figure 4.
60 μm thickness were pasted to inner surface of The fluid calibration results of first loading-un-
plastic tube in order to reduce friction between loading cycle indicated maximum non-linearity of
sand particles and plastic tube surface. A rigid 0.6% of FS (full scale) and 2.64% of FS for Type 1
wooden block followed by thick rubber pad was and Type 2 transducers, respectively.
placed on sand layer for proper transfer and uni- The calibration factors obtained from fluid cali-
form distribution of pressure between load cell bration of these two pressure cells are presented in
of triaxial apparatus and transducer, as shown in Table 3.
Figure 1. In case of in-soil calibration, the relation
Air pluviation was adopted for preparation of between applied pressure and measured strain
sand layer of 16 kN/m3 unit weight. Incremental is non-uniform for sand layer thickness of 0.25
displacements were applied manually to the triaxial DEPC and 0.5 DEPC. As the pedestal body was more
base, thereby monitoring applied stress on wooden rigid than diaphragm of EPC, arching of sand
block through load cell and induced strains in the might have caused non-uniform transfer of pres-
transducer. Strains were measured upto 50 kPa sure, in line with observations by Ingram (1968).
pressure. Results obtained from both fluid calibra- With increase in the sand layer thickness effect of
tion and in-soil calibration of both pressure sensors arching would have been reduced, thus allowing

259

ISGSR2013.indb 259 10/18/2013 9:39:56 AM

 Figure 5. Comparison of fluid and in-soil calibration
Figure 4. Typical fluid calibration results. results.

Table 4. Results of in-soil and fluid calibra-

Table 3. Fluid calibration factors obtained using in- tion of transducers.
house calibration device (Dave & Dasaka, 2012a).
Sand
Type 1 Type 2
Fluid 0.25DEPC DEPC 1.5DEPC
Best fit equation Y = 2.0534*X Y = 2.4672*X
Calibration factor 0.4868 0.4053 372 259 298 301
Coefficient of 0.9998 0.9999
determination (R2)
with fluid calibration factors, would understand
*Where, Y = measured strain (με); X = Applied pressure the earth pressures in sand by as much as 25%. As
(kPa).
the pressure sensors are routinely calibrated using
a fluid, either due to lack of information of their
intended use, or unavailability of in-soil calibra-
uniform transfer of pressure from load cell to EPC, tion devices, only fluid calibration factors are sup-
and thereby increasing output of EPC. Outputs plied to the user. These fluid calibration factors, if
using sand thicknesses of DEPC and 1.5DEPC were used, grossly underestimate the earth pressures on
almost similar and on higher side of that obtained the retaining walls. In the absence of in-soil cali-
for 0.25DEPC and 0.5DEPC. However, with further bration results, the earth pressure measurements
increase in thickness of sand layer beyond 1.5DEPC, in sand, employing fluid calibration results, need
EPC output was reduced, which may be attributed to be scaled up by 25%, to arrive at the precise
to stress dissipation inside the soil causing only a values.
portion of applied pressure to transfer to the dia- It is to be noted that the bias factors developed
phragm of the EPC. in this study are only applicable for the grain size
Figure 5 presents the fluid and in-soil calibra- distribution of the sand used in this study. Bias fac-
tion test results for Type 1 EPC. For any value of tors may vary depending on the grain size distribu-
applied pressure the induced strains in the EPC are tion of the material. Hence, further studies in this
higher in fluid calibration. direction are essential, involving different materials
Table 4 shows the pressure cell output corre- (sand, silt, and clay) with wide range of grain size
sponding to an applied pressure of 180 kPa, for distribution, to critically understand the influence
both fluid and in-soil calibration of EPC (Type 1). of material type and its grain size distribution on
It is evident from the above results that higher the bias factors.
calibration factor (≈180/372 = 0.4839) is obtained
for the EPC using fluid calibration, than that
obtained using in-soil calibration (in the range of 5 CONCLUSIONS
0.6962–0.8091, for sand thickness in the range of
10 mm–60 mm). Considering the fact that the Interpretation of earth pressures in sand using
optimum sand layer thickness is around 1.5DEPC, EPC, employing the fluid calibration results,
the bias factor for Type 1 and Type 2 EPC are should be handled carefully, as EPC responds
evaluated as 1.25. It means that these EPC, if used differently with fluid and sand. Following are

260

ISGSR2013.indb 260 10/18/2013 9:39:57 AM

 the major conclusions derived from the present Dave, T.N. & Dasaka, S.M. 2012c. In-house calibration
studies. of pressure transducers and scale effect on calibration
Optimum sand layer thickness to be placed factors. Geomechanics and Engineering—An Interna-
above the EPC during the calibration studies for tional Journal, Techno Press 5(1): 1–15.
Dunnicliff, J. & Green, G.E. 1988. Measurement of total
Type 1 and Type 2 EPC is obtained as 1.5DEPC, stress in soil. Geotechnical instrumentation for moni-
where DEPC is the diameter of the diaphragm of toring field performance, Wiley, New York, 165–184.
EPC. Hadala, P.F. 1967. The effect of placement method on the
For any value of applied pressure the induced response of soil stress gages. Technical Report No.
strains in the EPC are higher in fluid calibration 3–803, Army Eng. Waterways Expt. St., Vicksburg,
than that in the in-soil calibration. MS.
The bias factor for Type 1 and Type 2 EPC are Hvorslev, M.J. 1976. The changeable interaction between
evaluated as 1.25. soils and pressure cells. Tests and Reviews at the
Calibration factors should be obtained by in-soil Waterways Experiment Station, Technical Report
S-767, Army Eng. Waterways Expt. St., Vicksburg,
calibration of EPC, under the conditions similar to MS.
that prevail at the place of intended use for obtain- Labuz, J.F. & Theroux, B. 2005. Laboratory calibration
ing reliable results. of earth pressure cells. Geotechnical Testing Journal
28(2): 1–9.
Ramirez, A., Nielsen, J. & Ayuga, F. 2010. On the use
REFERENCES of plate type normal pressure cells in silos part 1:
Calibration and evaluation. Computers and Electron-
Askegaard, V. 1994. Applicability of normal and shear ics in Agriculture 71: 71–76.
stress cells embedded in cohesionless materials. Exper- Selig, E.T. 1980. Soil stress gauge calibration. Geotechnical
imental Mechanics 35(4): 315–321. Testing Journal 3(4): 153–158.
Chen, W. & Randolph, M. 2006. Measuring radial total Selig, E.T. 1989. In situ stress measurements, State-of-
stresses on model suction caissons in clay. Geotechnical the-Art of pavement response monitoring systems
Testing Journal 30(2): 1–9. for roads and airfields, U.S. Army Cold Reg. Res. and
Clayton, C.R.I. & Bica, A.V.D. 1993. The design Engg. Lab., Hannover, NH.
of diaphragm-type boundary total stress cells. Take, W.A. 1997. Lateral earth pressure behind rigid
Geotechnique 43(4): 523–536. fascia retaining walls, M.S. Thesis, The Uni. of New
Dave, T.N. & Dasaka, S.M. 2012a. Assessment of Brunswick, Canada.
portable traveling pluviator to prepare reconstituted Weiler, W.A. & Kulhawy, F.H. 1982. Factors affecting
sand specimens. Geomechanics and Engineering—An stress cell measurements in soil. Journal of Geotechni-
International Journal, Techno Press 4(2): 79–90. cal Engineering Divison 108(12): 1529–1548.
Dave, T.N. & Dasaka, S.M. 2012b. Universal calibration
device for fluid and in-soil calibration of pressure
transducers. Indian Geotechnical Journal—Springer
Publishers 42(3): 212–219.

261

ISGSR2013.indb 261 10/18/2013 9:39:58 AM

This page intentionally left blank
 Geotechnical Safety and Risk IV – Zhang et al. (eds)
© 2014 Taylor & Francis Group, London, ISBN 978-1-138-00163-3

Seismic damage occurrence probability in an existing estuary dam

T. Hara, M. Iwata, Y. Otake & Y. Honjo

Gifu University, Gifu, Japan

T. Kato, A. Nishida & H. Yukimoto

Japan Water Agency, Saitama, Japan

ABSTRACT: This paper presents an example of S-PSA, Seismic Probabilistic Safety Assessment,
applied to an existing estuary dam. Severe damage probabilities of respective parts constituting the dam,
which are estimated by reliability analysis with using geotechnical analysis (FEM) and statistical analysis
of uncertainties, are described. Finally, the seismic retrofit planning based on risk is discussed.

1 INTRODUCTION 2 TARGET STRUCTURE

Seismic design with considering occurrence prob- The target facility is an existing estuary dam to
ability of larger earthquakes than the design one supply city, industrial and agricultural water due
has been an important issue in Japan, because sev- to prevent running up of seawater to the river. The
eral large earthquakes caused tremendous damages facility length is 192.3 m and it consists of six gates
to human and social economy, such as Kobe quake (25 m in width, 7.3 m in height) and seven hanging
in 1995 and Tohoku quake in 2011. up structures of the gates. Figure 1 shows the exter-
Therefore, the authors have conducted a study nal appearance of the dam. The target structure in
on application of S-PSA, Seismic Probabilistic this study is the hanging up structures of the gates.
Safety Assessment (IAEA 1993, 2009), to com- This structure is built on the liquefiable ground
mon infrastructures except nuclear plants. In this during earthquake with 35 steel pipe piles (600 mm
paper, an example of S-PSA applied to an exist- in diameter, 9 mm in thickness, 25 m in length).
ing estuary dam is presented. Dynamic effective
stress FEM analysis is adopted in order to evaluate
the seismic response of the dam built on liquefi- 3 PROCEDURE FOR S-PSA
able ground during earthquake by using response
surface method (Honjo 2011) in this S-PSA. In Figure 2 shows the procedure for the S-PSA in this
the reliability analysis, following uncertainties are study. At first, the uncertainty of seismic intensity
considered. is set as a probabilistic density function from the
Response uncertainties: Earthquake intensity published probabilistic seismic hazard curve at the
based on probabilistic hazard curve at the site, site, which corresponds to residual service term of
reproducibility of the FEM analysis adopted in the target facility. Then, Response Surfaces (RS)
this study with respect to experimental responses
of the structure built on liquefied ground, and
reproducibility of the assumed response surface
with respect to the seismic responses estimated by
FEM analysis.
Resistance uncertainties: Strength estimation of
the structural members, bearing capacity of pile
foundation, spatial variation and transformation
error of geotechnical parameters.
Details of the S-PSA with using FEM analysis
as well as severe damage probabilities of the respec-
tive parts constituting the dam are introduced and
the seismic retrofit planning based on risk is dis-
cussed in this paper. Figure 1. External appearance of the target dam.

263

ISGSR2013.indb 263 10/18/2013 9:39:59 AM

 Figure 3. Probabilistic seismic hazard curve at the site.

Figure 2. Procedure for S-PSA.

of respective verification points are estimated

from sensitivity analyses with respect to contribu-
tion factors to each response of the points. At last,
damage occurrence probabilities of the verification
points are estimated by Monte Carlo Simulation
(MCS) with both uncertainties of response and
resistance. Safety verification is conducted from
the comparison with allowable damage occurrence
probability determined by one of recently built
similar structures, and seismic retrofit is studied
from the cost effectiveness of countermeasures,
which are estimated from viewpoint of both risk
reduction and cost.

4 UNCERTAINTY OF SEISMIC
INTENSITY Figure 4. Uncertainty of seismic intensity at the site.

4.1 Uncertainty of seismic intensity

Table 1. Seismic intensity uncertainty.
Residual service term of the target facility, “30-
year”, is determined as a basic term though, “50- Design term Mean (cm/s) SD (cm/s) Distribution
year” is also considered in this study from a concern
of the safety in case that the term is postponed. 30-years 34.9 19.4 Lognormal
The probabilistic seismic hazard curve cor- 50-years 42.6 20.2 Lognormal
responding to 30- and 50-year term at the target
site (J-SHIS 2010) is shown in Figure 3. Figure 4
and Table 1 present the probabilistic density func-
tion, the uncertainty of seismic intensity at the site, engineering basement defined by shear wave veloc-
assumed from the hazard curve. ity at each area of 250 m pitch, have been pub-
lished from the government in Japan. The seismic
wave caused by Tonankai-Nankai earthquake at
4.2 Basic seismic wave
the target site (COGJ 2006), shown in Figure 5,
Seismic waves with respect to several specific is adopted in this study as the basic seismic wave.
earthquakes, which are predicted to occur on the Although this wave, maximum acceleration and

264

ISGSR2013.indb 264 10/18/2013 9:39:59 AM

 Figure 5. Basically seismic wave.

velocity are 180 cm/s2 and 38 cm/s, is not so large,

it had been adopted for seismic verification of the
target facility until publishing the new wave in 2013
after occurrence Tohoku earthquake in 2011. The
seismic effect in this study is considered from the
basic seismic wave calibrated by the uncertainty of
seismic intensity.

5 RESPONSE SURFACE (RS) ESTIMATION

5.1 Response estimation method

Effective stress dynamic FEM analysis, FLIP (Ver.
6.0.6), is adopted in this study, in order to evaluate
the seismic response of the dam built on liquefi-
able ground during earthquake, by using response
surface method (Honjo 2011). Figure 6 shows
the FEM mesh of both river axis cross section
and dam axis one. Ground condition of top three
strata is presented on Table 2 and As2 stratum is the
liquefiable ground during earthquake.
The uncertainty of the FEM reproducibility
with respect to the experimental responses of struc-
ture built on liquefied ground during earthquake,
which were obtained from blind tests (JICE 2009),
is shown in Figure 7. Because of blind test, the
transformation error of the geotechnical parame-
ters used in the FEM is included in the uncertainty.
Although the experimental uncertainty was
obtained from FEM reproducibility with respect
to experimental results of river dike deformation,
it was referred to one of the response estimation of Figure 6. FEM mesh.
pile structures as a safety evaluation in this study.

5.2 Verification points Table 2. Ground condition.

A lot points have to be verified in normal seismic SPT-N value

design though, in this study, the minimum verifi-
cation points, which possess the possibility that Thickness (m) Mean SD
the response corresponding to increscent seismic
As2 14.5 15.2 4.4
intensity exceeds the limit state, are selected from
Ac3 7.0 12.4 3.0 (1.2)
sensitivity analyses of responses at each point. As3 10.0 33.4 10.4 (4.4)
The selected verification points are presented on
Table 3. ( ) in SD: converted ones for estimation of pile bearing
In this paper, because of paper limitation, safety capacity with using the proposed method by Otake &
assessments of gate pier shear force in the dam axis Honjo (2012).

265

ISGSR2013.indb 265 10/18/2013 9:40:01 AM

 Figure 7. Uncertainty of FEM reproducibility.

Table 3. Verification points.

Verification section
Figure 8. RS of maximum shear force of gate pier.
River axis Dam axis

Gate pier Curvature Verify N/A Table 4. Uncertainty of shear strength.

Shear force Verify Verify
Dam pier N/A N/A Mean SD Distribution
Pile Displacement Verify Verify
Entirely structural inclination Verify Verify 1.93 0.273 Normal

6 LIMIT STATE UNCERTAINTY

cross section with respect to 30-years term is only
introduced. The uncertainty of shear strength is presented
on Table 4, which was referred from the exist-
5.3 Response surfaces ing research on the reproducibility of the shear
strength estimated by Equation (2) to (4) with
It was confirmed that the uncertainty of seismic respect to experimental ones (Watanabe et al.
intensity is only influential one to the response of 2007).
the respective verification points from sensitivity
analyses with respect to uncertainties of seismic Ps Sc + Ss (2)
intensity, ground condition, deformation charac-
teristic of gate pier and pile. Therefore, Response
Sc c c ⋅ c e c pt ⋅ τ c ⋅ b d (3)
Surfaces (RSs) focused on only seismic intensity
are estimated. Equation (1) and Figure 8 show RS
of gate pier shear force in the dam axis cross section Aw sy d ⋅( + )
Ss = (4)
and its uncertainty of reproducibility with respect 1.15 ⋅ α
to the responses estimated by FEM analyses.
where, Ps = shear strength of reinforced concrete
T 1 2985 Vm + 559.95 (1) structure (N), Sc = shear strength boar by con-
crete (N), Ss = shear strength boar by reinforcing
where, T = maximum shear force of gate pier (kN), bar (N), cc, ce and cpt = compensation coefficients
Vm = maximum velocity of seismic wave (cm/s). concerning cyclic loading action, effective height

266

ISGSR2013.indb 266 10/18/2013 9:40:03 AM

 of the section (d) and, ratio of axial tensile where, δT, δV, δFEM, and δRS = random variables pre-
reinforcing bars’ area and sectional one, τc = mean sented on Table 5, Vm = maximum seismic velocity
shear stress boar by concrete (N/mm2), b = section (cm/s).
width (mm), Aw = area of shear reinforcing bars
(mm2), σsy = yield strength of shear reinforcing bar
(N/mm2), θ = angle of shear reinforcing bar and 8 SAFETY VERIFICATION
vertical axis (deg), α = pitch of reinforcing bars
(mm). 8.1 Allowable damage occurrence probability
Damage occurrence probability of the recently
built similar structures (Osumi et al. 2000, Unjo &
7 DAMAGE OCCURRENCE PROBABILITY Zhang 2009, Nakatani et al. 2009, Komori et al.
2011) is referred to specify the allowable ones for
7.1 Reliability analysis the target facility in this study.
Damage occurrence probability is estimated
by MCS with one million runs with respect to 8.2 Safety verification
the uncertainties of seismic intensity, repro- The comparison of the allowable damage occur-
ducibility of FEM with respect to experimen- rence probabilities and estimated ones of all veri-
tal responses of the structure built on liquefied fication points is presented on Tables 6 and 7. In
ground, reproducibility of RS with respect to these Tables, the values noted in the brackets are
the responses estimated by FEM, and the limit the damage probability in 50-year term. According
state uncertainties. to the results, shear collapse occurrence probabil-
ity of the gate column in dam axis cross section
7.2 Shear collapse of gate column (dam axis) is only not satisfied the allowable one. Therefore,
seismic retrofit for shear collapse of the gate col-
Equation 5 expresses the performance function. umn is studied based on risk in this paper.
The shear collapse occurrence probability of gate
column (dam axis) for 30-year term, 0.27%, is
shown in Figure 9. Table 6. Comparison of estimated damage occurrence
probabilities and allowable ones (dam axis, 30- and
1.2985 ⋅ δ v ⋅ Vm ⎞ 50-year term).
g δT T − ⎛ ⋅δ ⋅δ ≥ 0 (5)
⎝ +559.95 ⎠ FEM RS Damage probability (%)

Estimated Allowable Judge

Shear collapse 0.27 0.15 NG

of gate column (0.29)
Group pile 2.7E-3 0.35 OK
displacement (3.1E-3)
Residual inclination <1.0E-4 0.15 OK
of entire structure (<1.0E-4)

Table 7. Comparison of estimated damage occurrence

probabilities and allowable ones (river axis, 30- and
Figure 9. Shear collapse occurrence probability. 50-year term).

Damage probability (%)

Table 5. Random variables for estimation of shear col-
lapse occurrence probability (30-year term). Estimated Allowable Judge

Content Mean SD Distribution Flexural collapse 1.0E-4 0.15 OK

of gate column (2.0E-4)
δT Shear strength 1.93 0.27 Normal Shear collapse of 7.7E-3 0.15 OK
δV Maximum seismic 37.6 19.4 Lognormal gate column (0.01)
velocity (cm/s) Group pile 7.0E-4 0.35 OK
δFEM FEM 1.11 0.33 Lognormal displacement (9.0E-4)
reproducibility Residual inclination 1.0E-4 0.15 OK
δRS RS reproducibility 1.0 0.038 Normal of entire structure (1.0E-4)

267

ISGSR2013.indb 267 10/18/2013 9:40:05 AM

 9 SEISMIC RETROFIT STUDY

9.1 Risk estimation

Damage summation cost caused by shear collapse
of the gate column, which estimated by an assump-
tion of recovery term of 120 days, is presented on
Table 8. The risk with the damage, 113 million JPY,
is estimated as the production of damage cost and
occurrence probability of 0.27%.
Figure 10. Comparison of damage probabilities before
and after the execution of WRC.
9.2 Seismic countermeasures
Three countermeasures for seismic retrofit of the
Table 10. Cost effectiveness of respective counter-
gate column, which are Wrap Reinforced Concrete,
measures.
Wrap Steel Plate around the column (WRC, WSP)
and Spray Polymer Cement Mortal to the column WRC WSP SPM
(SPM), are firstly chosen. Then residual damage
probabilities after execution of each countermeas- B (million JPY) 101.9 105.0 96.4
ure are estimated as respective measure effects. (109.5) (113.0) (103.4)
Table 9 presents cost of each countermeasure and C (million JPY) 60 115 92
residual damage probability of respective coun- B/C 1.70 0.91 1.05
termeasures. Where, the reason why the residual (1.83) (0.98) (1.12)
damage probabilities of each countermeasure are
B and (B), B/C and (B/C): ones in 30- and (50)-year
smaller than the allowable one is to consider the
term.
minimum requirement of respective countermeas-
ure specifications. The comparison of damage
probabilities before and after execution of WRC is
the results, WRC is desirable one for the counter-
shown in Figure 10 for instance.
measure because of the highest cost effectiveness.

9.3 Countermeasure selection

10 CONCLUSIONS
Cost effectiveness (B/C) of respective countermeas-
ures is presented on Table 10. Where, Benefit (B) is
A concrete example of S-PSA application with
the difference of the damage risk before and after
using FEM analysis to an existing facility and
execution of each countermeasure. According to
seismic retrofit planning based on risk were intro-
duced in this paper. This paper can be concluded
as follows;
Table 8. Damage summation cost.
a. Application of S-PSA to common infrastruc-
Damage cost ture is not so difficult.
(million JPY) b. S-PSA with using FEM analysis is possible by
using response surface method.
Recovery cost 700 c. The uncertainties of FEM reproducibility with
City water damage 1,750
respect to experimental results, RS reproduci-
Industrial water damage 31,200
bility of responses estimated by FEM, and limit
Agricultural water damage 5,050
states as well as uncertainty of seismic intensity
Summation 38,700
are considered in the S-PSA.
d. Seismic retrofit planning based on risk with
using S-PSA result is effectiveness for counter-
Table 9. Cost and effectiveness of countermeasures. measure selection. This approach can be used
more effectively to make priority decision for
WRC WSP SPM seismic retrofit planning of multiple structures
despite its building region, structural type and
Cost (million JPY) 60 115 92 so on.
Damage probability 8.0E-3 0 (0) 0.022
after execution (%) (9.0E-3) (0.025) The authors are expecting to be applied S-PSA
and S-PRA to common infrastructures from
Prob. and (Prob.): ones in 30- and (50)-year term. the viewpoints of accountability to public and

268

ISGSR2013.indb 268 10/18/2013 9:40:06 AM

 reasonable use of limited public investment to Komori, N., Hoshikuma, J. & Sakai, J. 2011. A study on
seismic retrofit. evaluation of seismic limit state of reinforced concrete
bridge pier: Proc. of the 14th Symp. on Performance
Based Seismic Design on Highway Bridges, 329–336.
Nakatani, S., Shirato, M., Kono, T., Nakamura, Y.,
REFERENCES Nomura, T., Yokomaku, K. & Iochi, H. 2009. A study
on stability verification method of bridge foundation
Cabinet Office, Government of Japan (COGJ). 2006. based on performance based design concept: Techni-
Tonankai-Nankai seismic (acceleration) wave on engi- cal Note of Public Works Research Institute (PWRI),
neering basement (Vs = 400 m/s): www.bousai.go.jp/ No. 4136.
jishin/chubou/.html. Osumi, M., Unjo, S., Adachi, S. & Hoshikuma J. 2000.
Honjo, Y. 2011. Challenges in geotechnical reliability A study on seismic reliability evaluation of highway
based design: Proc. of the 3rd Int. Symp. on Geotech- bridges: proc. of the 4th Symp. on Seismic Design
nical Safety And Risk, 11–27, Munich, 2–3 June 2011. Based on Ductility Design Method for Bridges:
Germany: Bundesanstalt fur Wasserbau. 13–16.
International Atomic Energy Agency (IAEA). 1993. Otake, Y. & Honjo, Y. 2012. Verification of a simplified
Probabilistic safety assessment for seismic events. geotechnical reliability analysis scheme of spatial vari-
International Atomic Energy Agency (IAEA). 2009. ability based on local average of geotechnical param-
Evaluation of seismic safety for existing nuclear eters: JSCE Journal C 68 (3): 475–490.
installations. Unjo, S. & Zhang, K. 2009. A study on partial factor
Japan Institute of Construction Engineering (JICE). design for highway bridge seismic design: Report of
2002. Analysis method of seismic river dike deforma- road development research by government subsidy.
tion: JICE Document, Vol. 102001. Watanabe, H., Koga, H. & Nakamura, H. 2007. A study
Japan Road Association (JRA). 2012. Specifications for on installation of International standard of design
highway bridges Part V seismic design. on concrete structure: Report of general research by
Japan Road Association (JRA). 2012. Specifications for government subsidy.
highway bridges Part IV substructures.
Japan Seismic Hazard Information Station (J-SHIS).
2010. Seismic hazard data, www.j-shis.bosai.go.jp/
map/download.

269

ISGSR2013.indb 269 10/18/2013 9:40:07 AM

This page intentionally left blank
 Geotechnical Safety and Risk IV – Zhang et al. (eds)
© 2014 Taylor & Francis Group, London, ISBN 978-1-138-00163-3

Probabilistic analysis of dry soil mix columns

J. Huang, R. Kelly & S.W. Sloan

ARC Centre of Excellence for Geotechnical Science and Engineering, The University of Newcastle, Australia

ABSTRACT: Analytical probabilistic analysis and Monte Carlo simulation based on elasto-plastic
Finite Element Method (FEM) on dry soil mix columns are presented. It is shown that analytical method
is over conservative because it ignores the supports from adjacent columns. Probabilistic FEM analysis
can provide more accurate predictions, and thus lead to more economic designs. Probabilistic FEM analy-
ses show that the effects of adjacent columns can be destructive when applied load is close to the strength.
The reliability of the system of columns is analyzed by setting residual strength to zero. Results show that
close spacing has more safety margin than loose spacing.

1 INTRODUCTION cell approach is actually a One Dimensional (1D)

estimate. The advantage of this 1D estimate is that
Dry Soil Mix (DSM) columns are a method of it allows a straight forward probabilistic assessment
ground improvement used to strengthen and of column performance, which will be shown later
stiffen soft to firm cohesive soils having an und- in this paper. Its disadvantage is that the effects
rained shear strength of about 30 kPa or less. The from adjacent columns are ignored, which may lead
strength and stiffness of the columns is a function to over conservative design. In this paper we use the
of the ratio of water to cement and the quantity probabilistic finite element method (Griffiths et al.
of mixing that occurs. The constructed columns do 2011) to perform preliminary assessments for the
not have uniform strength within a single column effects of DSM column variability on embankment
or between columns due to the inherent variability settlement and column collapse performance.
of the ground along with non-uniform cement feed
rates. The coefficient of variation of strength can
range between 0.34 and 0.79 (Filz & Navin 2010). 2 DETERMINISTIC ANALYSIS
Experience shows that adjacent columns can have
very different strengths and stiffness even though Initially, a deterministic design was performed to
they are installed in the same ground using near calculate the spacing of columns required to resist
identical methods. Variability is more typically the embankment load. The DSM columns are
controlled by specifying tight acceptance crite- treated as an elastic material having a constant ratio
ria for column strength and stiffness along with of modulus to strength ( /Cu ) and have a range of
detailed quality control procedures in the con- strengths and centre to centre spacings in a square
struction documents. These requirements force the pattern. Assuming that the columns have a diam-
ground improvement contractor to install columns eter of 0.8 m, a length of 10 m, the embankment
with higher strength and stiffness than required by has a thickness of 3 m, the fill has a unit weight of
design in order to reduce the risk of performing 20 kN/m3, the columns have a stiffness to strength
rework to replace non-conforming columns. ratio of 200 and the spacing of the columns is con-
For a road or rail embankment the columns trolled so that the load applied to the columns is
are installed individually below the crest of the 75% of their strength (factor of safety is 1.33). By
embankment to control settlement and in panels using the unit cell method, various combinations
or grids below the batters to provide stability. For of column strength and spacing that satisfies the
settlement control the philosophy is the entire soil design criteria are summarised in Table 1. The
mass can be considered as having improved proper- combination of column strength and spacing are
ties even though the columns are installed at dis- then adopted in the probabilistic assessment.
crete intervals. This leads to design methods where In FEM analyses, soil fill is modelled by 20 nodes
equivalent strength and stiffness parameters for the brick elements. Soil fill is assumed to be elastic with
improved ground are computed using a unit cell Young’s modulus of 20 MPa and Poisson’s ratio of
approach and assuming equal strain conditions 0.3. Columns are modelled as spring elements and
apply throughout the soil (e.g, Kelly & Wong 2011). placed beneath every corner node of brick elements.
Since equal strain conditions are assumed, the unit One end of the springs is connected to the brick

271

ISGSR2013.indb 271 10/18/2013 9:40:07 AM

 Table 1. Spacing by FEM and 1D estimate.

Column Spacing (m) Spacing (m)

strength (kPa) by FEM by 1D

100 1.114 1.118

150 1.368 1.369
200 1.582 1.581
250 1.769 1.768
300 1.938 1.936
500 2.500 2.500

Figure 3. Deterministic settlements (column resistance

is 500 kPa, spacing is 2.5 m).

3 PROBABILISTIC ELASTIC ANALYSIS

The stiffness of columns E is assumed to be a log-

normally distributed random variable with coeffi-
cient of variation (VVE ) of 0.5. The probability that
S exceeds 0.0825 m can be estimated by 1D analy-
sis if the effects of adjacent columns are ignored.
Since the spacing was decided so that the settle-
ment is 0.075 m, we have
Figure 1. Deterministic settlements (column strength is WL
100 kPa, spacing is 1.114 m). = 0.075 (1)
μE A

where W = γHD2 is the applied load on the col-

umns, L is the length of the columns, μE is the
mean column stiffness and A is the cross sectional
area of the columns.
The probability that S exceeds 0.0825 m is (e.g.,
Griffiths et al. 2009)

P (S > ) = P ⎛⎝WL > 0.0825⎞⎠

EA
⎛ 0.075 μE ⎞
=P > 0.0825
⎝ E ⎠
⎛ μ ⎞
Figure 2. Deterministic settlements (column strength is =P E< E
⎝ 1.1⎠
250 kPa, spacing is 1.769 m).
⎛ ⎛ μE ⎞ ⎞
ln − μlln E ⎟
= Φ ⎜ ⎝ 1.1⎠
elements. The other end is fixed. Rollers are put at ⎜ ⎟
four sides. Since the performance of the column in ⎝ σ lln E ⎠
the middle is concerned, the number of columns is ⎛ ⎛ μE ⎞ 1 ⎞
decided so that the effect of boundary conditions on ⎜ ln − l μE + ln( + 2
)⎟
=Φ ⎝ 1.1⎠ 2 E
this column is negligible. Spacings shown in Table 1 ⎜ ⎟
are decided so that the settlement in the middle ( ) ⎜⎝ ln( + E 2 ) ⎟⎠
is 0.075 m. The spacing by FEM is very close to
⎛ ⎛ 1⎞ 1 ⎞
1D estimate, which means the boundary effects are
⎜ ln + ln(1 + VE 2 )⎟
negligible. Figures 1–3 show the settlement profiles =Φ ⎝1 .1⎠ 2
⎜ ⎟
in which soil settlements are included. It can be seen ⎜⎝ ln(1 +V
VE 2 ) ⎟⎠
that the settlement at the boundaries are smaller
than that in the middle. This is because of the reac- = 0.51
tive bending moments caused by the rollers. (2)

272

ISGSR2013.indb 272 10/18/2013 9:40:07 AM

 where Φ is the cumulative standard normal distri-
bution function.
In order to investigate the effects from adjacent
columns, two thousand Monte Carlo simulations
were carried out. Three typical simulations are
shown in Figures 4–6. The histograms of settle-
ments are shown in Figures 7–9.

Figure 7. Histogram of settlement (column strength is

100 kPa, spacing is 1.114 m).

Figure 4. A typical simulation (column strength is

100 kPa, spacing is 1.114 m).

Figure 8. Histogram of settlement (column strength is

250 kPa, spacing is 1.769 m).

Figure 5. A typical simulation (column strength is

250 kPa, spacing is 1.769 m).

Figure 9. Histogram of settlement (column strength is

500 kPa, spacing is 2.5 m).

The probability of settlement exceeds 0.0825 m is

compared to 1D estimate in Figure 10. It can be seen
from Figure 10 that the probability of the settlement
exceeds 0.0825 m is smaller than the 1D estimate. This
is because of the supports from adjacent columns.
As the spacing increases, this effect gets smaller. The
Figure 6. A typical simulation (column strength is probability is approaching 0.51 when the spacing is
500 kPa, spacing is 2.5 m). increasing as shown in Figure 10.

273

ISGSR2013.indb 273 10/18/2013 9:40:10 AM

 The above estimate ignores the effects of adja-
cent columns, which can be analysed by probabil-
istic FEM. Perfect plasticity is used in probabilistic
FEM analysis. For each simulation, if the column
in the middle collapses, the simulation is counted
as a failure, which is consistent with 1D estimation.
The collapse can be caused by the overloading of
the column, or by the loads redistributed by adja-
cent columns’ yielding. The probability of column
collapse is compared with 1D estimate in Table 2.
The results by FEM are significantly lower than
the 1D estimate. This is because of the perfect cor-
relation between column stiffness and strength, i.e.,
a weak column is loaded less than a strong column.
The above results showed that the effects of adja-
Figure 10. Probability of settlement exceeds 0.0825 m. cent columns are supportive. But this is not always
true. When safety margin is small (i.e., applied load
is close to mean resistance), the collapses of adja-
cent columns are more likely to cause the column
4 PROBABILISTIC ELASTO-PLASTICITY to fail. To demonstrate this, the spacing is adjusted
ANALYSIS so that applied load is equal to the mean resistance
(i.e., S = 0.1 m, FS = 1.0), and the probability of the
The strength of columns Cu is assumed to be a log- column collapse is shown in Table 3. It can be seen
normally distributed random variable with coeffi- from Table 3 that the probability of column col-
cient of variation (V
VCu ) of 0.5. The probability of lapses is much higher than the 1D estimate, which
column collapses can be estimated by 1D analysis. means that effects from adjacent columns are not
Since the spacing was decided so that the settle- supportive but destructive.
ment is 0.075 m, assuming column stiffness is 200 If perfect plasticity is used, the system of col-
times the column strength, we have umns is stable as long as the mean strength of col-
umns is larger than the applied load. The safety
WL margin of the column system is not clear. It is how-
= 0.075 (3)
200 μCu A ever, possible to access system reliability by using

W
= 1.5 μCu (4) Table 2. Probability of column collapse.
A
Column Probabilistic
If the Tresca failure criterion is used, the yield strength (kPa) FEM 1D estimate
stress is
100 1/2000 0.35
W 150 0.001 0.35
σ 0 75 μCu (5)
200 0.0015 0.35
2A
250 0.0065 0.35
The factor of safety is 1.333. The probability of 300 0.0165 0.35
column collapse is 500 0.0665 0.35

⎛
P 2Cu
⎝
W⎞
A⎠
P Cu( 0.75 μCu )
Table 3. Probability of column collapse.

= Φ⎜
(
⎛ ln 0.75 μC − μl
u
) Cu
⎞
⎟ Column Spacing FEM
⎜⎝ σ llnCu ⎟⎠ strength (kPa) (m) analysis 1D estimate

⎛ 1 ⎞ 100 1.286 0.8770 0.59

ln ( )+ ln(1 + VCu 2 )
= Φ⎜ ⎟
2 150 1.575 0.9075 0.59
⎜ ⎟ 200 1.817 0.9090 0.59
⎜⎝ ln( + Cu )2
⎟⎠ 250 2.031 0.9185 0.59
= 0.35 300 2.224 0.9175 0.59
500 2.869 0.9420 0.59
(6)

274

ISGSR2013.indb 274 10/18/2013 9:40:12 AM

 Table 4. Probability of system collapse. 5 CONCLUSIONS
Column Probability of By assuming equal strain condition, analytical
strength (kPa) Spacing (m) system collapse method is over conservative. It would lead to very
100 1.114 0.5e-3 tight acceptance criteria for column strength, and
150 1.368 0.55e-2 in terms leads to uneconomic design. Probabilistic
200 1.582 0.1075 FEM analysis can provide more accurate
250 1.769 0.3960 prediction. The results of probabilistic FEM anal-
300 1.938 0.7390 yses show that the effects from adjacent columns
500 2.500 0.9990 can be destructive. It will be pursued in a future
research to provide design charts for dry soil mix
columns based on probabilistic FEM analyses.

ACKNOWLEDGEMENT

The authors wish to acknowledge the support of

the Australian Research Council in funding the
Centre of Excellence for Geotechnical Science and
Engineering.

REFERENCES

Filz GM & Navin MP 2010. A Practical Method to

Account for Strength Variability of Deep-Mixed
Ground. GeoFlorida 2010: 2426–2433.
Griffiths DV, Dotson D and Huang J 2011. Probabilistic
Finite Element Analysis of a Raft Foundation
Supported by Drilled Shafts in Karst. Geo-Risk 2011:
Figure 11. Comparison between 1D estimate and FEM 233–239.
analysis. Griffiths DV, Huang JS and Fenton GA 2009. Influence
of Spatial Variability on Slope Reliability Using
2-D Random Fields. Journal of Geotechnical and
Geoenvironmental Engineering 135(10): 1367–1378.
zero residual strength for columns after yielding.
Kelly R & Wong P 2011. Lessons learnt from design
Table 4 shows the probability of system collapse. and construction of dry soil mix columns. Int.
The probabilities of column collapse by 1D esti- Conf. Advances in Geotechnical Engineering, Perth,
mate and FEM analyses are compared in Figure 11. Australia.
It can be seen that 1D estimate is conservative
when safety margin is large, but unconservative
when safety margin is small. The safety margin of
close spacing is larger than loose spacing.

275

ISGSR2013.indb 275 10/18/2013 9:40:14 AM

This page intentionally left blank
 Geotechnical Safety and Risk IV – Zhang et al. (eds)
© 2014 Taylor & Francis Group, London, ISBN 978-1-138-00163-3

Soft soil subgrade’s reliability and risk assessment under incomplete

probability conditions

C. Liu & M. Zhao

School of Civil Engineering, Hebei University of Technology, Tianjin, China

ABSTRACT: On the basis of the interval analysis, the non-probabilistic reliability analysis method
makes the reliability of slope stability and risk analysis more complete. Through the comparison
between the safety factor, probability and reliability indexes and the non-probabilistic reliability
index of dangerous sliding surface, the non-probabilistic reliability index’s advantage is proved
and a practical evaluation method is supplied for the reliability and risk assessment of geotechnical
engineering.

1 INTRODUCTION 2 THE INTRODUCTION TO RELIABILITY

ANALYSIS METHOD IN INTERVAL
There was a highway engineering project, which ANALYSIS
was totally collapsed through gravel pile reinforce-
ment. After the accident, we checked the coeffi- In coefficient of safety method, K means slope’s
cient of slope safety and probability reliability stability safety, which is the ratio between skid
index. The results show that the coefficient of resistance moment and sliding moment. Using
safety is 1.23 and the probability reliability index simple strip method, we can get the expression of
is 3.95, but the section and foundation around it safety coefficient K:
was stable because of the tamp treatment on it in n
fact.
M
∑( bi hi i i + ci li )
In the 1960s, the coefficient of safety method K= r = i =1 (1)
n
was accepted through long-term practice. But Ms
starting in the 1970s, slope engineering field ∑ γ bi hi i αi
i =1
began to accept the concept of uncertainty. Using
reliability method was proposed as a supplement. From formulas (1):
n
The use of probabilistic reliability method gradu-
ally became mature at the same time. It was found
Resistance R ∑( bi hi i i ci li ),
i =1
that this method also had certain limitations.
n
The nature of the soil has great variability. Soil
parameter statistics belongs to the small sample
Glide force S ∑ γ bi hi i αi
i =1
problem. So in subjective distribution assump-
tions, the results of probability reliability calcu- From probability reliability analysis’s state
lation may be distorted. Therefore, people put function:
n
forward to the non-probabilistic reliability analy-
sis method, which can further perfect the reliabil- Z R−S ∑( bi hi i i ci li )
i =1
ity theory. The uncertainty of evaluation is more
n
reasonable. The roadbed model is established on − ∑ γ bi hi i α i (2)
the bases of on finite element analysis software i =1
ABAQUS. This paper combines the finite element
numerical analysis with the limits equilibrium Because b, h, α are geometric parameters, γ is
method and conducts the probability and the physical parameter, we just put c, ϕ as variable. So
interval analysis on the soil shear strength param- the formula (2) can be simplified as follows:
eters and compared the differences between the n n
safety factor, probability and non-probability Z γ bh
bh α∑ i L ∑ ci γ bbh sin α (3)
indexes of reliability. i =1 i =1

277

ISGSR2013.indb 277 10/18/2013 9:40:15 AM

 Spreading formula (3) at mean: Non-probability reliability research calcula-
tion method: Taking parameters as basic interval
n n variables, if we know the parameter’s interval and
∑ (nγ bh ) ∑ (Lσ )
2 2
σz α 2
μϕ σ ϕ c (4) monotonic, we can get the upper and lower bound-
i =1 i =1 ary of limit state function:

But σ z = σ R 2 + σ S 2 (5) l
in G (δ1,δ 2 , ..., δ n )
δ i ∈Δ l
In the process of calculating the reliability Zu x G (δ1,δ 2 , ..., δ n )
index, we compare formula (4) with formula (5), δ i ∈Δ l

and then we find that the calculation of the resist-

ance R is considered, where the glide force R is not Then we can get the non-probability reliability
considered. From the code for investigation of geo- index by using of limit state function’s upper and
technical engineering, we can get that the value of c, lower boundary and approximation formula.
is got under the situation which is a normal distri-
= (Z + Z )
c u l
bution. If we get the distribution such as Figure 1 η=Z r
or Figure 2 (small sample), the sample parameters Z (Z u − Z l )
probability mode will indicate a larger deviation.
From the foregoing analysis, slope’s probabil-
ity reliability analysis method has limitations. In 3 CASE STUDY ANALYSIS
the 1990’s, the concept of non-probability reli-
ability was put forward to for the first time, and 3.1 Section numerical simulation
then people began to develop the slope stable non-
Building the model in ABAQUS, this is shown in
probability reliability analysis method researches
Figure 3.
on the basis of interval analysis method. Minghua
The embankment is filled with fly ash, and the
Zhao, Chong Jiang and Wengui Cao applied the
parameters are shown in Tables 1 and 2.
interval analysis non-probability reliability model
to the geotechnical engineering.
3.2 The safety coefficient and probabilistic
reliability analysis of section stability
Using strength subtraction calculation, the pro-
gram application result is convergent when K = 1.3,
which can be thought of as the slope has been dam-
aged, and the detail location of failure surface can’t
be got. Combined shear stress in Figure 4 (K = 1.2),
the biggest shear describes several connecting arcs,
as one of the most dangerous sliding surface, it
appears within the embankment and is below a
layer of soil. Then the stresses are taken from each
unit in sliding surface, which is shown in Table 3.

3.2.1 The safety coefficient and reliability index

Figure 1. Normal distribution scheme estimates (small of each unit in dangerous sliding surface
sample).
The function is

Z ( f Li + σ tan
)Li = cL t ϕ L i τ Li

Figure 2. Normal distribution scheme estimates (small Figure 3. The ABAQUS model of the section gravel
sample). pile roadbed.

278

ISGSR2013.indb 278 10/18/2013 9:40:17 AM

 Table 1. Soil parameters of embankment.

Layer Weight Cohesion Internal friction

Soil layer thinkness (m) (kN/m3) (KPa) angle (degree) ES1-2 (MPa)

Fly ash 4.05 14.00 10 25 10

Table 2. The pile parameters of soil layers and gravel pile.

Thickness of Weight Shear c Shear ϕ

Soil layer soil layer (m) (kN/m3) (KPa) (degree) ES1-2 (MPa)

Muddy silty clay 1 8.700 17.60 10.0 5.65 3.10

Sandy soil 5.100 19.00 0.0 20.1 19.93
Silt 1 6.510 20.30 20.0 18.7 8.70
Pravel pile Arrangement from Spacing (m) Length (m) Diameter (m) Weight (kN/m3)

Equilateral triangle 1.6 11.0 0.5 20.00

n
∑ τ f Li
i =1
K= n
= 181.02/89.32 = 2.03
∑ τ Li
i =1

n
∑ μz
i =1
β= = 22.93/6.66 = 3.44
n
Figure 4. Shear stress. ∑σ z 2

i =1

Division of grid, make each quadrilateral ele- For the first embankment layer τ f and τ along
ment approximation for 2 m length, pile and the sliding surface of the integral unit length, the
surrounding transverse length of 0.5 m units, the safety coefficient and probabilistic reliability index
soil elements between the pile lateral side length is can be calculated:
about 2.7 m. The maximum stress and minimum
stress of each unit are chosen from Figures 5 and 6. n
After calculation, the reliable indicators and the ∑ τ f Li
i =1
coefficients of safety of every unit in most likely K= n
= 752.16/666.84 = 1.13
damage surface of slope are listed in Table 3.
The Table 3 shows that the safety factor and prob- ∑ τ Li
i =1
ability reliability of embankment soil are large and
the shear of muddy silty clay layer unit is large, and
n
the safety factor through calculation of every unit
is close to 1 and the probabilistic reliability index is ∑ μz
i =1
close to zero, which tends to be damaged, and the β= = 85.56/42.88 = 2
n
probability reliable index of pile is not very large.
∑ σ z2
i =1
3.2.2 The safety coefficient and reliable index
of overall reliability in every soil layer
For embankment layers τ f and τ along the sliding 3.2.3 The safety coefficient and reliability index
surface of the integral unit length, the coefficient of embankment slope’s overall reliability
of safety and probabilistic reliability index can be We can get the coefficient of safety and probabil-
calculated: ity reliability index through integrating along the

279

ISGSR2013.indb 279 10/18/2013 9:40:20 AM

 Table 3. Probability index and safety coefficient.

μc μϕ μz σz Probability Safety
Soil layer Unit (KPa) (degree) (KN/m) (KN/m) index β coefficient K

Embankment 78 10.00 25.00 23.99 5.27 4.56 3.26

soil 77 10.00 25.00 22.30 6.05 3.69 2.09
88 10.00 25.00 22.03 7.33 3.00 1.77
89 10.00 25.00 23.39 7.72 3.03 1.79
Muddy silty clay 99 10.5 5.65 0.47 7.91 0.06 1.02
and gravel pile 3 (part 5) 2 45.00 −5.22 12.82 −0.41 0.93
100 10.5 5.65 0.16 6.05 0.03 1.01
4 (part 6) 2 45.00 6.35 16.18 0.39 1.08
131 10.5 5.65 0.00 6.42 0.00 0.99
5 (part 7) 2 45.00 19.43 16.20 1.20 1.31
161 10.5 5.65 −0.50 6.76 −0.07 0.99
5 (part 8) 2 45.00 25.06 14.71 1.70 1.51
330 10.5 5.65 −0.44 6.72 −0.07 0.99
5 (part 9) 2 45.00 32.96 14.21 2.32 1.84
388 10.5 5.65 −0.59 13.44 −0.04 0.99
389 10.5 5.65 −0.77 13.48 −0.06 0.99
410 10.5 5.65 3.50 8.51 0.41 0.99
431 10.5 5.65 5.16 8.13 0.63 1.09

n
∑ μz
i =1
β= = 177.26/44.91 = 3.95,
n
∑σ z 2

i =1

The coefficient of safety of embankment slope’s

overall reliability is 1.23 while the probabilistic reli-
ability index is 3.95 and the failure probability is
Figure 5. Maximum main stress.
less than 0.0001.

3.3 The non-probabilistic reliability analysis

of section stability
3.3.1 The non-probabilistic reliability
index of each unit in dangerous
sliding surface
Because the parameters of the fly ash embank-
ment are obtained from experience, the range of
parameters is shown in Table 4, if 0.2 is chosen as
the coefficient of variation value.
Figure 6. Minimum main stress. The function equation:
Z ( f Li + σ tan
)Li = cL t ϕ L i τ Li
whole dangerous sliding surface’s each unit length
about subgrade slope’s τ f and τ .
Taking a unit of embankment as an example:
n
σ = 15.66 KPa, τ = 5.3 KPa
∑ τ f Li
i =1
K= n
= 933.17/756.16 = 1.23 The shear strength standard values
∑ τ Li c =[ ]KPa, ϕ =[ ]°
i =1

280

ISGSR2013.indb 280 10/18/2013 9:40:23 AM

 Table 4. The range of parameters. Table 5. The non-probabilistic reliability index of unit
stability.
Shear strength index
Soil layer Unit σ (KPa) τ (KPa) η
Internal
Statistical Cohesion c friction angle Fly ash 78 15.66 5.30 3.29
Soil layer index (KPa) ϕ (degree) embankment 77 24.35 10.20 2.45
88 32.97 14.35 2.02
Fly ash Maximum 12 30 89 35.40 14.80 2.05
Minimum 8 20 Muddy silty 99 43.87 14.60 0.02
Muddy silty Maximum 14 8 clay 3 128.27 140.60 0.03
clay 1 Minimum 6 4.3 100 63.84 16.71 0.04
4 161.87 151.04 0.04
131 107.06 21.25 0.05
5 162.16 125.17 0.06
According to the optimization method: 161 135.86 24.27 0.06
5 147.17 98.93 0.05
Zu = 48.26 KN/m, Zl = 14.68 KN/m
/ 330 132.66 23.91 0.30
5 142.25 78.21 0.43
The non-probabilistic reliability index: 388 132.54 23.80 0.41
389 134.45 24.05 0.77
410 103.20 18.95 1.22
Z c Zu + Zl
η= = = 3.29 431 71.43 14.99 1.53
Z r Zu − Zl

Through calculation, the non-probability reli-

ability indexes of each unit in the most dangerous the non-probabilistic reliability index can be
sliding surface are shown in Table 5. calculated:
n
3.3.2 The non-probabilistic reliability index
Zc ∑ (Zu + Zl )dLLi
i =1
of soil layers η= = = 0.86
For embankment soil layers Zu and Zl along the Zr n

sliding surface of the integral unit length, the non- ∑ (Zu − Zl )dLLi
i =1
probabilistic reliability index can be calculated:

n 3.4 The stability analysis of the section

Z c ∑ (Zu + Zl )dLLi From each soil’s the coefficient of safety and reli-
i =1
η= = = 2.37 ability index, we know that the mucky silty clay
Zr n
∑ (Zu − Zl )dLLi layer is more likely to be destroyed after using
i =1 gravel pile to deal with the ground secsions. The
mucky silty clay layer’s coefficient of safety is
For the first embankment soil layer Zu and 1.13, its probability index is 2 and its destroying
Zl along the sliding surface of the integral unit probability is greater than 0.01, so the layer can
length, the non-probabilistic reliability index can not satisfy the demands. The most dangerous
be calculated: sliding surface’s non-probability reliability index
is the same as the coefficient of safety and prob-
n ability reliability index. The mucky silty clay lay-

Z c ∑ (Zu + Zl )dLLi er’s non-probability is 0.64, so its reliability is low

because its value is small. The whole slope’s safety
i =1
η= = = 0.64
Zr n coefficient is 1.23, probability reliability index is
∑ (Zu − Zl )dLLi 3.95, but the non-probability reliability index is
i =1 0.83. In fact, this section and the subgrade which
is through gravel pile reinforcement around it are
collapsing totally.
3.3.3 The non-probabilistic reliability index From the foregoing discussion, we can judge
of overall stability in embankment slope that the section is unstable, and the results of
For embankment soil layers Zu and Zl along non-probability reliability theory conform to
the sliding surface of the integral unit length, reality. If we use the probability reliability index,

281

ISGSR2013.indb 281 10/18/2013 9:40:26 AM

 with project case study, the calculation of the prob-
ability reliability index cannot accurately show the
stability of the section, but no matter how small
the samples are, they also have the range. So the
non-probabilistic reliability theory has certain
advantages in small sample situation.

REFERENCES
Figure 7. The relationship curve of k, β, η.
Junqi Zhang. Non-probabilistic reliability analysis of
slope stability. Hebei University of Technology, 2011.
we can not get the result, because it is influenced Minghua Zhao, Chong Jiang & Wengui Cao. 2008.
by parameter’s distribution. If the coefficient Non-probabilistic reliability analysis based on the the
of variation is bigger, the probability reliability interval theory retaining wall stability.The Journal of
index will decrease. So we can say that the non- Geotechnical Engineering 30(4): 467–472.
probability reliability theory is good for the slope Ruijiao Xin & Qiang Li. 2011. Structure interval fuzzy
engineering. stochastic finite element reliability analysis. Ha’erbin
University of Technology, 2011.
Shengfei Yu, Zhengzhou Chen & Mingrui Zhang. Slope
4 CONCLUSION stability analysis based on the range of uncertainty
analysis method. Journal of Engineering Geology,
1004-9665/2012/20(2)-0228.
Through the analysis of engineering examples, the Wengui Cao & Yongjie Zhang. 2007. Non-probabilistic
relationship between the three parameters is shown fuzzy reliability analysis of slope stability based on
in Figure 7. interval method. The Journal of Civil Engineering,
Figure 7 shows that the non-probabilistic reli- 2007.11.
ability indexes are similar to the safety coefficient Yong Liu & Huansheng Mu. 2011. Non-probabilistic
values. There are no clear provisions on the non- reliability analysis of highway roadbed slope stability
probabilistic reliability index, so the selecting on in soft ground conditions. Road Engineering 1002-
the value target of probabilistic reliability index 4786 17-0096-02.
can refer to the target value of safety coefficient. In Yingying Zhang. 2011. Non-probabilistic reliability
analysis based on particle swarm optimization and
addition, the quantity of soil parameters is small, its application in Slope Engineering. Zhongnan
so parameters distribution can only be assumed University, 2011.
to be normal distribution, and the coefficient of Zhengzhong Wang, Shengyuan Mou & Jun Liu. 2009.
variation is not accurate, when calculating the Calculated safety factor of slope aspect centrifugation.
probability of slope stability reliability. Combined Rock and Soil Mechanics, 2009.9.

282

ISGSR2013.indb 282 10/18/2013 9:40:28 AM

 Geotechnical Safety and Risk IV – Zhang et al. (eds)
© 2014 Taylor & Francis Group, London, ISBN 978-1-138-00163-3

Localised metamodelling techniques for geotechnical

reliability-based analysis

S.H. Marques, A.T. Gomes & A.A. Henriques

Department of Civil Engineering, Faculty of Engineering, University of Porto, Portugal

ABSTRACT: Metamodelling is actually a very interesting research area for the replacement of
simulation models when the trade-off between efficiency and accuracy is appropriate. Several approaches
have been proposed, and localised metamodelling techniques applied in reliability-based analysis may
involve the search for the most probable failure point. Response surface methodology has been suggested
for the approximation of functions, namely implicit limit state functions. The response surface is typically
constructed from a polynomial function fitted at a number of points, but no clear guidance has been
given for the effect of varying their location. Results for a geotechnical design example with correlated
nonnormal random variables are presented in the light of the quality of the design point approximation.
In particular, interpolation methodologies in conjunction with appropriate experimental design are
suggested for extracting nonlinear response surfaces whose roughness varies substantially over the input
domain, so that sequential algorithms may adapt to nonlinear features.

1 INTRODUCTION Monte Carlo simulation based on counting may be

used efficiently since the evaluation of the response
Metamodelling techniques are actually established surface function requires a reduced computational
to reduce the need to run expensive simulations, as effort. Alternatively, the first order reliability
stated by Can & Heavey (2011). In order to reduce method may be applied and corrections may be
computational costs in reliability analysis it has provided by the second order reliability method,
been suggested to utilise approximate response otherwise different reliability techniques.
functions, and one well established class of meth- This methodology gives an approximate closed
ods to deal with suitable approximations is the so form expression, based on the selected number
called response surface methodology (Bauer & of design of experiments, applied to generate the
Pula 2000, Tandjiria et al. 2000, Babu & Srivastava training data from the decision space of the prob-
2007, Massih & Soubra 2008, Mollon et al. 2009, lem, and is very useful when the response has to be
Lü & Low 2011), whose basic principles are also computed through a numerical procedure such as
described by Haldar & Mahadevan (2000). finite element analysis and the performance func-
The basic idea in utilising the response surface tion is implicit.
methodology is to replace the true limit state func- Two key issues of the response surface meth-
tion by an approximation whose functional values odology have been analysed in the literature: the
may be computed more easily. In general, a good polynomial degree of the response surface and the
physical knowledge of the system is very useful position of the sampling points. Whilst response
in deciding the appropriate order of the response surfaces with low degree polynomials are used in
surface approximation. The functions are typically practice to reduce the number of limit state func-
chosen to be first or second order polynomials, in tion evaluations, the approximation provided by the
fact higher orders require a high number of sup- response surfaces is by hypothesis only local to the
port points, to be obtained very close to the failure region where the sampling points are disposed, so
surface or limit state of interest. a good approximation implies that the coordinates
The response surface model should give the best of the most probable failure point, the design point,
possible fit to the collected data, and in general two have to be precisely discovered (Allaix et al. 2011).
different types of response surface models are distin- As a general rule, the approximate closed form
guished: regression models and interpolation mod- expression is valid only within the range of the
els (Bucher & Most 2008). By fitting the response values considered for the random variables, and
surface to a number of design of experiments of extrapolation beyond the considered range may
the true limit state function, the approximated not be accurate, furthermore, the approximation
limit state function is then constructed. After that, may be indeed inadequate for highly nonlinear

283

ISGSR2013.indb 283 10/18/2013 9:40:28 AM

 performance functions (Haldar & Mahadevan where n is the number of random variables; As is
2000). So, regarding nonlinear analysis, different the sth random variable; and a0, as, and an+s are
methodologies have been proposed. A very simple unknown coefficients to be determined.
iterative procedure for the search of the most proba- A brief explanation of the algorithm is summa-
ble failure point is detailed next considering that the rised as follows, for schemes of experimental design
approximate function is generated in the standard of axial type, other variants in Marques (2012):
normal space. For the purpose, the transformations
xi* = f(yi*) and f(xi*) = yi* derived by Equation 1 are [Step 1] Set the first iteration vector [p1; p2; …; pn]
further considered for representation of random in the standard normal space for instance
variables xi* with different distributions as a func- at zero, that means first iteration is cen-
tion of standard normal random variables yi*: tered in the median values for all the n
random variables;
xi * = FXi ⎡
⎣ ( yii )⎤⎦
* 1⎡
⎣ ( )
FXi xi * ⎤ = yi *
⎦
(1) [Step 2] Construct a (2n + 1) × (n) matrix P,
each column for each one of the n ran-
dom variables: first line is vector [p1 + 0;
where xi* = random variable in the original space; p2 + 0; … ; pn + 0]; second line is vector
yi* = random variable in the standard normal space; [p1 + h1; p2 + 0;…; pn + 0] where h1 is the
FXi = cumulative nonnormal distribution function; selected distance from the center point;
FXi −1 = inverse cumulative nonnormal distribution repeat the previous procedure for the next
function; Φ = cumulative standard normal dis- (n − 1) lines by summing to each one of
tribution function; and Φ−1 = inverse cumulative the other columns the selected number;
standard normal distribution function. (n + 2) line is vector [p1 − h1; p2 + 0; …;
The equivalent standard normal covariance pn + 0] where h1 is the selected distance
or correlation matrix may differ slightly from the from the center point; repeat the previ-
original, and for simplicity the unmodified matrix ous procedure for the next (n − 1) lines
has been frequently considered in the literature. by subtracting to each one of the other
A simplified approach suggested by Kiureghian & columns the selected number; concerning
Liu (1986) and described by Haldar & Mahadevan efficiency the minimum number of lines
(2000) may be considered for the equivalent stand- is considered; the use of additional lines
ard normal matrix computation. Considered the in regression analysis is indeed a possibil-
selected design example, the values of the correlation ity to be considered;
coefficients between the random variables are then [Step 3] Evaluate the performance function in
transformed by empirical relationship according the standard normal space by using the
to the detailed expressions, and so the transforma- points in each line of the constructed
tions are used for the computation of the resultant matrix P and construct the column vec-
equivalent standard normal correlation coefficient, tor R; if required use the bilateral trans-
when the random variables are normal and lognor- formation from Equation 1;
mal, or are both lognormal. Other transformations [Step 4] Construct a (2n + 1) × (2n + 1) matrix N
may be considered, as the proposed Nataf’s model by using a first unitary column vector,
further described by Haldar & Mahadevan (2000). next n columns from matrix P columns,
and next n columns from squared ele-
2 RESPONSE SURFACE METHODOLOGY ments in previous n columns;
ITERATIVE ALGORITHM [Step 5] Solve the equation of the approximation
function for the unknown coefficients by
The detailed response surface methodology itera- regression analysis expressed in Na = R,
tive algorithm approximates the limit state function where a is the column vector of the
by an explicit function of the random variables in unknown coefficients;
the standard normal space, and is improved by a [Step 6] Express the equivalent standard normal
very simple iterative procedure based on the imple- correlation matrix and the set of random
mentation of the first order reliability method algo- variables vector in the standard normal
rithm in the standard normal space. More detail in space, then run the first order reliability
Marques (2012). The expression considered for the method algorithm by using the generated
approximation function is described by Equation 2, approximate function in the standard
a second order polynomial with squared terms but normal space, and estimate the reliability
no cross terms, suitable for nonlinear analysis: index β and the corresponding coordi-
nates of the design point;
( )
n n
H ( As ) = a0 + ∑ (as As ) ∑ an + s ⋅ As2 (2) [Step 7] Set the next iteration vector [p1; p2;
s =1 s =1 …; pn] in the standard normal space at

284

ISGSR2013.indb 284 10/18/2013 9:40:28 AM

 the estimated coordinates of the design level is away. Considered the inclined eccentric load-
point and update matrix P; ing problem and the calculation model for bearing
[Step 8] Repeat steps 3 to 6 and verify the capacity, the performance function is described by
convergence to the selected tolerance the simplified Equation 3, detailed in Appendix 1:
criterion, for instance |βiteration i – βiteration
i-1| ≤ 5 × 10
−3
or 5 × 10−4, in which i is the
iteration number, as the methodology may
M (
f B B H H c w w cf f f q ) (3)

be applied beyond structural engineering; where the sum of B1 and B2 is the foundation width B;
further check the convergence process for H1 is the wall height; H2 is the foundation height; γc
the coordinates of the design point; is the unit concrete weight; ϕw is the friction angle of
[Step 9] If required, repeat steps 7 and 8 until the soil on the active and passive sides of the wall;
convergence to the selected tolerance γw is the unit soil weight on the active and passive
criterion, then express the final response sides of the wall; cf is the cohesion of the founda-
surface, and estimate the reliability index tion soil; ϕf is the friction angle of the foundation
β and the corresponding coordinates of soil; γf is the unit weight of the foundation soil; and
the design point; Monte Carlo simulation q is the variable load at ground surface. Other con-
based on counting is now efficient since sidered parameters, namely for the earth pressure
the evaluation of the response surface coefficients, are the soil-wall interface friction angle
function requires a reduced computa- on the active side of the wall δ w ϕ w; the soil-
tional effort and input random variables wall interface friction angle on the passive side of
are correlated in the standard normal the wall δwp = 0; and the soil-foundation interface
space; it is further noted that transforma- friction angle δf = ϕf; considered either a 50% reduc-
tions to the space of uncorrelated ran- tion for passive earth pressures.
dom variables may be considered. Table 1 summarises the description of basic
Thus, the efficiency and accuracy provided by input variables, with different distribution types.
different schemes of experimental design are ana-
lysed in the next sections with a design example for Table 1. Summary description of basic input variables.
a concrete gravity retaining structure, considered
correlated nonnormal random variables. Basic input
variables Distribution mv cv Statistics
3 DESIGN EXAMPLE B1 (m) Deterministic – – 1.75a
1.85b
The design example is referred to the concrete grav- 1.95c
ity retaining structure on a relatively homogeneous B2 (m) Deterministic – – 1.00
c − ϕ soil shown in Figure 1, wherein groundwater H1 (m) Deterministic – – 7.00
H2 (m) Deterministic – – 1.00
γc (kN/m3) Deterministic – – 24.00
ϕw (º) Lognormal 33.00 0.10 μ1 − 3.4915
σ1 − 0.0998
γw (kN/m3) Normal 18.80 0.05 –
cf (kN/m2) Lognormal 14.00 0.40 μ1 − 2.5648
σ1 − 0.3853
0.05a μ1 − 3.4645
σ1 − 0.0500
ϕf (º) Lognormal 32.00 0.10b μ1 − 3.4608
σ1 − 0.0998
0.15c μ1 − 3.4546
σ1 − 0.1492
γf (kN/m3) Normal 17.80 0.05 –
q (kN/m2) Weibull – – sh−3.25
sc−10.00

mv—mean value; cv—coefficient of variation.

μ1—log mean; σ1—log standard deviation; sh—shape;
sc—scale.
a
B1 corresponding to the coefficient of variation ϕf 0.05.
b
B1 corresponding to the coefficient of variation ϕf 0.10.
Figure 1. Concrete gravity retaining structure. c
B1 corresponding to the coefficient of variation ϕf 0.15.

285

ISGSR2013.indb 285 10/18/2013 9:40:29 AM

 Table 2. Coefficients of correlation between the random Table 3. Summary description of results for the bearing
variables. capacity model.

Correlation matrix for the bearing capacity model Coefficient Reliability Reliability
of variation index index
ρϕ ϕ ρϕwγw ρϕwϕf ρϕwq ρϕwcf ρϕwγf Model ϕf βFORM βMCS
w w
ργwϕw ργwγw ργwϕf ργwq ργwcf ργwγf
ρϕfϕw ρϕfγw ρϕfϕf ρϕfq ρϕfcf ρϕfγf Bearing 0.05 2.4842 2.5288
capacity 0.10 2.2248 2.2787
ρqϕw ρqγw ρqϕf ρqq ρqcf ρqγf
ρcfϕw ρcfγw ρcfϕf ρcf q ρcfcf ρcfγf 0.15 2.0410 2.0952
ργfϕw ργfγw ργf ϕf ργf q ργfcf ργfγf MCS results from 106 simulation steps.
1.0 0.5 0.9 0.0 0.0 0.5
0.5 1.0 0.5 0.0 0.0 0.9
0.9 0.5 1.0 0.0 0.0 0.5 variation for the friction angle of the foundation
0.0 0.0 0.0 1.0 0.0 0.0 soil. It is noted that the differences between the reli-
0.0 0.0 0.0 0.0 1.0 0.0 ability index obtained from both methodologies are
0.5 0.9 0.5 0.0 0.0 1.0 somewhat notorious due to the important nonlinear
behaviour presented by the bearing capacity model.
ρ—coefficient of correlation. The bearing capacity metamodels are built
according to different schemes of experimental
design of axial type, for a selected distance from
Correlation is the property in that, on average,
the center point of 0.01, 1, 2, 3 and 4. Grouped
two variables are linearly associated with one
these models, a comparative study about the con-
another, and the value of one provides information
vergence process is then summarised in Tables 4
on the probable value of the other. The strength of
and 5, for two convergence criteria and three dif-
the association is measured by a correlation coef-
ferent reference models corresponding to coeffi-
ficient in the range [−1,1], for the case that both
cient of variation ϕf 0.05, 0.10 and 0.15, considered
variables vary inversely or together exactly, respec-
the first order reliability method.
tively. The considered coefficients of correlation
According to Haldar & Mahadevan (2000), the
between the random variables are either presented
number of design sampling points should be kept
in Table 2, noted that the correlation matrices are
to a minimum to increase efficiency but must be
idealised with the main purpose of creating high
at least equal to the number of coefficients needed
correlation between some of the random variables.
to define the performance function, therefore effi-
It is further noted that the correlation between the
ciency is related to the number of design sampling
cohesion and the friction angle of the foundation
points and iterations, in other words the number
soil depends particularly on the material, and the
of design of experiments. Haldar & Mahadevan
question whether the two parameters are corre-
(2000) refer also that the efficient location of
lated is still not clearly decided in the literature, as
design sampling points around the center point
stated by Fenton & Griffiths (2003). According to
is essential for the accurate construction of the
Marques (2012), references for negative correlation
performance function, and increased accuracy is
are common but regarding a sensitivity analysis for
described in relation with an increased number of
the design example, the influence of negative cor-
design sampling points, with the discussion of dif-
relation on reliability is favourable and quite con-
ferent layouts. These issues are focused hereafter.
siderable when compared to the effects of positive
Thus, two convergence criteria are considered in
correlation. Thus, the hypothesis of uncorrelated
order to compare the accuracy of the models, given
random variables is probably cautious and further-
by the relative errors obtained by the first order
more, the characteristic values of soil properties are
reliability method, and also the efficiency from the
often calculated separately, neglecting the effects of
required number of iterations. It is noted yet that the
correlation, so both procedures are concordant.
percentage errors are currently referenced in the lit-
erature to Monte Carlo simulation based on count-
4 RESULTS AND DISCUSSION ing, although the reference is in fact the first order
reliability method, used in the iterative algorithm.
Considered the implementation of the foremost From the analysis of results, it is concluded that
detailed methodology, Table 3 presents the sum- the efficiency and accuracy of each model depend not
mary description of results for the bearing capacity only on the number of design sampling points and
model, included the reliability index obtained by the layout, but also on the distance from the center point
first order reliability method and the Monte Carlo that is considered to set the axial points, even whether
simulation, considered the different coefficients of the number of design sampling points and layout are

286

ISGSR2013.indb 286 10/18/2013 9:40:31 AM

 Table 4. Summary description of results for the bearing Table 5. Summary description of results for the bearing
capacity metamodels. capacity metamodels.

Coefficient of Model Model Coefficient of Model Model Model

variation ϕf 0.05 d0.01 d1 variation ϕf 0.05 d2 d3 d4

Tolerance criterion 5 × 10−3 5 × 10−3 Tolerance criterion 5 × 10−3 5 × 10−3 5 × 10−3

Number of iterations 4 4 Number of iterations 4 5 6
βFORM model 2.48 2.48 βFORM model 2.48 2.48 2.48
βFORM metamodel 2.48 2.48 βFORM metamodel 2.49 2.49 2.51
Relative error (%) 0.00 0.00 Relative error (%) 0.40 0.40 1.21
Tolerance criterion 5 × 10−4 5 × 10−4 Tolerance criterion 5 × 10−4 5 × 10−4 5 × 10−4
Number of iterations 4 5 Number of iterations 5 5 6
βFORM model 2.484 2.484 βFORM model 2.484 2.484 2.484
βFORM metamodel 2.484 2.484 βFORM metamodel 2.486 2.493 2.511
Relative error (%) 0.000 0.000 Relative error (%) 0.081 0.362 1.087

Coefficient of Model Model Coefficient of Model Model Model

variation ϕf 0.10 d0.01 d1 variation ϕf 0.10 d2 d3 d4

Tolerance criterion 5 × 10−3 5 × 10−3 Tolerance criterion 5 × 10−3 5 × 10−3 5 × 10−3

Number of iterations 4 4 Number of iterations 5 6 9
βFORM model 2.22 2.22 βFORM model 2.22 2.22 2.22
βFORM metamodel 2.22 2.22 βFORM metamodel 2.23 2.23 2.24
Relative error (%) 0.00 0.00 Relative error (%) 0.45 0.45 0.90
Tolerance criterion 5 × 10−4 5 × 10−4 Tolerance criterion 5 × 10−4 5 × 10−4 5 × 10−4
Number of iterations 5 5 Number of iterations 6 8 11
βFORM model 2.225 2.225 βFORM model 2.225 2.225 2.225
βFORM metamodel 2.225 2.225 βFORM metamodel 2.226 2.232 2.244
Relative error (%) 0.000 0.000 Relative error (%) 0.045 0.315 0.854

Coefficient of Model Model Coefficient of Model Model Model

variation ϕf 0.15 d0.01 d1 variation ϕf 0.15 d2 d3 d4

Tolerance criterion 5 × 10−3 5 × 10−3 Tolerance criterion 5 × 10−3 5 × 10−3 5 × 10−3

Number of iterations 4 5 Number of iterations 6 10 22
βFORM model 2.04 2.04 βFORM model 2.04 2.04 2.04
βFORM metamodel 2.04 2.04 βFORM metamodel 2.04 2.05 2.07
Relative error (%) 0.00 0.00 Relative error (%) 0.00 0.49 1.47
Tolerance criterion 5 × 10−4 5 × 10−4 Tolerance criterion 5 × 10−4 5 × 10−4 5 × 10−4
Number of iterations 4 5 Number of iterations 8 12 36
βFORM model 2.041 2.041 βFORM model 2.041 2.041 2.041
βFORM metamodel 2.041 2.041 βFORM metamodel 2.043 2.053 2.080
Relative error (%) 0.000 0.000 Relative error (%) 0.098 0.588 1.911

similar. When this distance is reduced in order to be reference models. The methodology is based on the
close to zero, the search for the design point in non- local approximation of the coordinates of the most
linear analysis is improved. It is also concluded that probable failure point, the design point, unlike the
as this distance becomes larger, the different models stochastic response surface methodology which
may present inaccurate results and important conver- may provide a general approximation statistically
gence problems, in particular for the tighter tolerance representative of the performance function.
criterion and the higher coefficient of variation for
the friction angle of the foundation soil.
The response surface methodology itera- 5 CONCLUSION
tive algorithm generates nonstochastic local
metamodels, in that they present statistical param- The design of civil engineering structures relies
eters more or less nonrepresentative of their heavily on models, abstract and simplified

287

ISGSR2013.indb 287 10/18/2013 9:40:31 AM

 representations of reality, yet accurate enough ⎛ c Nc ⋅ sc ic + γ w ⋅ H Nq ⋅ sq iq ⎞
to be used for design purposes and risk analysis. M =⎜ f ⎟⎠ ⋅ Be Evb
v
⎝ + 0.5 ⋅ γ f ⋅ Be Nγ ⋅ sγ iγ
Metamodelling encompasses the replacement of
simulation models when the trade-off between (4)
efficiency and accuracy is appropriate. In recent 2
⎛ ⎛ ϕf ⎞⎞
years, localised metamodelling techniques have
been suggested for the approximation of func-
Nq = tan 45 +
⎝ ⎝ 2 ⎟⎠ ⎟
⎠
exp
exp ( ( ))
f
(5)
tions, namely the response surface methodology,
particularly significant for the work with implicit
limit state functions. Nγ ( ( ) ) t (ϕ )
2 Nq ϕ f f (6)
Regarding these issues, results for a design
example with correlated nonnormal random
variables are presented. For the purpose it is con-
Nc (N ( ) − ) t ( )
q f f (7)
sidered an inclined eccentric loading problem
referred to a bearing capacity nonlinear calcula- sq = 1 (8)
tion model for a concrete gravity retaining struc-
ture, moreover a geotechnical reliability-based sγ = 1 (9)
analysis problem. The sensitivity of the bearing
capacity model when considered different coef- sc = 1 (10)
ficients of variation for the friction angle of the
foundation soil is then analysed, concluded that 2
the increased nonnormal behaviour of the bearing ⎛ Ehb ⎞
iq = ⎜1 − ⎟ (11)
capacity response correlates well with the higher
values of the parameter.
⎜⎝ Evb Be ⋅ c f ⋅ cot ( )
f
⎟⎠

Based on the degree of nonlinear behaviour and 3

on the efficiency and accuracy provided by differ- ⎛ Ehb ⎞
iγ = ⎜1 − ⎟ (12)
ent schemes of experimental design of axial type,
the convergence process is lastly discussed and the
⎜⎝ Evb Be ⋅ c f ⋅ cot ϕ f ( ) ⎟⎠
relative errors obtained by the first order reliabil-
ity method are presented. Considered the impor- 1 − iq
ic iq − (13)
tant nonlinear behaviour of the bearing capacity
model, results derived from the bearing capacity
Nc ⋅ tan ( )
f

metamodels built by the response surface meth- where the sum of B1 and B2 is the foundation
odology iterative algorithm show that improved width B; H1 is the wall height; H2 is the founda-
techniques for the search of the design point are tion height; γc is the unit concrete weight; ϕw is the
required, in particular if a tighter tolerance crite- friction angle of the soil on the active and passive
rion is considered. It is noted that the efficiency sides of the wall, expressed in degrees; γw is the unit
and accuracy of the iterative algorithm are very soil weight on the active and passive sides of the
sensitive to the distance from the center point that wall; cf is the cohesion of the foundation soil; ϕf is
is considered to set the axial points, so when this the friction angle of the foundation soil, expressed
distance is reduced in order to be close to zero, the in degrees; γf is the unit weight of the foundation
search for the design point in nonlinear analysis is soil; q is the variable load at ground surface; Nc,
improved. Nq, and Nγ are the bearing resistance factors; sc, sq,
Going back in history, several cases of instabil- and sγ are the foundation shape factors; ic, iq, and
ity have been reported in the literature regarding iγ are the load inclination factors; Be is the effec-
the use of reduced distances in schemes of experi- tive foundation width; Evb is the resultant vertical
mental design, but results for the selected design action for bearing capacity; and Ehb is the result-
example show a possibility for progression when ant horizontal action for bearing capacity.
considered the detailed response surface method-
ology iterative algorithm. It is noted lastly that the
axial type schemes are not always the most appro- REFERENCES
priate option.
Allaix, D.L., Carbone, V.I. & Mancini G. 2011. Advanced
response surface method for structural reliability. In:
APPENDIX 1 EQUATION (3) Proc. of the Thirteenth International Conference on
Civil, Structural and Environmental Engineering Com-
puting, Chania, Crete, Greece: Civil-Comp Proceedings
M (
f B B H H c w w cf f f q ) (3) 96, Paper 88, 1–14. Civil-Comp Press.

288

ISGSR2013.indb 288 10/18/2013 9:40:31 AM

 Babu, G.L.S. & Srivastava, A. 2007. Reliability analysis Kiureghian, A.D. & Liu, P.L. 1986. Structural reliability
of allowable pressure on shallow foundation using under incomplete probability information. Journal of
response surface method. Computers and Geotechnics Engineering Mechanics 112(1): 85–104.
34(3): 187–194. Lü, Q. & Low, B.K. 2011. Probabilistic analysis of
Bauer, J. & Pula, W. 2000. Reliability with respect underground rock excavations using response sur-
to settlement limit-states of shallow foundations face method and SORM. Computers and Geotechnics
on linearly-deformable subsoil. Computers and 38(8): 1008–1021.
Geotechnics 26(3–4): 281–308. Marques, S.H.M. 2012. Reliability-based analysis and
Bucher, C. & Most, T. 2008. A comparison of approxi- design in geotechnical engineering: Applications to
mate response functions in structural reliability Eurocodes: Thesis Project. Faculty of Engineering of
analysis. Probabilistic Engineering Mechanics 23(2–3): the University of Porto.
154–163. Massih, D.S.Y.A. & Soubra, A.H. 2008. Reliability-based
Can, B. & Heavey, C. 2011. Comparison of experimental analysis of strip footings using response surface
designs for simulation-based symbolic regression methodology. International Journal of Geomechanics
of manufacturing systems. Computers & Industrial 8(2): 134–143.
Engineering 61(3): 447–462. Mollon, G., Dias D. & Soubra, A.H. 2009. Probabilistic
Fenton, G.A. & Griffiths, D.V. 2003. Bearing-capacity analysis of circular tunnels in homogeneous soil using
prediction of spatially random c − ϕ soils. Canadian response surface methodology. Journal of Geotechnical
Geotechnical Journal 40(1): 54–65. and Geoenvironmental Engineering 35(9): 1314–1325.
Haldar, A. & Mahadevan, S. 2000. Reliability assessment Tandjiria, V., Teh C.I. & Low, B.K. 2000. Reliability
using stochastic finite element analysis. John Wiley & analysis of laterally loaded piles using response surface
Sons. methods. Structural Safety 22(4): 335–355.

289

ISGSR2013.indb 289 10/18/2013 9:40:35 AM

This page intentionally left blank
 Geotechnical Safety and Risk IV – Zhang et al. (eds)
© 2014 Taylor & Francis Group, London, ISBN 978-1-138-00163-3

Flood simulation considering probability of heavy rains

and uncertainty of soil properties of earth-fill dams

S. Nishimura & T. Shuku

Graduate School of Environmental Science, Okayama University, Okayama, Japan

H. Nagao
Naigai Engineering Corp., Kyoto, Japan

K. Fujisawa
Graduate School of Agriculture, Kyoto University, Kyoto, Japan

ABSTRACT: The risk of earth fills during heavy rains is evaluated in this research. The rainfall intensity
and the variability of the numerous soil parameters of the earth-fills related to soil erosion are dealt with as
probabilistic parameters. When the peak overflow head on a spillway bed becomes greater than the design
overflow head, overflow occurs. The peak overflow head, is determined within 72 hours of the rainfall,
and the various rainfall patterns are tested by the Monte Carlo method. Furthermore, the statistical values
of the parameters for the soil erosion are assumed, and the variability of the discharge hydrographs are
derived from the values. The discharge hydrographs are applied for the flood simulation as input waves. The
submerged area and the head of the flood discharge in the downstream area can be predicted. Considering
disaster loss, the risk to the downstream area of the earth-fill dams can be evaluated.

1 INTRODUCTION of overflow Pf is calculated from the number of the

event, hp < hd during the iterations of the Monte
Many earth-fill dams have been built for farm Carlo simulation. Peak overflow head hp is deter-
ponds in Japan. Some of the dams are getting old mined within 72 hours of the rainfall, and the vari-
and decrepit, and have weakened. Every year, a ous rainfall patterns are tested by the Monte Carlo
number of them are damaged by heavy rains and method. Furthermore, the statistical values of the
earthquakes, and in a few worst cases, the dams parameters related to the soil erosion are assumed.
are completely destroyed. To mitigate such disas- The variability of the discharge hydrographs are
ters, improvement work is conducted on the most derived from these statistical values. The discharge
decrepit earth-fill dams. Since there is a recent hydrographs are applied for the flood simulation
demand for low-cost improvements, the develop- as input waves. In this research, the shallow water
ment of a design method for optimum improve- equations are solved by the HLL Rieman solvers
ment work at a low cost is the final objective of (Toro 1999), and the submerged area and the head
this research. The previous study evaluated the of the flood discharge in the downstream area are
risk to earth fills during heavy rains and inves- predicted. From the estimated probability of over-
tigated the effect of the improvement according flow, and the probability of submerge in the down-
to the total cost (Nishimura et. al. 2009). In this stream are calculated. Considering the disaster loss,
research, the probability of heavy rains and the the risk of the submerge is evaluated. Considering
uncertainty of the soil properties are simultane- the improvement cost, including improvement of
ously considered. spillway, and the expected total cost is evaluated.
Firstly, the quasi-rainfall intensities are gener- By comparing the total costs between the original
ated from random numbers, based on the statisti- and improved states of the earth-fill dam.
cal rainfall model considering the sequence of the
rainfalls. Secondly, the inflow and discharge are
estimated considering the effect of the reservoir 2 QUASI-RAINFALL MODEL
storage. When the peak overflow head on the spill-
way bed, hp, becomes greater than the design over- In this research, the rainfall events continuing
flow head, hd, overflow occurs. Then, the probability for 72 hours are simulated based on the annual

291

ISGSR2013.indb 291 10/18/2013 9:40:35 AM

 maximum rainfall intensities obtained from the where z is standard normal random number gener-
rainfall data records in Okayama City, Japan for ated by using Box-Muller method (Rubinstein 1981).
a span of 45 years. A dam break almost happens Then, the normal random number Y is transformed
within 24 hours on an empirical basis. To cover all into the random number X which has the distribu-
cases, the longer consecutive rainfalls for 72 hours tion same as the actual rainfall using the following
are used. The cumulative distribution function Fk equation:
(x) of rainfall intensity x (mm/h), for k (= 1∼72)
hours after the rain starts, is determined with the X Fk−1 ( ( y)) (5)
mean rank method as follows:

Fk(x) = mk(x)/(N + 1) (1) If X is used directly as the quasi-rainfall, the

pattern causing overflows, may be fixed, because
where mk(x) is the number of rainfall intensities the cases of the heavy rains are very limited. To
after k hours that do not exceed x, and N denotes prevent it, a method that the rainfall intensity is
the number of years. Figure 1 shows an example of reduced or extended, keeping the shape of hyeto-
the cumulative distributions. graph is proposed.
Then, rainfall intensity x (mm/h) is transformed The Gumbel distribution is assumed for the
into random variable y following the standard nor- distribution of the total rainfall T (mm/72 hours)
mal distribution as of the annual maximum 72 hours rainfalls, F(T) is
determined as follows;
y 1
(Fk (x)) (2) 72
∑ xi T (6)
where Φ is the standard normal distribution i =1
function.
Then, the correlation coefficients ρij (i, j = 1,
F (T ) = exp − e ( y
) y = a (T − T ) (7)
2, …, 72) between probabilistic variable y after i
hours and j hours can be estimated by the follow- where xi is the intensity of the annual maximum
ing equation. 72 hours rainfalls after i hours, and a and T0 are the
parameters employed to adjust the observed data
45 to the theoretical distribution function (Iwai &
∑ ( yi μ) ( j μ ) Ishiguro 1970). Then, the random variable T′ is
I =1
ρij = (3) generated as total rainfall of the from Gumbel dis-
σ2 tribution F(T) shown in Figure 2.
Lastly, the adjusted rainfall intensity, X′(mm/h)
Here, the set of correlation coefficients is viewed is determined for each hour as following equation.
as a matrix, namely, R = [ρij]. Since R is positive
definite, the lower triangular matrix L, satisfied ⎛ 72 ⎞
with LLT = R, is obtained by the Choleky decom- X′
⎝
T′ ∑ Xi ⎠ X (8)
position. A normal random number Y can be pro- i =1
duced as follows:
An example of a generated 72 hours rainfall
Y = Lz (4) series is exhibited in Figure 3.

Figure 1. Cumulative probability distribution of rain- Figure 2. Cumulative distribution of maximum annual
fall intensity after 6 hours. continuous precipitation.

292

ISGSR2013.indb 292 10/18/2013 9:40:35 AM

 Table 1. Probability of overflow.

Area at full Water Probability

Earth-fill water level storage of overflow
dam Aw (m2) V (m3) Pf (%)

A 7700 18900 0.2935

B 3100 5000 0.3334
C 3400 10800 3.4728
D 3400 15700 2.7582

Figure 3. An example of quasi-precipitation.

The results for four dams are described in

3 EVALUATION METHOD FOR THE Table 1.
PROBABILITY OF OVERFLOW

At first, the quantity of inflow, discharge and stor- 4 FLOOD SIMULATION

age are calculated. The inflow equation is defined
as follows (JSIDRE 2002); 4.1 Determination of hydrograph

Qin f prA / ..6 (9) Once the event of the overflows happens, the flood
simulation must be done as a next step to estimate
the damage in the downstream area. The discharge
where Qin = inflow to the reservoir (m3/s), fp is
hydrographs are determined to be applied for
the peak runoff coefficient, r is the quasi-rainfall
the flood simulation as input waves. An example
intensity (mm/h) generated in chapter 2 and A is
of section of the embankments analyzed in this
the area of the basin (km2). Uniform random num-
study is shown in Figure 4. When the overflow
bers are used for fp in the 0.7 to 0.8 range (JSIDRE
occurs, the embankment is supposed to be failed
2002). The discharge equation for a rectangular
as in Figure 5, which shows a vertical section of an
weir as used in this study is;
embankment. This section of the flood way is to be
assumed to have a trapezoid shape. The overflow
Qout CBS h2 3
CB (10) hydrograph from broken section is determined by
the runoff discharge Qdis. is formulated as follows;
where Qout = discharge (m3/s), C = discharge
coefficient, Bs = width of spillway, and h = static
Qdiis = ggA3 ( ∂A / ∂h ) (14 )
or piezometric head on a weir referred to the weir
crest. The storage of water in the water reservoir U Qdiis / A (15)
Vr is estimated as follows;
dH / dt = − E /(1 n p ) (16 )
Vr Aw h (11)
E = α (τ − τ c ) (17 )
where Aw is area of water reservoir (km2) and h is τ ρC f U 2
( )
overflow head (m). The decreasing rate of the stor-
age V with the runoff is: Cf 2
/l {R exp((κ Ar − 1)}/
)}// s] ( )
Vr dt = Qin Qout
dV (12)
where A is the cross-sectional area of broken
The overflow head h is determined from section (m2), U is the flow velocity (m/s), H is the
Equation (12), and the maximum h within the overflow head on the bottom of the failed section
72 hours, is defined as the peak overflow head on (m), E is the erosion rate of embankment material
the spillway. When hp becomes greater than the (m/s), np is the porosity of the embankment, α is
design overflow head hd, the overflow occurs. Then, the erosion rate coefficient (m/s/Pa), τ is the shear
the probability of overflow is defined by Equation stress (Pa), τc is the critical shear stress (Pa), ρ is
(13) as the times of hp < hd in the iterations of the the water density (kg/m3), κ is Karman constant,
Monte Carlo simulation (Rubinstein 1981). R is the hydraulic radius (m), Ar is the constants
(=8.5) and κs is roughness height (m).To consider
uncertain material values, the variabilities of the
Pf ⎡ hd < hp ⎤ (13)
⎣ ⎦ several soil parameters of the earth-fills are dealt

293

ISGSR2013.indb 293 10/18/2013 9:40:37 AM

 ⎛ h⎞ ⎛ uh ⎞ ⎛ uh ⎞
U ⎜ h⎟ , F = ⎜ u 2 h + gh ⎟ ⎜
g / 2⎟ , G = ⎜
2
uvh
⎟
⎜ ⎟ ⎜ ⎟
⎝ vh ⎠ ⎜⎝ uvh ⎟⎠ ⎜⎝ v 2 h + gh2 / 2⎟⎠

(21)

Figure 4. A section of the embankments (site C). ⎛ 0 ⎞ ⎛ 0 ⎞

⎜ ⎟
S So + S f = ⎜⎜ ghSoox ⎟⎟ + ⎜ − ghS ffx ⎟ (22)
⎜⎝ ghSooy ⎟⎠ ⎜ − ghS ⎟
⎝ ffy ⎠

in which t is the time, x and y are the horizontal

orthogonal axes, h is the water depth, u and v are
flow velocities for x and y directions, g is the accel-
eration of gravity, and Sox and Soy are the inclina-
tions along the x and y axes of the river bottom, Sfx
and Sfy are the inclinations of friction. Inclination
Figure 5. Assumption of failure section with overflow. So is obtained from the height of the bottom, zb,
and the positive direction of inclination is defined
to be downstream, namely,
Table 2. Average and standard deviation of parameters.
∂zb ∂z
Β (m) κs (m) α (m/s/Pa) γ (rad) Sox = − , Soy = − b (23)
∂x ∂y
Average 5.9 0.09 5.50E-05 54.226
Standard deviation 1.18 0.018 1.10E-05 17.775 The inclinations of the friction are defined from
Manning’s formula,

with as probabilistic parameters. The normal ran- n2u u 2 + v 2 n2v u 2 + v 2

dom numbers are substituted for overflow length S fx = S fy = (24)
h4 / 3 h4 / 3
Β, κs and α. The angular degree of cross-section of
the failed section γ is also provided as a probabilis- in which n is Manning’s ratio. The equations are
tic variable (Francesco 2008). The average and the solved by the Finite Volume Method (Yoon &
standard deviation of each parameter are described Kang 2004) (FVM), employing two dimensional
in Table 2. Although, essentially, parameter τc is a rectangular cells. The FVM is the numerical
probabilistic parameter, it is dealt with as a deter- method based on the integral type equation, and
ministic parameter (= 0.01) for simplicity. Since the the analytical area is divided into finite number
assumption means that the strength of the earth- of cells. A group of cells objective for analysis is
fill against the erosion is not considered, it leads defined as a “control volume”. An example of the
the results of the flood simulation to the safe side. control volume is exhibited in Figure 6. The cell
Since the variability of np is small generally, it is objective for the calculation is defined as L, and
assumed to be constant of 0.30 in this study. At the surrounding cell are defined as R in the figure.
inflow cells of flood simulation, the flow velocity
U is given for the downstream direction, and the
overflow head is given as the water depth.

4.2 Flood simulation method

As basic governing equations, two-dimensional
shallow water equations are employed, in which
the flow velocity is assumed to be equally dis-
tributed along the vertical axis. Furthermore, the
incompressibility of water and the inclined bottom
are also assumed (Yoon & Kang 2004).
∂U ∂F ∂G
+ + =S (20)
∂t ∂x ∂y Figure 6. Typical control volume.

294

ISGSR2013.indb 294 10/18/2013 9:40:39 AM

 Following this definition, the governing equation 1
is derived as following equation; u*
2
( L + R ) + ghhL ghR (29)

( )
dU i 1 1 1
dt
=−
Ωi
∑ E ⋅ nij ΔΓ ij + Si (25) gh* =
2
ghhL + ghR +
4
( L − R )⋅n (30)
j =1

in which q = [ ] . Boundary conditions and

T
in which E = F + G, n is unit normal outward vector,
and ΔΓ is the length of the boundary. The numer- input cells are shown in Figure 7. In input cells,
ical flux E ⋅ n is calculated by the HLL Rieman h and v are obtained from hydrograph determined
solver (Toro 1999) expressed as follows; in 4.1.

⎧(E ) ⋅ n ( ≥ )
⎪ 5 RELIABILITY-BASED DESIGN
⎪ SR (E ) n − SL (E ) n
⎪
E n = ⎨ + SRSL ⎡⎣( U ) ( U )⎤⎦ The probability of submerge in a cell i per year is
⎪ SR − SL
(SL < < ) defined as:
⎪
⎪( ) ⋅ n ( ≤ )
NE
⎩ Pi ∑ Pf ,ii jjp ⋅ Psi jjp (31)
(26) jp =1

where UL and UR = reconstructions of U on the left in which the subscript jp corresponds to the event,
and right sides, respectively, and SL and SR = wave which is the combination of the failure dams, NE
speed estimates. is number of events, Pf is probability of overflow,
and Ps is the probability of submerge under the
⎧ ⎛ q ⋅ n − ghL ,⎞ condition that the overflow happens.
⎪min ⎜ *L ⎟ if both sides are wet The expected total cost within the lifetime
⎪⎪ ⎝u ghh* ⎠ period = t is given by the following equations.
SL = ⎨
⎪ q L ⋅ n − ghL if the right
i side is dry M
⎪ CT C + ∑ C f ,i E [ ni ] (32)
⎪⎩ q R n − 2 ghR if the left f side is dry
i
(27)
E [ni ]
⎧ ⎛ q ⋅ n − ghR ,⎞ ⎧ tI
⎪max ⎜ *R
⎪⎪ ⎝u ghh*
⎟ if both sides are wet
⎠
⎪∑ ⎡P (1 − PO ,i )
= ⎨ k =1 ⎣ O ,i
k −1
+ (tl { } (
k ) PI i ⎤⎦ )
SR = ⎨ ⎪
⎪ q L ⋅ n − ghL if the right
i side is dry ⎩tl ⋅ PI ,i (Improved)
⎪
⎪⎩ q R n − 2 ghR if the left f side is dry (33)
(28) in which M is number of cells, CT is the expected
total cost n is the frequency of overflows within a
lifetime span of t (years), PO and PI are the prob-
abilities of overflow a year corresponding to the
original and the improved states of the embank-
ment, respectively, C0 is the cost of the improve-
ment, and Cf is the damage loss due to flooding. PO
and PI are derived from Equation (31).
In this research, the improvement of the spill-
way is considered, and the improvement brings
about a drastic increase in the discharge ability of
the spillway. In Equation (32), improvement cost
C0 is zero for the original state of the embank-
ment, and it is assumed that when embankments
are failed due to overflows, that they will be
restored to the same level as the improved state of
Figure 7. Boundary condition and input cells. the embankment.

295

ISGSR2013.indb 295 10/18/2013 9:40:42 AM

 6 CASE STUDY

6.1 Outline of site

A group of earth-fill dams, A, B, C, and D, prescribed
in Figure 8, is analyzed in this study. These dams
were constructed in the valleys of a mountainous
area. Profiles of the dams are described in Table 1.
A section of the embankments of dam C is shown in
Figure 4 and similar to that of the other dams. The
broken sections of all ponds are same as in Figure 5.
Dams A and B and dams C and D are separated
by a ridge, and are not affected by each other, while
there are correlations for the probability of over- Figure 9. Hydrographs for scenario flood simulation
flow between A and B or C and D. Since dam B when dams C and D break.
is smaller than dam A, dam B is inevitably broken
following the breaking of dam A. If the upstream
dam, dam C, is broken, then dam D is inevitably
broken. Dam D never breaks without also break-
ing dam C, according to the calculation results of
the probability of overflow. Thus, dams C and D
are denoted together as dam CD. The probabilities
of overflow are expressed by Equation (34), con-
sidering the relationships among the dams.

Pf (A ∩ B ) ( ) (
0 Pf A ∩ B ≠ 0, Pf A ∩ B = 0 )
Pf (C ∩ D) 0 Pf (C D) = 0, Pf (C ∩ D) = 0 (34)

The symbols A, B, C, and D are the sets employed

to represent the events of breaking of each dams.

6.2 Result of flood simulation

The example of the discharge hydrographs from
the dams are shown in Figure 9. Figure 9 corre-
spond to the case in which dam CD breaks and

Figure 10. Head of flood discharge (cm).

the overflow depth and the velocity are shown. As

Manning’s ratio, the value, n = 0.026 is assumed.
The expected maximum head of the flood dis-
charge is described in Figures 10(a)–(c), which cor-
respond to the cases, in which dams A and B break,
CD breaks, and dams A, B, and CD break, respec-
tively. In the figure, submerged depths smaller than
Figure 8. Locations of earth-fill dams and geography. 1.0 cm are disregarded.

296

ISGSR2013.indb 296 10/18/2013 9:40:46 AM

 research. The generated rainfall events were
applied to calculate the inflow from the sur-
rounding basin into the reservoirs.
2. The probability of overflow caused by the heavy
rains has been calculated for several earth-fill
dams. The inflow and the discharge were esti-
mated considering the effect of the reservoir stor-
age. We assumed that overflows will occur when
the maximum overflow head on the spillway bed
becomes greater than the design overflow head.
The probability of overflow was then determined
by the Monte Carlo simulation. The probability
of overflow of a dam in the studied site is calcu-
lated to be 3.47% in the maximum case.
3. The variability of the several soil parameters of
the earth-fill dams have been dealt with as proba-
bilistic parameters to consider uncertain material
values for producing the discharge hydrograph.
Normal random numbers were assigned to the
overflow length, the roughness height and the
erosion rate coefficient. And the angular degree
of the cross-section of the failed section was also
Figure 11. Spatial distributions of risk. (Unit: 1,000 provided as a probabilistic variable.
JPY/m2). 4. The maximum head of submerge and the risk
in the downstream area were estimated. Finally,
the expected total costs were compared between
Table 3. Risks and expected total costs. current and improved states of the spillways,
State Improvement costs Total risk Total costs
and the effect of the improvement work has
been evaluated.
Current 0 335,055 335,055
Improved 199,200 114,970 314,170
ACKNOWLEDGMENT
Unit: 1,000 JPY.
This work was partly supported by JSPS
In comparison among Figures 10 (a), (b) and (c), KAKENHI (23248040).
the heads of discharge of (a) and (b) are greater than
(c), although the volume of discharged water is great-
est among the three cases. The reason is that the flow REFERENCES
velocity affects the results. This phenomenon means
Francesco M. 2008. Model for Predicting due to Earthen
that damage in the downstream area is not necessar- Dam Breaching I: Formulation and Evaluation, Jour-
ily correlated with the depth of submergence. nal of Hydraulic Engineering, ASCE 134: 1688–1969.
Figure 11 exhibits the spatial distributions of Iwai, S. and Ishiguro M. 1970. Applied hydro-statistics,
risk corresponding to the original and improved Morikita (in Japanese).
states of earth-fills. With improvement of the spill- Japanese Society of Irrigation, Drainage and Rural
way, the risk is reduced dramatically, since the prob- Engineering. 2002, Maintenance of irrigation tanks.
ability of overflow is predicted to be zero for Dam Guideline of land improvement designs (in Japanese):
CD. The total risk and the expected costs are given 157–162, 205–211.
Nishimura, S. and Mori, S. 2009. Risk evaluation and
in Table 3. The difference between the costs of the
reliability-based design of earth-fill dams for overflow
original and improved states means the effect of the due to heavy rains. Proc. of ICOSSAR2009: 246–251.
improvement works. The value is evaluated as: Rubinstein, R.Y. 1981. Simulation and the Monte Carlo
Method, John Wiley & Sons.
335,055 − 314,170 = 20,885 (1,000 JPY) Toro, E.F. 1999. Riemann Solvers and Numerical Methods
for Fluid Dynamics—A Practical Introduction—2nd
Edition, Springer: 315–331.
7 CONCLUSIONS Yoon, T.H. and Kang, S-K. 2004. Finite volume model
for two-dimensional shallow water flows on unstruc-
1. A generation method for quasi-rainfall, using tured grids. Journal of Hydraulic Engineering, ASCE
random numbers, has been proposed in this 130: 78–688.

297

ISGSR2013.indb 297 10/18/2013 9:40:48 AM

This page intentionally left blank
 Geotechnical Safety and Risk IV – Zhang et al. (eds)
© 2014 Taylor & Francis Group, London, ISBN 978-1-138-00163-3

Reliability analysis of 20-km river dike against liquefaction failure

Yu Otake, Yusuke Honjo & Yuichi Hiramatsu

Gifu University, Gifu, Japan

ABSTRACT: The authors proposed a reliability analysis method for long continuous structures
considering investigation sites to counter liquefaction failure. This method is characterized by continu-
ously calculating the failure probability caused by liquefaction at an arbitrary point and quantifying the
uncertainty concerning the investigation site. The purpose of this study is to verify its effectiveness based
on application to an actual river dike that suffered liquefaction failure due to the 2011 Earthquake off the
Pacific coast of Tohoku.

1 INTRODUCTION The main verification points are the possibility of

extracting the high-risk collapse zone, and con-
1.1 Purpose firming whether the investigation intervals and
quantity are adequate under the current investiga-
The authors have proposed a reliability analysis
tional status.
method for long continuous structures with con-
sideration for investigation sites against liquefac-
1.2 Target structure and soil conditions
tion failure using the liquefaction index PL-value
(Otake & Honjo 2012). This index is extremely The 20-km interval between STA 25 and 45 in a
simple, and enables calculation using the SPT first class river in Japan is targeted in this study.
N-value and determination of particle size. This “STA” represents the site from the river mouth,
method is characterized by continuously calculat- and is measured in units of 1 km.
ing the failure probability caused by liquefaction The soil in the levee body is sandy silt, and the
at an arbitrary point, and quantifying the uncer- dike is 3∼5 m high. The dike height increases as
tainty concerning the investigation site. you go downstream, and the rate of increase in
The purpose of this study is to verify its effec- height per STA is roughly constant.
tiveness based on application to an actual river dike Figure 1 shows a summary of the longitu-
which suffered liquefaction failure caused by the dinal soil profiles about the target river dike
2011 earthquake off the Pacific coast of Tohoku. and collapsed location. Table 1 shows the soil

Figure 1. Longitudinal soil profiles and site of collapsed location.

299

ISGSR2013.indb 299 10/18/2013 9:40:49 AM

 Table 1. Soil parameters of each layer.

Main soil SPT-N (mean) γ (kN/m2) Fc (%) D50 (mm) Vs (cm/s)

Bc Sandy silt 3.3 18 85 0.025 110

Bs Silt sand 65 0.350 110
Ac Silt 2.6 16 75 0.025 125
Sandy silt
Acs Silt sand 7.5 17 50 0.070 140
As Medium sand 18.3 18 10 0.350 150
Fine sand

Figure 2. Temporal history of strong movement. Figure 3. Collapse of river dike caused by liquefaction.

parameters for the dike body and alluvium. The 2 COLLAPSE RISK ASSESSMENT USING
river body soil is mainly sandy silt (Bc), but sandy PL-VALUE
earth filling has possibly accumulated under the Bc
in the former riverbed area. The liquefaction risk at an arbitrary point is
The alluvium layers under the dike body are evaluated using the liquefaction index i.e. PL-value,
roughly classified into two types, silt or sandy silt as proposed by SFHB2002 (JRA 2002). In fact,
(Ac) (which is a non-liquefiable layer) and medium we calculated the correlation between the PL-value
or fine sand (As) (which is liquefiable), and the total and collapse risk as follows:
thickness of the alluvium layer is between 10 and
1. Calculation of PL-value at all investigated sites.
25 m. Furthermore, diluvium, which is regarded as
2. Development of empirical fragility curve in the
engineering bedrock, has accumulated under these
form of a normal distribution function using a
alluvial strata.
two-value regression analysis based on informa-
The soil investigations mainly consist of a
tion about whether the site collapsed or not.
Standard Penetrating Test (SPT) and determination
of the particle size. There were 83 SPTs in total,
but only ten particle size determinations. 2.1 Calculation of PL-value
2.1.1 Calculation procedures
1.3 Profiles in strong movement and collapse PL-value at the investigation sites was basically
status of target river dike based on SFHB2002, but there were improvements
to the original method in order to take the long
Figure 2 shows the observed time history of accel- duration and local site of shaking amplification
eration at the site located on bedrock near the into consideration. The calculation procedures are
target river dike. The maximum acceleration of described based on notations in the improvement
this earthquake wave was moderate by Japanese points as follows.
standards, but it was characterized by a very long Liquefaction evaluation along the depth was
duration. calculated as follows.
The total duration was around 200 seconds,
and during this period, the main shaking occurred FL R /L
L (1)
between 90 and 145 seconds (i.e. for 55 s).
In addition, four points along the target river R cW RL (2)
dike interval collapsed after the shaking (Fig. 1).
A summary of the collapsed status is presented in where FL is the resistance ratio of liquefaction, R
Figure 3. All of the collapsed area was along the is the dynamic shear resistance ratio, L is the shear
former riverbed or natural levee, and was consid- stress load during strong movement, cw is the adjusted
ered to have been caused by liquefaction of earth coefficient of strong movement duration, and RL is
filling Bs, which resulted in leveling of ground on the dynamic triaxial resistance. RL is defined in Equa-
the flood plain. tion (3), and is correlated with SPT N-value.

300

ISGSR2013.indb 300 10/18/2013 9:40:49 AM

 RL 20

⎧⎪ 0.08822 Na /1.7 ( Na < 14 )

PL ∫0 ( FL )( − .5x )ddx (6)
=⎨
⎩⎪0.0882 Na /1 7 1 6 10 ( Na − 14 )
6 4.5
(14 a
) Here, x is the depth from the ground surface.
Note that when FL>1, FL is converted to 1.0.
(3)
Actually, we use the definition of Equation (7),
where, Na is the modified N-value for liquefaction since SPT-N is investigated each 1m over the
evaluation considering the effective overburden depth.
stress and fine particle fraction contents. 20
PL ∑( FLLii )( − 5xi ) (7)
Na c1N1 c2 (4) i= 1

2.1.2 Calculated PL-value

N1 170 N/(σv′ + 70 )
N/( (5) Figure 4 shows the calculated PL-values at all
investigation sites, with the upper figure indicating
c1 and c2 are adjusted coefficient correlations with the maximum acceleration at the ground sur-
fine particle fractions, and σV′ is over effective face based on response analysis. While calcu-
burden stress. cw is the adjusted coefficient for the lated maximum accelerations were observed to
duration of continuous shaking, which employs have relatively large variance, these tended to
1.0 for subduction zone earthquakes. decrease in the downstream part when viewed
Here, since the target earthquake lasted much macroscopically.
longer than typical subduction zone earthquakes, On the other hand, the lower figure shows PL
we use cw = 0.5 for the long-duration time subduc- (SHAKE) and PL (Original). These differ from the
tion zone earthquake as proposed by Yoshida et al. calculating method for L; here the former is when
(2009). using SHAKE, while the latter is in the case of
L is defined as the ratio between the shear stress simplified conventional equation in SFHFB. Here-
load and overburden stress. It is set as a standard after, in order to confirm the difference between
value for each of the three classified soil types for these indices, we continuously draw a parallel
ease of calculation. between PL (SHAKE) and PL (Original).
Here, we improved the calculation method for The vertical error-bar of PL-value in Figure 4
L for a more rational consideration of local site is observation error including the soil test and the
effects. That is to say, in order to evaluate the shear measurement of tri-axial test and SPT.
stress distribution at different depths during shak- In here, observation error of PL-value was cal-
ing, we employed one-dimensional equivalent lin- culated based on the assumption that the most
ear analysis using SHAKE (Schanable et al. 1972). part of regression error of Equation (3) is test and
Finally, the PL-value was calculated by weighted measurement error. Also, the regression error is
integration of FL at the depths based on Equation (6). modeled by the normal distribution, and the mean
Weight is set as a linear function of the depth. value is 0.0 and SD is 0.10 (Otake et al. 2012).

Figure 4. Results of ground response analysis in shaking and PL-value at all investigational sites.

301

ISGSR2013.indb 301 10/18/2013 9:40:51 AM

 From the above assumption, the error is assumed
as independent over the depth, it was calculated as
follows. It is different value in each investigation
site, but these are included the range of roughly
0.10–0.30 as COV.
20 20
⎛ 10 − 0.5xi ⎞
PL ∑ xi − ∑ ⎜⎝ Li ⎟⎠ Ri (8)
i=1 i=1

2
20
⎛ 10 − 0.5x ⎞
Var [ PL ] = ∑ ⎜⎝ L i ⎟⎠ Var[R L ] (9)
i=1 i

2.2 Empirical fragility curve

It is assumed the empirical fragility curve can be
expressed in normal distribution form, and devel-
oped as a function of the PL-value. Here, the
explained variables Y are 0, i.e. un-collapsed, and 1,
i.e. collapsed, and explanatory variables are calcu-
lated for PL at all investigation sites as mentioned
in subsection 2.1.2.
The likelihood function is expressed as

N
L ∏ [ FFRR(l PLi )]α i [ − FR(lln PLi )]1−α i (10)
i =1

where, F(.) represents the fragility curve for a

specific state of damage. The symbol of “ln” rep-
resents logarithm natural. And, PLi is the PL-value,
αi depends on whether or not there is damage
(0 or 1), and N is the total number of data (i.e.
investigation sites).
Figure 5. Histogram of PL in each case (fail and safe).
⎛ ln( μFFR / c ) ⎞
FR(ln( PLi )) = Φ ⎜ (11)
⎝ ξ ⎠⎟

Φ[.] is the standardized normal distribution

function. The two parameters c and ξ are com-
puted to satisfy the following equations to maxi-
mize ln L.

∂ ln L ∂ ln L
= =0 (12)
∂c ∂ξ

Figures 5 and 6 show the calculated empirical

Figure 6. Calculated empirical fragility curves.
fragility curves in each case (i.e. PL (SHAKE) and
PL (Original))
It can be seen that PL (SHAKE) is able to sepa- 3 LIQUEFACTION RISK ASSESSMENT
rate collapsed and non-collapsed sites than PL
(Original), so PL (SHAKE) is considered to be an 3.1 Random field modeling of PL-values
adequate parameter to evaluate the assessment of
liquefaction collapse risks for river dikes. From the The lnPL was modeled by Random Field (RF).
above, we chose the PL (SHAKE) as the adequate The mean value and variance of the RF were set
index for the risk assessment against liquefaction. as the mean value and variance of the calculated

302

ISGSR2013.indb 302 10/18/2013 9:40:53 AM

 lnPL respectively. Note that the mean value is 2.3
and variance is 0.692.
Then the auto-correlation function for modeling
the covariance of RF was analyzed. We classified
these data into two groups (i.e. Zone 1 from STA
25 to 35, and Zone 2 from STA 35 to 45) to con-
firm the effect of any difference in micro-topogra-
phy classification. Zone 2 has extremely complex
micro-topography composed mainly of the former
riverbed and a natural levee, whereas Zone 1 is
composed of a relatively simple micro-topography.
By considering the above, the identified results
for auto-correlation distances in the case of all
data, Zones 1 and 2 are shown respectively in these
figures. In conclusion, we chose the Gaussian
form and auto-correlation distances are identified
200 m in Zone 1, 100 m in Zone 2 respectively.

3.2 Liquefaction risk analysis

The probability of a river dike failing due to lique-
faction is expressed as:
Pf prob g ≤ )
prob( (13)
g F (ln PL ) − ln PL
FR (14)

where, FR is the fragility function and PL is the

PL-value. In such cases, FR and PL are assumed
to be independent and modeled by normal
distribution; Pf is expressed by Equation (15), and
can easily be calculated.
Figure 7. Identification of auto-correlation distance in
⎛ ⎞ case of PL (SHAKE).
⎛ 0 − μg ⎞ c − μln PL
Pf = Φ ⎜ ⎟ =Φ⎜ ⎟ (15)
⎝ σg ⎠ ⎜
⎝ ξ 2 + σ ln
2
PL
⎟
⎠ estimation error (i.e. statistical estimation error),
and the precision of fragility curves (i.e. model
where, Φ[.] is the standardized normal distribution error) at arbitrary points.
function. In Figure 9, the results in case of Zone 2 in
The continuous failure probabilities Pf along the same form are shown. The contribution of sta-
river dike are shown in Figure 8. The upper figure tistical estimation error in Zone-2 is very large all
shows the PL-value at arbitrary points, and the the points. The investigation in Zone-2 must be in
filled points show the calculated values using the shorter interval since the auto-correlation distance
investigation data at that point. is very small for being the former riverbed area.
The solid line is the mean value based on inter- The results of the analysis are believed to
polation using the Kriging method, and the dotted provide useful information to manage river dikes.
line shows a range of ±1σ (i.e. standard deviation). For example, the collapsed area (i.e. STA 31∼32,
And here, note that the hatched area is the col- STA 27.5∼34, STA 37.5∼34 and STA 39) charac-
lapsed area. Furthermore, the reason why the solid terized by small investigation intervals due to the
line does not pass through all data is because we focus on additional investigation after the shaking,
independently considered the observation error in the contribution of the model error indicates larger
Kriging. than 80%, and the statistical estimation error is
The middle figure shows the superimposed very small (i.e. less than 10%). This indicates that
failure probability. As shown in this figure, the owners should certainly plan to design counter-
proposed method is regarded as useful, since the measures due to this being a high-risk area.
collapsed sites (i.e. hatched area) are regarded as By contrast, in the case of large investigation
high-risk zones. intervals (e.g. STA 32∼36 and so on), statistical esti-
The lower figure shows the contribution of mation errors are a high contribution source (i.e.
uncertainty sources, in here separated PL-value about 40%). Fortunately, the area did not collapse,

303

ISGSR2013.indb 303 10/18/2013 9:40:56 AM

 Figure 8. Estimated PL-value, failure probability and contribution of uncertainties (Zone 1, STA 25-35).

Figure 9. Estimated PL-value, failure probability and contribution of uncertainties (Zone 2, STA 35-45).

but it was considered that additional investigations Otake, Y. & Honjo, Y. 2012. Reliability based design
should be conducted. on long irrigation channel considering the soil
investigation locations. Proceedings of Geo Congress
2012, 2836–2845.
Otake, Y., Honjo, Y. & Koike, K. 2012. Evaluation of
4 CONCLUSION soil liquefaction risk on linearly structure considering
the soil investigation locations, Japanese Geotechnical
Conducting countermeasures for a long continuous Journal 7(1): 283–293. (in Japanese).
structure such as this river dike is time-consuming Schnabel, P.B., Lysmer, J. & Seed, H.B. 1972.
and costly. The information provided here is very SHAKE-A computer program for earthquake response
useful for planning such a construction and deter- analysis of horizontally layered sites, Report EERC
mining its priorities. 72-12, Berkeley California, Earthquake Engineering
It should be pointed out that the owner could Research Center, University of California.
obtain more information on soil properties by Yoshida, N., Ohya, Y., Sawada, S. & Nakamura, S. 2009.
Simplified procedure for evaluating liquefaction
conducting additional soil investigations. The potential under ocean trench type long period earth-
results from this method provide very valuable quake, Journal of JAEE 9(3): 28–47. (in Japanese).
information in this regard.

REFERENCES

Japan Road Association 2002. Specifications For

Highway Bridges.
Japan Institute of Country-ology and Engineering (JICE)
2009. Manuals of River Structures. (in Japanese).

304

ISGSR2013.indb 304 10/18/2013 9:40:57 AM

 Geotechnical Safety and Risk IV – Zhang et al. (eds)
© 2014 Taylor & Francis Group, London, ISBN 978-1-138-00163-3

Effect of slenderness ratio on the reliability-based serviceability limit

state design of augered cast-in-place piles

S.C. Reddy & A.W. Stuedlein

Oregon State University, Corvallis, Oregon, USA

ABSTRACT: This study investigated factors that control the reliability of Augered Cast-In-Place (ACIP)
piles in predominately cohesionless soils under axial compression at the Serviceability Limit State (SLS).
A simple probabilistic hyperbolic model was used to account the uncertainty in the load-displacement
relationship using correlated bivariate curve-fitting parameters. Contrary to previous studies, the curve-
fitting parameters were found to be dependent on pile slenderness ratio (D/B) and the effect of D/B and
other pertinent variables (e.g., uncertainty in capacity, displacement) on SLS reliability was investigated
using a First-Order Reliability Method (FORM). The D/B ratio had a considerable effect on founda-
tion reliability, illustrating the importance of the dependence between the load-displacement behavior
(i.e. curve-fitting parameters) and pile geometry and stiffness. In general, the uncertainty in the capacity
model had a larger effect on reliability than that of the allowable displacement; the reliability index was
found to approach an upper bound limit regardless of the level of uncertainty in allowable displacement
and the pile capacity model.

1 INTRODUCTION Design [LRFD]). RBD procedures for Augered

Cast-In-Place (ACIP) piles (e.g., Stuedlein et al.
The behavior of geotechnical systems under load- 2012) are not yet accepted in codes.
ing is often difficult to predict due to the inher- Owing to the lack of model statistics for pile
ent heterogeneity of the geologic environment. displacement, foundation reliability at the SLS
Because the compositional and mechanical prop- is not as well understood compared to the ULS
erties of soils are variable, many parameters used (Phoon et al. 2006). In order to assess founda-
in geotechnical design are uncertain. Traditionally, tion reliability at the SLS, Phoon (2006) proposed
the uncertainty associated with many geotechnical a simple probabilistic hyperbolic model that
design parameters has been assessed jointly using accounts for the uncertainty in the entire load-
a deterministic global factor of safety; which is displacement relationship using a bivariate ran-
frequently based on engineering judgment and dom vector consisting of hyperbolic curve-fitting
experience. Reliability-Based Design (RBD) pro- parameters, which were found to be correlated
cedures can overcome many of the restrictions of and non-normally distributed. Phoon & Kulhawy
traditional design checks (e.g. Allowable Stress (2008) describe a translational model to incor-
Design [ASD]), and explicitly incorporate the porate the correlated random variables into reli-
uncertainty in the individual variables and their ability calculations using a database of 40 loading
potential correlation into the overall model. The tests on ACIP piles.
probability of failure for a prescribed limit state This study used an expanded database to inves-
that results thus allows a quantitative assessment tigate factors affecting the reliability of ACIP piles
of risk. As a result, RBD is quickly becoming the at the SLS. Contrary to Phoon & Kulhawy (2008),
preferred alternative as the demand for risk man- the hyperbolic model parameters were determined
agement in geotechnical engineering continues to to be dependent on the pile slenderness ratio. The
grow. dependence was removed by transforming the
Modern RBD codes, in which partial safety model parameters, which were then used to assess
factors are calibrated with respect to a specific foundation reliability for different pile geometries.
limit state (e.g. Ultimate Limit State [ULS], Serv- In order to determine the variables which govern
iceability Limit State [SLS]), are now mandated for reliability, a parametric study was conducted by
design of bridge foundation elements (e.g. American varying the mean and uncertainty of allowable
Association of State Highway and Transportation displacement, uncertainty of predicted resistance,
Officials [AASHTO] Load Resistance Factor and the slenderness ratio.

305

ISGSR2013.indb 305 10/18/2013 9:40:59 AM

 2 PROBABILISTIC HYPERBOLIC MODEL reported by Stuedlein et al. (2012), ten were col-
AT THE SLS lected by Park et al. (2012), three were reported by
Mandolini et al. (2002), and one loading test was
To ensure that a specified level of performance of selected from O’Neill et al. (1999). Table 1 shows
a structure is met, it is necessary to assess the likeli- the range of pile embedment depth, D, diameter,
hood of failure at both the ULS and SLS using B, slenderness ratio, D/B, average SPT-N along the
a consistent methodology. This study focuses on pile shaft, Navg, and QSTC.
reliability at the SLS, defined by one or more pre-
defined displacements that correspond to target
allowable loads. 4 RANDOMNESS OF THE HYPERBOLIC
The load-displacement behavior of ACIP piles MODEL PARAMETERS
is influenced by multiple sources of uncertainty
that can be implicitly accounted for by fitting In order for foundation reliability assessments to
load-displacement models to date from a load test be unbiased, k1 and k2 must be statistically inde-
database. In this approach, the aleatory and epis- pendent from other deterministic variables in the
temic uncertainties resulting from the uniqueness database (e.g., SPT-N and D/B). Based on the
of each load test and the error associated with Kendall’s tau test (Daniel 1990) and adopting a
measurements taken during testing are combined 5 percent significance level (α = 5), k1 and k2 are
together and statistically characterized. Although independent of average SPT-N with p-values =
a variety of functions can be used to model the 0.81 and 0.93, respectively. However, convincing
load-displacement relationship, a hyperbolic curve evidence (p-values < 0.05) suggested that both k1
was chosen herein in order to remain consistent and k2 were dependent on D/B. Figure 1 shows
with the work pioneered by Phoon (2006). The
hyperbolic curve is represented using the applied
load, Q, normalized by the slope-tangent capacity,
QSTC (Phoon et al. 2006):

Q y
= (1)
QSTC k1 k2 y

where y = pile head displacement, and k1 and k2

are fitted coefficients. The reciprocal of k1 and k2
is equal to the initial slope and asymptotic (or ulti-
mate) resistance, respectively. Model parameters
from the new data were calculated using ordinary
least squares regression, whereas the parameters
in the Chen (1998) and Kulhawy & Chen (2005)
database were obtained directly from Phoon &
Kulhawy (2008).

3 DATABASE

The expanded database included 87 load tests on

ACIP piles constructed in predominately cohe-
sionless soils. Forty loading tests were collected
by Chen (1998) and Kulhawy & Chen (2005), 23
were compiled by McCarthy (2008), ten were

Table 1. Range of variables observed in the database.

Navg QSTC
Variable D (m) B (mm) D/B (bl/0.3 m) (kN)
Figure 1. The dependence between slenderness ratio,
Minimum 7.5 300 20.0 4 367 D/B, and model parameters, (a) k1 and (b) k2 and the
Maximum 29.0 800 68.5 54 5300 corresponding Kendall tau correlation coefficients and
p-values.

306

ISGSR2013.indb 306 10/18/2013 9:40:59 AM

 moderately strong dependence between k1, k2, and (large) k2 indicates a slowly (quickly) decaying
D/B and the corresponding Kendall’s tau correla- function and a less (more) well-defined and larger
tion coefficient, ρτ, and p-value. (smaller) asymptote.
It is worthwhile to note that these correlations In order to perform unbiased reliability analyses
make physical sense in that a smaller k1 represents a at the SLS, the correlation between model param-
stiffer pile which corresponds to a smaller slender- eters and D/B must be considered. The dependence
ness ratio; whereas a smaller k2 indicates a larger of k1 and k2 on D/B was removed by transforming
ultimate resistance which likely relates to a larger D the model parameters to:
and D/B because of the narrow range of B in the
database. B
k1,t k1 (2a)
D
5 TRANSFORMATION OF THE MODEL
D
PARAMETERS k2,t k2 (2b)
B
In order to accurately model the uncertainty in the
load-displacement relationship for the assessment After transforming k1 and k2 to k1,t and k2,t, the
of foundation reliability at the SLS, the correlation Kendall’s tau correlation test between k1,t and k2,t,
between k1 and k2 must be considered (Phoon & and D/B produced p-values = 0.78, 0.56, indicat-
Kulhawy 2008). Figure 2a shows the inverse cor- ing no correlation. Similarly, the model parameters
relation between k1 and k2, and the corresponding remained independent of SPT-N following trans-
ρτ and p-value, where a large (small) k1 and small formation. Figure 2b illustrates the correlation
between k1,t and k2,t is slightly reduced but remains
valid after transformation.
To assess foundation reliability using the trans-
lational model approach described in Phoon &
Kulhawy (2008), the marginal distributions of k1,t
and k2,t must be determined. Figure 3a and b shows
the empirical, fitted normal, and fitted lognor-
mal marginal Cumulative Distribution Functions
(CDF) of k1,t and k2,t, respectively. Also shown is
the sample mean, ki ,t , standard deviation, σi,t, and
COVi,t, defined as the standard deviation divided
by the mean, of the model parameters.
Based on the Anderson-Darling goodness-of-fit
test (Anderson & Darling 1952) and α = 5 percent,
there was no evidence to reject the null hypothesis
of lognormality for k1,t and k2,t. Therefore, k1,t and

Figure 2. Correlation between model parameters (a) k1 Figure 3. Empirical, lognormal, and normal marginal
and k2 and (b) k1,t and k1,t and the corresponding Kendall cumulative distributions for the transformed hyperbolic
tau correlation coefficients and p-values. model parameters: (a) k1,t, and (b) k2,t.

307

ISGSR2013.indb 307 10/18/2013 9:41:00 AM

 k2,t were assumed to follow a lognormal distribution
for the purpose of assessing foundation reliability
at the SLS.

6 TRANSLATIONAL MODEL
FOR BIVARIATE PROBABILITY
DISTRIBUTIONS

The translational model approach for describing

the marginal distributions of k1,t and k2,t requires the
use uncorrelated standard normal random variables
Z1 and Z2 (mean = 0, standard deviation = 1).
Figure 4. Observed and simulated load-displacement
This study followed the basic procedure outlined curves using the translational model.
in Phoon & Kulhawy (2008). First, Z1 and Z2 are
converted into correlated random variables X1 and
X2 (Phoon & Kulhawy 2008): Figure 4 shows the fitted load-displacement
curves based on the observed loading tests and
X1 Z1 (3a) those generated using the procedure for simulat-
ing k1,t and k2,t outlined above. A sufficient number
X2 Z1ρln + Z2 1 − ρln
2
(3b) of deterministic values of D/B were used to back-
transform k1,t and k2,t into k1 and k2 in order to
where ρln is an equivalent-normal correlation adequately capture the uncertainty present in the
coefficient: observed load-displacement curves. In general, the
observed and simulated load-displacement curves
⎡ ⎤ are in good agreement, and the translational model
(eζ )(eζ
2 2
ln ⎢ ρ ,
− 2 ,t
− 1) + 1⎥ can be confidently used to assess foundation reli-
ρln = ⎣ ⎦ (4) ability at the SLS using the database herein.
ζ1, ζ 2,t

and where λ1,t, ζ1,t and λ2,t, ζ2,t are the approximate 7 RBD FOR THE SERVICEABILITY
lognormal mean and standard deviation of k1,t and LIMIT STATE USING A FIRST-ORDER
k2,t, respectively, and ρ is the standard product- RELIABILITY METHOD
normal correlation coefficient for two normally
distributed variables. The second moment statistics The SLS is reached when foundation displacement,
in Equation 4 were calculated as: y, equals or exceeds allowable settlement, ya. This
study followed the approach outlined in Phoon &
Kulhawy (2008), where the SLS can be evaluated
ζ i ,t ( σ i2,t / ki2,t ) (5a) using a performance function, P:

λi ,t ( i,t ) ζ i2,t (5b) P ya − y (Q ) (7)

The correlated, lognormal marginal distribu- Failure is defined as P ≤ 0, and the probability
tions of k1,t and k2,t were thus simulated using: of exceeding the SLS, pf, is:

(ζ1,t X1 + λ1,t ) pf P (P ≤ ) (8)

k1,t e (6a)

k2,t e
(ζ 2,t X 2 + λ2,t ) (6b)
By combining Equations 1, 7, and 8, and defin-
ing a deterministic mean global factor of safety,
FS, the probability of failure is:
In order to adequately reproduce the uncer-
tainty in the observed load-displacement curves, ⎛ ya 1 Q′ ⎞
k1,t and k2,t must be back-transformed into k1 and p f = Pr ⎜ < ⎟ (9)
k2. This study calculated k1 and k2 using determin- ⎝ 1
k k y
2 a FS
F Q′ p ⎠
istic values of D/B because the uncertainty asso-
ciated with pile geometry could not be evaluated where Q′ and Q′p correspond to the applied load
from the database. and predicted pile capacity, respectively. In order

308

ISGSR2013.indb 308 10/18/2013 9:41:03 AM

 estimate foundation reliability at the SLS, the reli-
ability index, β, defined as the number of standard
deviations between the mean of the multivariate
resistance distribution and the limit state surface,
was calculated as:

β = − Φ −1 ( ) (10)

where Φ−1 is the inverse standard normal function.

A First-Order Reliability Method (FORM)
was used to estimate foundation reliability at the
SLS. First, each random variable in the limit state
function (k1, k2, ya, Q′, Q′p) was transformed into a
standard normal variable, such that the difference
in magnitude of the random variables was elimi-
nated (Hasofer & Lind 1974). Then the probability
of failure was estimated by considering the area
beneath the multivariate distribution where P ≤ 0.
The FORM approach assumes that the limit state
function is linear at the failure point, and therefore
may not be appropriate for situations where pf is
large. However, this approach is considered suffi-
cient for most geotechnical applications where the
target probabilities of failure are very small.

8 FACTORS AFFECTING FOUNDATION

RELIABILITY AT THE SLS

This study assessed the factors which govern foun-

dation reliability at the SLS by calculating multi-
ple reliability indices using FORM. Each variable
in Equation 9 was assumed to follow a lognormal
distribution, whereas the second moment statis-
tics for k1,t and k2,t were obtained directly from the
database. The mean and COV of allowable dis-
placement was varied from 10 to 50 mm and 5 to
85 percent, respectively. The applied load and pre-
dicted pile capacity were assumed to be unit mean
variables, where COV(Q′) = 20 percent based on
recommendations from Paikowsky et al. (2004).
The COV of the predicted pile capacity was varied
from 5 to 85 percent, corresponding to different
capacity prediction methods with varying degrees
of uncertainty. A FS = 3 was selected based on that
commonly adopted in current practice (Phoon &
Kulhawy 2008). Slenderness ratios of 25 and 65
were selected in order to cover the range of D/B
Figure 5. The effect of COV(ya), COV(Q′p), and D/B on
values in the database and illustrate the effect of β for mean ya equal to (a) 10 mm, (b) 20 mm, (c) 30 mm,
pile geometry on β. (d) 40 mm, and (e) 50 mm using k1,t and k2,t developed
Figures 5a-e illustrate the effect of changing the herein.
mean ya, COV(ya), COV(Q′p), and D/B on founda-
tion reliability. Foundation reliability decreases regardless of the level of uncertainty in Q′p and
more rapidly for increasing uncertainty in Q′p when ya. The same general trend was observed at differ-
COV(ya) and COV(Q′p) are relatively small (5–45 ent levels of mean allowable displacement, where
percent) and ya > 20 mm. In general, COV(Q′p) the change in β was more prominent for larger ya.
has a larger effect on β as compared to COV(ya), Overall, β was larger for larger mean allowable

309

ISGSR2013.indb 309 10/18/2013 9:41:07 AM

 displacements if all other variables in the perform- REFERENCES
ance function remained constant. At large allowa-
ble displacements (i.e. ya = 50 mm), β was observed Anderson, T.W. & Darling, D.A. 1952. Asymptotic
to be largely insensitive to the level of uncertainty theory of certain goodness-of-fit criteria based on
in ya, compared to Q′p. This illustrates the advan- stochastic processes. The Annals of Mathematical
Statistics 23(2): 193–212.
tage of an accurate ACIP design methodology. At
Chen, J.R. 1998. Case history evaluation of axial behavior of
large allowable displacements, β approaches an augered cast-in-place piles and pressure-injected footings.
upper bound limit for each level of COV(Q′p) as M.S. Thesis. Cornell University. Ithaca, New York.
COV(ya) decreases. For a mean ya < 40 mm, β is Daniel, W.W. 1990. Applied nonparametric statistics.
smaller for larger D/B, whereas the opposite is true Boston: PWS-Kent.
for ya ≥ 50 as shown in Figure 5e. Thus, accounting Hasofer, A.M. & Lind, N. 1974. An exact and invariant
for the correlation between the hyperbolic model first-order reliability format. Journal of Engineering
parameters and D/B is critical when estimating the Mechanics 100(1): 111–121.
reliability of ACIP piles at the SLS. Kulhawy, F.H. & Chen, J.R. 2005. Axial compres-
In order to illustrate the effect of slenderness sion behavior of augered cast-in-place (ACIP) piles
in cohesionless soils. In C. Vipulanandan & F.C.
ratio on foundation reliability at the SLS, reli- Townsend (eds), Advances in Deep Foundations; Proc.
ability indices calculated herein may be com- of GeoFrontiers Congress (GSP 132), Austin, Texas,
pared to those reported in Phoon & Kulhawy 24–26 January 2005. Reston: ASCE.
(2008). Using the statistics for ya and Q′p recom- Mandolini, A., Ramondini, M., Russo, G. & Viggiani, C.
mended by a Phoon and Kulhawy (2008) and a 2002. Full scale loading tests on instrumented CFA piles.
D/B = 25, β = 2.214 (pf = 1.34%) is computed, In M.W. O’Neill & F.C. Townsend (eds), An Interna-
which is in good agreement with the previously tional Perspective on Theory, Design, Construction, and
reported value (2.210). However, for longer piles, Performance; Proc. of Deep Foundations (GSP 116),
say with a D/B = 65, the reliability index, β, equals Orlando, Florida, 14–16 February 2002. Reston: ASCE.
McCarthy, D.J. 2008. Empirical relationships between
1.774 (pf = 3.80%), a significantly different value load test data and predicted compression capacity of
than previously computed. augered cast-in-place piles in predominately cohesion-
less soils. M.S. Thesis. University of Central Florida.
Orlando, Florida.
9 SUMMARY AND CONCLUSIONS O’Neill, M.W., Vipulanandan, C., Ata, A. & Tan, F.
1999. Axial performance of continuous-flight-auger
This paper investigated the effect of varying model piles for bearing. Texas Dept. of Transportation report
statistics in reliability-based serviceability limit no. 7-3940-2, Center for Innovative Grouting Materi-
state design of ACIP piles installed in predomi- als and Technology: University of Houston.
nately cohesionless soils. First, a database con- Park, S., Roberts, L.A. & Misra, A. 2012. Design meth-
sisting of load tests conducted on ACIP piles in odology for axially loaded auger cast-in-place and
cohesionless soils was compiled, and the uncer- drilled displacement piles. Journal of Geotechnical and
Geoenvironmental Engineering 138(12): 1431–1441.
tainty in the entire load-displacement relationship Paikowsky, S.G. with contributions from Birgisson, B.,
was reduced to a correlated bivariate vector con- McVay, M., Nguyen, T., Kuo, C., Baecher, G., Ayyab, B.,
taining the hyperbolic model parameters. Contrary Stenersen, K., O’Malley, K., Chernauskas, L. &
to Phoon & Kulhawy (2008), both model param- O’Neill, M.W. 2004. Load and resistance factor design
eters were found to be correlated to pile slender- (LRFD) for deep foundations. NCHRP report no. 507,
ness ratio. Subsequent analyses used transformed Transportation Research Board: Washington, DC.
model parameters to avoid the undesirable effect Phoon, K.K. 2006. Serviceability limit state reliability-
of parameter dependence on geometric variables. based design. In M.L. Lin, C.T. Chin, H.D. Lin, Y.
The effect of varying the mean and uncertainty Honjo, K.K. Phoon (eds), New Generation Design
Codes for Geotechnical Engineering Practice; Proc.
of the allowable displacement and the uncertainty of Int. Symp., Taipei, Taiwan 2–3 November 2006.
of the capacity prediction method on the computed Singapore: World Scientific Publishing Company.
reliability index was assessed. In general, changing Phoon, K.K., Chen, J.R. & Kulhawy, F.H. 2006.
the uncertainty in Q′p had a larger effect on β com- Characterization of model uncertainties for auger
pared to ya. Overall, β was larger for larger mean cast-in-place (ACIP) piles under axial compression. In
allowable displacements when all other variables in R.L. Parsons, L.M. Zhang, W.D. Guo, K.K. Phoon, &
the performance function were unchanged. At larger M. Yang (eds), Foundation Analysis and Design—
allowable displacements, β was found to approach Innovative Methods (GSP 153). Reston: ASCE.
an upper bound limit and shown to be largely insen- Phoon, K.K. & Kulhawy, F.H. 2008. Reliability-based
design in geotechnical engineering. New York: Taylor &
sitive to the level of uncertainty in ya, compared to Francis.
Q′p. Because of the dependence of the model param- Stuedlein, A.W., Neely, W.J. & Gurtowski, T.M. 2012.
eters on pile stiffness and geometry, β was found to Reliability-based design of augered cast-in-place piles
be sensitive to D/B, and illustrates the importance of in granular soils. Journal of Geotechnical and Geoenvi-
accounting for this correlation in RBD. ronmental Engineering 138(6): 709–717.

310

ISGSR2013.indb 310 10/18/2013 9:41:09 AM

 Geotechnical Safety and Risk IV – Zhang et al. (eds)
© 2014 Taylor & Francis Group, London, ISBN 978-1-138-00163-3

Uplifit model for the Red River dikes of Vietnam

Pham Quang Tu
Delft University of Technology, The Netherlands
Hanoi Water Resources University, Vietnam

Bui Van Truong

Hanoi Water Resources University, Vietnam

P.H.A.J.M. van Gelder

Delft University of Technology, The Netherlands

ABSTRACT: The province of Thai Binh, an area with a large population and high socio-economic
status, is located at the downstream part of the Red River Delta. The area is protected by dike systems
whose safety has been upgraded significantly. While overtopping was an important failure mode in the
past, after raising crest levels, now piping and uplift phenomena are considered the dominant threats, hav-
ing led to several dike breaches in the Red river dike system. In order to address this issue and the large
uncertainties in commonly used uplift models, a series of field tests had been designed and carried out
on the Thaibinh formation (a regional top soil layer) for determining the critical uplift pressure. A special
structure is installed to increase water head artificially, by excavation through the top layers, and ground
water flow is concentrated by plastic sheet piles. Phenomena are monitored and measured in an excava-
tion, for thinning the thickness of impervious layers, during the test. The results cover critical head, uplift
displacement, time series and visual observations. In this paper, the authors will perform a multivariate
analysis for the field test results in order to develop statistical model for uplift mechanism. Furthermore, a
proposed model of uplift has also been generated including a calibration coefficient which represents for
uncertainties of both inherent and epistemic sources. The proposed model is also compared to test results
and statistical model to give a better understanding of this phenomenon. The findings of the research
might contribute to improving dike assessments in the Red River Delta.

1 INTRODUCTION visual observations. In this paper the author will,

first, briefly present the test results, then perform a
The province of Thai Binh, an area with a large multivariate analysis to elaborate a statistic model
population and high socio-economic status, is for the uplift mechanism. Furthermore, a proposed
located at the downstream part of the Red River model of uplift has been generated including a cal-
Delta. The area is protected by dike systems whose ibration coefficient. The model’s results are com-
safety has been upgraded significantly. While over- pared together to formulate a recommendation for
topping was an important failure mode in the past, the uplift model of the Red River Delta.
after raising crest levels, now piping and uplift phe-
nomena are considered the dominant threats, hav-
2 STUDY ON UPLIFT PHENOMENON
ing led to several dike breaches in the Red river dike
system. In order to address this issue and the large
2.1 Field tests of uplift on Thaibinh formation
uncertainties in commonly used uplift models, a
series of field tests had been designed and carried 2.1.1 Introduction
out on the Thaibinh formation (a regional top soil Number of phenomena and terms related to dike
layer) for determining the critical uplift pressure. A failure have been known in literatures, such as criti-
special structure is installed to increase water head cal gradient, piping, heave, sand boiled, uplift, etc.
artificially, by excavation through the top layers, To avoid confusion of those terminologies, it is
and ground water flow is concentrated by sheet necessary to remain them clearly.
piles. Phenomena are monitored and measured in Critical gradient is defined as ratio of differ-
an excavation, for thinning the thickness of imper- ent head and thickness of soil layers. It has been
vious layers, during the test. The test results cover commonly confused with another hydraulic gra-
critical head, uplift displacement, time series and dient which could be calculated from different

311

ISGSR2013.indb 311 10/18/2013 9:41:09 AM

 head of river and hinterland against the seepage
length from foreland to observation location, see
(Rijwaterstaat 1999, Sellmeijer 1988).
Uplift is a phenomenon that causes by high
piezo-metric level in aquifer (often happening for
river dike in flood season). The resulting cracks,
vertical displacement and failure of impervious
layers are common observed in combination with
this phenomenon, see Figure 2.
Heave (or sand boiled) is phenomenon which is
often visualized by sand or silt being carried out
from the aquifers, see Figure 3. Heave could be
understood as the first stage of piping under dikes,
or we could say if there were no mitigating meas- Figure 3. Heave observed during the field test, from Bui
ures, heave will lead to piping. (2009).
Piping is the most common failure mechanism
of river dikes, in which outflow discharge with There have been a number of studies on uplift
buoyant sand or silt increases continuously from and piping through years in Red river dikes of
beginning of this mechanism. Hypothetically, con- Viet Nam, for instance To (2002) carried out series
tinuous pipes are formed under dike embankment of laboratory test on the critical head of sand
due to backward erosion, see (Sellmeijer 1988, boil and uplift. The test results indicated that the
van Beek et al. 2012). If the pipe is fully formed, critical gradients is quite close to theoretical value,
enlargement of piping channel will elaborate and however small size of test equipments are one of
dike will collapse in hours. the main limitation of this research. Recently, Bui
(2009) reconsiders uplift mechanism in series of
field test on Thaibinh formation.
The objectives of these field tests are, first,
determine the critical head which leads to uplift of
the impervious layers in Red River dikes system in
Thai Binh; then, visualize the phenomena during
the tests to give a more apprehension about the
failure mechanisms. Related parameters for these
failure mechanisms will also be obtained for evalu-
ation and comparison the difference between the
test results and the theoretical model. In this study,
only uplift is considered experimentally and physi-
cally. The result is expected to give a better under-
standing for phenomena in Red River dikes.

2.1.2 Test set up and procedures

Critical head is created either by an excavation well
or by an injection well through impervious layers.
By doing so, water is supplied directly to aquifer,
Figure 1. Study location. see Figure 4. Plastic sheet piles are installed follow-
ing Ushape (see Fig. 4) to concentrate groundwater
flow, which is observed in locations along the open
direction of the sheet piles and forms 1D ground-
water flow model.
Observation phenomena, such as uplift and
heave, are monitored in an excavation which is
designed with different depths for thinning the
thickness of the impervious layers. Furthermore,
piezometric head of groundwater flow could be
measured from piezometers H1 – H3 in aquifer and
H4 – H7 in top layers during tests.
Water is pumped to the wells with an adjustable
discharge in order to keep water level in the well
Figure 2. Crack due to uplift, from Bui (2009). being constant in certain period of experiments.

312

ISGSR2013.indb 312 10/18/2013 9:41:09 AM

 Figure 5. Geotechnical conditions in field test loca-
tions, from Bui (2009).

Figure 4. Schematization of field test, (a) cross section,

(b) layout, adapted from Bui (2009).
Figure 6. Sudden raise of piezometric values in imper-
vious layers, from Bui (2009).
After that, piezometric heads are recorded in dif-
ferent locations, and the next increment of water
level could be started when the piezometer read-
ing being unchanged. The procedures repeat in
combination with the observation of phenomena
in the excavation. Uplift is defined to be failure
when crack are formed in blanket layers and water
is blown out of aquifers.
At the end of the test, the impervious layers are
failed including crack, sand boiled (or heave) and
water blown out of the aquifer. Ground conditions
could be seen in Figure 5.
2.1.3 Test results
Different heads corresponding to time are recorded
during experiments. Phenomena included out flow
discharge also be monitored. From Figure 7 it can
be seen that piezometric heads in aquifer increase
linearly, then it turns to non-linear relationship after
cracks were observed. Afterwards, heave expands Figure 7. Critical head during the field test (a) and
continuously with silt and sand being carried out of uplift displacement (b), from Bui (2009).
the aquifer. The tests have been stopped if the water
head in wells could not be remained constant due to In this analysis, only critical heads of uplift are
the limited supply of water and test apparatus. considered. Heave and piping are mentioned in
It also can be noted that, pore pressure exceeding other studies.
in impervious layers depends on its heterogeneous
characteristics rather than consolidation process.
2.2 Statistical model of uplift
In some cases, piezometric head raises suddenly
after a period of linear increments, see Figure 6. A From the results of test series, a statistical model
significantly heterogeneous characteristic could be is formulated for uplift mechanism. In this study,
seen that is the drought crack which could be visu- a regression analysis is performed to figure out the
alized to a depth of one meter or even deeper. relationship between critical gradient (or different
Uplift displacement is also measured, see head over thickness of soils) and soil’s proper-
Figure 7. The magnitude of vertical displacement ties. The approach has been mentioned in Seber &
varies from 7 to 25 mm in experiments. Wild (2003) and applied successfully in Sellmeijer,

313

ISGSR2013.indb 313 10/18/2013 9:41:11 AM

 de la Cruz, van Beek, & Knoeff (2011) & van der (Rijwaterstaat 1999). In Viet Nam, safety factor of
Zee (2011). For this purpose, a framework of 1.5–2.0 has also been recommended for design of
multivariate analysis has been developed for criti- berm in hinterland.
cal gradient from the field tests series. The linear From the field test results, it is suggested that the
regression could be expressed as below. critical gradients of uplift phenomenon are much
higher than which are expected. Therefore, a new
Y = βo + β1X1 + β2X2 + … + βnXn + ∈ (1) model to deal with this failure mechanism for the
study area has been implied. It can be seen form
in which Y is target regression function; Xi is Figure 8 that uplift zone has a width of L due to
noted for measuring variable; βi is unknown piezometric head increasing over weight of imper-
regression coefficient and ∈ is error of estima- vious layer. In equilibrium state, all the forces in
tion. Considering to field test results, a number vertical direction are vanished or balance, so from
of variables are chosen for analysis, then three Figure 9 we have:
following variables are considered for final calcu-
lation; they are soil density, cohesion and internal Hc P − 2((C P′′ φ) = 0 (5)
friction. A transformation of data to logarithm
domain is also applied, then we have following Or
relations.
⎛ ⎞
ρ d + 2 ⎜ C d + d K o ⋅ tanφ ⎟
2
β β2
⎛ ρ ⎞ 1⎛ C ⎞
βo ⎛ φ ⎞ hc (6)
icsta e ⎜ (2) ⎝ 3 ⎠
⎝ ρm ⎟⎠ ⎜⎝ Cm ⎟⎠ ⎜⎝ φ ⎟⎠
m
in which, Hc and hc are critical piezometric level and
Here, icsta is critical uplift gradient correspond- net critical head at the bottom of top layer; P = ρ′ ⋅ d
ing to the statistical model; ρ and ρm indicate the is effective weight of top layer; P′ is effective lat-
density of soil in each test and mean value of data eral earth pressure; d is thickness of top layer; C is
series respectively; C and Cm are cohesion of soil cohesion and φ is internal friction angle of soil; Ko
and mean value of cohesion series; φ and φm are is coefficient of at-rest earth pressure, which could
noted as internal friction angle of soil and mean be found in different forms in many soil mechanics
value of that data series. references, for instance (Das 2011, Verruijt 2010).
From the calculation results, we have the follow- There are many forms of critical heads in litera-
ing estimation of statistical model for uplift: tures, see (Terzaghi 1943, Tran et al. 1988, Ojha et al.
2001). The differences between these approaches
2.83 0.097 0.0258
⎛ ρ ⎞ ⎛ C ⎞ ⎛ φ ⎞ are basically the second term in Equation 6 which
icsta e −0.182 ⎜ (3)
⎝ ρm ⎟⎠ ⎜⎝ C ⎟⎠
m
⎜⎝ φ ⎟⎠
m
are treated elastically or plastically between rela-
tionship of stress and strain in soil, see (Chen

2.3 Physical-based model of uplift

Physically, uplift happens if piezometric head at
the bottom of impervious/top layer is larger than
its weight and other resistances. This phenomenon
has been investigated by many researchers through
century, for instance (Terzaghi 1943, Tran et al.
1988, Van et al. 2011). The basic equation for uplift
design can be defined: Figure 8. Uplift model for calculation.

ρsat − ρw
ico = (4)
ρw

where ρs is density of soil particles; ρw is density of

water; n is porosity of soil; ρsat is saturated density
of soil and ico is critical gradient of uplift.
The criteria of Tezaghi in Equation 4 assumes
that soil is non-cohesiveness, which is available until
now even different safety factor has been used in
practice, for instance critical gradient iocri = 0.3 − 0.8
in (USACE 2000) or factor of safety Fs = 1.5–2 in Figure 9. Uplift calculation diagram.

314

ISGSR2013.indb 314 10/18/2013 9:41:13 AM

 1975, Chen and Liu 1990) for more detail of these
behaviours. A calibration coefficient, θ, has been
introduced for the second term of Equation 6, in
which θ represents for uncertainties of both model
and soil parameters, see (van Gelder 2000, Vrou-
wenvelder and Steenbergen 2003, Kanning 2012)
for more detail of model factor. After validation,
we have the final form of critical water head for
uplift in our case study as follows:

⎛ ⎞
ρ d + θ ⎜ C d + d K o ⋅ tanφ ⎟
2
hcpro (7)
⎝ 3 ⎠

In this case, groundwater level is equal to the Figure 10. Critical gradient of uplift in different models.
level of hinterland and parameters of foreland’s
blanket layers are the same as that of hinterland; 3.2 Discussion
hcpro c is critical uplift head of the proposed model; From the observation during the tests, it can be
θ is the calibration coefficient. seen that uplift commences first at the weakest spot
Critical gradient of uplift can be calculated from in the impervious layers, then develops widely on
Equation 7, then we have: area if water pressure till increases continuously.
The weak spots mentioned here are cracks or het-
⎛ ⎞
= ρ ′ + θ ⎜ C + d K o ⋅ tanφ ⎟
hc pro 2 erogeneous structures in soils and the drought
icpro = (8) cracks are also supposed being a main disturbed
d ⎝ 3 ⎠
factor of the top layers in the depth from 0.8 to
Calibration coefficient can also be figured out by: 1.2 m, see (Bui 2009).
After crack, heave is also visualized in all test
3( cpro − ρ ′ ) locations and the moment of heave start is defined
θ= (9) as end of crack developing process. Time depend-
3C + d ⋅ K otanφ ent in test series may also be influenced to its
results. Unfortunately, all of these issues are not
By fitting icpro to uplift critical gradient from included in this study.
test’s data series, we could elaborated the values The statistical model fitted quite well with data
of θ. series of test, see Figure 8. However, it has not
shown the physical relationship between parameters
3 RESULTS AND DISCUSSION and its physical behaviours. It is clear from statisti-
cal analysis that, correlated coefficient of this model
3.1 Results has a negative trend (ρ = −5.3%) comparing to posi-
tive trend of two other models (ρ = 8.9–16.9%). It
Both uplift and heave are observed during the seem unreasonably that different heads increase
increasing of water level indicating for flood water lead to decline of critical gradient, following the
level. Crack was initiated first in weak spot of the statistical model. So that, in practice it might be
blanket, then appears in many spots with sand or incorrect if we apply the statistical model without
silt flowing out in a certain area. understanding about its physical behaviours.
There are total of fifty eight tests performs on It also can be seen from Figure 10 that the pro-
Thaibinh formation for uplift and heave phenom- posed model has the smallest variation comparing
ena. The plotting of the final results of critical gra- to others models. Calibration coefficient has been
dient of uplift could be found in Figure 10. assumed to be normal distributed and parameters
The critical uplift gradients are found to be nor- are mean and standard deviation by μθ = 0.5519
mal distributed of mean and standard deviation and σθ = 0.746 respectively. In practice, it is recom-
equivalent to 0.836 and 0.105 respectively. From mended that θ should be chosen by 0.5 for engi-
the theoretical Equation 4, the so-called theoreti- neering purposes.
cal critical gradient has a value of 0.758. Whereas, In term of comparison between statistical to pro-
in the proposed model, critical gradient could be posed model, the probability if actual critical gradi-
formulated in range of normal distribution with ent is larger than that of theoretical model could be
parameters N(0.8301;0.0497). It is figured out that calculated, for instance P (i icpro ) P (i .758)
statistical model fits quite well to normal distribu- The final results show that exceeded probability is
tion with parameters of N(0.8323;0.066). P (icpro ico ) = 92. % for the proposed model and

315

ISGSR2013.indb 315 10/18/2013 9:41:16 AM

 Chen, W.-F. 1975. Limits analysis and soil plasticity.
Elsevier scientific publishsing company.
Chen, W.-F. & Liu, X. 1990. Limits analysis in soil
mechanics. Elsevier scientific publishsing company.
Das, B.M. 2011. Principles of foundation engineering, sev-
enth edition. Cengage Learning.
Kanning, W. (2012). The weakest link—Spartial variability
un the piping failure mechanism of dikes. Ph. D. thesis,
Technology University of Delft, The Netherlands.
Ojha, C., Sigh, V., & Adrian, D. 2001. Influence of poros-
ity on piping models of levee failure. Journal of Geo-
technical and Geoenvironmental Engineering 127(12),
1071–1074.
Rijwaterstaat (1999). Technical report on piping.
Technical report.
Figure 11. Probabilistic distribution of uplift models.
Seber, G.A.F. & Wild, C. (2003). Nonlinear regression.
John Wiley & Sons.
Sellmeijer, J.B. 1988. On the mechanism of piping under
is P (icsta ico ) = 86.8 % for the statistical model, impervious structures. Ph. D. thesis, Civil Engineering
whereas that is P (ictest icc ) = 77. 6% for series of and Geosciences of Technical University of Delft.
field test data, see Figure 11 for more detail. Sellmeijer, H., de la Cruz, J.L., van Beek, V., & Knoeff,
H. 2011. Fine-tuning of the backward erosion piping
model through small-scale, medium-scale and ijkdijk
4 CONCLUSIONS experiments. European Journal of Environmental and
Civil Engineering, 1139–1154.
It is clear from the above results that critical gra- Terzaghi K. 1943. Theoretical Soil mechanics. John
dients, which lead to uplift failure in Thaibinh Wiley & Sons.
formation of Red River dikes, are significant high To, X.V. 2002. Study on the influence of seepage effect
comparing to current understanding of this phe- of formations in dike stability, application for a Red
River dike section. Ph.D. thesis, Ha Noi University of
nomenon. Uplift failure initiates at the weakest Mining and Geology, Viet Nam.
spot, then expands to many spots in large area if Tran, V.T., Tran, M.L., & Dang, Q.T. 1988. Stability of
piezometer levels still increase continuously. impervious layer in red river dike of ha noi. Water
Field test data has been fitted with both statis- Resources Bulletin, Ministry of Water Resources,
tical and physical-based models. Calibration coef- Vietnam 261(3+4), 40.
ficient for supposed model (in Equation 8) has USACE. 2000. Design and construction of levees.
been recommended by 0.5 in practical application. Engineering manual 1110-2-1913, Army Corps of
It is expected that the results could be applied in Engineer.
flood defence assessment for the Red river dikes on Van, M., Koelewijn, A., & Barends, F. 2011. Uplift phe-
nomenon: Model, validation and Design. Stichting
Thaibinh formation. Deltares, Delft, The Netherlands.
Limitation: Dimensions of excavation in the van Beek, V., de Bruijn, H., Knoeff, J., Bezuijen, A. &
field test are the same for all test series, so there is Frster, U. 2012. Levee failure due to piping: A full-
no influence of that has been taken in to account. scale experiment. Reliability Engineering and System
Safety, 141–150.
van der Zee, R. 2011. Influences of sand characteris-
ACKNOWLEDGEMENTS tics on the piping process. Master’s thesis, Technical
University of Delft.
This paper is funded by the Project of Upgrading van Gelder, P. 2000. Statistical Methods for Risk-Based
the Training Capacity in Coastal Engineering Design of Civil Structure. Ph.D. thesis, Technology
of the Water Resources University of Vietnam University of Delft, The Netherlands.
from the Royal Netherlands Embassy in Hanoi Verruijt, A. 2010. Soil mechanics. Delft University of
of Vietnam, and CICAT and the Department of Technology.
Hydraulic Engineering of TU Delft. The authors Vrouwenvelder, A. & Steenbergen H. 2003. Theorie-
handleiding pc-ring, versie 4.0. deel b: Statistische
wish to thank for their support. modellen. Technical report.

REFERENCES

Bui, V.T. 2009. Study on the failure of under seepage

of lower Red river dike in Thai Binh province and
field experiment-based analysis the design solutions.
Ph. D. thesis, Hanoi university of Mining and
Geology, Vietnam.

316

ISGSR2013.indb 316 10/18/2013 9:41:18 AM

 Geotechnical Safety and Risk IV – Zhang et al. (eds)
© 2014 Taylor & Francis Group, London, ISBN 978-1-138-00163-3

Effect of rainfall on the reliability of an infinite slope

J. Yuan, I. Papaioannou, C.M. Mok & D. Straub

Engineering Risk Analysis Group, Technische Universität München, Munich, Germany

ABSTRACT: Rainfall is one of the most common factors triggering landslides, since infiltration of
water into the soil has a significant impact on pore water pressure buildup that affects slope stability. In
this study, the influence of the wetting front development on the reliability of an infinite slope is analyzed.
The failure condition of the slope is expressed in terms of the factor of safety. Rainfall infiltration is
simulated by a time-dependent model, based on the Green and Ampt assumptions. The vertical variability
of the saturated hydraulic conductivity and soil strength parameters are modeled as random fields. The
reliability of the slope is evaluated by Monte Carlo simulation. A numerical example demonstrates the
influence of the vertical variability on the analysis results.

1 INTRODUCTION Monte Carlo simulation, the statistics of the

wetting front development and the slope factor of
Rainfall has a considerable influence on landslide safety as well as the probability of failure of the
occurrence (Polemio 1997, Corominas et al. 2002, slope are estimated. Moreover, the influence of
Zêzere et al. 2008). In wet seasons, intense rainfall the variability of the soil properties on the analysis
events often cause shallow slope failure. In such results is investigated.
cases, the failure surface is approximately parallel to
the ground surface and hence an infinite slope model
2 TIME-DEPENDENT MODEL
can be used to represent the slope failure mecha-
nism (Wu & Abdel-Latif 2000, Santoso et al. 2011).
2.1 Stability of infinite slope
Since surficial infiltration is predominately driven
by gravitation, the infinite slope is often simplified Consider a slice of an infinite slope, as shown in
in practice as a column subject to vertical infiltra- Figure 1. Assuming that the water table and poten-
tion, unless topography is very steep. It should be tial slip surface are parallel to the ground surface,
noted that some authors assumed that water perco- and according to the linear Mohr-Coulomb failure
lates perpendicular to the sloping surface (Chen & criterion, the factor of safety of the slip surface is
Young 2006, Muntohar & Liao 2010). defined as (Griffiths et al. 2011):
In this paper, rainfall infiltration is modeled by
applying the Green and Ampt assumptions for one- (γ H cos 2 β u )tan ϕ ′ + c ′
dimensional vertical subsurface flow (Green & Ampt FS = (1)
γ H sin β cos β
1911). Hence, it is assumed that the wetting front
represents a clear border between the wetted and dry
(in-situ condition prior to rainfall) soil zones. Fur- in which, γ is the average unit weight of the soil
ther assuming that the water table is parallel to the mass above the slip surface; H is the depth of
ground surface, the infiltration model can be com- the slip surface; β is the slope inclination; u is the
bined with the classical equation for infinite slope pore water pressure at the slip surface; ϕ ′ is the
stability analysis to evaluate the time-dependent
behavior of the factor of safety of the slope.
Soil property values tend to vary in space, even
within a lithologic layer that appears to be homo-
geneous (Phoon & Kulhawy 1999). We model
the inherent vertical variability of the strength
parameters and saturated hydraulic conductivity
as one-dimensional statistically homogeneous ran-
dom fields. Through discretization of the random
fields, the soil column is represented by a multilay-
ered system whose layers have uniform parameters Figure 1. Typical unit slice from an infinite slope.
described by correlated random variables. Through

317

ISGSR2013.indb 317 10/18/2013 9:41:19 AM

 effective friction angle at the slip surface; c ′ is causes a well-defined wetting front (see Fig. 2).
the effective cohesion at the slip surface. Above the wetting front, the soil is fully saturated
while below the wetting front it continues to have
2.2 Stochastic model of soil parameters its initial moisture content.
The soil column is divided into a number of
The stability of the slope depends on the shear equal-thickness layers with varying saturated
strength of the soil at the slip surface, expressed hydraulic conductivities as discussed in Section 2.2.
in terms of the effective strength parameters c ′ We approach the problem by assuming that the
and ϕ ′. Moreover, the infiltration process, which wetting front development takes place in a step-
affects the pore water pressure u during the wetting wise manner, whereby the wetting front advances
front development and the rise of the water table, by one layer in each computational step. That is, in
is governed by the vertical saturated hydraulic con- step j the wetting front is located at the bottom of
ductivity K of the soil. In this study, we account the jth layer. For the sake of simplicity, the initial
for the variability of the tangent of the friction moisture content θ0, soil porosity η and suction
angle and the saturated hydraulic conductivity in heat at wetting front S are considered to be con-
the depth direction through a one-dimensional stants within the unit slice.
random field modeling. This study focuses on short and intense rainfall
The parameters are assumed to be log-normally events, during which the wetting front might not
distributed (Gelhar 1986). We model the natural reach the water table. The infiltration process can
logarithms of the parameters as homogeneous be divided into two distinct phases (see Fig. 3):
Gaussian random fields with the following expo- vertical infiltration during the rainfall period and
nential autocorrelation coefficient function: plug flow driven by the saturated hydraulic con-
ductivity of the soil until formation of the new
− 2rτ
ρ (τ ) = e (2) ground water table.
During the rainfall period [Phase I, see Fig. 3(a)],
where r is the correlation length; τ is the separation the hydraulic gradient i from the ground surface to
between two locations in the vertical direction. The the wetting front at layer j along the vertical direc-
correlation length is a measure of the spatial varia- tion is obtained by application of Darcy’s law:
bility of the random field. A correlation length that z S
is much larger than the depth of the soil slice implies i= (3)
a uniform soil profile; in this case, the soil param- z
eters can be modeled by random variables. In the where z is the depth of the wetting front at the
extreme case where the correlation length is close to bottom of the layer j; S is the suction head at the
zero, the values of the soil parameters at each loca- wetting front. Note that Equation (3) neglects
tion become independent from each other. the influence of the ponding water depth. We can
The random fields are discretized by the mid- then evaluate the infiltration rate fin as follows:
point method (Der Kiureghian & Ke 1988). The fin Ke ⋅ i (4)
unit slice is divided into a number of equal lay-
ers; the randomized properties are assumed to be where Ke is the effective vertical saturated hydrau-
constant within each layer, represented by their lic conductivity, namely the harmonic mean of the
values at the midpoints of the layers. Since homog- vertical saturated hydraulic conductivities at the
enous fields are assumed, the mean u and standard soil layers within the wetted zone (Freeze & Cherry
deviation μ of the random variables corresponding 1979):
to the discrete layers are constant over the entire
field. The entries of the correlation matrix of the
logarithms of the soil properties at each layer
are evaluated with Equation (2) where τ is the dis-
tance between each pair of midpoints. Simulation
of the soil parameters is performed by simulating
the joint normal random variables of all layers and
taking the exponential of the resulting samples. It is
noted that the cross-correlation coefficient between
the two parameters is set to zero in this study.

2.3 Infiltration analysis

We utilize the Green and Ampt assumptions to
simulate the infiltration process (Green & Ampt
1911). That is, we assume that vertical infiltration Figure 2. Green and Ampt model on the unit slice.

318

ISGSR2013.indb 318 10/18/2013 9:41:20 AM

 stops [see Fig. 3(b)]. Equation (10) can be evaluated
numerically over the computational layers as
1
t Δz ⋅ Δθ ⋅ ∑ (11)
fj

where Δz is the thickness of each layer; fj is the

flow rate at the jth computational step of either the
Figure 3. Infiltration process. infiltration or the plug flow phase depending on
whether the time t is smaller or greater than the dura-
tion of the rainfall event T. Through Equation (11)
j we obtain a mapping between the cumulative time
Ke = (5)
1 and the wetting front development.
∑ l =1 K
j

l
2.4 Pore water pressure distribution
in which Kl is the saturated hydraulic conductivity The pore water pressure varies with time and space
of each wetted computational layer. within one unit slice. In this study, emphasis is on
We assume that the rainfall supply is larger than the pore water pressure within the wetted zone and
the infiltration capacity. Therefore the cumulative below the water table.
infiltration will be governed by the infiltration Within the wetted zone, the spatial variability of
capacity and can be obtained by the saturated hydraulic conductivity may introduce
F z ⋅ Δθ (6) large hydraulic gradients that will result in vari-
able pore water pressure. Employing the plug flow
in which Δθ is the change in moisture content, assumption given by the Green and Ampt infiltra-
given as follows: tion model, we assume that the effective flow rate
of the wetted zone f computed by Equations (4)
Δθ η − θ0 (7) and (9) equals the flow rate at each wetted layer
(Liu et. al. 2008), i.e.
where η is the porosity of the soil and θ0 is the
initial moisture content (see Fig. 2). Noting that f1 f2 = … = fl f (12)
fin ddF /ddt , we can obtain the cumulative infiltra-
tion time as follows: where fl is the flow rate at wetted layer l, computed
by
t z Δθ
∫0 dt = ∫0 fin
dz (8)
Δhl
fl K l ⋅ il Kl ⋅ (13)
Δz
The second phase that follows the rainfall event
is governed by plug flow. It is assumed that the wherein il is the hydraulic gradient of layer l; Δhl
suction at the bottom of the wetted zone equals the change of hydraulic head within the lth layer.
the one at the top [see Fig. 3(b)]. Therefore, the Combining Equations (12) and (13), we obtain:
hydraulic gradient equals unity and the plug flow f
rate fp equals the harmonic mean of the saturated Δhl Δz (14)
hydraulic conductivities of the layers correspond- kl
ing to the wetted zone:
At the bottom of lth wetted layer, the hydraulic
fp Ke head hl is obtained by summing the incremental
(9) heads Δhk of each computational layer k ≤ l
Therefore, the time needed for the wetting front
to reach the bottom of each layer is obtained as hl h − ∑ k l Δhk (15)
follows:
in which h0 is the initial hydraulic head at the top of
⎧ z Δθ
the wetted zone, i.e. h0 = 0 during Phase I [Fig. 3(a)]
⎪ ∫0 f dz , t T
and h0 = −z0 − S during Phase II [Fig. 3(b)], where
⎪ in
t=⎨ (10) z0 is the initial elevation head. Since the hydrau-
⎪T +
z Δθ
⎪⎩ ∫z L f p dz , t T lic head hl consists of the pressure head ψl and the
elevation zl, the pressure head at the bottom of zl th
layer is evaluated as:
wherein T is the duration of the rainfall event; L
is the length of wetting front exactly after rainfall ψ l = hl zl (16)

319

ISGSR2013.indb 319 10/18/2013 9:41:21 AM

 with a 5 m deep layer of sandy soil with zero
cohesion c′, subjected to a 24 h intense rainfall.
The slope angle is taken as β = 18°. The tangent
of the friction angle tan ϕ ′ and the saturated
vertical hydraulic conductivity K are modeled as
lognormal random fields (Table 1). The correla-
tion length for the two parameters is identical and
we assume a zero cross-correlation. The random
fields are discretized into 100 layers. The water
Figure 4. Hydrostatic pressure. table is at a depth of 1.5 m below the ground
surface. The parameter of the Green and Ampt
model for sandy soil are taken from Rawls (1982)
At the wetting front, the pressure head equals as given in Table 2. The number of Monte Carlo
the suction head, i.e. ψj = −S. The pore water pres- samples is set to 104, which was found to give
sure can be computed as ul = γw ⋅ ψl, where γw is the acceptable coefficient of variations of the prob-
unit weight of water. ability estimates for the cases considered.
Below the water table, the soil is subjected to
hydrostatic pressure u = γwz* = γwzwcos2β, where z* 3.1 Influence of the correlation length
is the projection of the equipotential line in vertical The analysis is performed for two values of the cor-
direction; zw is the depth of the soil layer under the relation length r, that is 2 m and 0.5 m. Figure 5
ground water table (see Fig. 4). The hydrostatic water presents the development of the mean depth of
pressure builds up when the ground water elevation the wetting front with time. The time period of
rises as the wetting front reaches the water table. interest is 240 hr (10 days) after the start of the
rainfall event.
2.5 Reliability analysis Figure 5 shows that in the case of a smaller cor-
Slope failure occurs when the factor of safety FS is relation length, the water flow will move slower
less than unity. Hence, the probability of failure of towards the water table. This can be explained by
the slope is defined as the fact that a small correlation length implies a
large variability within the wetted zone. The flow
Pf FS < 1]
P [ FS (17) within the wetted zone is driven by the lower values
of K, as evident from the use of the harmonic mean
In this study, we estimate Pf by Monte Carlo of the saturated hydraulic conductivity in Eq. (5);
simulation. Each realization of the random fields the higher the variability, the larger the likelihood
results in different values of the strength param- of occurrence of low values.
eters and the saturated hydraulic conductivity at Figure 6 shows the mean factor of safety of the
each soil layer. Based on the latter parameter value, slope with time. As also discussed in (Griffiths et al.
we compute the pore water pressure distribution 2011), the initial mean factor of safety is smaller
for each time step and each layer, as discussed in for a smaller correlation length. A large correlation
Section 2.4. We then substitute the strength param- length implies a uniform soil profile and thus the
eters and pore water pressure into Equation (1) most critical potential slip surface is at the base of
to obtain the factor of safety. Note that for each the slope. On the other hand, a small correlation
time step we evaluate the factor of safety for each
potential slip surface, corresponding to the bot-
tom of each layer. The overall factor of safety Table 1. The values and distributions of the input
corresponds to the minimum value among all slip variables.
surfaces. The probability of failure for each time
step is computed as the number of samples Nf Variable Distribution Mean μ CV*
for which FS is less than one divided by the total
tanϕ ′ Lognormal 0.7002 0.1
number of samples N:
K [m/h] Lognormal 3.6 × 10−3 2
Nf
Pf ≈ (18) *CV stands for coefficient of variation (= σ/μ).
N
Table 2. Parameters for Green and Ampt model.
3 EXAMPLE
Variable η θ0 S [m]
We are interested in the slope stability for short
and intense rainfall events. We consider a slope Value 0.437 0.102 0.1734

320

ISGSR2013.indb 320 10/18/2013 9:41:25 AM

 of shallow slope failure decreases. Slope failure is
now driven by the formation of the new ground
water elevation that takes place when the wetting
front reaches the water table. This happens faster
in the case of a larger correlation length and hence
the decrease of the mean factor of safety is larger
in this case. A small correlation length implies
that layers with low hydraulic conductivity exist
in most realizations and these layers imped down-
ward movement of infiltration water.
Figure 7 depicts the transient evolution of the
Figure 5. Development of the mean wetting front depth. point-in-time probability of failure for the two
considered correlation lengths. The initial prob-
ability of failure is higher when the correlation
length is smaller, which corresponds to a smaller
mean factor of safety. For this case, the probabil-
ity of failure increases during the rainfall event,
only to decrease again until the wetting front
approaches the ground water table where the
probability rises once more. The case correspond-
ing to a larger correlation length presents a dif-
ferent behavior: During the rainfall event, shallow
slope failure will mostly occur at the wetting front
Figure 6. Influence of the correlation length on the fac- as there is no localized layer with lower relative
tor of safety. hydraulic conductivity value that impedes infil-
tration and leads to pore-pressure building up
within the wetting front. The wetting front will
length implies vertically variable parameters and move faster towards the water table and hence the
weak layers are likely to be present in most Monte likelihood that the water table is reached within
Carlo realizations. The most-critical potential slip one realization is higher. This results in a smooth
surface will likely occur in a weak layer, which increase of the probability of failure in terms of
will be present at a different location in each time since the failure will be governed by the rise
realization. of the hydrostatic pressure.
Since the pore water pressure increases due to
the wetting front development, the mean factor of 3.2 Influence of the coefficient of variation
safety then decreases during the infiltration proce-
We now look at the influence of the coefficient
dure for both cases. A faster decrease is observed
of variation of the hydraulic conductivity on the
for the case of the larger correlation length, since
stability and probability of slope for the case of
the water flow moves faster towards the water table,
a small correlation length (0.5 m); the coefficient
which causes a rise of the hydrostatic pressure.
of variation is varied from 1 to 3. Figures 8 and 9
During the rainfall event period, the mean factor
show the change of the mean factor of safety and
of safety decreases rapidly and it reaches a tran-
the probability of failure with time.
sient lower value when rainfall stops. The reduc-
Increasing the coefficient of variation will
tion is higher for the case with a small correlation
increase the variability within the wetted zone,
length, which implies that the likelihood of shallow
slope failure is higher. A small correlation length
introduces changes between large and small values
of the hydraulic conductivity occurring at short
distances within the wetted zone. Higher hydraulic
gradients occur across the layers with low hydrau-
lic conductivity values causing higher pressure to
be built up immediately above such layers within
the wetted zone. Therefore, a smaller correlation
length will increase the likelihood of slip surfaces
at intermediate layers above the wetting front.
After the rainfall event, the water flow trans-
forms to plug flow [Phase II in Fig. 3(b)] and the Figure 7. Influence of the correlation length on the
mean factor of safety increases, since the likelihood probability of failure.

321

ISGSR2013.indb 321 10/18/2013 9:41:25 AM

 of variation will favor shallow slope failure due
to the decrease of the infiltration rate and the
development of large pressure gradients within
the wetted zone. In the cases with large correla-
tion lengths and small coefficient of variation,
the reliability is dominated by the rise of the
hydrostatic pressure due to the fast wetting front
development.

REFERENCES
Figure 8. Influence of the coefficient of variation of
the saturated hydraulic conductivity on the factor of Chen L. & Young M.H. 2006. Green-Ampt infiltration
safety for r = 0.5 m. model for sloping surface. Water Resources Research
42(7): W07420.
Corominas, J., Moya, J. & Hürlimann, M. 2002. Landslide
rainfall triggers in the spanish eastern pyrenees. Proc.
4th EGS Plinius Conference. Mallorca, Spain.
Der Kiureghian, A. & Ke, J.-B. 1988. The stochastic finite
element method in structural reliability. Probabilistic
Engineering Mechanics 3(2): 83–91.
Freeze, A.R. & Cherry, J. A. 1979. Groundwater. Upper
Saddle River, NJ: Prentice Hall.
Gelhar L.W. 1986. Stochastic subsurface hydrology from
theory to application, Water Resources Research 22(9):
135–145.
Green, W. H. & Ampt, G. A. 1911. Studies on soil physics,
Figure 9. Influence of the coefficient of variation of part I—the flow air and water through soils. Journal
the saturated hydraulic conductivity on the probability of Agricultural Science 4(1): 1–24.
of failure for r = 0.5 m. Griffiths, D.V., Huang, J. & Fenton, G.A. 2011.
Probabilistic infinite slope analysis. Computers and
Geotechnics 38(4): 577–584.
which will cause the wetting front to move slower Liu J. T., Zhang J. B. & Feng J. 2008. Green–Ampt Model
towards the water table and therefore failure will for Layered Soils with Nonuniform Initial Water Con-
be governed by shallow slip surfaces. Moreover, tent Under Unsteady Infiltration. Soil Science Society
a large coefficient of variation leads to steeper of America Journal 72(4): 1041–1047.
changes between small and large values of the Muntohar, A.S. & Liao H. 2010. Rainfall infiltration:
hydraulic conductivity within the wetted zone; infinite slope model for landslides triggering by rain-
storm. Natural Hazards 54(3): 967–984.
hence the pore water pressure at layers above the
Phoon, K.K. & Kulhawy, F.H. 1999. Characterization
wetted zone increases, which leads to a larger prob- of geotechnical variability. Canadian Geotechnical
ability of shallow slope failure and smaller corre- Journal 36(4): 612–624.
sponding mean factors of safety. Polemio, M. 1997. Rainfall and Senerchia landslides,
southern Italy. Proc. 2nd COBRAE, Rio de Janeiro.
Rawls, W.J., Brakensiek, D.L. & Saxton, K.E. 1982. Esti-
4 CONCLUSION mation of soil water properties. Transactions of the
ASAE 25(5): 1316–1320 & 1328.
Santoso, A.M., Phoon, K., & Quek, S. 2011. Effect of
In this paper, we developed a simplified time-
1D Infiltration Assumption on Stability of Spatially
dependent model to study the reliability of slopes Variable Slope. Geo-Risk 2011: Risk Assessment and
subjected short-term and intense rainfall events. Management. 704–711.
The model incorporates infinite slope stability Wu, T.H. & Abdel-Latif, M.A. 2000. Prediction and
analysis, one-dimensional infiltration analysis mapping of landslide hazard. Canadian Geotechnical
based on the Green and Ampt assumptions as Journal 37(4): 579–90.
well as a random field modeling of the effective Zêzere, J.L. Trigo, R.M. & Fragoso, M., Oliveira, S.C. &
friction angle and the saturated hydraulic con- Garcia, R.A.C. 2008. Rainfall-triggered landslides in
ductivity of the soil. An example demonstrated the Lisbon region over 2006 relationships with the
north Atlantic Oscillation. Natural Hazards and Earth
the influence of the correlation length and coef-
System Science 8(3): 483–499.
ficient of variation of the hydraulic conductivity
on the development of the factor of safety and
probability of failure with time. It is shown that
a small correlation length and high coefficient

322

ISGSR2013.indb 322 10/18/2013 9:41:26 AM

 Geotechnical Safety and Risk IV – Zhang et al. (eds)
© 2014 Taylor & Francis Group, London, ISBN 978-1-138-00163-3

Influence of particle transport on slope stability under rainfall

infiltration

L. Zhang & L.L. Zhang

Department of Civil Engineering, Shanghai Jiaotong University, Shanghai, China

ABSTRACT: Internal erosion and transport of fine particles are among the possible consequences of
the infiltration of rainwater into a slope. The transport of fine particles may lead to local variation of
soil porosity and hydraulic properties, which affects the stability of a soil slope under rainfall infiltration.
In this study, the governing equation for transport of fine particles based on the conservation of mass
of moving particles is coupled with the governing equation of transient seepage analysis for unsaturated
soils. An internal erosion law is used to describe the relationship between the rate of eroded fine particles
and the hydraulic gradient. The effect of internal erosion on change of porosity and saturated perme-
ability is also considered. The influence of transport of fine particles on slope stability under infiltra-
tion is investigated by a finite-element analysis. The numerical results show that the rainfall infiltration
can induce reduction of the density of fine particles in the shallow depth of soil slope. This can further
increase the permeability and water seepage in the slope and the slope stability is reduced consequently.
A parametric study is also conducted to investigate the effect of saturated coefficient of permeability on
water seepage and slope stability.

1 INTRODUCTION of internal erosion and particle transport on

unsaturated soil slope under rainfall infiltration
Slopes in natural terrain are usually composed with have not been investigated. In this study, a cou-
mixed coarse and fine particles due to deposition pled model of seepage and internal erosion in
or weathering process. In-situ full scale and labora- unsaturated soil is developed. The variation of
tory reduced scale model tests showed that under porosity due to particle transport is considered.
rainfall infiltration internal erosion may occur in A parametric study is conduced to investigate
zones with dramatic changes of seepage gradient the effect of saturate coefficient of permeability
and fine particles were observed to be transported on pore water pressure distribution and slope
from the upper and surface part of the slope to stability.
the toe of slope (Hu et al. 2005, Jian et al. 2005,
Xu et al. 2006). When transport of fine particles
occurs in a slope, the porosity and permeability 2 SLOPE STABILITY ANALYSIS BASED
may increase due to loss of fine particles. This may ON COUPLED ANALYSIS OF SEEPAGE
further lead to variation of water flow or move- AND INTERNAL EROSION
ment of wetting front and affect the slope stability
during infiltration. 2.1 Mass-balance equations of internal erosion
Some researchers investigated particle trans-
port in porous media. Vardoulakis et al. (1996) The soil is regarded as a mixture of solid and liq-
established a one-dimensional model of internal uid which consists of the soil skeleton phase, water
erosion based on mass conservation law of mul- phase and liquefied particle phase. Liquefied parti-
tiphase fluid. Sterpi (2003) developed a constitu- cles are fine particles which are scoured by internal
tive equation for internal erosion based on soil erosion from the soil skeleton and can move freely
column tests and performed numerical analyses with water.
for internal erosion in saturated soils. Dahaghi The liquefied fine particles satisfy mass con-
et al. (2011) described variation of permeability servation equation (Cividini & Gioda 2004,
coefficient considering transport and deposi- Vardoulakis & Papamichos 2005):
tion of fine particles based on the conservation
of mass. These studies focus on internal erosion ∂ρtr
+ div( ρtrvi ) = qer (1)
in fully saturated soil. However, the influences ∂t

323

ISGSR2013.indb 323 10/18/2013 9:41:27 AM

 where ρtr = density of liquefied fine particles; 2.3 Constitutive equation of internal erosion
vi = velocity of water flow in i direction; qer = volume The constitutive equation of internal erosion
flux of fine particles eroded. As vi is equal to qi/n, describes the process in which the water
where qi is Darcy velocity in i direction and n is phase transforms the fine particles into liquefied
porosity, the above equation can be written as particle phase by erosion (Cividini & Gioda
follows: 2004):
∂ρtr ∂ ⎛ qx ⎞ ∂ ⎛ qy ⎞ qer (t,v ) β erv[ ρ f (t
(t )
+ ⎜ ρtr ⎠⎟ + ⎜ ρtr ⎟ = qer (2) f ∞ (v )] (8)
∂t ∂x ⎝ n ∂y ⎝ n⎠
where ρf = density of fine particle in soil; ρf∞ =
The mass conservation of soil skeleton phase ultimate or long term density of fine particle;
can be expressed as (Vardoulakis & Papamichos βer = a parameter of erosion.
2005): The ultimate or long term density of fine parti-
cle is a function of the flow velocity:
∂ρsk
+ div( ρskvisk ) = − qer (3)
∂t ⎪⎧ ρ ( ρ f 0 − ρ * f ∞ )v / v* ≤ v(t ) ≤ v*
ρ f ∞ (v ) = ⎨ *f
where ρsk = density of soil skeleton; visk = velocity ⎩⎪ ρ f α er llog((v / v* ) v* ≤ v(t )
of soil skeleton. The density of soil skeleton ρsk is (9)
equal to (1-n)ρs, where ρs is the density of soil par-
ticles. The velocity of soil skeleton phase can be where v* = initial flow velocity; ρf0 = initial density
assumed to be zero: of fine particles; ρ*f∞ = the long term density cor-
responding to v*; αer = a parameter of erosion.
visk = 0 (4)
2.4 Hydraulic functions of unsaturated soil
Therefore, the governing equation of porosity
can be obtained as follows: In an unsaturated soil, water content and perme-
ability are related with matric suction. The relation
∂n qer (5) between volumetric water content θw and mat-
= ric suction is named as Soil Water Characteristic
∂t ρ s
Curve (SWCC). The relation between the coeffi-
cient of permeability and matric suction is called
2.2 Governing equation of unsaturated flow the permeability function. In this study, the Van
Genuchten model is adopted:
The flow of water through soil is governed by the
mass conservation law and Darcy’s law. According ⎧ 1
to the Darcy’s law, the fluid flow can be written θ − θr ⎪ mw
uw < 0
in terms of hydraulic conductivity and hydraulic Sr = w = ⎨ ⎡1 + [ ua − uw ]n w ⎤
θs − θr ⎪ ⎣ w ⎦
gradient as follows: 1 uw ≥ 0
⎩
qi = ki∇H (6) (10)

where Sr = degree of saturation; θs = saturated

where ki = hydraulic conductivity in i direction; volumetric water content; θr = residual volumet-
qi = Darcy velocity; H = total head. ric water content; αw = a curve fitting parame-
Based on the mass conservation law and the ter inversely related to the air-entry value; nw =
Dary’s law, the governing equation for two-dimen- a curve-fitting parameter related to the pore size
sional water flow in an unsaturated soil can be distribution; mw = 1–1/nw; ua = pore-air pres-
written as (Fredlund & Rahardjo 1993): sure; uw = pore-water pressure; (ua − uw) = matric
suction.
∂ ⎛ ∂H ⎞ ∂ ⎛ ∂H ⎞ ∂θ w In this paper, the unsaturated coefficient of
⎜ kx ⎟+ ky = (7)
∂x ⎝ ∂x ⎠ ∂y ⎜⎝ ∂y ⎟⎠ ∂t permeability k is described as an exponential
function of the matric suction (Gardner 1958):
where kx = hydraulic conductivity in x direc-
tion of the flow; ky = hydraulic conductivity in k ks exp[− α w (ua uw )] (11)
y direction of the flow; θw = volumetric water
content. where ks = saturated coefficient of permeability.

324

ISGSR2013.indb 324 10/18/2013 9:41:27 AM

 Based on the Carman-Kozeny equation, the
saturated coefficient of permeability is defined as
a function of porosity:

n3 (1 − n0 )2
ks = × ks 0 (12)
(1 − n ) 2
n03

where n0 = initial porosity; ks0 = initial saturated

coefficient of permeability corresponding to n0.

2.5 Numerical implementation

In this study, a finite element model for the coupled Figure 1. Typical infinite slope under rainfall
governing equations [Eqs. (2), (5), and (7)] is devel- infiltration.
oped in the commercial multiphysics modeling
finite element program, COMSOL. In COMSOL,
the Galerkin method is used to discretize the suction and is close to the effective friction of soil
PDEs. A nonlinear differential algebraic equation at low suction (Cho & Lee 2002).
solver IDA which was created by the Lawrence
Livermore National Laboratory (Hindmarsh et al.
2005) is used as a time integrator for the differen- 3 AN ILLUSTRATIVE EXAMPLE
tial algebraic system, in which the backward differ-
entiation formulas are used to discretize the time 3.1 Finite element model
derivative terms. As the time stepping scheme is
implicit, the damped Newton method (Deuflhard Figure 2 shows the finite element model of a hypo-
1974) is then used to solve the resulting nonlinear thetical slope with a slope angle β of 35°. The slope
equations. Within each step of the Newton itera- length L is 17.5 m and the height H is 6 m. The
tion, the most recently updated nodal values of the slope is composed of residual soils. The finite ele-
dependent variables are used to compute nonlinear ment model is composed of 420 quadrilateral ele-
coefficients. The nonlinear iteration at each time ments. The initial groundwater table is along BC.
step is continued until the numerical solutions sat- The initial density of liquefied fine particles is
isfy prescribed convergence criteria. zero.

2.6 Slope stability analysis 3.2 Boundary conditions

Since slope failures induced by rainfall infiltration The boundary conditions and initial conditions for
are usually shallow, the use of infinite slope stability seepage and internal erosion are defined as follows.
analysis for the evaluation of rainfall-induced land- The velocity of liquefied particle phase flow is zero
slides is justified (Zhang et al. 2010, Santoso et al. along the AB, BC and AD boundaries. CD bound-
2011). The method used in traditional infinite slope ary is free for flow of liquefied particle.
analysis must be modified to take into account the For rainfall infiltration, AB, BC and CD are
variation of the pore water pressure profile that fixed water level. Assume the rainfall intensity is
results from the infiltration process. q, the infiltration flux along the boundary AD is
For an infinite slope with seepage parallel to the defined as follows.
slope surface, the safety factor for the slip surface
at depth h (Fig. 1) is: flux = mN (uw )q + mb (uw )Rb (− uw / γ w ) (14)

where flux = infiltration flux; mN, mb = comple-

c′ tan φ ′ hpγ w φb
Fs = + − (13) mentary smoothing functions (Fig. 3); Rb = exter-
γ th β cos β tan β γ th β nal resistance; q = rain intensity. When the pore
water pressure along the slope surface AD is
where β = slope angle; γt = total unit weight of the negative, the infiltration flux is equal to rainfall
soil; c′ = effective cohesion; φ′ = effective friction intensity q. When the pore water pressure along
angle; hp = pore pressure head; γw = unit weight of the slope surface AD is positive, the infiltration
water; φb = angle indicating the rate of increase in flux is negative to guarantee no accumulation of
shear strength related to matrix suction. The value water available on top surface (Chui & Freyberg
of φb is generally dependent on the range of matric 2009).

325

ISGSR2013.indb 325 10/18/2013 9:41:30 AM

 4 RESULTS AND DISCUSSION

4.1 Effect of internal erosion

To analyze the influence of internal erosion on
the slope stability, the results of coupled seepage
with internal erosion are compared with the results
of seepage analysis without considering internal
erosion.
As shown in Figure 4, after 2 hours of rainwater
infiltration, the depth of the wetting front is 1.4 m
when considering the internal erosion. The depth
of the wetting front is 1.2 m when neglecting the
internal erosion. After 6 hours of rain, the depth
of the wetting front is 3.5 m with internal erosion
Figure 2. Finite element model of the illustrative and is about 3.0 m without internal erosion. After
example. 10 hours of rainfall, the slope is fully saturated
with internal erosion.
Figure 5 shows the variation of fine particle con-
tent and porosity along cross-section X-X′. It can
be seen that the decrease of fine content in soil and
the increase of porosity corresponds to the advance
of wetting front. Internal erosion occurs mainly
in the shallow depth of the slope. The concentra-
tion of fine particles increases with depth and
slightly decreases around the groundwater table.
Correspondingly, the porosity decreases with depth
Figure 3. Function mN and mb.
and slightly increases around the groundwater table.
The increase of porosity around ground water table
is mainly because when the soil is fully saturated
Table 1. List of parameters for the case study.
the flow velocity is large enough to transport the
Parameters Value Parameters Value fine particles based on the constitutive model of
transport in Eq. (9). Therefore, the transport of
ρf0/ρt 0.2 θr 0 fine particles can occur in both the unsaturated
ρ*f∞/ρt 0.193 θs 0.395 zone within wetting front and the saturated zone.
ρs/ρt 1.184 ks0 (m/s) 2 × 10−5 Figure 6 shows the safety factor of slope along
v* (m/s) 3.6 × 10−7 kx (m/s) ky various depth of slip surface. It can be seen that the
αer 4.76 αw 0.2
βer 6.95 × 10−3 nw 1.35
g (m/s2) 9.81 β (°) 35
n0 0.25 φ′ (°) 34
Rb 103 c′ (kPa) 2
q (m/s) 2 × 10−5 φb (°) 25

3.3 Soil parameters

Assume the content of fine particles in the residual
soil is 20%. If sufficient time is allowed, the erosion
process will completely remove the fine particles
(ρf0→ρ*f∞). In fact, the theoretical time necessary
to complete the erosion becomes exceedingly large
when the gradient value is within the range met in
practical applications (Sterpi 2003). So ρ*f∞ is taken
about 95% of ρf0, 19.3%. The density of soil parti-
cles is 1.184 times the total density ρt. The values of
parameters are summarized sin Table 1 (Santoso Figure 4. Variation of pore-water pressure profile along
et al. 2011, Cividini & Gioda 2004). cross-section X-X′ (ks0 = 2 × 10−5 m/s).

326

ISGSR2013.indb 326 10/18/2013 9:41:32 AM

 soil slope reaches an unstable state more rapidly
when considering the internal erosion. After 6 hours
of rainwater infiltration, the depth of slip surface for
a slope without erosion is around 1.5 m. However,
for a slope with erosion, the critical slip surface is
around 2.5 m. It implies that the consequence of a
landslide may be more severe with erosion effect.

4.2 Effect of ks0

To investigate the effect of saturated coefficient of
permeability, a parametric study with different val-
ues of initial saturated coefficient of permeability
ks0, 5 × 10−6, 2 × 10−5 and 8 × 10−5 m/s is conducted.
The rainfall intensity is the same value, 2 × 10−5 m/s
for the three cases.
Figure 7 shows that before 6 hours of rainfall
the internal erosion has slight effect on pore water
pressure profile and advance of wetting front for
the case with ks0 equal to 5 × 10−6 m/s. After 6 hours,
the advance of wetting front becomes faster as time
increases. Figure 8 illustrates the safety factor pro-
files with different times. It also shows that after
6 hours the effect of internal erosion on slope sta-
bility becomes more significant.
Figure 9 presents the pore water pressure profiles
when ks0 is equal to 8 × 10−5 m/s. It can be seen that as
the saturated permeability is greater than the rainfall
intensity, the slope will not be fully saturated. The
internal erosion and transport of fine particles have
little effect on pore water pressure distributions.
Comparing Figure 4 with Figures 7 and 9, it
can be seen that when the rainfall flux is equal or
greater than the saturated coefficient of perme-
Figure 5. Variation of (a) fine particle content ρf/ ability (q/ks ≥1), the influence of internal erosion
ρf0 and (b) porosity with time along cross-section X-X′ on water infiltration and slope stability is more
(ks0 = 2 × 10−5 m/s). significant.

Figure 6. Profile of safety factor at different times

(ks0 = 2 × 10−5 m/s). Figure 7. Pore-water pressure profiles (ks0 = 5 × 10−6 m/s).

327

ISGSR2013.indb 327 10/18/2013 9:41:33 AM

 the influence of internal erosion on water infil-
tration and slope stability is more significant.

REFERENCES

Cho, S.E. & Lee, S.R. 2002. Evaluation of surficial sta-

bility for homogeneous slopes considering rainfall
characteristics. Journal of Geotechnical and Geoenvi-
ronmental Engineering 128(9): 756–763.
Chui, T.M. & Freyberg, D.L. 2009. Implementing hydro-
logic boundary conditions in a multiphysics model.
Journal of Hydrologic Engineering 14(12): 1374–1377.
Cividini, A. & Gioda, G. 2004. Finite element approach to
the erosion and transport of fine particles in granular
soils. International Journal Geomechanics 3(4): 191–198.
Dahaghi, A.K. Gholami, V. & Moghadasi, J. 2011.
Figure 8. Profiles of factor of safety (ks0 = 5 × 10−6 m/s). A novel workflow to model permeability impairment
through particle movement and deposition in porous
media. Transport in Porous Media 86(3): 867–879.
Deuflhard, P. 1974. A modified Newton method for
the solution of ill-conditioned systems of nonlin-
ear equations with application to multiple shooting.
Numerische Mathematik 22(4): 289–315.
Fredlund, D.G. & Rahardjo, H. 1993. Soil mechanics for
unsaturated soils. Wiley, New York.
Gardner, W.R. 1958. Steady state solutions of the
unsaturated moisture flow equationwith application
to evaporation from a water table. Soil Science 85(4):
228–232.
Hindmarsh, A.C. Brown, P.N. Grant, K.E. Lee, S.L.
Serban, R. & Shumaker, D.E. 2005. Woodward CS.
SUNDIALS: Suite of Nonlinear and Differential/
Algebraic Equation Solvers. ACM Transactions on
Mathematical Software 31(3): 363–396.
Hu, M.J. Wang, R. Meng, Q.S. & Liu, G.S. 2005. Research
on erosion process and features of loose gravelly soil
slope. Rock and Soil Mechanics 26(11): 1722–1726
(in Chinese).
Jian, B.T. Lu, X.B. Wang, S.Y. Chen, X.Q. & Cui, P. 2005.
Figure 9. Pore-water pressure profiles (ks0 = 8 × 10−5 m/s). The movement of fine grains and its effects on the
landslide and debris flow caused by raining. Chinese
Journal of Underground Space and Engineering 1(7):
1014–1016 (in Chinese).
5 CONCLUSIONS Papamichos, E. & Vardoulakis, I. 2005. Sand erosion with
a porosity diffusion law. Computers and Geotechnics
In this study, a coupled model of seepage and inter- 32(1): 47–58.
nal erosion in unsaturated soil is established to inves- Santoso, A.M. Phoon, K.K. & Quek, S.T. 2011. Effects
tigate internal erosion on water infiltration and slope of soil spatial variability on rainfall-induced land-
stability. The major conclusions are following: slides. Computers and Structures 89(11–12): 893–900.
Sterpi, D. 2003. Effects of the erosion and transport of
1. The soil porosity increases with the advance of fine particlesdue to seepage flow. International Journal
water front and process of internal erosion. The Geomechanics 3(1/2): 111–122.
advance of wetting front in the slope is more Vardoulakis, I. Stavropoulou, M. & Papanastasiou, P.
rapid and the slope reaches an unstable state 1996. Hydro-mechanical aspects of the sand production
more quickly due to internal erosion. problem. Transport in Porous Media 22(2): 225–244.
Xu, J.C. & Shang, Y.G. 2006. Influence of permeability
2. With the effect of internal erosion, the slip sur-
of gravel soil on debris landslide stability. Chinese
face of failure is deeper, which implies that the Journal of Rock Mechanics and Engineering 25(11):
consequence of a landslide may be more severe 2264–2271 (in Chinese).
with erosion. Zhang, L.L. Zhang, J. Zhang, L.M. & Tang, W.H. 2010.
3. When the rainfall flux is equal or greater than the Stability analysis of rainfall-induced slope failure:
saturated coefficient of permeability (q/ks ≥1), a review. Geotechnical Engineering 164(5): 299–316.

328

ISGSR2013.indb 328 10/18/2013 9:41:35 AM

 7 Reliability- and risk-based monitoring and site investigation

ISGSR2013.indb 329 10/18/2013 9:41:37 AM

This page intentionally left blank
 Geotechnical Safety and Risk IV – Zhang et al. (eds)
© 2014 Taylor & Francis Group, London, ISBN 978-1-138-00163-3

Effect of inadequate site investigation on the cost and time

of a construction project

A.H. Albatal
Arab Academy for Science, Technology and Maritime Transport, Cairo, Egypt

H.H. Mohammad
Zagazig University, Zagazig, Egypt

M.E. Abd Elrazik

Arab Academy for Science, Technology and Maritime Transport, Cairo, Egypt

ABSTRACT: The appropriate information or data is the keystone of any successful design. Site
investigation is one of the first steps of any construction projects. The purpose of a soil subsurface inves-
tigation is to provide data concerning the engineering properties of the soil for the proper design and
safe construction of a project. The site investigation phase of any geotechnical design plays a vital role to
provide the geotechnical engineer by the most appropriate data to ensure that the design data represent
the investigated soil. Inadequate soil investigation may contribute to either a significantly over designed
foundations, that is not cost-effective, or an under designed foundations, which may lead to potential
failures. Insufficient geotechnical investigation is one of first sources of projects’ delays, disputes, claims,
and projects’ cost overruns. This paper aims to focus on the impact of varying the scope of the site investi-
gation process, on the financial risk of construction projects. Another goal is to compare the cost of extra
site investigation with the repairing or reconstruction cost result from improper site investigation. The
results of limited site investigation scope are clearly shown in a recycling factory at the 6th of October city
governorate, Egypt, which suffer from cost overrun and long delays. The apparent savings due to making
inadequate site investigation leads to cost overrun by about 64.2% of the project cost.

1 INTRODUCTION Several studies have been published over the last

30 years or so that clearly demonstrate that, in civil
Site investigation is normally carried out prior to engineering projects, the largest element of finan-
the commencement of design of any project. Due cial and technical risk usually lies in the ground
to lack of or inadequacy of guide/code require- (National Research Council 1984, Institution of
ment regarding the extent as well as quality of Civil Engineers 1991, Whyte 1995).
site investigation work, geotechnical failures often Goldsworthy et al. (2007) defined the financial
occurred. These failures sometime lead to cata- risk as the total cost, which includes costs associ-
strophic disaster and imposed serious threat to ated with undertaking the site investigation, con-
public safety (Moh, 2004). structing, and any works required to rehabilitate
Baecher & Christian (2003) divided the charac- the failure. Goldsworthy et al. (2004) mentioned
terization of ground conditions into two phases. that the risk of a foundation failure is heavily
First is a preliminary investigation or desk study, dependent on the quantity and quality of infor-
which involves collecting information about the mation obtained from a geotechnical site investi-
regional geology and geological history. The second gation aimed at characterizing the underlying soil
phase is a site investigation designed to obtain data conditions. Project risk is a measure of the poten-
based on detailed measurements of soil properties. tial inability to achieve overall project objectives
As a result, the geotechnical data obtained from within defined cost, time schedule, quality, environ-
limited characterization of ground conditions can mental impact and technical constraints and can
be both inadequate and/or inappropriate. This be estimated as the combination of the probability
situation can lead to failure and a high level of of a risk event occurring and its consequences for
financial and technical risk (Institution of Civil project objectives (Carlsson, 2005).
Engineers 1991, National Research Council 1984, Jaksa et al. (2005) suggested that the site inves-
Temple & Stukhart 1987). tigations that inadequately quantify the variability

331

ISGSR2013.indb 331 10/18/2013 9:41:37 AM

 of the ground can result in three possible cost Unforeseen ground conditions can, and often do,
outcomes: raise the costs by 10% or more (Paul et al. 2002).
a. The foundation is underdesigned as a result on
an overly optimistic geotechnical model, and
2 CASE STUDY APPLICATION
hence fails to comply with the design criteria,
which can ultimately lead to some level of struc-
Al-Ertikaa Factory is a recycling factory at the site
tural distress.
of Low Cost Housing project at the 6th October
b. The foundation is overdesigned as a conse-
and includes the following buildings:
quence of a pessimistic geotechnical model and/
or inherent conservatism in the design process. a. Administration building
c. Unforeseen conditions require substantial b. Restaurant hall
changes to the foundation system, which also c. Two residential buildings for technicians and
result in construction delays. labors
d. Factory building.
It is clear that over the last 30 years geotech-
nical investigation prices have been driven down, The main philosophy of the low cost project is
with the scope often being governed by minimum to economize the used construction material. Each
cost and time of completion (Institution of Civil of the two residential buildings consists of number
Engineers 1991). As a consequence, the Institution of adjacent units. Each unit consists of a room,
of Civil Engineers concluded that: “You pay for a hall, kitchen and bathroom. Each building consists
site investigation whether you have one or not.” of one floor (ground floor) only. The buildings
The UK public accounts committee states that structural system consists of wall bearing founded
the average cost increase for major road projects on reinforced concrete strip footings. Walls are
(1988–1989) was 28% over original tender prices. constructed of red bricks and limestone blocks.
The main reason was judged to be the undertak- The foundations are strip reinforced concrete over
ing of larger and more complex schemes that plain concrete. The depth of the foundation is less
involved greater risks, particularly with ground than 1 m below the ground surface. Information
conditions. from the site indicates that there is a replacement
Despite the level of sophistication available for fill placed below the foundations. The thickness of
the determination of risk and uncertainty asso- the replacement fill is about 1.5 to 2.0 m.
ciated with ground work operations, a review of
5000 industrial building projects by the National
2.1 Project site investigation scope
Economic Development Office in the UK
(NEDO, 1983) showed that 37% of the projects Since the philosophy of the low cost housing is
suffered delays due to ground related problems. to construct the building with minimum cost, the
When 8000 commercial buildings were examined, project owner, consultant and contractor agreed
50% of the samples were found to have suffered to reduce the number of site investigation tests
unforeseen ground difficulties (NEDO, 1988). The as much as possible. Site investigation boreholes
financial scale of the problem was confirmed by have been taken randomly in the site regardless
the National Audit Office (1994), in a report that the building locations. The contractor adopted
recorded 210 premature failures during construc- the knowledge of the geology of the area and the
tion works, and that geotechnical failures were a knowledge of the soil interpretation from previ-
major concern. Alhalby & Whyte’s (1994) research ous site investigation in the project to predict the
concluded that “90% of risk to projects originates stratigraphy and properties of the soil in the fac-
from unforeseen ground conditions which could tory location.
often have been avoided by adequate and full site
investigation”.
2.2 Problem
The cost of site investigations in relation to the
total project cost is small. Typical values in build- Directly after finishing the factory construction,
ings projects are between 0.05% and 0.20% of total cracks started to appear. The damaged struc-
project cost, or 0.5% and 2.0% of foundations tures are the two residential units for technicians
cost, and the typical values are between 0.20% and and labors. The rest of the structures suffer no
1.50% of total project cost or 1.0% and 5.0% of observed damage or cracks. Some of the cracks
foundations cost in roads projects. Site investiga- are dangerous especially at the middle units where
tions should be continued until the ground condi- concentrations of the cracks are to the extent that
tions are known well enough for work to precede the ceramic are cracked and some ceramic slabs
safely. Although a doubling in site investigations were fallen down. Figure 1 shows some cracked
costs can add 1.0% to the total project cost. elements.

332

ISGSR2013.indb 332 10/18/2013 9:41:37 AM

 inundation of 2700 kPa. These numbers indicate
high intrinsic expansiveness and swelling poten-
tial of the clay. The clay exists in the site from a
depth of about 1 m and extends down to a depth
of 4 to 5.5 m below the ground surface. This means
that the swelling clay exists under the replacement
fill that is under the foundations. Planting green
areas and trees adjacent to the damaged buildings
introduced water to the subsurface formation. The
fact that replacement fill is sand which is perme-
able material facilitated the seepage of water to the
swelling clay under the replacement fill. The swell-
ing of the clay caused differential vertical displace-
ment that caused distress to the walls and domes
of the buildings.

2.3 Problem corrective action

The corrective action for these problems was as
following.
a. Soil investigation was made to identify the cause
of the buildings cracks.
b. Removing the trees adjacent to the buildings to
a distance of 1.50 m from the buildings. Includ-
ing changing the manholes locations and sewer
lines paths.
c. Performing plastic sheets U-shape protection to
prevent irrigation water from leaking to the soil
under the buildings.
d. Replanting the landscape area and provide it by
appropriate sprinkler irrigation system.
e. Monitoring recording of vertical displacements
at Elevation Reference Points for a period of
two months.
f. Fixing structural damages and replacing dam-
aged parts. This including fixing the building
cracks.
It should me mentioned that by removing the
trees and protecting the building underlain soil
from any source of water there is no need to make
any adjustment for the soil. The building has small
weight and then the stresses on the soil are low
stresses.

2.4 Rehabilitation cost for these problems

The rehabilitation cost for the items that per-
formed in the site were as in Table 1. These costs
have been obtained from the consultant engineer
Figure 1. Buildings cracks. in the site.

2.5 Typical site investigation cost

According to the consulting services report,
the damage reason was due to presence of swell- According to literature, the typical site investi-
ing clay under the replacement fill under the gation cost in residential buildings projects is
foundations. The clay has free swell values up to between 0.05% and 0.20% of total project cost.
220% and one dimensional swelling pressure upon Accordingly, site investigation cost for this project

333

ISGSR2013.indb 333 10/18/2013 9:41:37 AM

 Table 1. Rehabilitation cost for the damaged buildings. construction contracts. In it turns the construction
contracts should be coinciding with the country
Item Rate Price law. On the other hand, the laws do not specify
No* Unit (LE) Quantity (LE) who is responsible for taking the site investigation,
a L.S. 8,750 but most of laws identify who is the responsible
b m2 150 450 67,500 for any unforeseen site conditions. By default,
d m2 15 700 10,500 the responsible for the unforeseen site condition
c m2 250 450 112,500 is the one who is also responsible for investigate
e Day 150 60 9,000 these conditions.
f L.S. 25,000
Total rehabilitation cost 233,250 3.1 Responsibility of site investigation in the
Egyptian law
*From section 2.3.
Responsibility of site investigation in the Egyptian
law could be concluded from the following
articles.
should be between (LE 182), and (LE 726). While
According to the Egyptian code of practice for soil • Egyptian Civil Code: Article 147
mechanics, design and execution of foundations The contract makes the law of the parties. It can be
(2001), the boreholes number for such type of con- revoked or altered only by mutual consent of the
structions is to be at least two boreholes if the area parties or for reasons provided for by law.
less than 300 m2 and to a depth of 10 m, one extra When, however, as a result of exceptional
borehole for each (300 m2 to 500 m2). Since the area and unpredictable events of a general character,
of each building is about 222 m2, then the number the performance of the contractual obligation,
of boreholes should be four boreholes for the two without becoming impossible, becomes excessively
buildings. The cost of the four boreholes is about onerous in such way as to threaten the debtor with
(LE 3550). Accordingly, the site investigation cost exorbitant loss, the judge may according to the
for the two buildings must be at least (LE 3550). circumstances, and after taking into consideration
the interests of both parties, reduce to reasonable
limits, the obligation that has become excessive.
2.6 Comparing the rehabilitation cost with
Any agreement to the contrary is void.
rehabilitation and construction cost
• Egyptian Civil Code: Article 651
The rehabilitation cost is (LE 233,250) which rep- The engineer and contractor are jointly and sever-
resent 65.7 times the typical site investigation cost. ally responsible for a period of ten years for the
While the total cost for the buildings including total or partial demolition of constructions or
landscape was (LE 363,150). By comparing the other permanent works erected by them, even if
rehabilitation cost to the construction cost it could such destruction is due to a defect in the ground
be concluded that the rehabilitation cost represents itself, and even if the master authorized the erec-
64.2% of the total construction cost. tion of the defective construction, unless, in this
case, the constructions were intended by the par-
ties to last for less than ten years.
2.7 Time delay due to the problem
The warranty imposed by the preceding para-
The project contract duration was 12 months. graph extends to defects in constructions and erec-
Due to the cracks, the project delivery time was tions which endanger the solidity and security of
3 months more than the scheduled time. This the works.
means that the delay represents 25% of the project The period of ten years runs from the date of
total duration. delivery of the works.
This article does not apply to the rights of
action which a contractor may have against his
3 CONTRACTUAL AND LEGAL ASPECTS sub-contractors.
OF SITE INVESTIGATION • Article 80 in Code number 98, 1989
The contractor is responsible for checking the
Unforeseen site conditions may have an impact nature of the work including any tests required
on time and cost of construction project. for ensuring the validity of the specifications,
Sometimes they may prevent the contractors form drawings and designs. The contractor shall give
performing the contractual obligations, and other notice to the owner as soon as practicable, and
times they only make it harder or more costly will be responsible for all the specifications, draw-
to perform the contract. The responsibility of ings and designs contents as like he prepare them
the site investigation should be identified in the himself.

334

ISGSR2013.indb 334 10/18/2013 9:41:38 AM

 Egyptian law is clear about bearing the In the guiding contract for design and execution
contractor the whole responsibility of any unfore- (by the owner funding), the ministerial decree
seen conditions. There is no argument about the number 246 for the year 1999, article number 4–11
contractor responsibility of any unforeseen con- under title “under-ground unexpected conditions”,
ditions and then by default the soil investigation. the following:
Even though the Egyptian law makes the consult-
• If the contractor faced any under-ground
ant jointly with the contractor responsible for any
unexpected conditions and he thinks that these
defective construction, but according to the code
conditions could not be discovered by an experi-
number 89, 1989 the contractor is responsible for
enced contractor, the contractor should inform
any required test to check the work nature even if
the owner to inspect these conditions. After
the owner or the consultant made these.
inspection and exploration, the owner should
agree on or decide to:
3.2 Contractual aspects of site investigations 1. What is the time extension is the contractor
deserve?
The contract is the main reference in case of any
2. What is the extra cost that should be added to
disputes between the contract parties. Article 147
the contract price due to these conditions?
of the Egyptian Civil Code, states: “The contract
makes the law of the parties”. So, it is important After deciding, the owner should inform the
to prepare the contracts in a way that ensure the contractor what is his decision.
responsibility of the site investigations. There It is obvious that the guiding decrees are compat-
is number of ministerial decrees that suggested ible with the Egyptian law. In both the contractor
comprehensive contracts between the owner and is the responsible for making the site investigation,
the contractor, and the owner and the consultant. while the consultant is the one who responsible for
These contracts are guiding contracts, and the supervising the site investigations. Consultant and
parties have the option of follow them or choosing contractor both are responsible for making a deci-
their own contract form. The article in these con- sion if more site investigation is required. If any
tracts concerning site investigation or unforeseen unforeseen site investigation faced, the owner and
site conditions are as following. the consultant have the upper hand of deciding
It is mentioned in the Ministerial decree number whether if the contractor deserving to be paid for
222 for the year 1994 about the general conditions overcoming these conditions or not. In the other
for the construction contracts the following. hand the contractor and the consultant are both
responsible for any consequences for the unfore-
• Article 43. Boreholes and investigations: if the
seen site condition. This makes sense because the
engineer (The consultant) or the contractor
owner usually does not have experience about the
found that more soil investigation is required,
right procedures to construct his structure. But
then the engineer should mandate the contrac-
such conditions make the contractor constrained
tor to make the required soil investigations. The
to the consultant and owner willing to pay.
extra soil investigation is to be considered as an
extra work except if this work has been listed in
the Bill of Quantities.
4 CONCLUSION
• Sub-article 24/2. Site inspection and preview: if the
contractor faced any natural or artificial obstacles
The results of the analyses conducted in this
which may affect the project time and duration,
research, show that the inadequacy of the site
he should inform the engineer (consultant). If the
investigation represents major factor on the con-
consultant is convinced that these conditions
struction cost and duration Due the inadequate
could not be discovered by an expert contrac-
site investigation, the extra cost represents 65.7
tor, he should review the contractor requests to
times the required site investigation cost. While the
determine the extra cost that the contractor paid
delay time due to the inadequacy of site investiga-
to face any circumstances that were not expected
tion represents 25% of the project total duration.
during tendering period. The owner should pay
The best procedure to avoid these problems is
these expenses, and the contractor should be
by making a proper site investigation. The proper
given an extra time to overcome these obstacles.
site investigation comes in two phases. In the first
In the guiding contract for studies and design- phase, the owner should make a preliminary site
ing, it is mentioned in the addendum 2 of the min- investigation before the bidding process. Before
isterial decree number 221 for the year 1994, it is submitting his proposal, the contractor should
one on the consultant scope of work to supervise take into consideration, based on the preliminary
the soil inspection and experiments, and study and site investigation, whatever if more site investiga-
evaluate the soil reports. tion is required or not. If more site investigation is

335

ISGSR2013.indb 335 10/18/2013 9:41:38 AM

 required, the scope of the extra site investigation Goldsworthy, J.S., Jaksa, M.B., Fenton, G.A.,
should be identified. The expenses of the site Griffiths, D.V., Kaggwa, W.S and Poulos, H.G.,
investigation are minor comparing with the conse- 2007. Measuring the Risk of Geotechnical Site
quences that might happen if the site investigation Investigations. Proceedings of the Geo-Denver 2007
Symposium. K.K. Phoon et al., Eds., ASCE.
ignored. Institution of Civil Engineers, 1991. Inadequate Site
It is very important to put a clause or clauses Investigation. Thomas Telford, London.
in the contract to specify who is responsible for Jaksa, M.B., Goldsworthy, J.S., Fenton, G.A.,
the unforeseen site conditions. The party who is Kaggwa, W.S., Griffiths, D.V., Kuo, Y.L. and
responsible for these conditions will tend to make Poulos, H.G., 2005. Towards Reliable and Effective Site
a proper site investigation to reduce the risk of Investigations. Geotechnique Journal 55: 109–121.
unforeseen site condition problems. It should be Moh, Z.C., 2004. Site Investigation and Geotechnical
taken to the consideration that the responsible Failures. Proceeding of International Conference
here is only responsible for the cost and time delay. on Structural and Foundation Failures, August 2–4,
2004, Singapore.
While the consequences of unforeseen site condi- National Audit Office 1994. Quality Control of Roads
tions in case of partially or total failure is restricted and Bridge Construction. HMSO Publications, July
in the law of each country. According to the 1994, UK.
Egyptian law, the responsibility of any destruction National Economic Development Office (NEDO),
due to the inadequate site investigation is held by 1983. Faster Building for Industry: Building E.D.C.
the consultant and the contractor. So, they should National Economic Development Office, London.
convince the owner to make a proper site investiga- National Economic Development Office (NEDO), 1988.
tion if the owner underestimated its usefulness. Faster Building for Commerce. HMSO, London.
National Research Council, 1984. Geotechnical Site
Investigations for Underground Projects. US National
Committee on Tunneling Technology, Vol. No. 1,
REFERENCES National Academy Press, Washington.
Paul, T., Chow, F. and Kjekstad, O., 2002. Hidden
Alhalaby, N.M. and Whyte I.L., 1994. The Impact of Aspects of Urban Planning- Surface and Underground
Ground Risks in Construction on Project Finance. Development. Thomas Telford Publishing, London.
Risk and Reliability in Ground Engineering Temple M.W.B. and Stukhart G., 1987. Cost
Conference, London, November 11–12, 1993. Effectiveness of Geotechnical Investigations. Journal
Baecher, G.B. and Christian, J.T., 2003. Reliability and of Management in Engineering 3(1), pp. 8–19.
Statistics in Geotechnical Engineering. John Wiley & Whyte, I.L., 1995. The Financial Benefit from Site
Sons Ltd., Chichester, England. Investigation Strategy. Ground Engineering 10,
Carlsson, M., 2005. Management of Geotechnical Risks 33–36.
in Infrastructure Projects. Licentiate Thesis, Division
of Soil and Rock Mechanics, Royal Institute of
Technology, Stockholm, Sweden.
Goldsworthy, J.S., Jaksa M.B. Kaggwa, W.S.,
Fenton, G.A., Griffiths, D.V. and Poulos H.G. 2004.
Cost Of Foundation Failures Due To Limited Site
Investigations. Proceeding of International Conference
on Structural and Foundation Failures, August 2–4,
2004, Singapore, pp. 398–409.

336

ISGSR2013.indb 336 10/18/2013 9:41:39 AM

 Geotechnical Safety and Risk IV – Zhang et al. (eds)
© 2014 Taylor & Francis Group, London, ISBN 978-1-138-00163-3

Probabilistic characterization of Young’s modulus using Markov Chain

Monte Carlo simulation

Z. Cao
State Key Laboratory of Water Resources and Hydropower Engineering Science, Wuhan University, Wuhan,
P.R. China

Y. Wang
Department of Civil and Architectural Engineering, Shenzhen Research Institute, City University of Hong Kong,
Hong Kong, China

ABSTRACT: This paper presents a Markov Chain Monte Carlo Simulation (MCMCS)-based approach
for probabilistic characterization of the undrained Young’s modulus, Eu, of clay, which utilizes both
the prior knowledge and project-specific SPT data to generate a large number of equivalent samples
of Eu for its probabilistic characterization. The proposed approach combines the prior knowledge and
project-specific SPT data systematically under a Bayesian framework and allows general choices of
realistic prior knowledge (e.g. an arbitrary histogram type of prior distribution). Equations are derived
for the proposed approach, and a sensitivity study is performed to explore the effects of prior knowledge
on probabilistic characterization of soil properties. It is shown that the proposed equivalent sample
approach integrates the information provided by different types of prior knowledge with project-specific
information in a rational manner and improves significantly probabilistic characterization of soil
properties by incorporating consistent prior knowledge.

1 INTRODUCTION of geotechnical structures. This problem is further

complicated by fact that geotechnical site investiga-
In the past few decades, probability analysis and tion is a multi-step process that relies on both site
Reliability-Based Design (RBD) of geotechni- information available prior to the project (namely
cal structures (e.g. foundations and retaining “prior knowledge”, such as engineering experience
walls) have attracted considerable attention. This and judgment) and site observation data obtained
has led to several probability analysis methods from test borings, in-situ tests (e.g. Standard Pen-
(e.g. Christian et al. 1994, El-Ramly et al. 2005, etration Test (SPT)), and/or laboratory tests (e.g.
Wang et al. 2011b, Wang 2012) and RBD codes Mayne et al. 2002, Cao & Wang 2013, Wang & Cao
(e.g. Phoon et al. 1995, Honjo et al. 2010, Wang 2013). It remains a challenging task for geotechni-
et al. 2011a, Wang 2011) in geotechnical engineer- cal engineers to integrate systematically the prior
ing to deal rationally with various geotechnical- knowledge and site observation data in a probabi-
related uncertainties, such as the uncertainty in listic manner.
soil properties. These probability analysis meth- To address these challenges, this paper presents
ods and RBD codes usually require probabilistic a Markov Chain Monte Carlo Simulation
estimations of soil properties as input, such as the (MCMCS)-based approach for probabilistic char-
statistics (e.g. mean, standard deviation, and/or acterization of the undrained Young’s modulus,
lower 5% quantiles) and probability distributions Eu, of clay, which utilizes both the prior knowledge
(e.g. Probability Density Function (PDF) and and project-specific SPT data to generate a large
Cumulative Distribution Function (CDF)) of soil number of equivalent samples of Eu for determin-
properties. This poses a challenge in the implemen- ing the statistics and probability distributions of Eu.
tation of these probabilistic analysis methods and The proposed approach combines the prior knowl-
RBD codes because the number of soil property edge and project-specific SPT data systematically
data obtained during geotechnical site investiga- under a Bayesian framework, in which the prior
tion is generally too sparse to generate meaning- knowledge is reflected by the prior distribution. It
ful statistics and probability distributions of soil allows general choices of realistic prior distribu-
properties for probability analysis and/or design tions (e.g. uniform prior distribution, triangular

337

ISGSR2013.indb 337 10/18/2013 9:41:39 AM

 prior distribution, an arbitrary histogram type of between the Eu and the N values measured during
prior distribution, and normal prior distribution). SPT tests. Consider, for example, an empirical
The paper starts with development of the Bayesian regression between the Eu measured by pressurem-
framework for integrating the prior knowledge eter tests and SPT N values (Kulhawy & Mayne
with project-specific test data and derivation of the 1990, Phoon & Kulhawy 1999b):
PDF of Eu using the integrated information, fol-
lowed by generating a large number of equivalent Eu / pa .3 N 0.63 (2)
samples of Eu for its probabilistic characterization
using MCMCS. Then, a sensitivity study is per- where pa is the atmospheric pressure (i.e. 0.1 MPa).
formed to illustrate the flexibility of incorporating In a log-log scale, Equation 2 can be rewritten as:
various different types of realistic prior knowledge
into the proposed approach and to explore the
effects of prior knowledge on probabilistic charac- ξ u ε (3)
terization of soil properties.
in which ξ = ln(N) = the logarithm of SPT N value;
ln(Eu) = the logarithm of Eu; a = 1.587, b = −1.044,
2 BAYESIAN FRAMEWORK and ε is a Gaussian random variable with a zero
mean and a standard deviation σε = 1.352 (Phoon &
The properties of geotechnical materials are inher- Kulhawy 1999b). The last term ε represents a mod-
ently variable because of various factors (e.g. eling scatterness or transformation uncertainty
parent materials, weathering and erosion processes, associated with the regression equation. Since Eu is
transportation agents, and conditions of sedimen- a lognormal random variable, ln(Eu) in Equation 3
tation, etc.) during their formation process (e.g. is a normal random variable, and it can be written
Phoon & Kulhawy 1999a). Consider, for example, as (e.g. Ang & Tang 2007)
the undrained Young’s modulus, Eu, within a clay
layer. To model explicitly its inherent variability, ln( u) N σNz (4)
Eu is represented by a lognormal random vari-
able with a mean μ and standard deviation σ. The
in which z is a standard Gaussian random variable;
model parameters μ and σ are unknown and need
to be determined during geotechnical site investi-
μ N l μ − σ N2 /2 and σ N l ( + (σ / μ )2 ) are
the mean and standard deviation of ln(Eu), respec-
gation based on both prior knowledge and project-
tively. Combining Equations 3 and 4 leads to:
specific test results (e.g. SPT N values). Within
a Bayesian framework, the updated knowledge
about μ and σ is reflected by their joint posterior ξ μN + ) + aσ N z + ε (5)
distributions based on prior knowledge and site
observation data (e.g. Ang & Tang 2007, Wang & When the inherent variability is assumed to be
Cao 2013) independent of the transformation uncertainty
(i.e. z is independent of ε), ξ is a Gaussian random
P ( μ ,σ | Data ) K ( Data | μ ,σ )P ( μ ,σ )
KP (1) variable with a mean of aμN + b and standard
deviation of a 2σ N2 + σ ε2 . The project-specific SPT
in which K = (∫μ,σP(Data|μ,σ)P(μ,σ)dμdσ)−1 is a data (i.e. Data = {ξi = ln(Ni), i = 1,2, …, ns}) can
normalizing constant that does not depend on μ be considered as ns independent realizations of the
and σ; Data = {ξi = ln(Ni), i = 1, 2, …, ns}, where Gaussian random variable ξ. The likelihood func-
ξi, i = 1, 2, …, ns, denote ns SPT N values obtained tion for the project-specific SPT data is therefore
within a clay layer in a log scale; P(Data|μ,σ) is the expressed as (e.g. Ang & Tang 2007, Wang & Cao
likelihood function reflecting the model fit with 2013)
Data; P(μ,σ) is the prior distribution of μ and σ
that reflects the prior knowledge on μ and σ in the P ( Data | μ,σ )
absence of Data. Calculations of the likelihood ns ⎧ ⎡ ⎤ ⎫
2
1 ⎪ 1 ⎢ξi (aaμ N b )⎥ ⎪
function and prior distribution are presented in the =∏ exp ⎨− ⎬
following two subsections, respectively. 2π a 2σ N + σ ε2 ⎪ 2 ⎢⎣ a σ N + σ ε ⎥⎦ ⎪
2 2 2 2
i =1
⎩ ⎭
2.1 Likelihood function (6)
The undrained Young’s modulus Eu can be esti-
2.2 Prior distribution
mated from SPT tests (Kulhawy & Mayne 1990,
Phoon & Kulhawy 1999b). The Eu value of the The prior distribution can be simply assumed
tested soil is obtained by means of regression as a joint uniform distribution of μ and σ with

338

ISGSR2013.indb 338 10/18/2013 9:41:39 AM

 respective minimum values of μmin and σmin and Combining Equations 1 and 8 leads to:
respective maximum values of μmax and σmax, and it
is expressed as (e.g. Ang & Tang 2007) P ( Eu | Data, )
= K ∫ P ( Eu | μ ,σ )P ( Data | μ ,σ )P ( μ ,σ )d μdσ
μ ,σ
P( μ,σ ) =
⎧ 1 for μ [μ min , μ max] (9)
⎪
⎨( μ max min )( max min ) and σ [σ m
min ,σ mmax ] where P(Eu|μ,σ) = conditional PDF of Eu for a
⎪ 0 otherwise given set of model parameters (i.e. μ and σ). Since
⎩
Eu is a lognormal random variable, P(Eu|μ,σ) is
(7) given by a lognormal distribution with a mean μ
and standard deviation σ (e.g. Ang & Tang 2007,
Note that only the possible ranges (i.e. μmin and Wang & Cao 2013). Equation 9 is a product of
μmax, σmin and σmax) of the model parameters are the normalizing constant K and the integral term
needed to completely define a uniform prior dis- defined as I = ∫μ,σ P(Eu|μ,σ)P(Data|μ,σ)P(μ,σ)dμdσ,
tribution herein. This requires relatively limited and it gives the PDF of Eu for a given set of prior
prior knowledge (e.g. reasonable ranges of soil knowledge (i.e. Prior) and project-specific SPT
properties of interest), which is commonly avail- data (i.e. Data). In the next subsection, the Markov
able in geotechnical literature (e.g. Kulhawy & Chain Monte Carlo Simulation (MCMCS) method
Mayne 1990, Phoon & Kulhawy 1999a,b). The (e.g. Robert & Casella 2004) is used to generate a
approach proposed in this paper is general and sequence of Eu samples whose limiting stationary
equally applicable for more sophisticated types distribution is the PDF of Eu (i.e. Eq. 9).
of prior distributions (e.g. triangular prior dis-
tribution, an arbitrary histogram type of prior
distribution, and normal prior distribution), 3.1 Markov Chain Monte Carlo Simulation
which of course require relatively informative MCMCS method is a numerical process that simu-
prior knowledge as justifications. Effects of dif- lates a sequence of samples of a random variable
ferent prior knowledge are further discussed in (e.g. Eu) as a Markov Chain with the PDF of the
Section 4. random variable (e.g. Eq. 9 for Eu) as the Markov
Chain’s limiting stationary distribution (e.g. Beck &
Au 2002, Robert & Casella 2004). The states of the
3 PROBABILITY DENSITY FUNCTION Markov Chain after it reaches stationary condition
OF UNDRAINED YOUNG’S MODULUS are then used as samples of the random variable
with the target PDF. It provides a feasible way to
In this study, the undrained Young’s Modulus is generate samples from an arbitrary PDF, particu-
modeled by a random variable Eu, which follows a larly when the PDF is complicated and is difficult
lognormal distribution with a mean μ and stand- to express analytically or explicitly.
ard deviation σ. Both prior knowledge and project- In this study, the Metropolis-Hastings (MH)
specific test data (e.g. SPT N values) are used to algorithm (e.g. Hastings 1970, Beck & Au 2002)
estimate the distribution model parameters μ and is used in MCMCS to generate totally nMCMC
σ in geotechnical site investigation. For a given set number of the Eu samples from Equation 9. The
of prior knowledge (i.e. Prior) and project-specific Eu Markov Chain starts with an arbitrary ini-
SPT data (i.e. Data), there are many sets of pos- tial state, Eu,1. Then, a candidate sample, Eu*,m ,
sible values of μ and σ. Each set of μ and σ has m = 2, 3, …, nMCMC, for the m-th state Eu,m of the
its corresponding plausibility (or occurrence prob- Markov Chain is generated from the proposal
ability), which is defined by a joint conditional PDF Eu*,m | Eu,m − 1 ) of the Markov Chain
PDF P(μ,σ|Data, Prior). Using the conventional based on its previous state (m – 1)-th, Eu,m−1 (e.g.
notation of Bayesian framework, P(μ,σ|Data, the initial state Eu,1 for Eu,2). Herein, the proposal
Prior) is simplified as P(μ,σ|Data) and is given PDF f Eu*,m | Eu,m − 1 ) is taken as a Gaussian PDF,
by Equation 1. Using the Theorem of Total Prob- which is centered at the previous state Eu,m−1 and
ability (e.g. Ang & Tang 2007), the PDF of the has a Coefficient of Variation (i.e. COV) equal
undrained Young’s modulus Eu for a given set of to the mean COV of the prior knowledge (e.g.
prior knowledge and project-specific SPT data is COV = (σmax + σmin)/(μmax + μmin) for the uniform
expressed as: prior given in Eq. 7). The chance to accept the
candidate sample Eu*,m as the Eu,m depends on
P ( Eu | Data, ) the “acceptance ratio”, ra. Using Eu*,m , Eu,m−1, the
proposal PDF, and the PDF of Eu (i.e. Eq. 9), the
= ∫ P ( Eu | μ ,σ )P ( μ ,σ | Dataa, )d μdσ (8)
μ ,σ acceptance ratio for Eu*,m is calculated. Details of

339

ISGSR2013.indb 339 10/18/2013 9:41:42 AM

 the calculation of ra are found in Wang & Cao (2013). obtained from the clay site of the US National
After ra is calculated, a random number u is gen- Geotechnical Experimentation Sites (NGES) at
erated from a uniform distribution with a range Texas A&M University (Wang & Cao 2013). It
from zero to one. When u is less than ra, Eu*,m is was shown that, based on the limited SPT data
accepted as Eu,m, i.e. Eu,m Eu*,m . Otherwise, Eu*,m and relatively uninformative prior knowledge (i.e.
is rejected, and Eu,m is taken as equal to the previ- reasonable ranges of soil parameters reported in
ous state Eu,m−1, i.e. Eu,m = Eu,m−1. Starting from the the literature), the equivalent sample approach
initial sample Eu,1, the procedure described above provides reasonable estimates of the statistics and
is repeated nMCMC − 1 times to generate nMCMC − 1 probability distributions of Eu.
samples of Eu, i.e. Eu,m, m = 2, 3, …, nMCMC. This
leads to a Markov Chain that is comprised of
nMCMC Eu samples (including the initial sample). 4 SENSITIVITY STUDY ON DIFFERENT
Finally, the Eu samples obtained after the Markov PRIOR KNOWLEDGE
Chain reaches its stationary condition are consid-
ered as appropriate samples for probabilistic char- The proposed approach provides engineers with
acterization of Eu. flexibility to use different types of prior distribu-
tions, which reflect the prior knowledge realisti-
cally. To illustrate such flexibility and explore the
3.2 Equivalent samples effects of prior knowledge, a sensitivity study is
Equation 9 shows that the PDF of Eu contains performed in this section using simulated SPT
information from both project-specific site obser- data that is generated from Equation 5 with
vation data (i.e. Data) and prior knowledge (i.e. μN = 2.23 and σN = 0.39 (i.e. μ = 10.0 MPa and
Prior). The information from these two different σ = 4.0 MPa) and four different sets (i.e. prior
sources is integrated probabilistically in a rational knowledge I, II, III, and IV) of prior knowledge
manner. The MCMCS samples that draw from the shown in Figure 1.
Eu PDF, therefore, contain the integrated informa-
tion of both project-specific site observation data
and prior knowledge. When the project-specific
data is limited, the MCMCS samples mainly reflect
the prior knowledge. As the amount of project-
specific data increases, the effect of the project-
specific data on the MCMCS samples gradually
increases.
More importantly, a large amount of Eu samples
can be generated conveniently by MCMCS. From a
statistical point of view, these MCMCS samples are
equivalent to those Eu data that are measured physi-
cally from laboratory or in-situ tests (e.g. pressurem-
eter tests). Therefore, this large amount of equivalent
samples can be analyzed statistically, using conven-
tional statistical methods, to estimate the required
statistics (e.g. mean, standard deviation, and/or
lower 5% quantiles) and probability distributions
(e.g. PDF and CDF) of Eu (or ln(Eu)) in probabilistic
analysis and RBD of geotechnical structures.
The proposed approach integrates probabilisti-
cally the prior knowledge (e.g. previous engineer-
ing experience including regression equations) and
project-specific test data (e.g. SPT test data) and
transforms the integrated information into a large
number, as many as needed, of equivalent samples.
This allows meaningful statistics of soil properties
to be obtained using conventional statistical analy-
sis. The proposed approach effectively tackles the
difficulty in generating meaningful statistics from
the usually limited number of soil property data
obtained during geotechnical site investigation. It
has been illustrated using a set of real SPT data Figure 1. Prior knowledge used in the sensitivity study.

340

ISGSR2013.indb 340 10/18/2013 9:41:45 AM

 As shown in Figure 1, the prior knowledge I is
a uniform prior distribution with μmin = 5.0 MPa,
μmax = 15 MPa, σmin = 0.5 MPa, and σmax = 13.5 MPa.
Prior knowledge I is consistent with the typical
ranges of undrained Young’s modulus reported
in the literature (e.g. Kulhawy & Mayne 1990,
Phoon & Kulhawy 1999a, b), and it is used as the
baseline case in this sensitivity study. The prior
knowledge II and III follow a triangular distribu-
tion and an arbitrary histogram type of distribu-
tion, respectively. They have the same ranges of
both μ and σ values as the prior knowledge I.
Note that, compared with the prior knowledge I,
the prior knowledge II and III have relatively large
Figure 3. Scatter plot of the equivalent samples for
PDF values allocated close to the true values (i.e.
undrained Young’s modulus.
10MPa and 4 MPa, respectively) of μ and σ that
are indicated by dashed lines in Figure 1a and 1b,
respectively. Therefore, the prior knowledge II and distributions of Eu (or ln(Eu)) can be estimated
III are slightly more informative and confident through conventional statistical analysis. For
than the prior knowledge I, and it is slightly more example, the estimations (i.e. μN* and σN*) of μN
consistent with the true values of μ and σ . The and σN are obtained from conventional mean and
prior knowledge IV is the Gaussian best fit of the standard deviation equations, and they are 2.26
prior knowledge III, and information provided by and 0.54, respectively. Such a procedure described
the prior knowledge III and IV are more or less above is repeatedly performed for each combina-
the same. tion of a set of simulated data (10 sets in total)
In this sensitivity study, 10 sets of simulated and a set of prior knowledge (4 sets in total).
SPT data with 10 SPT N values in each set (i.e. data This leads to 40 sets of μN* and σN* with 10 sets
quantity ns = 10) are generated using Equation 5. for each set of prior knowledge. In addition, for
For example, Figure 2 shows a set of the simulated each set of the simulated SPT data, the Eu values
SPT data. Note that in practice the actual values of are also directly calculated using the regression
soil properties are unknown, and they are estimated model given by Equation 2 in the absence of prior
through prior knowledge and project-specific test knowledge. This leads to another 10 sets of μN*
results. and σN*.
Each set of simulated data is combined with
the four different sets (i.e. prior knowledge I, II,
4.1 Effects on the mean of ln(Eu)
III, and IV) of prior knowledge shown in Figure 1
to generate 30,000 equivalent samples of Eu using Figure 4a shows values of μN* obtained from the
the proposed approach, respectively. For exam- equivalent sample approach using prior knowl-
ple, Figure 3 shows 30,000 equivalent samples of edge I, II, III, and IV by open squares, crosses,
Eu generated using the set of SPT data shown in open triangles, and open circles, respectively. It
Figure 2 and prior knowledge I. Using the 30,000 also includes the μN* values directly estimated
equivalent samples, the statistics and probability from the regression model (i.e. Eq. 2) without
prior knowledge by solid squares. For compari-
son, Figure 4a also shows the true value (i.e. 2.23)
of μN by a dashed line. In general, open squares,
crosses, open triangles, and open circles (i.e. the
results estimated from equivalent samples) plot
more closely to the dashed line than solid squares
(i.e. the results directly estimated from the regres-
sion model without prior knowledge). The μN*
values estimated from the equivalent samples are
generally better estimations of the true value than
those directly estimated from the regression model
without prior knowledge. By incorporating prior
knowledge, the equivalent sample approach signif-
icantly improves the estimation of μN in this study.
Figure 4a also shows that the μN* values obtained
Figure 2. A set of simulated SPT data. using prior knowledge II, III, and IV (see crosses,

341

ISGSR2013.indb 341 10/18/2013 9:41:47 AM

 circles, respectively) are slightly better than those
obtained using prior knowledge I (see open squares),
because the information on σN provided by prior
knowledge II, III, and IV is slightly more consistent
with the true value of σN and more informative than
that provided by prior knowledge I (see Fig. 1).

5 SUMMARY AND CONCLUSIONS

This paper presented a Markov Chain Monte

Carlo Simulation (MCMCS)-based approach for
probabilistic characterization of the undrained
Young’s modulus, Eu, of clay. The proposed
approach integrates the information provided by
the prior knowledge and project-specific SPT data
systematically under a Bayesian framework and
transforms the integrated information into a large
number of equivalent samples of Eu for its probabi-
listic characterization using MCMCS. It effectively
tackles the difficulty in generating meaningful
statistics from the usually limited number of soil
property data obtained during geotechnical site
investigation. Equations were derived for the pro-
posed equivalent sample approach. The proposed
approach provides engineers with flexibility to use
Figure 4. Effects of prior knowledge. different types of realistic prior distributions, such
as uniform prior distribution, triangular prior
distribution, an arbitrary histogram type of prior
open triangles, and open circles, respectively) are distribution, and normal prior distribution. Such
slightly better than those obtained using prior flexibility was illustrated through a sensitivity
knowledge I (see open squares). It is not surprising study on prior knowledge. It was shown that the
to see this because the prior knowledge II, III, and proposed equivalent sample approach combines
IV are slightly more consistent with the true value the information provided by different types of
of μN and more informative than the prior knowl- prior knowledge with project-specific information
edge I, as shown in Figure 1. in a rational manner and improves significantly
probabilistic characterization of soil properties by
incorporating consistent prior knowledge.
4.2 Effects on the standard deviation of ln(Eu)
Figure 4b shows values of σN* obtained from the
equivalent sample approach using prior knowledge ACKNOWLEDGEMENT
I, II, III, and IV by open squares, crosses, open tri-
angles, and open circles, respectively. It also includes The work described in this paper was supported
the σN* values directly estimated from the regression by a grant from the Research Grants Council of
model (i.e. Eq. 2) without prior knowledge by solid the Hong Kong Special Administrative Region,
squares. For comparison, Figure 4b also shows the China (Project No. 9041550 (CityU 110210)),
true value (i.e. 0.39) of σN by a dashed line. Similar a grant from City University of Hong Kong
to μN* values, the σN* values estimated from the (Project No. 7002695), and a grant from National
equivalent samples (see open squares, crosses, open Natural Science Foundation of China (Project No.
triangles, and open circles for prior knowledge I, II, 51208446). The financial supports are gratefully
III, and IV, respectively) are generally better estima- acknowledged.
tions of the true value than those directly estimated
from the regression model without prior knowledge REFERENCES
(see solid squares). By incorporating prior knowl-
edge, the equivalent sample approach significantly Ang, A.H.-S. & Tang, W.H. 2007. Probability Concepts
improves the estimation of σN in this study. In addi- in Engineering: Emphasis on Applications to Civil and
tion, the σN* values obtained using prior knowledge Environmental Engineering. New York: John Wiley
II, III, and IV (see crosses, open triangles, and open and Sons.

342

ISGSR2013.indb 342 10/18/2013 9:41:48 AM

 Beck, J.L. & Au, S.K. 2002. Bayesian updating of Phoon, K.K., Kulhawy, F.H. & Grigoriu, M.D. 1995.
structural models and reliability using Markov chain Reliability-based Design of Foundations for Trans-
Monte Carlo simulation. Journal of Engineering mission Line Structures, Report TR-105000, Electric
Mechanics 128(4): 380–391. Power Research Institute, Palo Alto.
Cao, Z. & Wang, Y. 2013. Bayesian approach for Phoon, K.K. & Kulhawy, F.H. 1999a. Characterization
probabilistic site characterization using cone of geotechnical variability. Canadian Geotechnical
penetration tests. Journal of Geotechnical and Journal 36(4): 612–624.
Geoenvironmental Engineering 139(2): 267–276. Phoon, K.K. & Kulhawy, F.H. 1999b. Evaluation of geo-
Christian, J.T., Ladd, C.C. & Baecher, G.B. 1994. technical property variability. Canadian Geotechnical
Reliability applied to slope stability analysis. Journal Journal 36(4): 625–639.
of Geotechnical Engineering 120(12): 2180–2207. Robert, C. & Casella, G. 2004. Monte Carlo Statistical
El-Ramly, H., Morgenstern, N.R. & Cruden, D.M. 2005. Methods, Springer.
Probabilistic assessment of stability of a cut slope in Wang, Y. 2011. Reliability-based design of spread foun-
residual soil. Geotechnique 55(1): 77–84. dations by Monte Carlo simulations. Geotechnique
Hastings, W.K. 1970. Monte Carlo sampling methods 61(8), 677–685.
using Markov chains and their applications. Wang, Y., Au, S.K. & Kulhawy, F.H. 2011a. Expanded
Biometrika 57: 97–109. reliability-based design approach for drilled shafts.
Honjo, Y., Kikuchi, Y. & Shirato, M. 2010. Development Journal of Geotechnical and Geoenvironmental Engi-
of the design codes grounded on the performance-based neering 137(2): 140–149.
design concept in Japan. Soils and Foundations 50(6): Wang, Y., Cao, Z. & Au, S.K. 2011b. Practical reliability
983–1000. analysis of slope stability by advanced Monte Carlo
Kulhawy, F.H. & Mayne, P.W. 1990. Manual on estimating simulations in spreadsheet. Canadian Geotechnical
soil properties for foundation design. Report EL 6800, Journal 48(1): 162–172.
Electric Power Research Institute, Palo Alto, 306p. Wang, Y. 2012. Uncertain parameter sensitivity in Monte
Mayne, P.W., Christopher, B.R. & DeJong, J. 2002. Carlo simulation by sample reassembling. Computers
Subsurface investigations—Geotechnical site and Geotechnics 46: 39–47.
characterization. No. FHWA NHI-01-031, Federal Wang, Y. & Cao, Z. 2013. Probabilistic characterization
Highway Administration, U.S. Department of of Young’s modulus of soil using equivalent samples.
Transportation, Washington D.C. Engineering Geology 159: 106–118.

343

ISGSR2013.indb 343 10/18/2013 9:41:48 AM

This page intentionally left blank
 Geotechnical Safety and Risk IV – Zhang et al. (eds)
© 2014 Taylor & Francis Group, London, ISBN 978-1-138-00163-3

Site investigation approaches for the proposed Ukhuu Khudag

to Gashuun Sukhait Railway, South Gobi desert, Mongolia

A.D. Mackay
Nishimatsu Construction Company Limited
Formerly SMEC Asia Ltd.

N.R. Wightman
Sinclair Knight Merz (HK) Limited
Formerly SMEC Asia Ltd.

ABSTRACT: Due to the current resource demand from the People’s Republic of China and the
availability of these resources in the Gobi desert, Mongolia, there is a need to construct a railway to
provide an efficient haulage system from the Gobi desert mines to the People’s Republic of China. The
proposed railway currently runs for about 225 kilometres within Mongolia from Ukhuu Khudag to the
People’s Republic of China border crossing facilities at Gashun Sukhait. As the Oyo Tolgoi open cast
mine will also come into operation in the near future provision for haulage of these mined resources needs
to be made; the alignment passes in close proximity to this location accordingly. As part of the feasibil-
ity design a preliminary site investigation was carried out, including a limited desk study, a walkover
survey, a topographic survey and a basic ground investigation, including boreholes and trial pit explo-
ration stations. The ground investigation stations were located according to geotechnical construction
risks assessed from the desk study, topographic survey and walkover survey information; at this pre-
liminary stage these included stream and river crossings, areas with considerable volumes of cut and fill,
material suitability for ballast and embankment construction and areas with near surface groundwater.
The ground conditions along the alignment included superficial deposits; alluvium, cemented alluvium,
talluvium and aeolian (wind-blown) deposits; and solid geology; sedimentary (coal measures, shale and
sandstone) and igneous (tuff, basalt, andesite and granite) rock. This paper outlines the methods adopted
for carrying out the site investigation, which included considerations for the past environmental and geo-
logical history when assessing the physical characteristics relevant to construction.

1 INTRODUCTION the geological conditions in the Gobi desert in the

vicinity of the proposed alignment at that time.
The Ukhuu Khudang (UKh) to Gashuun Sukhait The approach to the SI considering the implica-
(GS) Railway will run about 225 kilometres across tions of the environmental and geological history
the Gobi desert, Mongolia, from UKh to the is provided.
north-west to the GS border crossing facilities to
the south east. To ensure efficient operation strin-
2 SETTING
gent geometrical limitations (gradient and curva-
ture) are required for the adopted rail alignment.
2.1 Location
Given the topographic variability of all available
rail corridors across the Gobi desert considerable The existing railway alignment is located in the
volumes of cut and fill will be required. A prelimi- south Gobi, Omnigovi province, Mongolia. Refer
nary Site Investigation (SI), including a desk study, to Figure 1 for the general location in Mongolia.
topographic survey and walkover survey data was The proposed alignment runs from the
carried out to address the construction risks. At north-west, Chainage (Ch.) 0 from UKh, to the
the preliminary stage the risks included the stream south-east, up to Ch. 225 at GS. Refer to Figure 2
and river crossings, which were generally located for the proposed alignment across the Gobi desert.
at topographic low points requiring considerable The Gobi desert terrain typically comprises sandy
volumes of embankment fill, material suitability and rocky soils often with sparse vegetation and a
for ballast, areas with near surface groundwater thin overlay of wind-blown deposits, comprising
and deep cuts. This paper provides a background finer grained soil and dust particles. The proposed
to the railway requirements and an overview to rail alignment descends from an altitude of about

345

ISGSR2013.indb 345 10/18/2013 9:41:48 AM

 Mongolian to PRC rolling stock. Refer to Figure 2
for the locations. The maximum rail gradient for
loaded trains is generally 0.6%, dependent on the
type of locomotive used. At the Balloon Loop area
the and marshaling yards the gradient is restricted
to zero. The minimum horizontal curvature is
10km radius respectively.
To accommodate the geometrical limitations of
the railway alignment and the abrupt changes in
the topographic profile along all proposed align-
ments variable embankments and cutting heights
were required ranging to a maximum of about
Figure 1. UG-GS Rail location (Google Earth, 2011). 17 m vertically. Other important construction con-
siderations include:
Drainage—surface drainage running into
embankments require culvert installation to allow
unimpeded flow beneath the alignment. Flow run-
ning into cuttings required suitable surface drain-
age to redirect the flow. Due to the size of the
Undaii river crossing, a bridge with piled founda-
tions was being considered.
Structural—foundations for telecom towers and
towers; major culverts and accessories for the sid-
ings and marshaling yards and accessory buildings.
Suitable concrete mixes and reinforcement needed
consideration for the aggressive soil conditions.
Environmental—animal crossings, including
major underpasses or overpasses to allow unim-
peded animal migration across the Gobi desert.
Structures to accommodate oil change sump pits,
latrines and groundwater abstractions were also
required.
Traffic—maintenance road running alongside
the alignment for its entire length. Traffic crossings
are required at regular intervals.
Earthworks—ballast (250 mm thickness) is
required for the entire alignment. The embank-
ment construction required a sub-base layer
Figure 2. Rail alignment, Gobi desert (Google Earth, (250 mm thick) and embankment layer (variable).
2011).
As the marshaling yards are level a significant lev-
eled terrace platforms were required to place the
1500 m above sea level (asl) at UKh decreasing Yards. Passing sidings were proposed at approxi-
to 900 m asl towards the southernmost portion. mately every 2 kms.
Prominent features along the alignment include
Refer to Figure 3 for the major railway
the Galbin Gobi drainage basin, the major Undaii
components.
river crossing, traversing the alignment at its lowest
level, and the Tsagaan Hud mountain range cross-
ing the southernmost portion of the alignment 2.3 Geological overview
(refer to Fig. 2).
The South Gobi region has undergone continental
accretion and basin range crustal extension. The
2.2 Rail considerations
area is dominated by east to west trending mountain
The commencement of the alignment, at UKh, is ranges cutting the alignment, typically corresponding
located near the Tavan Tolgoi coking coal open with dyke intrusions and faults. This suggests crustal
excavation mine. The GS border crossing facilities movement from both the north and south directions
at the end of the alignment is located about 10km in the Gobi area. A major fault, referred to as the
from the People’s Republic of China (PRC) bor- Navin Sukhait fault, bounds the intrusive rock and
der and will facilitate transfer of the coal from the sedimentary deposits; basalts and coal measures,

346

ISGSR2013.indb 346 10/18/2013 9:41:48 AM

 Figure 4. Mongolia geological map.
Figure 3. Rail alignment requirements.

conglomerates and sandstones, towards the northern

portion of the alignment. A granitic batholith
intrusion, which surfaces in areas with the granitic
tor terrain, underlies the entire alignment at depth.

3 GROUND CONDITIONS

3.1 Approach to the site investigation

An initial Ground Investigation (GI) was car-
ried out during 2009 and comprised boreholes
(BH) positioned at about every 500 m along the
proposed alignment. This prescriptive approach
typically follows the traditional Russian SNiP code
standards (Fookes et al. 2005), which is still often
used to specify GIs in Mongolia. Generally as the
geological descriptions used in Russian codes can-
not be easily used to estimate geotechnical param-
eters for design input the data obtained was limited
value for the engineering design (Mackay 2010). Figure 5. SI using the “Total Engineering Geology”
To allow rapid feedback for the design prepara- approach (Baynes et al. 2007).
tion the “Total Engineering Geology” approach
(Fookes 1997) to the SI was adopted. The premise
of this approach is “that at any location the geology undulating plains, locally exposed rock, typically
is the sum of its history”. The aim was to maximize associated with abrupt changes in gradient, and
the amount of useful information that could be granitic tor landscapes. Refer to Figure 6 for the
obtained during the initial literature search and walk- terrain evaluation and Plates 1 to 3 for the terrain
over survey to prepare a more efficient and focused encountered.
GI. Reference conditions, which group geological
materials into similar engineering characteristics
3.3 Ground investigation
interpreted from SI, were estimated and refined as
more information became available. This approach is Following the desk study and walkover survey
compared to the traditional SI approach in Figure 4 the GI was estimated using a risk based approach
(Baynes et al. 2005, Mackay et al. 2008). from the effects of adverse ground and groundwa-
ter conditions on the construction as outlined in
Table 1.
3.2 Terrain evaluation
As indicated from the risk assessments a sim-
The terrain encountered along the alignment plified, preliminary GI was carried out including
varies from undulating plains, soft hillocks, flat exploratory stations, pits, BHs and rock samples.

347

ISGSR2013.indb 347 10/18/2013 9:41:50 AM

 All risks referenced in Tables 1 and 2 involved
further detailed walkover surveys to complement
the existing walkover survey and the GI. Due to
the nature of the ground, including cemented soils,
which were highly sensitive to disturbance and
addition of water, true undisturbed sampling was
difficult to achieve.

3.4 Findings
The major geological units, including areas of
exposed solid ground beneath the superficial
deposits, generally corresponded with the terrain
evaluation as follows:
Aeolian deposits—generally covering the entire
area with localized increases in thickness in zones
of “small hillocks” and in areas of flatter land
Talluvium—typically standing at shallow gradi-
ents beneath the rock outcrops
Alluvium and lake deposits—extensive areas of
flat ground typically in the vicinity of major rivers
Sedimentary rock (sandstone, shale and coal
measures)—typically undulating terrain
Igneous rock (dyke intrusions; basalt and
andestite)—associated with abrupt changes in
topography and surface exposure
Igneous rock (batholitic intrusions; granite)—
“tor” landscape and surface exposure
The general findings of the SI, are presented in
Table 2 and Figure 7.

4 INTERPRETATION

A major concern was the adequacy of the ground

bearing capacity to support the embankment, pat-
icularly placement on the recent deposits (recent
to 66 Ma) identified in Table 2. The total engineer-
ing geology approach was used to assess the effects
from the recent geological and environmental his-
tory on the engineering properties of the ground
conditions.

4.1 Weathering
The relative effects of the extent of weathering in
Figure 6. Terrain evaluation (a) and Plates 1–3 (b–d). different environments are presented in Figure 8
(Fookes 1997).
As shown weathering in desert environments,
The BHs were advanced using undisturbed such as the Gobi desert, is minimal. Relatively com-
sampling and Standard Penetration Testing tech- petent ground with an increased bearing capacity
niques and bulk samples and in-situ testing within can therefore be expected near surface. As presented
the pits. Subsequent laboratory testing included in Plate 5 the most recent deposits, referenced Ap-La
particle size distribution, atterberg limit and Q111-N (Table 2), located further to the north-west,
aggregate testing. The GI specification followed showed the weaker superficial deposits to have a
international British and Eurocode standards (BS minimal thickness and be present over more compe-
1999) as required. tent rock, as shown by the difficulty in excavation.

348

ISGSR2013.indb 348 10/18/2013 9:41:52 AM

 Table 1. Ground/groundwater risk based GI evaluation.

Risks Effects on SI method Mitigation

Surface/ Embankment/ Exploratory stations Suitable culvert embankment/cutting protection

sub-surface cutting stability (pits and boreholes) geotextile/gravel layer suitable area
water bearing capacity in-situ permeability tests identification/limit fines content in earthworks
earthworks piezometric monitoring
Talluvium Loose Exploratory stations Alignment adjustment slope support
cutting stability (pits/boreholes, BHs)
Lake deposit/ Bearing capacity Exploratory stations Remove if underlying rock present leave in place if
crust (pits/BHs) underlying soil present
avoid diversion of surface/subsurface water flow
Tectonic Cut slope Exploratory stations Avoid fines placement in embankment
embankment particle size distribution add suitale factor for slope/bearing capacity
foundations (laboratory tests) design

Table 2. Ground conditions encountered in the SI.

Geological period Ground condition

Recent (66 Ma to now) Alluvium/lake deposits (gravel, clay and sand (key Ap-La Q111-N); talluvium;
alluvium/lake deposits (gravel, clay and sand (key L-aN2) and alluvium/lake
deposits (cemented granular sand and clay). (key P3).
Cretaceous (66–145 Ma) Lake-alluvial deposits (key LaK2) and coal measures (“aleurolite”—siltstone
and coal measures)
Jurassic (145–200 Ma) Intrusive rocks (basalt, andesite and granite)
Carboniferous (300–350 Ma) Shale with basalt and andesite intrusions
Devonian (350–417 Ma) Sandstone with tuff and basalt and andesite intrusions

water flow than present, a variety of resistant gravel

fragments, transported by past water flow and
derived from source areas from a greater distance
than the nearby mountain ranges, and the effects
of weathering, which indicated effects of a different
past weathering environment which may have been
tropical (see Fig. 8). The greater extent of weath-
ering, shown by the relative ease of excavation in
Plate 6, indicates the relative extent of weathering.
Due to the potential reduced bearing capacity at
a greater depth consideration for excavation and
replacement by selected backfill was needed.

4.2 Cemented superficial deposits

In general the majority of the Gobi desert had a
highly competent upper crust, termed “duricrust”,
Figure 7. General locations of the ground conditions. which is typical of many desert environments and
has a high bearing capacity (refer to Fig. 9).
Duricrust is formed as a result of cement-
In contrast the older “recent” deposits, referenced ing minerals coming out of solution during the
P3 (Table 2), located further to the south west in the evaporation of the near surface groundwater flow.
vicinity of the Galbin Gobi drainage basin towards Notwithstanding due to its recent precipitation the
the south east indicated influence from different cementing agents are highly prone to re-solution
environmental effects including the extent of the and are easily dissolved in the presence of water.
drainage basin, which accommodated a greater Furthermore weaker material, which remained

349

ISGSR2013.indb 349 10/18/2013 9:41:54 AM

 Figure 10. Plate 7: Cemented soil with overlying
Aeolian deposits.

5 CONCLUSIONS

The desk study and walkover identified risks,

physical properties and engineering implications
at an early stage in the potential development of
the railway alignment by interpreting the physi-
cal characteristics of the ground as a sum of its
history. In particular the technique assisted in esti-
mating the bearing capacity to place embankments
across superficial deposits.

ACKNOWLEDGEMENTS

The authors wish to thank Ronald Chan in

assistance in preparing this paper. The opinions
Figure 8. Weathering effects in different environments expressed in this paper are those solely of the
(a) and Plates 5 and 6 (b & c). authors and not of any other party.

REFERENCES

Baynes, F.J., Fookes, P.G. & Kennedy, J.F. 2005. The total
engineering geology approach applied to railways in
the Pilbara, Western Australia. Bulletin of Engineering
Geology and the Environment 64: 67–94.
British Standards (BS), 1999. Code of practice for site
investigations. London: British Standards Institu-
tion (BSI).
Fookes, P.G. 1997. The 1st glossop lecture, geology
for engineers: the geological model, prediction and
performance. Quaterly Journal of Engineering Geology
30(4): 293–424.
Fookes, P.G., Hearn, G.J. & Tollin, B. 2005. Aligning oil
Figure 9. Geotechnical properties. and gas pipelines through the Makharov Mountains,
Sakhalin Islands, Russia. In M. Sweeney (ed.), Int.
Conf. on Terrain and Geo-hazards Challenges Facing
On-shore oil and gas pipelines, London, 2–4 June
uncemented was typically present at depth beneath 2004.
the duri-crust. Refer to Plate 7 on Figure 10. Mackay, A.D. 2010. The role of the civil engineer in the
The thickness of the duricrust and its vicinity to changing demand for space in HK. In HKIE Civil
ephemeral stream course was an important consid- Division Conference 2010: infrastructure solutions for
eration for estimating the bearing capacity. tomorrow, Hong Kong, 12–14 April 2010.

350

ISGSR2013.indb 350 10/18/2013 9:41:54 AM

 Geotechnical Safety and Risk IV – Zhang et al. (eds)
© 2014 Taylor & Francis Group, London, ISBN 978-1-138-00163-3

Bayesian model updating of a tunnel in soft soil with settlement

measurements

I. Papaioannou, W. Betz & D. Straub

Engineering Risk Analysis Group, Technische Universität München, Munich, Germany

ABSTRACT: Monitoring is an essential element of modern tunneling construction. The most common
monitoring method is measuring displacements, for example convergence of the tunnel opening or surface
settlements. Measurement outcomes can be used to update the knowledge on material properties of the soil
or other parameters that enter numerical models of the structural behavior of the tunnel. In probability
theory, this process can be formalized in the concept of Bayesian updating. In this paper, we apply the
Bayesian concept to update the numerical model of a tunnel in soft soil conditional on settlement meas-
urements. The tunnel is constructed by means of the conventional tunneling method and modeled with 2D
finite elements applying the stress reduction method. We assume that settlement measurements are taken
at full excavation and utilize the measurements to update the material properties of the soil as well as the
the relaxation factor of the stress reduction method. Updating is performed by means of BUS, a recently
proposed method for Bayesian updating of mechanical models with structural reliability methods.

1 INTRODUCTION sampling is used to sample directly from the pos-

terior distribution, thus bypassing the solution of
In tunneling design, engineers establish numeri- the aforementioned integral (Gelman 2004). An
cal models of the tunnel excavation and conduct alternative approach is based on interpreting the
structural analyses to predict the stresses and updating problem as a structural reliability problem
deformations for the considered designs. However, (Straub and Papaioannou 2013). This approach,
there is significant uncertainty in the choice of the termed BUS, applies methods originally developed
model parameters. Uncertainties may be related to for structural reliability analysis to obtain samples
the inherent spatial variability of the mechanical from the posterior distribution.
properties of the soil but also to the application of In this paper, we apply BUS to learn the model
dimensionally reduced models to represent complex parameters of a tunnel in soft soil using settle-
phenomena. A proper assessment of the safety and ment measurements. The tunnel is constructed by
serviceability of the structural design involves the the conventional tunneling method. We model the
modeling of the uncertainties by use of probabilis- tunnel in 2D using nonlinear plain-strain finite ele-
tic models and the evaluation of the structural reli- ments and the 3D arching effect is approximated
ability against the respective design requirements. by application of the stress reduction method.
During the tunnel construction process, meas- Using assumed settlement measurements at full
urements of physical quantities such as defor- excavation, we update the material properties of
mations and stresses are typically conducted. the soil as well as the relaxation factor of the stress
Measurements can be used to compare predictions reduction method.
of the numerical model with the actual structural
behavior, to verify the reliability of the struc-
2 MODEL DESCRIPTION
tural design as well as to update the probabilistic
description of the parameters of the numerical
2.1 Mechanical model
model. The latter is formalized in the concept of
Bayesian updating. Thereby, a prior probabilistic A conventional driven tunnel with a horse-shoe
model is updated with new data and information shaped profile is considered in this study (see
to a posterior probabilistic model. Fig. 1). The problem is modeled in the SOFiSTiK
Bayesian updating requires the solution of a Finite Element (FE) software package (SOFiSTiK
potentially high-dimensional integral to obtain the AG 2012), using 2D plain strain finite elements.
posterior distribution of the model parameters. The numerical model has a width of 80 m and a
Commonly, Markov Chain Monte Carlo (MCMC) height of 26 m. The FE mesh is shown in Figure 2.

351

ISGSR2013.indb 351 10/18/2013 9:41:57 AM

 friction angle: 35°, cohesion: 200 kPa. Due to the
much larger stiffness of the limestone compared to
the stiffness of the overlaying materials, only 3.8 m
of this layer are modeled.
The height of the tunnel above the limestone
layer is 6.2 m. Consequently, the tunnel is located
in a depth of 16 m below the ground surface. At the
intersection of the second and the third layer, the
Figure 1. Ground layers considered in the model. tunnel has a width of 9.16 m. In the vicinity of
the tunnel the Keuper marl is reinforced with nails.
This is modeled by increasing the cohesion in the
affected region (see Fig. 1) by 25 kPa. Moreover,
the tunnel is located above the groundwater level.
The shotcrete lining is modeled using linear beam
elements with a normal stiffness of 10.5 GN and a
flexural rigidity of 26.78 MNm2.

2.2 Prior probabilistic model

Figure 2. Finite element mesh.
The cover layer and the limestone layer are con-
sidered as deterministic in the analysis. Since the
In this study, we are interested in surface settle- cover layer is a man-made fill, we assume that its
ments over the tunnel center line (point A in Fig. 1). soil properties are well-known, and the associated
The excavation process is modeled by application uncertainties are small compared to the uncertain-
of the stress reduction method (Panet and Guenot ties in the material description of the Keuper marl
1982). In this method, a prescribed fraction β of layer and hence can be neglected. The limestone
the initial stress is left inside the tunnel as a support layer is also modeled as deterministic, because, due
pressure to approximately account for the three- to its large stiffness, the contribution of this layer
dimensional arching effect. This support pressure to the surface settlements is negligible. The prob-
is then removed after installation of the lining. The ability distributions describing the uncertainties in
parameter β is termed relaxation factor. the material parameters of the Keuper marl layer
The model consists of three different ground are listed in Table 1. We assume that the stiffness
layers; the layers are illustrated in Figure 1. The modulus for primary loading E50 equals the oedo-
cover layer is a man-made fill and has a depth metric stiffness modulus Eoed. We also consider a
of 5.4 m. Heavily weathered soft rock known as correlation of 0.7 between Eoed and the elastic mod-
Keuper marl forms the second layer. The thickness ulus Eur. The friction angle and the cohesion are
of this layer is 16.8 m. We adopt a hardening plas- assumed to have a negative correlation of −0.5.
ticity soil model (SOFiSTiK AG 2012) to describe In conventionally driven tunnels, there is usually
the material behavior of the first two layers. This a large uncertainty in the choice of the relaxation
material model allows for a realistic description of factor β ∈ [0, 1] of the stress reduction method
the stiffness and hardening behavior of soft soil (Möller 2006). In this study β is modeled as a beta-
in settlement analysis (Möller 2006). The mate- distributed random variable (see Table 1).
rial properties of the cover layer are as follows:
elastic modulus for unloading-reloading: 30 MPa,
Poisson’s ratio: 0.2, specific weight: 20 kN/m3, Table 1. Prior distribution of the parameters of the
friction angle: 25°, cohesion: 10 kPa, oedometric Keuper marl layer.
stiffness modulus: 10 MPa, stiffness modulus for
primary loading: 10 MPa. The exponent in the Parameter Distribution Mean CV
hardening law is selected as 0.5 for the first and the
second layer. The angle of dilatancy is assumed as Relaxation factor β Beta (0.0, 1.0) 0.5 10%
zero, corresponding to a non-associated flow rule. Elast. mod. Eur [MPa] Lognormal 80.0 32%
Oedometr. mod. Lognormal 30.0 32%
The soil parameters of the Keuper marl layer are
Eoed [MPa]
modeled as random and their prior probabilis-
Poisson’s ratio v Beta (0.0, 0.5) 0.2 15%
tic description is discussed in Section 2.2. Strong
Friction angle ϕ [°] Beta (0.0, 45.0) 20.0 15%
limestone constitutes the bottom layer. The Mohr-
Cohesion c [kPa] Lognormal 25.0 30%
Coulomb law is applied for this layer. The mate-
Specific weight γ Lognormal 24.0 5%
rial properties are: Young’s modulus: 575 MPa, [kN = m3]
Poisson’s ratio: 0.2, specific weight: 23 kN/m3,

352

ISGSR2013.indb 352 10/18/2013 9:41:57 AM

 A reliability assessment of the tunnel was pre- correlated Gaussian fields. Due to the form of
sented in Ranjan et al. (2013). Therein, a two-step their cross-correlation function, the latter fields
approach is adopted. In a first step, the reliability can be transformed to independent Gaussian
analysis was performed applying the First Order fields by performing the Cholesky decomposition
Reliability Method (FORM) [e.g. see (Der of the 2 × 2 correlation matrix, whose off-diagonal
Kiureghian 2005)]. As a byproduct of the FORM, terms express the correlation of the two fields at
the influence coefficients provide information on the the same location. The underlying independent
sensitivity of the reliability in terms of the random Gaussian fields are discretized by application of the
variables. This information was used to identify the Karhunen-Loéve expansion (Ghanem and Spanos
random variables with the highest influence that, in 1991). That is, each field is represented as a trun-
a second step, are modeled as random fields. cated series of products of the eigenfunctions of its
Figure 3 depicts in a pie graph the squared autocorrelation function and independent random
influence coefficients obtained by the FORM. It is variables. Each random field is dicretized with 100
observed that the variable with the largest influence random variables. Therefore the total number of
is the oedometric stiffness modulus Eoed followed random variables in the problem is 205.
by the relaxation factor β. Based on this result,
we account for the inherent spatial variability of
Eoed. Since Eoed is strongly correlated with the elas- 3 BAYESIAN UPDATING WITH
tic modulus Eur, the spatial variability of the lat- STRUCTURAL RELIABILITY
ter parameter is also modeled. This is achieved by METHODS (BUS)
modeling the two parameters as cross-correlated
homogeneous random fields. The joint distribution Let X denote the n-dimensional random vector rep-
of the two fields at each pair of locations is mod- resenting the uncertain model parameters discussed
eled by the Nataf distribution (Der Kiureghian in Section 2.2. Also, let f(x) be the prior joint Prob-
and Liu 1986) with lognormal marginals according ability Density Function (PDF) of X. Assume that
to Table 1. The spatial variability depends only on a measurement uA,m of the surface settlement (point
the separation in horizontal and vertical direction A in Fig. 1) is made at full excavation. The meas-
between two locations, Δx and Δy. The following urement is subjected to an additive error ∈ which
exponential autocorrelation coefficient function is is described by a normal PDF f∈ with zero mean
chosen for both random fields: and standard deviation σ∈. The measurement infor-
mation can be described by the event Z = {uA,m −
⎛ Δx Δ y ⎞ uA(x) = ∈}, where uA(x) is the surface settlement
ρ ( x, Δy ) exp ⎜ − ⎟ (1) evaluated by the FE program for a realization x of
⎝ lx ly ⎠
the random vector X. The corresponding likelihood
function can be expressed as follows:
where lx = 20 m and ly = 5 m denote the correlation
lengths in horizontal and vertical direction, respec- L f∈[u
[ uA,m uA ( x )]
tively. The cross-correlation coefficient function is:
1 ⎛ (uA,m uA )2 ⎞
Δ ) ρc ρ( Δx, y ) (2) = p⎜ −
exp ⎟ (3)
cross ( σ∈ 2σ ∈2
2 ⎝ ⎠
where ρc = 0.7 denotes the correlation of Eoed and
Eur at the same location. The posterior joint PDF of X conditional on the
Since the random fields have the Nataf dis- measurement event Z can be obtained by applica-
tribution, they can be expressed as functions of tion of Bayes’ rule:

L f( )
f Z) = (4)
∫R n L f ( )dx

The evaluation of the n-fold integral in the denom-

inator of Equation (4) is computationally demand-
ing; this has motivated the application of MCMC
algorithms for sampling directly from the posterior
f(x|Z) (Gilks et al. 1998, Gelman 2004). Here, we
apply an alternative approach, termed BUS, which
uses methods originally developed for structural
Figure 3. Squared influence coefficients obtained by reliability analysis to obtain samples from the pos-
FORM. terior distribution (Straub and Papaioannou 2013).

353

ISGSR2013.indb 353 10/18/2013 9:41:58 AM

 The method is based on the algorithm for reliability for the random field representation of the soil
updating developed in Straub (2011) and applied in properties.
Papaioannou and Straub (2012) to the reliability
updating of geotechnical structures. 3.1 SubS-based BUS
The BUS approach introduces the following
limit state function: The SubS method, originally developed in Au &
Beck (2001), evaluates the probability Pr(Ze) of the
h( , u0 ) u0 1
[cL( )] (5) event Ze = {H(u) ≤ 0} as a product of larger condi-
tional probabilities. This is achieved by expressing
where u0 is the outcome of a standard normal the event Ze as the intersection of M intermediate
random variable U0, Φ−1(.) is the inverse of the events that are nested, i.e. it holds Z1 ⊃ Z2 ⊃ … ⊃
standard normal cumulative distribution function ZM = Ze. The events {Zi, i = 1, …, M} are defined
and c is a positive constant chosen to ensure that as Zi = {H(u) ≤ bi}, where b1 > b2 > … > bM = 0. The
cL(x) ≤ 1. It is shown in Straub and Papaioan- probability Pr(Ze) is then expressed as
nou (2013) that the posterior PDF f(x|Z) is pro-
M
portional to the prior PDF f(x) conditional on
the event Ze = {h(x, u0) ≤ 0}. Hence, solving the
Pr( e) ∏ Pr(
P( i | Zi − 1 ) (6)
i=1
updating problem becomes equivalent to solving
the structural reliability problem of estimating the where Z0 denotes the certain event and Pr(Zi|Zi−1)
probability Pr(Ze). is the probability of the event Zi conditional on the
It should be noted that the constant c has con- occurrence of the event Zi−1. The values bi can be
siderable influence on the efficiency of the BUS chosen adaptively, such that the estimates of the
approach, since its value is directly proportional conditional probabilities correspond to a given
to the probability Pr(Ze) (Straub and Papaioan- value p0.
nou 2013). A large value of Pr(Ze) is beneficial The probability Pr(Z1|Z0) = Pr(Z1) is computed
for most structural reliability methods. Therefore, by applying crude Monte Carlo simulation. To
c should be chosen a large as possible, while still estimate the conditional probabilities {Pr(Zi|Zi−1),
ensuring that cL(x) ≤ 1. For the likelihood func- j = 2, …, M}, we need to obtain samples of U
tion of Equation (3), the optimal choice is c = [max conditional on the occurrence of the events {Zi−1,
f∈(∈)]−1 = σ ∈ 2π . j = 2, …, M}. Assume that at each subset level i, J
For most reliability methods, it is convenient to samples {u(j), j = 1, …, J} of U conditional on Zi−1
transform the problem from the original random are available. The threshold bi is set as the (1 − p0)-
variable space to a space of independent standard percentile of the samples; the samples u(j) for which
normal random variables. Since the distribution H(u(j)) ≤ bi are then used as seeds for the simulation
of the random vector X is described by the Nataf of samples conditional on Zi by application of an
model, such a transformation [U1, …, Un] = T(X) is MCMC algorithm (Papaioannou et al. 2013).
straightforward (Der Kiureghian and Liu 1986). For Bayesian updating, we are interested in
The limit-state function h(x, u0) can be expressed in obtaining samples conditional on Ze. Therefore, we
the transformed space as H(u) = h[T−1(u1, …, un); u0], add one final step, which is to obtain K such sam-
where T−1 denotes the inverse transformation and u ples trough MCMC starting from the samples gen-
∈ Rn+1 is the outcome of U = [U0, U1, …, Un]. erated at the last subset level M for which H(u(j)) ≤ 0.
The formulation of the updating problem in These samples are then transformed to the original
terms of the limit-state function of Equation (5) space as { ( ) ((uu1( ) , , un( ) ),
) k , , K } in
allows for the application of a variety of structural order to obtain samples from the posterior dis-
reliability methods for estimation of the posterior tribution f(x|Z). In this study, the parameters of
PDF. Application of crude Monte Carlo method the SubS algorithm for Bayesian updating are set
will lead to a rejection-acceptance scheme, with as follows: p0 = 0.1; number of samples per level
Pr(Ze) being the acceptance probability. However, J = 1000; number of target samples K = 1500.
this approach becomes very inefficient for small
Pr(Ze), which correspond to cases where the pos-
terior distribution differs considerably from the 4 RESULTS AND DISCUSSION
prior. In the following section, we discuss the
application of BUS in conjunction with Subset We consider a measurement outcome of
Simulation (SubS), which is an adaptive Monte uA,m = 20 mm. The prior mean of uA(X) is 10.5 mm,
Carlo method for estimating small probabilities. which indicates that the prior model underestimates
The SubS is especially efficient in high dimensional the measured surface settlement. The updating was
problems, as is the case in the present application performed for two different values of the standard
where a large number of random variables is used deviation σ∈ of the measurement error: 1 mm and

354

ISGSR2013.indb 354 10/18/2013 9:42:00 AM

 2 mm. Figures 4 and 5 show the posterior sample Table 2. Statistics of the posterior distribution of the
means of the oedometric stiffness modulus Eoed random variables of the Keuper marl layer.
and the elastic stiffness modulus Eur, respectively.
The samples statistics of the remaining material σ∈ = 1 mm σ∈ = 2 mm
parameters of the Keuper marl layer are shown in Parameter Mean CV Mean CV
Table 2.
Looking at the results for σ∈ = 2 mm, one can Relaxation factor β 0.42 8.3% 0.44 8.4%
observe that the posterior mean of Eoed at the ele- Poisson’s ratio v 0.19 14.5% 0.19 15%
ments within and around the tunnel is smaller than Friction angle ϕ [°] 18.5 17.2% 18.9 17.2%
the prior. Note that Eoed is the parameter with the Cohesion c [kPa] 29.04 30.5% 26.3 29.3%
highest influence on the tunnel’s reliability (see Specific weight γ 24.2 5% 24.4 5.2%
Fig. 3). Away from the tunnel, the posterior mean [kN/m3]
of Eoed increases and the value of its prior mean is
reached at the upper left and right corners of the
computational domain. Similar results are obtained
for Eud, which is highly correlated with Eoed. Due to the symmetry of the problem, it is expected that the
spatial distribution of the posterior means of both
Eoed and Eud be symmetric about the vertical cen-
tral axis. This result is evident in the area close to
the tunnel, however away from the tunnel one can
observe what seems like local random fluctuations
from the expected result. This effect is attributed
to sampling error and is related to the fact that in
the areas away from the tunnel the influence of the
values of Eoed and Eud on the surface settlements is
minor. That is, a large number of combinations of
material values in these areas can justify the meas-
urement outcome, which requires a large number
of samples for the SubS algorithm to account for
all the possible combinations.
The effect of the measurement is also evident in
the posterior mean of the relaxation factor β of
the stress reduction method, which is decreased
compared to its prior. Its posterior Coefficient of
Variation (CV) is also smaller than its prior, which
Figure 4. Posterior mean of the oedometric stiffness reflects the impact of the measurement on the
modulus Eoed of the Keuper marl layer. variable. Moreover, the initial stress for the stress
reduction is computed based on the elements cor-
responding to the tunnel whose posterior means
are much lower than their priors. This reveals the
influence of the 2D modeling of the arching effect
by the stress reduction method, which is consistent
with the fact that most settlements will take place
in the excavation phase, i.e. before the installation
of the lining.
The friction angle φ is slightly influenced by the
measurement; its posterior mean is decreased com-
pared to the prior however its coefficient of vari-
ation is somewhat increased which indicates that
the impact of the measurement on φ is small. The
mean of the cohesion c is increased reflecting its
negative correlation with the friction angle, while
the influence of the measurement on the rest of the
parameters is negligible.
The results for the case where the standard
deviation of the measurement is decreased, i.e.
Figure 5. Posterior mean of the elastic stiffness modu- when σ∈ = 1 mm, confirm the high impact of the
lus Eur of the Keuper marl layer. measurement outcome on the stiffness variables

355

ISGSR2013.indb 355 10/18/2013 9:42:01 AM

 Figure 6. Posterior PDF of the relaxation factor β for σ∈ = 1 mm.

Eoed, Eud and the relaxation factor β. In this case, that models the 3D arching effect of the stress
which implies higher information content of the distribution.
measurement, the posterior mean of β is further
decreased and the weak zone around the tunnel
with low values of Eoed and Eud is increased. REFERENCES
Figure 6 demonstrates the influence of the prior
knowledge on the relaxation factor β. Therein, the Au, S.K. & Beck, J.L. 2001. Estimation of small failure
prior and posterior PDFs of β are plotted for two probabilities in high dimensions by subset simulation.
different assumed prior coefficients of variation Probabilist. Eng. Mech. 16(4): 262–277.
(10% and 20%) and for σ∈ = 1 mm. It is observed Der Kiureghian, A. 2005. First- and second-order reli-
ability methods. In E. Nikolaidis, D.M. Ghiocel,
that as the prior knowledge on beta decreases, i.e. and S. Singhal (Eds.), Engineering Design Reliability
as its prior coefficient of variation increases, the Handbook. Boca Raton, FL: CRC Press.
influence of the measurement becomes higher. Der Kiureghian, A. & Liu, P.L. 1986. Structural reliabil-
Comparing the posterior PDFs for the two cases, ity under incomplete probability information. J. Eng.
we see that the same measurement information Mech.-ASCE 112(1): 85–104.
leads to much lower values of β when a larger Gelman, A. (2004). Bayesian data analysis. Boca Raton,
prior coefficient of variation is assumed. This FL: Chapman & Hall/CRC.
result further highlights the influence of the 2D Ghanem, R. & Spanos, P. 1991. Stochastic Finite
modeling of the arching effect on the surface set- Elements—A Spectral Approach. Berlin: Springer.
Gilks, W.R., S. Richarson & D.J. Spiegelhalter (Eds.)
tlements. Moreover, it shows how the confidence 1998. Markov chain Monte Carlo in practice. Boca
on the prior assumption can influence the updat- Raton, FL: Chapman & Hall/CRC.
ing results that may provide a basis for further risk Möler, S. 2006. Tunnel induced settlements and structural
and reliability assessment. forces in linings. Ph.D. thesis, Institut für Geotechnik,
Universität Stuttgart.
Panet, M. & Guenot, A. 1982. Analysis of convergence
5 CONCLUSION behind the face of a tunnel. In Proc. International
Conference on Tunnelling ’82, London. IMM.
In this paper, we performed Bayesian updating Papaioannou, I., Betz W., Zwirglmaier K. & Straub D. 2013.
MCMC algorithms for subset simulation. Manuscript,
of the parameters of a 2D numerical model of a Engineering Risk Analysis Group, TU München.
tunnel in soft soil, conditional on settlement meas- Papaioannou, I. & Straub D. 2012. Reliability updating
urements. We applied BUS, a recently proposed in geotechnical engineering including spatial variabil-
method for Bayesian updating with structural reli- ity of soil. Comput. Geotech. 42: 44–51.
ability methods, combined with subset simulation, Ranjan, R., Betz W., Papaioannou I. & Straub D. 2013.
an adaptive Monte Carlo method that is able to A two-step approach for reliability assessment of a
handle efficiently problems with a large number tunnel in soft soil. In Proc. 3rd International Confer-
of random variables. The results demonstrate the ence on Computational Methods in Tunnelling and Sub-
influence of the accuracy of the measurement surface Engineering EUROTUN: 2013, Bochum.
SOFiSTiK AG. 2012. SOFiSTiK analysis programs ver-
device as well as the prior knowledge of the uncer- sion 2012. Oberschleißheim: SOFiSTiK AG.
tain parameters on their posterior distributions. It Straub, D. 2011. Reliability updating with equality infor-
was shown that the highest impact of the measure- mation. Probabilist. Eng. Mech. 26(2): 254–258.
ment fell on the stiffness moduli and the relaxa- Straub, D. & Papaioannou I. (under review 2013). Bayesian
tion parameter of the stress reduction method updating of mechanical models. J. Eng. Mech.-ASCE.

356

ISGSR2013.indb 356 10/18/2013 9:42:03 AM

 Geotechnical Safety and Risk IV – Zhang et al. (eds)
© 2014 Taylor & Francis Group, London, ISBN 978-1-138-00163-3

A role of monitoring to reduce the uncertainty in the performance

of pile foundations

Satoko Ryuo
Railway Technical Research Institute, Tokyo, Japan

Yoshiharu Asaka
Shimizu Corporation, Tokyo, Japan

Kenichi Soga
University of Cambridge, Cambridge, UK

ABSTRACT: In typical conventional foundation design, the inherent variability of soil properties,
model uncertainty and construction variability are not modeled explicitly. A main drawback of this is that
the effect of each variability on the probability of an unfavorable event cannot be evaluated quantitatively.
In this paper, a method to evaluate the uncertainty-reduction effect on the performance of a vertically-
loaded pile foundation by monitoring the pile performance (such as pile load testing or placing sensors in
piles) is proposed. The effectiveness of the proposed method is examined based on the investigation of a
120-pile foundation placed on three different ground profiles. The computed results show the capability
of evaluating the uncertainty-reduction effect on the performance of a pile foundation by monitoring.

1 INTRODUCTION in predicting the overall differential settlements.

However, the effectiveness of pile monitoring on
Most pile foundations are designed in a determin- the performance of a pile group foundation from
istic manner. That is, the inherent variability of the viewpoint of improving the accuracy of the
soil properties, model uncertainty and construc- design was not addressed.
tion variability are not modeled explicitly; instead In this paper, a procedure to evaluate the
some factors of safety are applied to account for uncertainty-reduction on the performance of a
the uncertainty and variability. In such conven- vertically-loaded pile foundation by monitor-
tional design method, a main drawback is that ing is developed in the following manner. First,
the effect of a specific variability on the probabil- at the locations of the borehole, the ultimate
ity of an unfavorable event cannot be evaluated resistance of a “virtual” pile is calculated using the
quantitatively. A more rational design method that borehole information and widely-used calculation
evaluates the uncertainties for a pile group founda- formulae. Using this data, the performance of piles
tion is required. at actually installed locations is evaluated using
The performance of a vertically-loaded pile the Kriging method. The uncertainty is modeled
that considers the variation of soil properties was stochastically by considering the variability of soil
investigated by Phoon et al. (1990), Quek et al. parameters and construction. The uncertainty of
(1992), Fenton and Griffiths (2003), Haldar and monitored piles is also modeled. Then, the uncer-
Babu (2008) and Wang et al. (2011). In general, tainty of the whole pile group foundation that
these studies do not model the variability of soil consists of both normal and monitored piles is
properties in three dimensional conditions and modeled. To demonstrate the effectiveness of the
do not evaluate the uncertainty for a pile group proposed evaluation method, a hypothetical case
foundation. of an end-bearing 120-pile foundation placed on
Monitoring of a vertically-loaded pile is becom- three different ground profiles is considered. The
ing increasingly popular in order to find the pile computed results show good agreement with typi-
performance directly and to reduce uncertainly of cal rules of thumb, indicating that the developed
the performance. Leung et al. (2011) proposed a method is capable of evaluating how the uncer-
method to decide the optimum locations of moni- tainty on the performance of a piled foundation
toring points based on decreasing the uncertainty can be reduced by monitoring.

357

ISGSR2013.indb 357 10/18/2013 9:42:04 AM

 2 A PROCEDURE TO EVALUATE The ultimate base resistance is computed by the
THE UNCERTAINTY REDUCTION following formulae.

2.1 Uncertainty of individual pile Rb,cal b Ab (4)

A vertically-loaded pile is modeled by shaft resist-
qb Nccu (for clayey soils) (5)
ance and base resistance. The uncertainty of an
individual pile performance can then be defined qb Nqσ υ′ (for sandy soils) (6)
by the ultimate shaft resistance and ultimate base
resistance. In this paper, widely-used calculation
formulae (ICE(2012)) are applied to estimate where Ab = base area of pile; and Nc, Nq = coeffi-
the ultimate shaft resistance and ultimate base cients of bearing capacity. The value of Nc is typi-
resistance. cally 9 and Nq is a function of internal friction angle
The ultimate shaft resistance is given as (Berezantzev et al. (1961)). Apart from the variance in
follows. base area Ab, the uncertainty in ultimate base resist-
ance is assumed to originate from undrained shear
L strength for clayey soils, and effective overburden
Rs,cal π ∫0 τ sdz (1) pressure and internal friction angle for sandy soils.
In this paper, the uncertainty of soil parameters
τs α cu (for clayey soils) (2) and construction are characterized by a normal
distribution. Table 1 shows the variation coeffi-
τs σ h′ tan δ σ υ′ tan δ ′ cient of soil parameters used for estimating pile
= βσ υ′ ( il ), (3) ultimate capacity. Based on these results, the varia-
tion coefficient of ultimate capacity resulting from
where D = iameter of pile; L = thickness of layer uncertainty in soil parameters is assumed to be
which corresponds to length of pile calculat- 0.3 for both shaft resistance and base resistance. It
ing shaft resistance; τs = shaft resistance per should be noted that the formulae for calculating
unit area; α = empirically-determined reduc- the resistance values have also uncertainty and that
tion coefficient; cu = undrained shear strength there is no appropriate reference about the degree
(cohesion); σ υ′ = effective stress in horizontal direc- of uncertainty for the formulae themselves. In this
tion; σ υ′ = effective stress in vertical direction (effec- study, the variation coefficient of ultimate resist-
tive overburden pressure); K = coefficient of lateral ance resulting from the uncertainty in the formulae
pressure; δ ′ = friction angle between pile shaft and is arbitrarily set to 0.3. Further study is needed to
ground; and β = K × tan δ ′. Apart from the variance examine this uncertainty.
in diameter D, the uncertainty of the ultimate shaft Using the values presented above, the magni-
resistance is assumed to originate from layer thick- tude of uncertainty (i.e. variation coefficient) in
ness and undrained shear strength for clayey soils, the ultimate shaft resistance and ultimate base
and layer thickness, effective overburden pressure, resistance of a pile located at a borehole location
coefficient of lateral pressure and friction angle for is given in Table 2. The uncertainty of the ultimate
sandy soils. shaft resistance is 0.42, whereas the uncertainty

Table 1. Soil parameters published in the literature.

Soil parameter Variation coefficient range References

Weight density 0–0.15 Harr (1984), Lacasse and Nadim (1997), Phoon
and Kulhawy (1999a)
Undrained shear strength 0.1–0.70 Harr (1984), Cherubini (1997), Lacasse and Nadim
(UU, UC) (1997), Phoon and Kulhawy (1999a, 1999b),
Duncan (2000)
Internal friction angle 0.02–0.15 Harr (1984), Phoon and Kulhawy (1999a, 1999b)
SPT N-value 0.15–0.5 Harr (1984), Barker et al. (1991), Phoon and
Kulhawy (1999a)
Shear modulus or 0.15–0.65 Phoon and Kulhawy (1999a)
Young’s modulus
DMT Horizontal stress index 0.2–0.6 Phoon and Kulhawy (1999a)
Compression index 0.1–0.37 Harr (1984), Duncan (2000)

358

ISGSR2013.indb 358 10/18/2013 9:42:04 AM

 Table 2. Magnitude of uncertainty and its factors for virtual pile.

Set values of variation coefficients

Resistance Variability factors at borebole locations Remarks

Ultimate Layer thickness Known At borehole locations

shaft Soil parameter
resistance Clayey soil 0.3 Shear strength
Sandy soil 0.3 Unit weight, coefficient of lateral
pressure, and friction angle
Calculation formula 0.3 Model uncertainty and
and construction construction variability
Overall rating 0.42 = (0.32 + 0.32)0.5
Ultimate Upper layer thickness Known At borehole locations
base Soil parameter
resistance Clayey soil 0.3 Shear strength
Nq of sandy soil 0.3 Unit weight and friction angle
Calculation formula 0.3 Model uncertainty and
and construction construction variability
Overall rating 0.42 = (0.32 + 0.32)0.5

Table 3. Magnitude of uncertainty and its factors for monitored pile.

Set values of variation coefficients

Resistance Variability factors at borehole locations Remarks

Ultimate shaft Sensor accuracy, variation 0.05 Overall rating

of deformation modulus
Ultimate base resistance Sensor accuracy 0.05 Overall rating

of the ultimate base resistance is also 0.42. These 1. First, using the borehole data, the thickness of soil
values of uncertainty approximately coincide with layers and the soil parameter values are estimated
the variation coefficients of resistance calculated by the Kriging method at all pile locations.
from field tests in previous studies (ex. Barker et al. 2. At borehole locations, the ultimate shaft resist-
(1991), Okahara et al. (1993), Kay (1993), Archi- ance and ultimate base resistance of a “virtual”
tectural Institute of Japan (AIC) (2001) ). Hence pile are calculated. Also, based on the thickness
it is concluded that the developed procedure for of soil layers and the soil parameter values eval-
calculating uncertainty is reasonable. uated from Step (1) above, the ultimate shaft
For a pile that is directly monitored by instru- resistance and ultimate base resistance of each
mentation or if a pile load test was carried out, the pile in the actual foundation are calculated by
uncertainty of the “monitored” pile is governed by using each borehole data.
the accuracy of measurement devices, the parame- 3. A pile or piles that will be monitored are ran-
ters used in conversion of measurement and physi- domly selected as a given monitoring scenario.
cal quantity such as Young’s modulus of concrete The mean value of the ultimate shaft resist-
or steel. In this paper, for monitored piles, the vari- ance and base resistance obtained from Step
ation coefficients of ultimate shaft resistance and (2) has a variability characterized by a normal
ultimate base resistance are assumed as 0.05 (see distribution and the 95% confidence bounds
Table 3). These values are smaller than the values is set as 0.05 for monitored piles and 0.42
assigned for non-monitored piles, which are given for “virtual” piles. Then, 1000 data sets of
in Table 2. shaft resistance and base resistance are ran-
domly produced for both monitored piles and
“virtual” piles.
2.2 Uncerainty of a whole pile group foundation
4. From each of 1000 data sets, the ultimate shaft
The uncertainty of a whole pile group foundation resistance and base resistance for each pile in the
that consists of normal and monitored piles can be actual foundation are estimated by the Kriging
evaluated by the following steps. method.

359

ISGSR2013.indb 359 10/18/2013 9:42:06 AM

 5. From the 1000 data set of ultimate shaft resist-
ance and ultimate base resistance evaluated at
the actual pile locations, the respective mean
values are adopted as the ultimate shaft resist-
ance and ultimate base resistance of the actual
piles. Then the sum of the two resistance val-
ues is the ultimate overall resistance of the
piles. Also the uncertainty value of each pile is
derived from the mean of Kriging variance and
the sum of the uncertainty values is called the
“uncertainty” of the whole piled foundation.
Steps 3 to 5 are repeated by changing the pile(s)
that will be monitored. The appropriate monitor- Figure 1. Layout of piles and boreholes and soil profile.
ing locations can then be decided from the degree
of reduction in the overall uncertainty.
two layers with abrupt change in top layer thickness
2.3 Evaluation criterion for the uncertainty- due to faulting (i.e. buried terrace structure). For
reduction by monitoring each ground profile, a clayey soil layer is underlain
by a sandy soil layer (see Figure 2).
From the procedure described in the previous sec- Eight boreholes are deployed. Tables 4 to
tion, the ultimate resistance and the uncertainty of 6 show the soil parameters and the resistance val-
a whole piled foundation that consists of normal ues of a virtual pile placed at the location of each
and monitored piles are calculated. In this study, borehole. In the following discussion, the varia-
the evaluation criterion to select the most suitable tion coefficient of cu and qb are assumed to be 0.3
pile(s) for monitoring is given as follows. as the uncertainty of soil parameters. Then, the
Evaluation criterion: Find the location of moni- uncertainty of ultimate resistances of each bore-
toring point(s) that minimize(s) the following hole shown in Tables 4 to 6 is modeled by assuming
equation, their variation coefficient as 0.42 and the distribu-
tion as a normal distribution.
Kriging varianceof pile i
∑ USRand UBRof pile i
the number of piles.
i∈all pile 3.2 Computed results for one monitored pile

The estimated value is called “average of varia- The results for the cases with one monitored pile are
tion co-efficient” in this paper. As a separate study, shown in Figures 3 to 5 for the three different soil
other evaluation criteria for selecting the moni- profiles. In these figures, the effectiveness of moni-
toring point(s) to reduce the uncertainty of piled toring for the whole piled foundation is defined
foundation are considered and the results will be by the difference between the average of variation
reported elsewhere in the future. coefficient for whole pile foundation calculated
in one monitored pile situation and that without
monitoring. The differential value is computed at
each monitored pile location and then a contour
3 CASE STUDY
plot showing the spatial distribution of the differ-
ential value is generated. The location that gives the
3.1 Condition for calculation
maximum value is considered to be the ideal place
A piled foundation which has 120 single piles to monitor. Also the maximum value itself gives
placed at 7-meter intervals is considered as a hypo- the value of monitoring; that is, the larger the value
thetical case study. The plan area is composed of is, the uncertainty reduces more by monitoring.
7-span-49 meter × 14-span-98 meter as shown in Figure 3 shows the result for Case 1. In the case
Figure 1. Each pile is cast-in-place pile with diam- of horizontally-layered conditions, the pile located
eter of 0.8 m and its length coincides with the thick- away from the borehole, especially the one near the
ness of clay as shown in Tables 4 to 6. The ultimate center of the piled foundation, will be ideal in terms
resistance is given by the shaft resistance and the base of reducing the uncertainty of the foundation.
resistance considering the soil conditions. Three soil The result for Case 2 is shown in Figure 4. It indi-
profiles are considered; (i) Case 1: two horizontally cates that a pile located away from the boreholes
layered conditions, (ii) Case 2: two layers with an and on the left hand side of the ground that has a
inclined interface, in which the slope angle changes moderate interface slope angle would be effective
at the centerline of the foundation, and (iii) Case 3: one for monitoring to reduce the uncertainty.

360

ISGSR2013.indb 360 10/18/2013 9:42:06 AM

 Table 4. Soil parameters and resistance for Case 1.

BH1 BH2 BH3 BH4 BH5 BH6 BH7 BH8

cu of clay (kPa) 50 50 55 55 50 50 45 45
qb at toe (kPa) 12000 11700 11700 12300 12000 12300 12000 12000
Thickness of clay = pile length (m) 20 19.5 19.5 20.5 20 20.5 20 20
Diameter of model pile, D (m) 0.8 0.8 0.8 0.8 0.8 0.8 0.8 0.8
Adhesion factor of shaft resistance, α 0.6 0.6 0.6 0.6 0.6 0.6 0.6 0.6
Ultimate shaft resistance, USR1 (kN) 1508.0 1470.3 1617.3 1700.2 1508.0 1545.7 1357.2 1357.2
Ultimate base resistance, UBR (kN) 6031.9 5881.1 5881.1 6182.6 6031.9 6182.6 6031.9 6031.9
USR1 + UBR(kN) 7539.8 7351.3 7498.3 7882.9 7539.8 7728.3 7389.0 7389.0

Table 5. Soil parameters and resistance for Case 2.

BH1 BH2 BH3 BH4 BH5 BH6 BH7 BH8

cu of clay (kPa) 50 50 55 55 50 50 45 45
qb at toe (kPa) 10200 9900 10800 12300 12000 13800 15000 15000
Thickness of clay = pile length (m) 17 16.5 18 20.5 20 23 25 25
Diameter of model pile, D (m) 0.8 0.8 0.8 0.8 0.8 0.8 0.8 0.8
Adhesion factor of shaft resistance, α 0.6 0.6 0.6 0.6 0.6 0.6 0.6 0.6
Ultimate shaft resistance, USR1 (kN) 1281.8 1244.1 1492.9 1700.2 1508 1734.2 1696.5 1696.5
Ultimate base resistance, UBR (kN) 5127.1 4976.3 5428.7 6182.6 6031.9 6936.6 7539.8 7539.8
USR1 + UBR(kN) 6408.8 6220.3 6921.6 7882.9 7539.8 8670.8 9236.3 9236.3

Table 6. Soil parameters and resistance for Case 3.

BH1 BH2 BH3 BH4 BH5 BH6 BH7 BH8

cu of clay (kPa) 50 50 55 55 50 50 45 45
qb at toe (kPa) 12000 11700 11700 12300 12000 17700 18000 18300
Thickness of clay = Pile length (m) 20 19.5 19.5 20.5 20 29.5 30 30.5
Diameter of model pile, D (m) 0.8 0.8 0.8 0.8 0.8 0.8 0.8 0.8
Adhesion factor of shaft resistance, α 0.6 0.6 0.6 0.6 0.6 0.6 0.6 0.6
Ultimate shaft resistance 1, USR1 (kN) 1508 1470.3 1617.3 1700.2 1508 2224.2 2035.8 2069.7
Ultimate base resistance, UBR (kN) 6031.9 5881.1 5881.1 6182.6 6031.9 8897 9047.8 9198.6
USR1 + UBR (kN) 7539.8 7351.3 7498.3 7882.9 7539.8 11121.2 11083.5 11268.3

In Case 3, the result is given in Figure 5. Results new information about ground profile is found
show that monitoring in the space enclosed by to be similar to the existing ground profile, this
BH6, BH7 and BH8 where the clayey soil layer is reduces the uncertainty in the ground profile. This
relatively thicker will not be effective in reducing would increase the precision of fitting function
the uncertainty. The most effective location for to the ground profile, which in turn reduces the
monitoring is at the left hand side of the ground uncertainty of the whole foundation.
in which the soil profile is uniform. Another effec-
tive monitoring area is the space enclosed by BH4,
3.3 Computed results for multi-monitored pile
BH5 and BH6 where the thickness of the clayey
soil layer is variable. Based on the contour plots, the location(s) of
The efficient locations obtained from the calcu- the monitoring are determined and the effect of
lations are consistent with typical rules of thumb. monitoring number in reducing the uncertainty
According to Case 2 and 3, the effective area for was examined. The number of monitoring was
monitoring is located at the ground that has a increased up to ten and a relationship between
moderate interface slope angle and that has uni- monitoring number and the effect is shown in
form soil profile in large part. This is because when Figure 6. The average of variation coefficient for

361

ISGSR2013.indb 361 10/18/2013 9:42:07 AM

 Figure 2. Pattern diagram of Cases 1 to 3.

Figure 6. Effectiveness of monitoring at several loca-

tions for each ground profile.

the case of no monitoring pile is 0.123 for Case 1,

0.134 for Case 2, and 0.155 for Case 3. The value of
Case 1 with uniform soil conditions is smaller than
the values of the other cases, showing the effect
of soil conditions on uncertainty as expected. In
all cases, the degree of uncertainty decreases with
increasing number of monitoring. For example,
Figure 3. Effectiveness of monitoring based on Evalua- in the case of ten monitored piles placed at well-
tion critetion 1 (Case 1: Horizontally-layered, one moni- chosen locations, the uncertainty value is half of
tored pile). that of the non-monitored pile foundation case.

4 CONCLUSIONS

In this paper, a procedure to evaluate the

uncertainty-reduction effect of pile monitoring on
the performance of a pile foundation is presented.
The variability of the single-loaded pile behavior
was modeled by assuming their shaft resistance
and base resistance as 0.42 and the distribution as
a normal distribution. When the pile is monitored
(or a pile test is carried out at this location), it was
assumed that the variability would reduce to 0.05.
The uncertainty of a whole piled foundation that
consists of monitored and non-monitored piles was
Figure 4. Effectiveness of monitoring based on Evalu- then evaluated. To demonstrate the effectiveness
ation critetion 1 (Case 2: Dipping bearing stratum, one
monitored pile).
of the proposed evaluation method, a hypothetical
case of 120-pile foundation placed was examined
for three different ground profile conditions. The
results show that the developed method gave good
agreement with typical rules of thumb and could
provide quantitative data to make more rational
decision in carrying out pile monitoring or pile
testing.

REFERENCES

Architectural Institute of Japan (AIJ) (2001).

Recommendations for Design of Building foundations.
Architectural Institute of Japan, (in Japanese).
Barker, R.M., Duncan, J.M., Rojiani, K.B.,
Figure 5. Effectiveness of monitoring based on Evaluation Ooi, P.S.K.,Tan, C.K. & Kim, S.G. 1991. Manuals for
critetion 1 (Case 3: Buried terrace, one monitored pile). the design of bridge foundations. NCHRP Report 343.

362

ISGSR2013.indb 362 10/18/2013 9:42:07 AM

 Berezantzev, V.G., Khristoforov, V. & Golubkov, V. 1961. Leung, Y.F., Soga, K. & Klar, A. 2011. Multi-objective
Load bearing capacity and deformation of piled foun- foundation optimization and its application to pile
dations. Proc. 5th ICSMFE 2 2: 11–15. reuse. Geotechnical Special Publication 211: 75–84.
Cherubini, C. 1997. Data and considerations on the Okahara, M., Kimura, Y., Ochiai, H. & Matsui, K. 1993.
variability of geotechnical properties of soils. In Statistical characteristics of bearing capacity of single
Proceedings of the international conference on safety piles. In proceedings of the ninth us-japan bridge engi-
and reliability, ES-REL, Volume 97: 1583–1591. neering workshop, Volume 3230.
Duncan, J.M. 2000. Factors of safety and reliability in Phoon, K.K. & Kulhawy, F.H. 1999a. Characterization
geotechnical engineering. Journal of Geotechnical of geotechnical variability. Canadian Geotechnical
and Geoenvironmental Engineering 126(4): 307–316. Journal 36(4): 612–624.
Fenton, G.A. & Griffiths, D.V. 2003. Bearing-capacity Phoon, K.K. & Kulhawy, F.H. 1999b. Evaluation
prediction of spatially random c-ϕ soils. Canadian of geotechnical property variability. Canadian
geotechnical journal 40(1): 54–65. Geotechnical Journal 36(4): 625–639.
Haldar, S. & Babu, G.L.S. 2008. Reliability measures for Phoon, K.K., Quek, S.T., Chow, Y.K. & Lee, S.L. 1990.
pile foundations based on cone penetration test data. Reliability analysis of pile settlement. Journal of
Canadian Geotechnical Journal 45(12): 1699–1714. Geotechnical Engineering 116(11): 1717–1734.
Harr, M.E. 1984. Reliability-based design in civil engi- Quek, S.T., Chow, Y.K. & Phoon, K.K. 1992. Further
neering. Department of Civil Engineering, School of contributions to reliability-based pile-settlement
Engineering, North Carolina State University. analysis. Journal of geotechnical engineering 118(5):
Institution of Civil Engineers (ICE) 2012. ICE Manual 726–741.
of Geotechnical Engineering, Vol. II. Geotechnical Wang, Y., Au, S.K. & Kulhawy, F.H. 2011. Expanded
Design, Construction and Verification, Edited by reliability-based design approach for drilled shafts.
Burland, J., Chapman, T., Skinner, H. & Brown, M. Journal of geotechnical and geoenvironmental engineer-
Kay, J.N. 1993. Probabilistic design of foundations and ing 137(2): 140–149.
earth structures. Proceedings of Probabilistic methods
in geotechnical engineering, Edited by K.S. Li & S.C.R.
Lo. A.A.: 49–62.
Lacasse, S. & Nadim, F. 1997. Uncertainties in character-
ising soil properties. Publikasjon-Norges Geotekniske
Institutt 201: 49–75.

363

ISGSR2013.indb 363 10/18/2013 9:42:09 AM

This page intentionally left blank
 Geotechnical Safety and Risk IV – Zhang et al. (eds)
© 2014 Taylor & Francis Group, London, ISBN 978-1-138-00163-3

Using head monitoring for reliability updating of levees

T. Schweckendiek
Deltares, Unit Geo-Engineering, Delft, The Netherlands
Department of Hydraulic Engineering, Delft University of Technology, Delft, The Netherlands

ABSTRACT: Backward erosion piping is an important internal erosion failure mechanism in deltaic
areas. Piping is very sensitive to ground conditions, especially to the permeability of aquifers under levees.
The uncertainties in these ground properties can be reduced by monitoring the hydraulic head in the piping-
sensitive aquifer in order to update the probability distributions of the basic random variables influencing
the geo-hydrological conditions such as the hydraulic resistance of the aquifer. This paper demonstrates how
Bayesian posterior analysis can be used to incorporate head monitoring information in order to update the
reliability with respect to uplift, heave and piping for river levees. The paper starts by recapping background
on uplift, heave and piping reliability, before we discuss how the equality type of information involved can
be used in posterior analysis by following the approach proposed by Straub (2011). The workings of the
theory are illustrated by an example considering uplift (of a blanket layer) only using a simplified ground-
water model for the sake of illustration. Sensitivity analysis on the example shows that the effect of reliability
updating can be considerable for large prior uncertainties and for observations which are rather unlikely
given the prior distributions of the random variables (i.e. far from the expected value of the measurement).

1 INTRODUCTION judgment, the incorporation of which in reliability

analysis is not straightforward.
Flood defense levees are crucial for flood protec- In order to provide a more rigorous approach
tion, especially in riverine and deltaic areas. One for reliability updating with such head monitor-
of the most critical failure modes of levees is back- ing data, this paper explores the possibilities and
ward internal erosion or piping. In the VNK2 effects of applying Bayesian Updating. By work-
project (Jongejan et al. 2013) which is analyzing ing with Monte Carlo analysis and exact methods
the reliability of all primary flood defenses in The without approximations in the methods them-
Netherlands, piping was identified as a main con- selves, we will show how not only the reliability is
tributor to the probability of flooding. updated but also how the posterior distributions of
Such reliability analyses of flood defenses work the basic random variables change with respect to
with prior probabilities of relevant parameters their priors. The latter allows more detailed inter-
such as the ground conditions, which are based on pretation of the results and the plausibility of the
sparse site investigation data and/or expert judg- observations.
ment. The prior uncertainties in ground properties The paper starts by recapping background on
are relatively large compared to other civil engi- uplift, heave and piping reliability, before we dis-
neering materials like concrete or steel. But these cuss how the equality type of information involved
uncertainties can be reduced by gathering and can be used in posterior analysis by following the
incorporating additional information. approach proposed by Straub (2011). The workings
One source of information is monitoring of the of the theory are illustrated by an example consider-
hydraulic head in an aquifer underlying a levee and ing uplift (of a blanket layer) only using a simplified
a blanket layer. It is common engineering practice groundwater model for the sake of illustration.
to instrument levees, where the geo-hydrological
conditions are rather uncertain, with observation
2 PLIFT, HEAVE AND PIPING
wells to observe the response of the hydraulic head
RELIABILITY
to changes in the river water level. This allows us to
get an impression of not only the hydraulic resist-
2.1 Physical processes
ance of the aquifer but also on the seepage length
and other geo-hydrological parameters. However, The conceptual model underlying piping reli-
the interpretation of such data is usually done ability modeling in Dutch practice is illustrated in
by interpolation, regression analyses or expert Figure 1. The typical situation addressed here is a

365

CH46_7-6.indd 365 10/18/2013 6:07:29 PM

 Figure 1. Phases of the piping process including uplift and heave.

relatively low permeability levee with low perme-

ability blanket on the land-side. High floods gen-
erate high pore pressures in the relatively highly
permeable aquifer underneath (Fig. 1, stage a).
When and where the pressure exceeds the weight of
the blanket (i.e. uplift, stage b), the latter may rup-
ture causing accelerated upward ground water flow.
If the flow exceeds a critical rate (i.e. heave, stage
c), sand particles start to erode. Subsequently, the
internal erosion progresses in upstream direction Figure 2. Definition of the damping factor for estimat-
(i.e., towards the river) forming erosion channels— ing the head at the critical exit point.
so-called “pipes” (stage d). If the piping process
does not come to a halt, the levee can be “under-
mined” and collapse (stages e, f).
Note that the same process is often also called to surface level), the latter to be calculated by
under-seepage or backward erosion in the literature. groundwater flow analysis or to be estimated from
For sake of conciseness we will refer to the whole measurements. For the sake of illustration, in the
process as piping throughout the paper. example elaborated below we will apply the follow-
ing simplified approach using a damping factor λ
for predicting the head difference in the aquifer at
2.2 Limit states the potential exit point (see Fig. 2):
Below we present the commonly used limit state
Δφ = λ(h − hp) (2)
functions for uplift heave and piping. For uplift,
the head difference at the bottom of a (potential)
exit point needs to exceed the weight exerted by the where h [m] is the (river) water level (all absolute
blanket: levels are with respect to MSL = mean sea level)
and hp [m] is the phreatic level in the polder (i.e.
⎛γ ⎞ landside of the levee). The damping factor λ can
gu ( x ) mu d ⎜ ssat − 1⎟ − Δφ (1) be based on expert judgment or groundwater
⎝ γw ⎠ flow analysis and is supposed to include model
uncertainty.
where mu is a model factor addressing the uncer- While uplift is concerned with the breaching of
tainty in the critical uplift head difference, d [m] the impervious boundary, heave is concerns the
is the blanket thickness, γsat [kN/m3] is the satu- start of erosion; the resistance against heave is
rated volumetric weight of the blanket, γw [kN/m3] expressed in terms of a critical exit gradient ic:
is the volumetric weight of water and Δφ [m] is
the head difference at the exit point (with respect gh(x) = ic − Δφ/d (3)

366

ISGSR2013.indb 366 10/18/2013 9:42:09 AM

 For piping we consider the average gradient- terms of the definition of conditional probability
based equilibrium model by Sellmeijer (1988), as in Benjamin and Cornell (1970):
which was revised recently (Sellmeijer, 2012) by
recalibrating the coefficients in the critical piping P (F | ) P (F ) / P (ε ) (8)
head difference Hc [m]. The performance function
is given by: where ε is the evidence, which in our application
means that we monitor or measure the head φ
gp(x) = mpHc − (h − hp − 0.3 d) (4) at the exit point with measurement error em [m],
resulting in the measured value of the head φm [m]:
where mp is a model uncertainty factor.
ε {φ + −φ = } (9)
2.3 Reliability analysis
Let the failure event Fi for failure mode i be defined The well-known issues with this equality type
as the limit state function gi assuming negative of information are that the numerator in Eq. (8)
values. The probability of failure is then given by: cannot be determined by standard reliability anal-
ysis techniques and the denominator is zero by
definition. Using the approach proposed by Straub
P ( Fi ) = P ( gi ( ) < 0) = ∫g ( fX dx (5) (2011), however, we can re-formulate the observa-
)< 0
tion space ε in terms of the following equivalent
i

inequality domain:
where X is the vector of random variables and fX(x)

{ }
its (prior) Probability Density Function (PDF).
Fragility curves are a frequently used alternative εe ⎡ fe (φm − φ ( ))⎤⎦ < (10)
⎣ m
way of expressing the uncertainty in the aggre-
gated resistance R with respect to a dominant load where u is a standard Normal distributed random
variable. In other words, R contains all variables variable, Φ −1 is the inverse cumulative standard
in X except the dominant load variable, which for Normal distribution function, fem is the PDF of
river levees is the water level h), providing the con- the measurement error and c is a positive constant
ditional probability of failure with respect to h: to be chosen such that the term in square brackets
is always between zero and one. For the deriva-
P ( Fi | h ) ∫g ( f ( r))dr (6) tion and further details of this method we refer to
i h) < 0 R
Straub (2011).
As explained in the previous section, failure This transformation allows us to compute the
can only occur if uplift precedes heave and heave updated probability of failure using standard reli-
precedes piping, because piping can only develop ability analysis methods by evaluating the follow-
if the groundwater can flow upwards through the ing expression:
ruptured blanket (i.e. uplift) sufficiently to move
sand particles vertically (i.e. heave). In reliability P (F ∩ ε e ) ∫F ∩ε fX ϕ u )du
d dx
terms this relationship can be described by means P (F | ε ) = = e
(11)
P (ε e )
of a parallel system, for which the failure set is ∫ε e
fX ϕ u )du dx
given by:
where ϕ(u) is the standard normal PDF. Notice
F {gu < } ∩ {gh < } ∩ {g p < } (7) that this way the posterior probabilities of
failure for the separate failure mechanisms
Note that the prior probability of failure and the uplift, heave and piping can be computed as well
fragility curves as defined above can be determined as the posterior (parallel system) probability
per failure mode individually (i.e. Fu, Fh and Fp) as of failure for the whole failure mode (i.e.
well as for the “system” failure event F. P(Fu∩Fh ∩Fp|ε)).

3 RELIABILITY UPDATING
4 EXAMPLE: UPLIFT
This section provides the theory and definitions
4.1 Limit state
used in this study for reliability updating with infor-
mation from pore pressure monitoring. We apply In the following we provide a numerical elabora-
Bayes’ rule (Bayes, 1763) to obtain the posterior tion for uplift failure as defined in section 2.2. The
(updated) probability of failure, here expressed in limit state function for this example is obtained by

367

ISGSR2013.indb 367 10/18/2013 9:42:11 AM

 inserting the load term (Eq. 2) in the uplift limit Table 2. Annual prior and posterior reliability.
state function (Eq. 1) resulting in
Probability of failure Reliability index
⎛ γ ssat ⎞
gu ( x ) mu d
⎝ γw ⎠
(
λ h − hp ) (12) Prior
Posterior
P(Fu) = 6.6 ⋅ 10−2
P(Fu|ε) = 4.8 ⋅ 10−3
β = 1.50
β = 2.59

Equating to zero and solving for the water level

gives the critical uplift water level hc,u [m], which
will is the aggregated resistance and will be used for
illustration of results in terms of fragility curves:

⎛ γ ssat ⎞
hc,u mu d /λ hp (13)
⎝ γw ⎠

Notice that the cumulative distribution Fhc,u is

the uplift fragility curve. Figure 3. Water level distribution vs. prior and poste-
rior uplift fragility curve at the exit point.
4.2 Input parameters and prior analysis
The example parameters as presented in Table 1
are fictitious but realistic. The annual prior prob-
ability of uplift failure is P(Fu) = 6.6 ⋅ 10−2 obtained
by Monte Carlo simulation, which corresponds to
a reliability index of β = 1.50.

4.3 Posterior analysis

We suppose that monitoring during a 100-year
water level of h′ = 3.9 m has resulted in a measured
head in the aquifer at the dike toe (i.e. exit point)
of φm = 2.3 m. The duration of the reached flood
level is assumed sufficient for a near steady-state
pore pressure response in the aquifer, which means
that the applied limit state models are applicable
(i.e. steady state assumption in the groundwater Figure 4. Posterior probability densities of selected
basic random variables.
model). Furthermore, the measurement is assumed
to be unbiased with an error em with zero mean and
standard deviation equal to 0.1 m. the probability of failure to decrease by an order
Since the failure models are computationally of magnitude.
inexpensive, we can obtain exact results by using Notice that even though the prior and poste-
Monte Carlo simulation as summarized in Table 2. rior probability are not extremely low, the pos-
Incorporating the monitoring observation causes terior analysis takes a large number of Monte
Carlo-realizations to converge. The reason is the
Table 1. Input parameters uplift example. numerator term in Bayes’ rule, P(Fu ∩ ε) which can
be orders of magnitude lower than P(Fu|ε) itself.
Parameters (mean and
Xi Distribution type standard deviation)
The large number of realizations is not so much
a problem of computation time but a memory
h Gumbel μ = 2.67, σ = 0.38 issue, especially if one wants to retain the values of
hp Normal μ = 0.3, σ = 0.1 all random variables for subsequent analysis and
d Lognormal μ = 3.0, σ = 0.5 interpretation of results.
mu Lognormal μ = 1.0, σ = 0.1 Figure 3 shows that the fragility curve of the
γsat Normal μ = 20.0, σ = 1.0 critical water level hc,u changes significantly as a
λ Lognormal μ = 0.8, σ = 0.1 result of the posterior analysis. Besides fragility
em Normal μ = 0.0, σ = 0.1 curves and load effect, also the posterior probabil-
γw Deterministic 10 ity densities of some basic random variables can be
obtained as shown in Figure 4. Notice that only hp

368

ISGSR2013.indb 368 10/18/2013 9:42:14 AM

 and λ, the variables influencing the measurement,
are updated. While the probability density of hp
only slightly shifts to the left, the we see a signifi-
cant shift and considerable decrease of spread in
the distribution of λ. Figure 5 also shows that after
updating hp and λ are correlated.

4.4 Sensitivity analysis

In the following we investigate the sensitivity of
the outcomes to the measured head at the potential
exit point (φm), the observed water level at which Figure 7. Sensitivity of the posterior annual reli-
the measurement was obtained (h’) and the meas- ability with respect to the measurement error σe,m (for a
urement error (em). measured head of φm = 2.3 m and a 100-year flood level
Figure 6 is a contour plot showing the pos- h′ = 3.9 m).
terior reliability index for combinations of the
observed water level (h′) and the measured head
(φm) for a given measurement error of σe,m = 0.1m. considered range of the measured head, the poste-
The posterior probability of failure increases rior probability of failure varies by several orders
with the measured head. For the (relatively large) of magnitude.
For a given value of the measured head, vary-
ing the water level illustrates that the higher the
observed water level, the greater the impact on
the posterior reliability. The decrease of the reli-
ability index is roughly a linear function of both,
φm and h′.
Figure 7 clearly demonstrates that the updat-
ing effect vanishes with increasing measurement
error.

5 CONCLUSION

The reliability updating method for equality type

of information proposed by Straub (2011) enables
us to do posterior analysis for structural reliability
problems where we have direct measurements of
one or more system characteristics (i.e. variables
involved in the performance function).
Figure 5. Prior and posterior joint posterior probability
This paper has discussed how head monitoring
densities of λ and hp.
information can be used to update the reliability
regarding uplift, heave and piping (i.e. under-
seepage) for levees using Straub’s approach. The
(simplified) example on uplift has demonstrated
the workings of the method and illustrated that
the effect of updating (i.e. the change from prior to
posterior probability of failure) can be consider-
able, if the prior uncertainties are rather large and
the observation was rather unlikely given the prior
probability distributions.
An extension of the presented posterior analysis
to pre-posterior and decision analysis is discussed
in Schweckendiek (2013a); a more extensive case
study on including uplift, heave and piping can be
found in Schweckendiek (2013b), which includes
Figure 6. Posterior annual reliability as a function of more sophisticated limit state models, system reli-
the measured head φm and the observed water level h′ (for ability aspects and treatment of (discrete) stratifi-
a measurement error of σe,m = 0.1 m). cation scenarios.

369

ISGSR2013.indb 369 10/18/2013 9:42:16 AM

 REFERENCES Schweckendiek, T. 2013b. On reducing piping
uncertainties—a bayesian decision approach. PhD
Bayes, T. 1763. An essay towards solving a problem in the thesis, Delft University of Technology (in preparation).
doctrine of chances. Philosophical Transactions of the Sellmeijer, J., et al. 2012. Fine-tuning of the piping model
Royal Society of London, 53, 370–418. through small-scale, medium-scale and experiments.
Benjamin, J. & Cornell, C. 1970. Probability, statistics and European Journal of Environmental and Civil Engi-
decision for civil engineers. New York: McGraw-Hill. neering, Special Issue devoted to “Erosion in geome-
Jongejan, R., et al. 2013. The VNK2-project: a fully chanics” (in press).
probabilistic risk analysis for all major levee systems Sellmeijer, J. 1988. On the mechanism of piping under
in The Netherlands. IAHS Publication, 357: Floods: impervious structures. PhD-thesis, Delft University of
From Risk to Opportunity. Technology.
Schweckendiek, T. 2013a. Reliability updating and deci- Straub, D. 2011. Reliability updating with equality
sion analysis for head monitoring of levees, Georisk information. Probabilistic Engineering Mechanics 26
7(2), in press. (2011), 254–258.

370

ISGSR2013.indb 370 10/18/2013 9:42:17 AM

 8 Performance evolution and risk assessment of high slopes
in hydropower engineering in China

ISGSR2013.indb 371 10/18/2013 9:42:17 AM

This page intentionally left blank
 Geotechnical Safety and Risk IV – Zhang et al. (eds)
© 2014 Taylor & Francis Group, London, ISBN 978-1-138-00163-3

Numerical simulation of blasting induced damage of high rock slope

Y.G. Hu, W.B. Lu, M. Chen & P. Yan

State Key Laboratory of Water Resources and Hydropower Engineering Science, Wuhan University,
Wuhan, Hubei, China
Key Laboratory of Rock Mechanics in Hydraulic Structural Engineering of Ministry of Education,
Wuhan University, Wuhan, Hubei, China

ABSTRACT: The capability of predicting Blasting Induced Damage (BID) in the surrounding rock mass
is important to excavation of high rock slopes. A tensile-compress damage model linked to the computer
code LS-DYNA through the user subroutine interface was used by introducing the consideration of com-
press damage and modifying the method of definition of the elastic constants based on the existed blasting
damage models. To verify the rationality of the proposed model, based on the case study of the excavation
of high rock slope at the Xiluodu Hydropower Station in Sichuan province of China, spatial distribution
of blasting induced damage zone at the 640 m berm were determined with cross-hole sonic test. Numerical
simulations of BID at the 640 m berm were implemented with tensile-compress model and other four
existing blasting damage models. The damage zone obtained by the modified tensile-compressive damage
model agreed with observations better than other four existing blasting damage models.

1 INTRODUCTION the underground excavation. So for the huge

high rock slope, the requirement of stability and
During a blasting excavation, damage is induced safety in long term makes the investigation of the
by a combination of the effects of the blasting load EDZ characteristics is important and absolutely
and stress redistribution (Martino & Chandler necessary.
2004). Investigation of damage characteristics is Researchers often get the characteristics of
most important to restrict the damage extension. EDZ with the experiment and the engineering
The Excavation induced Damage Zone (named detection. But the method of full size experiment
EDZ) is generally defined as the zone beyond the is very expensive and time-consuming. The engi-
boundary where the rock has been considerably neering detection is after the event in the engineer-
damaged or disturbed (Malmgren et al. 2006). ing, which is not good for the control of EDZ.
Based on the AECL’s Underground Research Numerical method, derived from sound mechani-
Laboratory, Read (2004) have got progress in cal principles and validated against experimental
the aspect of formation mechanism, mechanical data, indicates a promising approach to reveal
property and detection technique for the damage the EDZ formation process. For this case, various
zone in the surrounding rock mass. In support of damage models have been suggested to study the
the excavation for the permanent shiplock of Three damage of rock mass under blasting. For instance,
Gorges Project, Sheng et al. (2002) investigated Grady & Kipp (1980), Taylor (1986) and his co-
the deformation and the damage in the remain- workers, Yang et al. (1996) and Liu & Katsabanis
ing rock with sonic wave detection and numerical (1997) developed several widely used continuum
simulation. However, compared to the study on the damage models to describe the dynamic damage
EDZ of underground openings, fewer attentions process of rock mass in tension. The damage vari-
are paid to the formation and spatial distribution able is introduced to describe the changing state of
of EDZ of high rock slope. a material. However, the damage variable is estab-
The excavation of high rock slope is one of the lished based on the tensile strain and compressive
biggest problems in the construction. Goodman & damage is neglected in these damage models. Thus
Kieffer (2000) pointed out that slopes and under- far, application of an appropriate damage model
ground openings have distinctly different environ- to simulate the different damage pattern is rarely
ments and the slope environment is generally less found.
secure than the underground environment. It is In the present study, based on the blasting exca-
believed that the excavation damage zone associ- vation for the rock mass high slope of Xiluodu
ated with the surface excavation would be larger hydropower station, the outline of blasting induced
and more complicated than that associated with damage zone of the berm at the elevation 640 m

373

ISGSR2013.indb 373 10/18/2013 9:42:17 AM

 was determined with methods of cross-hole sonic 2/3
test firstly. Then a modified damage model is pre- 1 ⎛ 20 K IC ⎞
a= (3)
sented by incorporating the compressive damage 2 ⎜⎝ ρcεmax ⎟⎠
into the exiting damage model. The damage of the
2
berm at the elevation 640 m was subsequently sim- 5 ⎛ K IC ⎞ m
Cd k εv (4)
ulated, to verify the accuracy of the model, four 2 ⎜⎝ ρcεv max ⎟⎠
kinds of existed blasting damage models were used
to be compared. where ρ is the mass density, c is the wave speed, and
KIC is the fracture toughness of the material, εv max
is the maximum volumetric strain rate experienced
2 TENSILE-COMPRESS DAMAGE MODEL by the representative volume element at fracture.
Kuszmaul (1987) considered the material over-
The investigation of blasting damage models is rep- lap of high density fracture zone, the decrease
resented by America Sandia National Laboratory. caused by damage was introduced into the activ-
To predict the damage and failure, the dynamic frac- ity ratio, and the expression of the crack density
ture was regarded as a continuum process of damage parameter was modified as:
cumulation and the relationship between the dam-
2
age scalar D and the crack density was established. 5 ⎛ K IC ⎞ m
Cd k εv (1 − Dt ) (5)
2 ⎜⎝ ρcεmax ⎟⎠
Several blasting damage models were put forward
to describe the degradation of rock stiffness and
strength caused by cracks growth under blasting
An expression based on the percolation theory
(Grady & Kipp 1980, Taylor et al. 1986, Kuszmaul
(Englman & Jaeger 1987):
et al. 1987, Thorne et al. 1990, Yang et al. 1996).
But in the existed blasting damage model, the dam-
⎛ 16 ⎞
age scalar D was defined by the tensile strain, while μ μ exp − βCd (6)
the compressive damage was neglected. It is known ⎝ 9 ⎠
that a compressive damage zone was formed by the
blasting load in rock mass near the charge hole. The Inspection of Eq. (4) suggests that the tensile
next section describes a tensile-compressive damage damage can be defined by
model in which the compressive damage was intro- 2
duced and a modified method of definiting the elas- 16 (1 − )
Dt Cd (7)
tic constants was proposed. 9 (1 − 2 μ )
Based on the research of Taylor et al. (1986), the
effective bulk modulus of the cracked solid is On the other hand, when the rock material is
in compression, the von Misses yield condition is
K 16 (1 − μ ) expressed as:
= 1− Cd (1)
K 9 (1 − 2 μ )
Φ Λ 2 − σ y2 = 0 (8)
where K and μ are the original bulk modulus 3
where Λ = sij sij is the von Misses equivalent
and Passion’s ratio for intact material, the barred 2
stress.
quantities represent the corresponding degraded
Based on the coupling principle of strain-rate
constants for damaged material, with Cd being the
effect in the RDA model (Furlong et al. 1990), the
crack density parameter.
compressive damage Dc is expressed as:
Following Grady & Kipp (1980), the crack
density parameter is assumed to be proportional λW p
to the product of N, the number of cracks of per D c = (9)
unit volume, is expressed as a Weibull statistical 1 − Dt
distribution function activated by the bulk strain
εv = (εx + εy + εz)/3, according to where λ is the sensitivity constant of Dc which
is taken to be equal to 1.0 × 10−3 kg/J in this
N k ( v )m (2) study (Sun 2002). While the plastic work Wp is
computed by
where k and m are the material constants. The aver-
∫ iijj d ε ij
p
age crack radius and the crack density parameter Wp (10)
are estimated from the nominal fragment radius for
dynamic fragmentation in a brittle material (Grady There are the tensile damage scalar and the
1983) as Eqs. (6) and (7) compress damage scalar in the model at the same

374

ISGSR2013.indb 374 10/18/2013 9:42:17 AM

 time. Introducing the symbol D into expressing well-known general purpose commercial FEM code
the damage scalar and considering the biggest that has been developed for simulating the nonlin-
damage effect, the damage scalar can be get ear dynamic response of structures. It has several
constitutive models such as *MAT_BRITTLE_
D max( Dt ,Dc ) (11) DAMAGE (Govindjee et al. 1995) and the H-J-C
model (Holmquist & Johnson 1993) designed for
In the exiting blasting damage models, the dynamic damage analysis of brittle rock. What is
degraded shear modulus G and bulk modulus K important, it provides user subroutine interface to
for a material point can be written as implement the userdefined model (LSTC 2010).

E E ( − D) (12)
3 SONIC WAVE TEST AND BLASTING
G G ( − D) (13) INDUCED DAMAGE DISTRIBUTION

The above method of determining the elastic To fully understand the damage distribution in
constants could not meet the mathematic relation the remaining rock mass near the berm, the cross-
between the elastic constants. Normally, the elastic hole sonic wave test was employed to determine
constants have the following relations: the extent of damage zone surrounding the berm
at the 640 m elevation of the Xiluodu high rock
E 2(1 μ )G (14) slope. The arrangement of the testing holes was
E 3(1 2 μ )K (15) shown in Figure 1. Three groups measuring results
were obtained in this test. One group of them was
The method of determining the elastic constants in the left and other two groups were in the right.
was modified in this study, the degraded bulk mod- The change of acoustic velocities and the damage
ulus and Poisson’s ratio were definited as the same depths are listed in Table 1.
as the origin method. According to the results of sonic wave test, the
acoustic velocities in the remaining rock mass near
K K ( − D) (16) the berm range from 1500 m/s to 5800 m/s. The
acoustic velocities decreases form the outer flank
⎛ 16 ⎞ to the inside of the berm. The minimum acoustic
μ μ exp − βCd (17)
⎝ 9 ⎠ velocity was found at the outer flank of the berm,
which means the rock mass in this part is the most
Other elastic constants were determined by the serious damaged.
mathematic relation: The outline of EDZ for rock mass near the
berm was showed in Figure 2. In the vertical
3(1 − 2 ) direction the damage depth ranges from 1.0 m to
G K (18)
2(1 + μ ) 4.0 m and decreases nonlinearity from the outer
flank to inside of the berm. In the horizon direction
9KG
E= (19)
3K G

The constitutive relations recording the damage

effect could be definited with Hooke’s law of incre-
ment as follow:

dσ ij Kdd ε kkδ iij

K Gdeij (20)

Based on the above analysis, a tensile-compress Figure 1. Schematic illustration of sonic wave test holes
damage model was established. LS-DYNA is a at the EL.640 m berm.

Table 1. The result statistics of sonic wave test.

Damage before Damage after

Section No. Rock mass Vp blasting blasting PPV

XZ640BH III1 3200∼5500 0∼0.50 1.00∼4.0 18.3

XYB640B III2 1500∼4200 0.00 0.25∼1.10 8.7
XYB640C III1 2500∼5800 1.50∼3.25 1.75∼4.50 9.7

375

ISGSR2013.indb 375 10/18/2013 9:42:21 AM

 Figure 3. Damage distribution of tensile-compress
Figure 2. Extent of damage zone in remaining rock at damage model.
the EL.640 m berm.

Table 2. Material constants and damage constants.

Density Elasticity Poison’s Dynamic tensile Damage Damage Damage

(kg/m3) modulus (GPa) ratio strength (MPa) constants (k) constants (m) KIC constants (λ)

2530 25 0.228 2 2.3 × 1024 7 0.92 0.0001

the damage radius is between 1.2 m and 3.0 m, and tensile-compress damage model for the rock mass
decreased nonlinearity from the top to the bottom of basalt were listed in Table 2. The material con-
of slope. The maximum vertical damage depth and stants such as elastic modulus, Poisson’s ratio et al
the maximum horizontal damage extent are found were determined according to the geological data
to be at the outer flank of the berm. and field experiments. To simulate borehole blast-
It can be seen that damage degree and extent of ing, a radial pressure was applied to the elements on
rock mass near the berm is more serious and much the borehole wall. The amplitude and the duration
larger than other parts of slope. There are three of the pressure wave can be determined according
induced factors for this: firstly, the rock mass near to the method proposed by Lu et al. (2011), the
the outer flank of the berm was affected by buffer density of explosive is 1000 kg/m3 and the detona-
blasting of former bench most seriously; secondly, tion velocity is 3200 m/s. To prevent the effects of
there are two free surfaces at the top of slope, reflected waves at the edge of the region, the trans-
which the enlarge degree of stress redistribution mitting boundaries were adopted in numerical sim-
and induce a more larger damage extent; thirdly, in ulation. Damping in the model is done through the
the current bench blasting excavation, the combina- Rayleigh classical approach by making the damp-
tion of effect the shock wave and detonation gas is ing matrix equal a linear combination of the mass
another main damage induced factor. To guarantee and stiffness matrix.
the safety of the slope, the reinforcement measure
should be carried out on the top of slope in time. 4.2 Numerical results
Figure 3 shows the damage distribution character-
istics of the former bench and presplitting blasting
4 NUMERICAL SIMULATION OF
of the current bench.
BLASTING INDUCED DAMAGE ZONE
It can be seen that the damage depth caused by
the former excavation bench is about 0.6 m. The
4.1 Numerical model
damage degree decreases from the outer flank to
In the present study, 3D numerical simulation of the inner side of the berm. After the presplitting
blasting induced damage was implemented with the blasting of underlayer excavation, a columnar
computer code LS-DYNA. The height of blasting damage zone was formed, the degree of dam-
excavation layer is 15 m depth. Diameter of the blast age decreases with the distance to the blast hole.
holes is 110 mm. The excavation region can be divided The rock mass close to the blast hole which was
into many same sections. One section contains five crushed by the shock wave was critical destroyed
presplit holes, two buffering holes and one produc- for the damage scalar equals 1.0, then the shock
tion hole. Considering the limit of mesh number and wave changed into stress wave and the damage
computation time, one section was selected to estab- degree decreases to about between 0.2 and 0.8. The
lish the numerical model. The number of nodes is maximum damage extent is located in the top face
566016 and the element number is 541400. of the blast hole.
Assuming to be isotropic and homogenous To verify the accuracy of the tensile-compress
material, the damage constants used in the blasting damage model, other four exiting blasting

376

ISGSR2013.indb 376 10/18/2013 9:42:24 AM

 damage models were used to make a comparison.
Yang et al. (1996) establishd a continuum dam-
age constitutive model which was widely used in
the numerical simulation of dynamic damage.
LS-DYNA has several damage models designed
for studying dynamic damage, the material Type
96 can be used to simulate the large strains, high
strain states and high pressures to which the rock
mass may be subjected in blasting. Especially, the
accumulative plastic strain was used to describe Figure 5. Damage distributions of four existing blast-
the degree of damage which is quite different from ing damage models.
other blasting damage models. Considering that
the tensile-compressive damage model was modi-
fied based on the TCK model and KUS model, so site measurement results was shown in Figure 5.
the comparison between the MAT96, the YANG The damage threshold Dcri is set as 0.19.
model, the TCK model, the KUS model and the It could be obviously seen that the result of the
tensile-compressive damage model was carried out tensile-compress damage model agrees with the test
in the next paragraphs. result well. The extents of the tensile-compress model,
The YANG model, the TCK model and the KUS YANG model, the TCK model and the KUS model
model were implemented into LS-DYNA though are bigger than that of sonic wave test. Respectively
its user subroutine interfaces with FORTRAN. For Errors of these models are 0.09 m, 0.123 m, 0.15 m
the MAT96, the viscosity constant and the frac- and 0.17 m. The result of the MAT96 is smaller than
ture toughness constant which were determined as that of test value and the error is as big as 0.35 m.
140.35 NM−1/2 and 0.723 MPaS−1 according to the It could be seen that the result of tensile-compress
keyword user’s manual of LS-DYNA. The calcula- damage model agreed with the observations better
tion procedure was the same as that of the tensile- than other damage models.
compress damage model. Figure 4 lists the results
of four existing damage models.
It can be seen that the distributions of the dam- 5 CONCLUSIONS
age zone for these damage models are overall
identical, but the extents are different. The dam- In the present study, the spatial distribution of
age radius of YANG model is the largest and the EDZ near the EL.640 m berm of Xiluodu high
TCK model takes the second place, while the KUS rock slope was determined by cross-hole test. The
model and the tensile-compress model are smaller. results reveal that an approximation columnar
The damage zone of MAT96 is the least. damage zone was formed near the excavation face.
Based on the cross-hole sonic wave result of The extent of EDZ and damage degree at the top
the berm at the EL.640 m of Xiluodu as shown of slope is much larger and more serious than other
in Figure 2, the quantitative accuracy of numerical parts. The horizontal damage depth decreases non-
simulation results could be confirmed. The com- linearly from the top to the bottom of slope, the
parison between these blasting damage models and vertical damage depth increases non-linearly from
the inner side to the outer flank of the berm, the
maximum horizontal damage radius and the maxi-
mum vertical damage depth are found to be at the
outer flank of the berm.
Based on the exiting blasting damage models,
the damage pattern of compress was introduced
and the method of determining the material elas-
tic constants was modified, and then a modified
tensile compress damage model was implemented
with the user subroutine interface of LS-DYNA.
The damage effect induced by blasting excavation
of the ET.640 berm of Xiluodu high rock slope
was simulated by using the modified model and
other four blasting damage models. Compared
with other widely used four blasting damage mod-
els, the result of the tensile-compressive damage
Figure 4. Damage distributions of four existing blast- model agreed with the observations better than
ing damage models. four blasting damage models.

377

ISGSR2013.indb 377 10/18/2013 9:42:25 AM

 However, the numerical simulation done in this larger strains, high strain rates and high pressure. In
paper is just a preliminary work. The isotropic and N. Jackson, S. Dickert (eds); 14th International Sym-
homogeneous damage model for the rock mass posium on Ballistics, Quebec, 26–29 September 1993.
is employed in the numerical simulation for the Kuszmaul, J.S. 1987. A new constitution model for
fragmentation of rock under dynamic loading. In
rock damage under blasting, while the anisotropic L. William, D.D. Richard (eds.); Proc. of 2nd Int. Symp.
and inhomogeneous of the rock mass in reality on Frag. by Blastin, Keystone, 23–26 August 1987.
are ignored. The site measurement of EDZ and Li, H.B., Xia, X. & Li, J.C. 2010. Rock damage control in
numerical simulations done in the present study bedrock blasting excavation for a nuclear power plant.
still provide a good reference for the similar high Int. J. Rock Mech. Min. Sci. 48(2): 210–218.
rock slope excavation engineering. Liu, L.Q. & Katsabanis, P.D. 1997. Development of a
continuum damage model for blasting analysis. Int. J.
Rock Mech. Min. Sci. 34(2): 217–231.
ACKNOWLEDGEMENTS LSTC, 2010. LS-DYNA theoretical manual. Livermore:
Livermore Software Technology Corporation.
LSTC, 2010. LS-DYNA keyword user’s manual, Version
This work is supported by Chinese National 970. Livermore: Livermore Software Technology
Science Fund for Distinguished Young Scholars Corporation.
(51125037), Chinese National Programs for Funda- Lu, W.B., Yang, J.H. & Chen, M. 2011. An equivalent
mental Research and Development (973 Program) method for blasting vibration simulation. Simulation
(2011CB013501), Chinese National Natural Modeling Practice and Theory 19(9): 2050–2062.
Science Foundation (51279146 and 51179138) and Lu, W.B., Yang, J.H., Yan, P., Chen, M., Zhou, C.B.,
the Fundamental Research Funds for the Central Luo, Y. & Jin, L. 2012. Dynamic response of rock
Universities (2012206020205). The authors wish to mass induced by the transient release of in-situ stress.
Int. J. Rock Mech. Min. Sci. 53: 129–141.
express their thanks to all supporters. Ma, G.W. & An, X.M. 2008. Numerical simulation of
blasting-induced rock fractures. Int. J. Rock Mech.
Min. Sci. 45(6): 966–975.
REFERENCES Malmgren, L., Saiang, D., Töyrä, J. & Bodare, A. 2006.
The excavation disturbed zone (EDZ) at Kiirunavaara
Chen, S.H., Wang, G.J., Zhou, H., Wang, W.M. & mine, Sweden-by seismic measurements. Int. J. Rock
Zou, L.C. 2012. Evaluation of excavation-induced Mech. Min. Sci. 61(1): 1–15.
relaxation and its application an arch dam founda- Martino, J.B. & Chandler, N.A. 2004. Excavation-induced
tion. International Journal For Numerical And Analyti- damage studies at the Underground Research Laboratory.
cal Methods In Geomechanics 36(2): 166–181. Int. J. Rock Mech. Min. Sci. 41(8): 1413–1426.
Englman, R. & Jaeger, Z. 1987. Theoretical aids for the Read, R.S. 2004. 20 years of excavation response studies
improvement of blasting efficiencies in oil shale and at AECL’s Underground Research Laboratory. Int. J.
rocks. Report of Soreq Nuclear Research Center. Rock Mech. Min. Sci. 41(8): 1251–1275.
Furlong, J.R., Davis, J.F. & Alme, M.L. 1990. Modeling Sheng, Q., Yue, Z.Q. & Lee, C.F. 2002. Estimating the
the dynamic load/unload behavior of ceramics under excavation disturbed zone in the permanent shiplock
impact loading. Final Report for U.S. ARMY slopes of the Three Gorges Project, China. Int. J. Rock
RESEARCH OFFICE. Mech. Min. Sci. 39(2): 165–184.
Goodman, R.E. & Kieffer, D.S. 2000. Behavior of rock Sun, Y.X. 2002. Studies on penetration resistance of
slopes. Journal of Geotechnical and Geoenvironmental concrete. Hefei: University of Science and Technology
Engineering 126(8): 675–684. of China.
Govindjee, S., Kay, G.J. & Simo, J.C. 1995. Anisotropic Taylor, L.M., Chen, E.P. & Kuszmaul, J.S. 1986. Micro-
modeling and numerical simulation of brittle dam- crack induced damage accumulation in brittle rock
age in concrete. Int. J. Numer. Methods Eng. 38(2): under dynamic loading. Computer Methods in Applied
3611–3633. Mechanics and Engineering 55(3): 301–320.
Grady, D.E. & Kipp, M.E. 1980. Continuum modeling Thorne, B.J., Hommert, P.J. & Brown, B. 1990. Experimen-
of explosive fracture in oil shale. Int. J. Rock Mech. tal and computational investigation of the fundamental
Min. Sci. 17(3): 147–157. mechanisms of cratering. In Proc. of 3rd Int. Symp. on
Grady, D.E. 1983. The mechanics of fracture under high- Frag. by Blasting, Brisbane, 26–31 August 1990.
rate stress loading. In Z.P. Bazant (ed.), Preprints Wang, Z.L., Wang, J.G. & Li, Y.C. 2006. Numerical study
of the William Prager symposium of mechanics of on craters and penetration of concrete slab by ogive-
geomaterials: rocks, concrete and soils: 149–188. nose steel projectile. Computers and Geotechnics 34(2):
Evanston: Northwestern University. 1–9.
Hao, H., Wu, C.Q. & Zhou, Y.X. 2002. Numerical analy- Wang, Z.L., Li, Y.C. & Wang, J.G. 2007. Numerical simu-
sis of blast-induced stress waves in a rock mass with lation of tensile damage and blast crater in brittle rock
anisotropic continuum damage models part 1: equiva- due to underground explosion. Int. J. Rock Mech.
lent material property approach. Rock Mechanics and Min. Sci. 44(5): 730–738.
Rock Engineering 35(2): 79–94. Yang, R., Brwden, W.F. & Katsabanis, P.D. 1996.
Holmquist, T.J. & Johnson, G.R. 1993. A computa- A new constitutive model for blast damage. Int. J.
tional constitutive model for concrete subjected to Rock Mech. Min. Sci. 33(3): 245–254.

378

ISGSR2013.indb 378 10/18/2013 9:42:25 AM

 Geotechnical Safety and Risk IV – Zhang et al. (eds)
© 2014 Taylor & Francis Group, London, ISBN 978-1-138-00163-3

Slope reliability analysis using a non-intrusive stochastic finite

element method

S.H. Jiang & D.Q. Li

State Key Laboratory of Water Resources and Hydropower Engineering Science, Wuhan University,
Wuhan, P.R. China

ABSTRACT: Reliability analysis of slope stability has received considerable attention. This paper aims
to propose a Non-Intrusive Stochastic Finite Element Method (NISFEM) for slope reliability analy-
sis with an implicit performance function. This procedure is formulated on the basis of the Stochastic
Response Surface Method (SRSM) and the deterministic finite element method. The probabilistic analy-
sis of slope stability is decoupled with the deterministic finite element analysis. A global sensitivity analy-
sis based on Sobol’s indices is carried out to identify the key factors influencing the factor of safety of a
slope. A practical rock slope example is presented to demonstrate the capability of the proposed method.
The results indicate that the proposed method can effectively evaluate the slope reliability with an implicit
performance function with a sufficient accuracy. The internal friction angles of the fault f42-9 and the class
III2 rock mass have a significant effect on the stability of the Jinping left abutment slope.

1 INTRODUCTION Simulation (MCS) can be used for such purpose.

However, it is prohibitively expensive because a
It is widely recognized that there are many uncer- large number of finite element analyses of slope
tainties in the analysis of slope stability because stability will be needed especially for the case of
of inadequate information of site characteriza- small probability level (e.g. Tamimi et al. 1989).
tion and inherent variability and measurement Recently, the Response Surface Methods (RSM)
errors in geological and geotechnical parameters have also been applied for slope reliability analysis
(Phoon & Kulhawy 1999, Baecher & Christian (Wong 1985, Cho 2009). However, an iterative cal-
2003). Traditional slope stability analysis using the culation of factor of safety was indispensable in
factor of safety method cannot effectively account the RSM for reliability analysis, making it difficult
for these uncertainties in a quantitative way. In to combine with the commercial finite element
order to remove such limitations, a new perspective codes. Therefore, many attempts are made to solve
concerning reliability-based method is developed. the slope reliability problems using the optimiza-
On the other hand, the finite element method and tion methods (Low & Tang 1997, Tang et al. 2012,
finite difference method have been extensively used Luo et al. 2012). However, the process of searching
for slope stability analysis (Matsui & San 1992, global optimization could be prohibitively expen-
Zou et al. 1995, Kim & Lee 1997, Farias & Naylor sive due to a substantial number of slope stability
1998, Griffiths & Lane 1999). In comparison with analysis required. Although the Stochastic Finite
the factor of safety method, these methods can Element Method (SFEM) can also be used for this
effectively represent the slope actual stress state, purpose (Ishii & Suzuki 1987, Ghanem & Spanos
meet the displacement compatibility condition 1991), it requires significant modification of exist-
and account for the effect of deformation of soil ing finite element codes underlying the determinis-
and rock masses on the slope stability. In this case, tic finite element model for slope stability.
the factor of slope safety obtained from finite The main objective of this study is to propose
element analysis cannot be explicitly expressed a non-intrusive stochastic finite element method
as a function of input parameters. The computa- for slope reliability analysis with an implicit per-
tional effort for complex slope stability problem formance function. This method is formulated
is always expensive. In this respect, both First on the basis of the Stochastic Response Surface
Order Reliability Method (FORM) and Second Method (SRSM) (Li et al. 2011, Mollon et al.
Order Reliability Method (SORM) are not applied 2011) and the deterministic finite element method.
directly due to the implicit performance function For illustration, the finite element stress-based
(Der Kiureghian et al. 1987). The Monte Carlo method (Farias & Naylor 1998) using the software

379

ISGSR2013.indb 379 10/18/2013 9:42:26 AM

 SIGMA/W and SLOPE/W is selected to compute Save the deterministic slope stability model file
the factor of slope safety. The SRSM is employed as one source FEM-FS *.xml input file, which
to perform the probabilistic analysis of slope can be directly viewed via the text editor.
stability. A practical rock slope example is pre- 3. The Hermite Polynomial Chaos Expansion
sented to illustrate the proposed method. (PCE) is adopted as a meta-model to approxi-
mate the implicit relationship between the
factor of slope safety FS and input uncertain
2 NON-INTRUSIVE STOCHASTIC FINITE parameters (Li et al. 2011, Mollon et al. 2011,
ELEMENT METHOD Mao et al. 2012):
A flowchart summarizing the links between the
( )+ ∑ ai i Γ 2 ( )
n n i1
probabilistic analyses and deterministic Finite FS (ξ) = a0 Γ 0 + ∑ i1 1 2
Element Analysis (FEA) of slope stability is pre- i1 = 1 i1 1 i2 = 1

sented in Figure 1. Several major steps of the

∑ ∑ ∑ ai i i Γ 3 (ξi )
n i1 i2

Non-Intrusive Stochastic Finite Element Method + 1 2 3 1

ξi2 , ξi3
i1 = 1 i2 = 1 i3 = 1
(NISFEM) are outlined as follows.
( )
n i1 i2 in −1

1. Identify the random variables and estimate their + ⋅ ⋅ ⋅ + ∑ ∑ ∑ ⋅ ⋅ ⋅ ∑ ai1 i2 , …,, in Γn ξi , ξi2 , …, ξin
i1 = 1 i2 = 1 i3 = 1 in = 1
means, Coefficients Of Variation (COV), distri-
bution types and correlations between them for (1)
the considered slope reliability problem.
2. Construct the FEA model of slope stability with where FS(ξ) is the vector of the factor of safety;
the mean values of the input random variables a = (a0, ai1, ai1i2, ⋅⋅⋅) are the unknown coefficients
using the software SIGMA/W and SLOPE/W. to be estimated, in which n is the number of ran-
dom variables representing the uncertainties of
the model inputs, and i1 + i2 + ⋅⋅⋅ + in ≤ n; ξi = (ξi1,
ξi2, ⋅⋅⋅, ξin) is the vector of independent stand-
ard normal variables resulting from the trans-
formation of the input uncertain parameters,
Xi; Γn(ξi1, ξi2, ⋅⋅⋅, ξin) is the multivariate Hermite
polynomials of degree n. (Li et al. 2011, Mollon
et al. 2011).
4. Generate and select Nc sets of collocation
points in the independent standard normal U
space, ξ = (ξ1, ξ2, …, ξNc) based on the linearly
independent principle. The linearly independent
Probabilistic Collocation Method (PCM) can
yield accurate results with much less computa-
tional cost. The number of runs of slope stability
model equals to the number of unknown coef-
ficients in the PCE of order p, Nc, Nc = (n + p)!/
(n! × p!). The reader is referred to Mao et al.
(2012) for details.
5. Map the selected collocation points ξ in the
independent standard normal U space into the
collocation points X in the physical space. If
the input random variables are correlated non-
normal variables, the Nataf transformation
(Nataf 1962) can be used to transform ξ into X.
Then, the Nc sets of collocation points X can be
obtained and taken as input parameters for the
deterministic FEA model of slope stability.
6. Replace the corresponding mean values of
input random variables in the source FEM-FS
*.xml file generated in step (2) with the Nc sets
of the collocation points X. Then Nc different
new FEM-FS *.xml input files can be gener-
Figure 1. Flowchart of the NISFEM for slope reliabil- ated. Thus, no programming effort is required
ity analysis. to modify the existing finite element codes in

380

ISGSR2013.indb 380 10/18/2013 9:42:26 AM

 contrast to the popular spectral stochastic finite
element method (Ghanem & Spanos 1991).
7. Run the SIGMA/W and SLOPE/W with each
new input file generated in step (6) to perform
deterministic FEA of slope stability. Note that
these input files can be executed automatically
with the help of the WinbatchTM software. Once
the solution process is finished, the correspond-
ing result files are obtained. Then, post-process-
ing can be carried out to extract the factor of
slope safety FS, FS = (FS1, FS2, …, FSNc).
8. Establish a linear algebraic system of equations
in terms of the unknown coefficients a accord-
ing to Eq. (1) with the ξ and FS, and solve it to
obtain the unknown coefficients of PCE.
9. Once the coefficients in the PCE is obtained, the
implicit performance function can be explicitly
expressed by a meta-model for slope reliability
Figure 2. A deeply cut V-shaped valley at the dam site
analysis, G(ξ) = FS(ξ) − 1. The probability of of Jinping I hydropower station.
slope failure and the corresponding reliability
index can be estimated for the performance
function involving the output responses repre-
sented by the Hermite PCE. The first four sta-
tistical moments and Sobol’s indices can also
be directly calculated using the coefficients of
PCE. In this study, the direct MCS with a sam-
ple size of 105 is used for such purpose. It should
be pointed out that the evaluation of the factor
of slope safety does not involve finite element
model runs, but only the evaluation of simple
algebraic expressions, which is much more com-
putationally efficient.

3 RELIABILITY ANALYSIS OF JINPING

SLOPE IN CHINA WITH AN IMPLICIT
PERFORMANCE FUNCTION

Jinping I hydropower station is located at the

upstream of Yalong River, between Muli and
Yanyuan counties in Sichuan Province, China. The
double curvature concrete arch dam is 305 m in
height upon completion (Wu et al. 2008). The dam Figure 3. The FEM model for the section II1-II1 of
site is located in a typical deeply cut V-shaped valley natural slope at the Jinping left abutment.
and the left abutment slope is very high and steep,
as shown in Figure 2. The lamprophyre dike X,
faults f5, f8 and f42-9 and some other weak geological slope stability model considered, the specific unit
structural features are found in the left abutment weights γi, Young’s modulus Ei and Poisson’s ratios
slope. Besides, the relief fractures develop widely in νi of rock masses and structural planes as well as
the subsurface of slope, even up to 100∼200 m in the parameters related to geometry are treated as
depth. It can be observed that the geological con- deterministic quantities, their values are listed in
ditions at the dam site are extremely complicated, Table 1. The cohesions and internal friction angles
which lead to poor stability conditions of the left of the materials of some weak structural planes
abutment slope. and rock masses closely related to the slope stabil-
A typical section, Section II1-II1, of the left abut- ity are identified and treated as random variables.
ment slope shown in Figure 3 is selected for relia- The statistical parameters of 10 random variables
bility analysis. Note that the slope being considered of slope materials are summarized in Table 2. ci and
here is the natural slope before excavation. For the φi (i = 1, 2, …, 5) denote the cohesions and internal

381

ISGSR2013.indb 381 10/18/2013 9:42:27 AM

 Table 1. The mechanical parameters of rock masses and weak structural planes in the Jinping
left abutment slope stability model.

Materials γi (kN/m3) Ei (GPa) νi ci (kPa) φi (°)

Class II rock mass 27 23.5 0.225 2000 53.47

Class III1 rock mass 27 10.5 0.25 1500 46.94
Class III2 rock mass 27 5.5 0.275 900 45.57
Class IV1 rock mass 27 2.5 0.3 600 34.99
Class IV2 rock mass 27 1.9 0.3 400 30.96
Lamprophyre dike X 25 1.0 0.2 20 16.7
Fault f42-9 25 0.45 0.25 20 16.7
Fault f5 25 0.45 0.35 20 16.7
Fault f8 25 0.45 0.35 20 16.7

Table 2. Statistics of random variables in the Jinping as well as the topography of the computational
left abutment slope stability model. domain, the Finite Element Method (FEM) (Zou
et al. 1995, Kim & Lee 1997) is used to calculate the
Random factor of safety of the left abutment slope because
Materials variables Mean COV
this approach can account for the influence of the
Lamprophyre dike X c1 (kPa) 20 0.25 deformation of rock masses on the slope stability.
φ1 (°) 16.7 0.15 Theoretically, a Three-Dimensional (3D) model
Fault f42-9 c2 (kPa) 20 0.3 for the slope stability problem considering spatial
φ2 (°) 16.7 0.2 variability of rock masses should be employed.
Class III2 rock mass c3 (kPa) 900 0.15 For simplicity, a finite element model based on
φ3 (°) 45.57 0.08 2D plane strain analysis is built in the SIGMA/W
Class IV1 rock mass c4 (kPa) 600 0.18 with the means of random variables, as shown in
φ4 (°) 34.99 0.10 Figure 3. The analyzed domain is 420 m in height
Class IV2 rock mass c5 (kPa) 400 0.2 from the elevation 1650 m to the elevation 2070
φ5 (°) 30.96 0.12 m and 384.7 m in length of the direction perpen-
dicular to the river from the centerline of rive to
the within slope. The gravity loads induced by the
self-weight of rock masses are applied. For illustra-
friction angles of five types of materials, respec- tive purposes, a conventional elastic and perfectly
tively. The means of the shear strength parameters plastic model based on the Mohr-Coulomb failure
are determined based on field test, laboratory test criterion is adopted to represent the stress-strain
supplemented with engineering judgment, such as behavior of rock masses and structural planes. The
direct shear test and triaxial test (Wu et al. 2008). finite element model consists of three-node trian-
The Coefficients of Variation (COV) of the shear gular and four-node quadrilateral hybrid elements.
strength parameters are adopted from the litera- There are totally 14363 zones and 14113 grid points
ture (Phoon & Kulhawy 1999, Baecher & Christian in the mesh. With regard to the boundary condi-
2003, Tang et al. 2012). Additionally, all random tions, the horizontal movement on the vertical left
variables are assumed to be characterized statisti- boundary of the mesh is restrained, whereas the
cally by a lognormal distribution. The slope sta- base of the mesh is not allowed to move in both the
bility is evaluated under two working conditions, horizontal and the vertical directions. The initial
namely natural condition and rainfall condition. stress field is obtained using the SIGMA/W. Then,
For the natural condition, the groundwater table is it is imported into the SLOPE/W for slope stability
below the slip surface of the slope. If the ground- analysis. Based on the geological investigation, it
water table is above the slip surface of the slope, can be expected that a potential deep sliding may
the resulting load should be taken into account, occur along the lamprophyre dike X and fault f42-9,
which corresponds to the rainfall condition. The and shear off superficial rock masses in accord-
pore-water pressure ratio Ru = 0.1, a ratio of the ance to the path highlighted with a red dashed line
pore-water pressure to the weight of overburden in Figure 3. For the left abutment slope with the
rock masses is used to account for the effect of prescribed slip surface, the factors of slope safety
rainfall on the stability of the left abutment slope. obtained from the FEM are 1.083 and 0.981 for
In order to model the structures and the gen- the natural and rainfall conditions, respectively,
eralization of geomechanical model of the slope which are consistent with 1.110 and 1.014 using

382

ISGSR2013.indb 382 10/18/2013 9:42:29 AM

 the Morgenstern-Price method. These results fur-
ther demonstrate that the FEM is an effective tool
for slope stability analysis.
As mentioned in Section 2, the NISFEM can
effectively decouple the probabilistic analysis
with the deterministic FEA of slope stability. It is
employed to analyze the considered slope reliabil-
ity problem with an implicit performance function.
Firstly, the order of PCE is determined through
comparing the differences in the probability of fail-
ure between two consecutive order PCEs. Table 3
shows the probabilities of slope failure obtained
from the 2nd and 3rd order PCEs for the natural
and rainfall conditions. The results of the direct
MCS with a sample size of 105 are also provided in
Table 3, which is used for the benchmark purpose. Figure 4. Comparison among Sobol’s indices of input
Note that both the computational accuracy and random variables for two working conditions.
cost associated with the NISFEM increase with
the order of PCE. There is a slight difference in
the probabilities of failure obtained from the 2nd
conditions are plotted in Figure 4. It is observed
and 3rd order PCE. However, the number of finite
that the internal friction angle φ2 of the fault f42-9 is
element model runs for the latter is about four
the most significant random variable for the slope
times larger than that for the former. In addition,
stability under two working conditions, followed
the probabilities of slope failure obtained from the
by φ3, and the other random variables have a slight
2nd order PCE match well with those obtained
influence on the slope stability because their Sobol’
from the direct MCS. The computational cost for
indices are below 0.06.
the former is just 1/1515 of the latter. As for the
example considered, a 2nd order PCE with less
computational cost is adopted for slope reliability
4 CONCLUSION
analysis and subsequent sensitivity analysis. Addi-
tionally, the probability of failure under the rain-
This paper has proposed a non-intrusive stochastic
fall condition has been greatly increased to 61%,
finite element method for analyzing slope reliabil-
thus some reinforcement measures should be taken
ity problems with implicit performance functions.
to improve the slope stability.
A practical rock slope example is investigated to
A global sensitivity analysis based on Sobol’s
demonstrate the validity of the proposed method.
indices (Mao et al. 2012) is conducted herein to
Several conclusions can be drawn from this study.
determine the contribution of each random vari-
able to the variation of factor of slope safety. The 1. A non-intrusive stochastic finite element method
corresponding results for the natural and rainfall for slope reliability analysis is proposed. The pro-
posed method does not require the user to modify
existing deterministic finite element codes, which
Table 3. Comparison of the reliability results from the are used as “black boxes”. Moreover, the probabi-
NISFEM and MCS for the left abutment slope. listic analysis and the deterministic finite element
pf ε = |pf − pf,MCS|/
analysis are decoupled. The proposed method
Conditions Methods (%) pf,MCS × 100 (%) provides a practical tool for reliability problems
requiring complex finite element analysis.
Natural 2nd order PCE 13.20 0.99 2. The non-intrusive stochastic finite element
(66 FE runs) method can efficiently evaluate the slope reli-
3rd order PCE 13.17 0.77 ability with an implicit performance function.
(286 FE runs) It can reduce the number of calls to the deter-
MCS 13.07 – ministic finite element model substantially and
(105 FE runs) is much more efficient than the direct Monte
Rainfall 2nd order PCE 60.64 0.85 Carlo simulations. For the studied example, the
(Ru = 0.1) (66 FE runs) 2nd order PCE requiring 66 runs of the deter-
3rd order PCE 61.03 0.21 ministic finite element model can produce suf-
(286 FE runs)
ficiently accurate reliability results.
MCS 61.16 –
(105 FE runs) 3. The results of sensitivity analysis based on
Sobol’s indices indicate that the internal friction

383

ISGSR2013.indb 383 10/18/2013 9:42:29 AM

 angle of the fault f42-9 is the most significant ran- Li, D.Q., Chen, Y.F., Lu, W.B. & Zhou, C.B. 2011.
dom variable for the slope stability, followed by Stochastic response surface method for reliability anal-
that of the class III2 rock mass, and the other ysis of rock slopes involving correlated non-normal
random variables have a slight influence on the variables. Computers and Geotechnics 38(1): 58–68.
Low, B.K. & Tang, W.H. 1997. Efficient reliability
slope stability because their Sobol’ indices are evaluation using spreadsheet. Journal of Engineering
below 0.06. Therefore, the fault f42-9 and the Mechanics 123(7): 749–752.
class III2 rock mass should be given consider- Luo, X.F., Li, X., Zhou, J. & Cheng, T. 2012. A Kriging-
able attention in the following slope excavation based hybrid optimization algorithm for slope reliabil-
and reinforcement. ity analysis. Structural Safety 34(1): 401–406.
Mao, N., Tamara, A.B. & Soubra, A.H. 2012. Probabilis-
tic analysis and design of strip foundations resting on
ACKNOWLEDGMENTS rocks obeying Hoek-Brown failure criterion. Interna-
tional Journal of Rock Mechanics and Mining Sciences
49(1): 45–58.
This work was supported by the National Basic Matsui, T. & San, K.C. 1992. Finite element slope sta-
Research Program (973 Program) (2011CB013506), bility analysis by shear strength reduction technique.
the Academic Award for Excellent Ph.D. Soils and Foundations 32(1): 59–70.
Candidates Funded by Ministry of Education Mollon, G., Dias, D. & Soubra, A.H. 2011. Probabilistic
of China (5052012206001) and the Fundamental analysis of pressurized tunnels against face stability
Research Funds for the Central Universities using collocation-based stochastic response surface
(2012206020201). method. Journal of Geotechnical and Geoenvironmen-
tal Engineering 137(4): 385–397.
Nataf, A. 1962. Détermination des distributions de prob-
abilité dont les marges sont données. Comptes Rendus
REFERENCES de l’Académie des Sciences 225: 42–43.
Phoon, K.K. & Kulhawy, F.H. 1999. Characterization
Baecher, G.B. & Christian, J.T. 2003. Reliability and sta- of geotechnical variability. Canadian Geotechnical
tistics in geotechnical engineering. New York: John Journal 36(4): 612–624.
Wiley and Sons. Tamimi, S., Amadei, B. & Frangopol, D.M. 1989. Monte
Cho, S.E. 2009. Probabilistic stability analyses of slopes Carlo simulation of rock slope reliability. Computers
using the ANN-based response surface. Computers and Structures 33(6): 1495–1505.
and Geotechnics 36(5): 787–797. Tang, X.S., Li, D.Q., Chen, Y.F., Zhou, C.B. & Zhang, L.M.
Der Kiureghian, A., Lin, H.Z. & Hwang, S.J. 1987. 2012. Improved knowledge-based clustered partition-
Second order reliability approximations. Journal of ing approach and its application to slope reliability
Engineering Mechanics 113(8): 1208–1225. analysis. Computers and Geotechnics 45: 34–43.
Farias, M.M. & Naylor, D.J. 1998. Safety analysis using Wong, F.S. 1985. Slope reliability and response surface
finite elements. Computers and Geotechnics 22(2): method. Journal of Geotechnical Engineering 111(1):
165–181. 32–53.
Ghanem, R. & Spanos, P.D. 1991. Stochastic finite ele- Wu, F.Q., Qi, S.W., Song, S.W. & Gong, M.F. 2008.
ment: a spectral approach. New York: Springer- Deformation and stability study of complicated high-
Verlag. steep rock slopes—take Jinping I Hydropower Station
Griffiths, D.V. & Lane, P.A. 1999. Slope stability analysis for an example. Beijing: Science Press.
by finite element. Geotechnique 49(3): 387–403. Zou, J.Z., Willians, D.J. & Xiong, W.L. 1995. Search for
Ishii, K. & Suzuki, M. 1987. Stochastic finite element critical slip surfaces based on finite element method.
method for slope stability analysis. Structural Safety Canadian Geotechnical Journal 32(2): 233–246.
4(2): 111–129.
Kim, J.Y. & Lee, S.R. 1997. An improved search strategy
for the critical slip surface using finite element stress
fields. Computers and Geotechnics 21(4): 295–313.

384

ISGSR2013.indb 384 10/18/2013 9:42:29 AM

 Geotechnical Safety and Risk IV – Zhang et al. (eds)
© 2014 Taylor & Francis Group, London, ISBN 978-1-138-00163-3

A novel bond contact model for rock and its calibration

T. Jiang, M.J. Jiang, H. Chen, F. Liu & Z.M. Shi

Faculty of Civil Engineering, Department of Geotechnical Engineering, Tongji University, Shanghai, China
Key Laboratory of Geotechnical and Underground Engineering of Ministry of Education, Tongji University,
Shanghai, China

ABSTRACT: A novel bond contact model for rocks was proposed, based on a series of laboratory tests
on bonded granules idealized by two glued aluminum rods. The model was then implemented into a two-
dimensional distinct element code to numerically carry out the uniaxial and biaxial compression tests on
Lac du Bonnet granite. The results were compared to the experimental data and the DEM simulations
using an existing Bonded-Particle Model (BPM). The results demonstrate that the proposed model can
reproduce the strength envelope of Lac du Bonnet granite adequately. The slope of the strength envelope
obtained from our simulation is higher than that obtained with the BPM due to higher peak shear and
rolling strength in the proposed model, and that the ratio of shear and rolling failure to tension failure is
also higher.

1 INTRODUCTION This paper aims to propose a new bond contact

model based on the experiments on the analogous
The Distinct Element Method (DEM), originally cemented samples. This model was then imple-
developed by Cundall (1971), is a numerical simu- mented into a two-dimensional distinct element
lation technique based on the discrete mechanics. code in order to numerically carry out a series of
The original application of the DEM was in the compression tests. The proposed model was cali-
fields of granular material and blocky rock sys- brated by comparing the DEM simulation results
tems (Cundall & Strack 1979). Nowadays, DEM to the published experimental data.
is widely used due to its simple principle in many
aspects of geomechanics, such as granular mechan-
2 A BOND CONTACT MODEL FOR ROCK
ics (Thornton 2000), constitutive models for gran-
ular material (Jiang et al. 2005) and anisotropy
2.1 A conceptual bond contact model
of soil (Anandarajah 2000). In addition, DEM is
also employed to model rock behavior by treating Jiang et al. (2006) proposed a bond contact model
crystalline rocks as cemented granular materials. through theoretical derivation by assuming that
The Bonded-Particle Model (BPM) proposed by the bond contact width is continuously distributed
Potyondy & Cundall (2004) was widely applied to with the normal/tangential basic elements. The
simulate mechanical behavior of rocks in the past mechanical responses are illustrated in Figure 1.
decades. Although the BPM is able to reproduce Figure 1a presents the normal mechanical
most mechanical behavior of Lac du Bonnet gran- response. In both tensile and compressive direc-
ite in DEM uniaxial compression tests, the mac- tions, the normal contact force Fn between two
roscopic responses of biaxial compression tests particles increases linearly with the normal dis-
with established microscopic parameters do not placement un, while in the tensile direction, after
match the experimental data (Cho et al. 2007). In Fn reaches the tensile bond strength Rt, the bond
order to improve the BPM, Wang & Tonon (2009) is broken and Fn is abruptly reduced to zero.
developed a distinct element code to model the Figure 1b presents the tangential mechanical
mechanical behavior of Lac du Bonnet granite in response. The tangential force Fs increases linearly
triaxial compression; Christian et al. (2011) pro- with the tangential displacement us until the shear
posed a progressive failure model for DEM and strength Rs is reached, then the bond breaks and Fs
reproduce many features of rock failure observed drops to the residual frictional shear strength. In
in published experiments. However, the bond con- Figure 1c, the mechanical response in the rolling
tact models proposed by most researchers neces- contact direction is similar to that in the tangential
sarily contain hypotheses lacking experimental direction, the moment M increases linearly with
verification. the rolling rotation angle θ first, and then it drops

385

CH49_8-3.indd 385 10/18/2013 6:08:23 PM

 cemented samples based on typical interaction
relationships between particles.

2.3 Strength envelope

2.3.1 Strength envelope for cemented sample
with bond thickness of 0.6 mm
Figure 3a and b presents the strength envelopes
obtained from the shear tests under different
normal forces and the rolling tests under differ-
ent normal forces on the cemented samples with
bond thickness of 0.6 mm respectively. The solid
lines and dash lines shown in Figure 3 repre-
sent the peak strength and the residual strength,
respectively. Figure 3 shows that the peak shear
and rolling strengths depend on the applied normal
force. With increasing normal force Fn, the peak
shear strength Rs and the peak rolling strength Rr
increase first, and then decrease until the compres-
sive strength Rc is reached, where the peak strength
envelopes intersect with the residual strength enve-
lopes. When Fn exceeds Rc, bond is broken and the
residual strength increases linearly as the normal
force increases. The strength envelopes are formu-
lated as follows:

⎧ Rc
Figure 1. Schematic illustration of the mechanical ⎪μb R R
responses of the bond contact model: (a) normal contact ⎪ c t
⎪ ⎡ ⎛ R Rt ⎞ s ⎤
f
model; (b) tangential contact model; (c) rolling contact Rs = ⎨ ⋅ (Fnb Rt ) ⋅ ⎢1 + gs ⋅ ⎜ ln c ⎥ , Fn ≤ Rc ()
model (Jiang et al. 2006, 2012a). ⎪ ⎢⎣ ⎝ Fnb Rt ⎟⎠ ⎦⎥
⎪F , Fn > Rc (2)
⎪ n b
to the residual strength when the rolling bond ⎩
strength Rr is reached.
⎧1 Rc
⎪ 6 β br R R ⋅ ( Fnb Rt )
2.2 Experimental setup ⎪ c t
⎪ ⎡ ⎛ R Rt ⎞ r ⎤
f
In order to validate the contact law of bonded gran- Rr = ⎨ ⋅ ⎢1 + gr ⋅ ⎜ ln c ⎥, ≤ (3)
⎝ Fnb Rt ⎟⎠ ⎥⎦
n c
ules as well as verify the mechanical responses of ⎪ ⎢⎣
the model, Delenne et al. (2004) first presented an ⎪1
experimental investigation on mechanical behavior ⎪ Fn β br , n > c ( 4)
⎩6
of cemented granules by performing simple tests
on a pair of aluminum rods glued together with
where Rt is tensile strength; μb is friction coeffi-
epoxy resin. Taking the normal force and the bond
cient of cement; βb is rolling resistance coefficient
material into account, Jiang et al. (2012a, b) con-
of cement; r = 2r1r2/(r1 + r2) is the common radius
ducted a series of simple and complex loading tests
of two contact particles with radii of r1 and r2;
on a pair of aluminum rods glued with a rock-like
material, cement, in addition to epoxy resin.
The analogous cemented sample adopted in
the experiment is shown in Figure 2. The alumi-
num rods glued together with cement are 12 mm
in diameter and 50 mm in length. The bond of this
sample is 0.6 mm thick and 3 mm wide. Five dif-
ferent loading paths (tension, compression, shear
under different normal forces, rolling under dif-
ferent normal forces and shear-rolling tests under
different normal forces) were applied to charac- Figure 2. Analogous cemented sample used in experi-
terize the mechanical behavior of the analogous ments (Jiang et al. 2012a, b).

386

ISGSR2013.indb 386 10/18/2013 9:42:29 AM

 where knp is inter-particle normal contact stiffness,
knb is inter-bond normal contact stiffness.
When the normal force is tensile, only inter-
bond contact is present and the strength envelopes
are therefore the same as expressed in Equations
(1) and (2). While the normal force is compressive,
the peak shear and rolling strength can be regarded
as a combination of inter-bond and inter-particle
strength, and can be formulated as follows:

⎧ Rc
⎪ μb R R ⋅ ( Fnb Rt )
⎪ c t
⎪⎪ ⎡ ⎛ R Rt ⎞ s ⎤
f
Rs = ⎨ ⋅ ⎢1 + gs ⋅ ⎜ ln c ⎥ + μ p Fnp Fn Rccbr (7 )
⎝ Fnb Rt ⎟⎠ ⎥⎦
np
⎪ ⎢⎣
⎪
⎪ μ p Fnp
n b Fnnb , Fn Rccbr (8)
⎪⎩

⎧1 Rc
⎪ 6 β b r R R ⋅ ( Fnb Rt )
⎪ c t
Figure 3. Strength envelopes for cemented samples with ⎪ ⎡1 + gr ⎤
bond thickness of 0.6 mm in terms of: (a) shear strength ⎪ ⎢ f ⎥ 1
Rr = ⎨ ⋅ ⎢ ⎛ Rc Rt ⎞ r ⎥ + β p Fnp , 0 < n ≤ cbr (9 )
vs. normal force; (b) rolling strength vs. normal force. ⎪ ⎢⋅ ⎜⎝ ln F ⎟ ⎥ 6
⎪ ⎣ nb Rt ⎠ ⎦
⎪1 1
⎪ 6 β p np + 6 ⋅ β b rFnnb
and fs, gs, fr, gr are fitting parameters that control , Fn > Rccbr (10 )
the envelope shapes, whose values are 0.986, 2.15, ⎩
0.761 and 3.055 respectively.

2.3.2 Strength envelope for cemented sample

with zero bond thickness
Note that the real microstructure of Lac du Bonnet
granite consists of mineral grains and micro voids
between mineral grains, and the distance between
two contact particles is so small that it can be approx-
imated as zero (Jiang et al. 2006, Lan et al. 2010).
Thus the contact characteristics are similar to that of
the cemented sample with zero bond thickness.
Due to the poor flowability of cement, it is much
more difficult to prepare cemented samples with
zero bond thickness in laboratory. Thus, the strength
envelopes for cemented sample with zero bond
thickness were obtained from the theoretical analy-
sis based on the experimental results of cemented
samples with bond thickness of 0.6 mm. The com-
plete strength envelope for cemented sample with
zero bond thickness is presented in Figure 4.
The contact in cemented samples with zero
bond thickness is assumed to be composed of
inter-particle contact and inter-bond contact, and
they transmit the applied normal compressive
force in parallel mode. Thus, the inter-particle con-
tact force Fnp and inter-bond contact force Fnb can
therefore be calculated by:
Figure 4. Strength envelopes for cemented samples
Fnp Fn ⋅ knp knp + knb ) (5) with zero bond thickness in terms of: (a) shear strength
vs. normal force; (b) rolling strength vs. normal force; (c)
Fnb Fn ⋅ knb knp + knb ) (6) shear strength vs. rolling strength.

387

ISGSR2013.indb 387 10/18/2013 9:42:31 AM

 where μp is inter-particle friction coefficient; βp 3 CALIBRATION OF THE BOND
is inter-particle rolling resistance coefficient; CONTACT MODEL
Rcbr is critical normal force for cemented sam-
ples with zero bond thickness which can be In order to calibrate the proposed model, the bond
expressed as: contact model comprising the strength envelopes
for cemented samples with zero bond thickness
knb knp has been implemented into a two-dimensional dis-
Rcbr = Rc tinct element code using C++ language. A series of
knb uniaxial compression tests and biaxial compres-
(11)
sion tests were then simulated and the results were
Note that when Fn is compressive but less than compared with the experimental data obtained by
Rcbr, if the strength envelopes are exceeded, the Martin (1993) and Martin & Chandler (1994).
bond breaks and the shear force or rolling resist-
ance drops to their residual strengths. When Fn 3.1 Generation of DEM rock sample
exceeds Rcbr, the bond is already broken and the
peak strength envelopes coincide with the residual A great number of simulations were conducted to
strength envelopes. find the appropriate microscopic parameters of mate-
In addition, if shear, rolling and normal force rial and the particle size distribution. They were even-
are present simultaneously, the strength envelope tually determined, as shown in Table 1 and Figure 5.
would be a three-dimensional surface, which is The multilayer undercompaction method pro-
different with the curved lines shown in Figure 4a posed by Jiang et al. (2003) was employed for its
and b. The relationship between the peak shear capability of controlling homogeneity and density.
strength and the peak rolling strength obtained The total number of particles in each numerical sam-
from the shearing-rolling tests under specific nor- ple is 10000, which ensures that the macro mechani-
mal force exhibits elliptical shape as presented cal behavior will not significantly changes with more
in Figure 4c, and the corresponding theoretical particles. After the consolidation process, the DEM
equation is: rock sample was generated, as shown in Figure 6.
Note that the bonds are represented as solid lines
connecting the corresponding particle centers.
Fs2 M 2
+ =1 (12)
Rs2 Rr2 Table 1. Microscopic parameters used in DEM
simulations.

2.4 Comparisons between the BPM Parameter Value

and the proposed model
Particle density ρs (kg/m3) 2700
Concerning the strength envelopes in the BPM, Initial void ratio e 0.20
the peak shear strength keeps constant firstly, Normal stiffness of particles kn(N/m) 1.8 × 1011
and then increases linearly with increasing nor- Tangential stiffness of particles ks(N/m) 9.47 × 1010
mal force, while the peak rolling strength increases Tensile strength of bonds Rt (N) 6.5 × 104
linearly with increasing normal force. In the pro- Compressive strength of bonds Rc (N) 8.0 × 107
posed model, the peak shear and rolling strengths Inter-particle friction coefficient μp 1.0
increase as the normal force increases in parabolic Inter-particle rolling resistance coefficient βp 1.5
mode first, and then increase linearly with increas- Friction coefficient of bonds μb 0.5
ing normal force. Thus, under the same com- Rolling resistance coefficient of bonds βb 0.5
pressive normal force, the peak shear and rolling
strengths in our proposed model are higher than
that in the BPM.
In addition, bond failure modes in each model
are different, namely tension and shear failure
(for which peak tensile or shear strength has been
exceeded respectively) in the BPM, while ten-
sion failure (for which tensile strength has been
exceeded), shear and rolling failure (for which
either the peak strength envelopes shown in
Figure 4a and b has been exceeded) and compres-
sion failure (for which the critical normal force Rcbr
has been exceeded) in the proposed model.
More details about the BPM can be referred to Figure 5. Particle size distribution used in the DEM
Potyondy & Cundall (2004). analyses.

388

ISGSR2013.indb 388 10/18/2013 9:42:33 AM

 3.2 Simulation of the uniaxial compression test
The uniaxial compression test was simulated
with the micro parameters in Table 1. Figure 7
presents the stress-strain and the number of broken
bonds-strain relationships. The peak strength
shown in Figure 7 is about 200 MPa, which is in
consistent with the published experimental data
(Martin 1993). Figure 7 also shows that, with
increasing axial strain, the axial stress increases lin-
early up to the peak value, where the slopes of the
number of broken bonds curves change obviously.

3.3 Simulation of the biaxial compression test

The biaxial compression tests were performed under
confining pressures of 1 MPa, 5 MPa, 10 MPa, 20
MPa and 30 MPa. Figure 8 illustrates the corre-
sponding stress-strain and number of broken bonds-
strain relationships. Figure 8 shows that the axial
stress increases linearly as the axial strain increases
up to the peak value, where the slopes of the number
of broken bonds curves change obviously. Figure 9
presents the strength envelopes obtained by labo-
ratory tests, DEM simulations with the proposed
model and with the BPM, whose micro parameters
are chosen by a calibration process with Lac du
Bonnet granite. It can be indicated that the peak
axial stress increases linearly with increasing confin-
ing pressure, but the test results using the proposed
model are more consistent with experimental data
than the BPM, which cannot provide a high peak
stress when confining pressure is high.
Figure 10 presents the ratio of shear and rolling-
induced bond failure to tension-induced bond fail-
ure at the peak axial stress under different confining
Figure 8. Stress-strain relationship accompanied with
number of broken bonds-strain relationship obtained
from DEM biaxial compression tests under confin-
ing pressure of: (a) 1 MPa; (b) 5 MPa; (c) 10 MPa;
(d) 20 MPa; (e) 30 MPa.

Figure 6. A DEM rock sample and generated bonds.

Figure 9. Strength envelopes obtained from DEM

numerical simulations and laboratory tests.

pressures. Figure 10 indicates that with increasing

confining pressure, the ratio of shear and rolling
Figure 7. Stress-strain relationship accompanied with failure to tension failure in the proposed model
number of broken bonds-strain relationship obtained increases in a parabolic mode, while in the BPM,
from DEM uniaxial compression test. although the number of shear failure and tension

389

ISGSR2013.indb 389 10/18/2013 9:42:34 AM

 REFERENCES

Anandarajah, A. 2000. On influence of fabric anisotropy

on the stress–strain behaviour of clays. Computers and
Geotechnics 27(1): 1–17.
Cho, N., Martin, C.D. & Sego, D.C. 2007. A clumped
particle model for rock. International Journal of Rock
Mechanics & Mining Sciences 44(7): 997–1010.
Christian, E., Robert, S. & Peter, E. 2011. A discrete ele-
ment model to describe failure of strong rock in uniax-
ial compression. Granular Matter 13(4): 341–364.
Cundall, P.A. 1971. A computer rock model for simulat-
Figure 10. Ratio of shear (and rolling) failure to tension ing progressive large scale movements in blocky rock
failure at the peak axial stress under different confining systems. In Proc. Int. Symp. on Rock Fracture(ISRM):
pressures. vol. 1, 11–8. France: Nancy.
Cundall, P.A. & Strack, O.D.L. 1979. A discrete element
model for granular assemblies. Geotechnique 29(1):
failure both increase with increasing confining 47–65.
pressure, the ratio of shear failure to tension failure Delenne, J.Y., Youssoufi, M.S.E., Cherblanc, F. &
increases slowly and remains low under high con- Benet, J.C. 2004. Mechanical behaviors and failure of
fining pressure. In addition, although the tensile cohesive granular materials. International Journal for
strength keeps constant in our DEM simulations, Numerical and Analytical Methods in Geomechanics
the shear strength and rolling strength depending 28(15): 1577–1594.
on the normal force increase with increasing con- Jiang, M.J., Harris, D. & Yu, H.S. 2005. Kinematic
fining pressure. Moreover, the shear strength and models for non-coaxial granular materials: part II:
evaluation. International Journal for Numerical and
rolling strength under a specific normal force in
Analytical Methods in Geomechanics 29(7): 663–689.
the proposed model are much higher than that in Jiang, M.J., Konrad, J.M. & Leroueil, S. 2003. An effi-
the BPM, whose shear strength is independent of cient technique for generating homogeneous speci-
normal force. Thus, the strength envelope obtained mens for DEM studies. Computers and Geotechnics
from our simulations can provide a higher slope, 30(7): 579–597.
which is more approximate to the experimental Jiang, M.J., Sun, Y.G., Li, L.Q. & Zhu, H.H. 2012a.
data shown in Figure 9. Contact behavior of idealized granules bonded in
two different interparticle distances: An experimental
investigation. Mechanics of Materials 55: 1–15.
4 CONCLUSIONS Jiang, M.J., Sun, Y.G., Xiao, Y. 2012b. An experimen-
tal investigation on the mechanical behavior between
A series of experiments were conducted on a pair cemented granule. Geotechnical Testing Journal 35(5):
of cemented aluminum rods in laboratory in order 678–690.
to simulate the mechanical behavior of rock via Jiang, M.J., Yu, H.S. & Harris, D. 2006. Bonds rolling resist-
ance and its effect on yielding of bonded granulates by
DEM. Based on the experimental data, a bond
DEM analyses. International Journal of Numerical and
contact model was developed and implemented Analytical Methods in Geomechanics 30(8): 723–761.
into a distinct element code. Simulations of uniax- Lan, H., Martin, C.D., & Hu, B. 2010. Effect of hetero-
ial and biaxial compression tests were carried out geneity of brittle rock on micromechanical extensile
with this code, and the results were compared to behavior during compression loading. Journal of Geo-
the published experimental data of Lac du Bonnet physical Research 115(B01202): 1–14.
granite for model calibration. The results show Martin, C.D. 1993. The strength of massive Lac du
that the proposed bond contact model is able to Bonnet granite around underground openings. Canada:
model a high slope of strength envelope that is National Library of Canada.
Martin, C.D. & Chandler, N.A. 1994. The progres-
more approximate to the experimental data, and
sive failure of Lac du Bonnet granite. International
videlicet, the proposed bond contact model was Journal of Rock Mechanics and Mining Sciences &
successfully calibrated. Geomechanics Abstracts 31(6): 643–659.
Potyondy, D.O. & Cundall, P.A. 2004. A bonded-
particle model for rock. International Journal of Rock
ACKNOWLEDGMENT
Mechanics & Mining Sciences 41(8): 1329–1364.
Thornton, C. 2000. Numerical simulation of deviatoric
The research was funded by China National Funds shear deformation of granular media. Géotechnique
for Distinguished Youth Scientists with Grant No. 50(1): 43–53.
51025932, and the Major Project of Chinese National Wang, Y. & Tonon, F. 2009. Modeling Lac du Bonnet gran-
Programs for Fundamental Research and Develop- ite using discrete element model. International Journal of
ment (973 Program) with Grant No. 2011CB013504. Rock Mechanics & Mining Sciences 46(7): 1124–1135.
All of these supports are greatly appreciated.

390

ISGSR2013.indb 390 10/18/2013 9:42:36 AM

 Geotechnical Safety and Risk IV – Zhang et al. (eds)
© 2014 Taylor & Francis Group, London, ISBN 978-1-138-00163-3

Study on cracking risk of Jinping high and steep slope

P. Lin & X.L. Liu

State Key Laboratory of Hydroscience and Engineering, Tsinghua University, Beijing, China

ABSTRACT: By employ physical experiment and numerical simulation, this study focusses on cracking
risk, the cracking stability and the reinforcement effectiveness of the high and steep left hand bank of the
Jinping dam. The study results relating to the major slope at Jinping showed that: (1) The bending and
toppling cracks in the unreinforced zone of the slope as well as strong relaxation and tension cracks were
extended simultaneously. The rock cracking and slope failure controlling factors could be directly seen in
the structurally weak rock mass. (2) The controlling failure mode for the Jinping steep and high left bank
slope consists of the integral catastrophic instability mode, whereby the slope energy is totally dissipated
and the slope destabilizes at a limit state. The slope cracking factor of safety and the integral stability fac-
tor of safety satisfy the stability requirements. (3) The reinforcement has been effective in keeping cracks
closed in the rock, maintaining the slope stability, and increasing the slope’s inherent safety factor.

1 INTRODUCTION failure due to the unloading of the rock mass.

Such activity constitutes a hidden danger to the
Slope cracking, failures occur throughout the safety of the dam.
world and contribute to economic losses and The Jinping hydropower station is located at a
casualties. The impact of slope failures is often sharp bend in the Yalongjiang River at Jinping,
undervalued. Every year, the United States in the middle reaches, in Sichuan province, P.R.
experiences more than $1 billion in damage and China. The slope of the left hand bank at the
approximately 50 deaths (David 2008); world- Pusiluogou dam site is steep and over 1000 m
wide, slope failures cause hundreds of billions of high. The slope consists of relatively hard mar-
dollars in damage and hundreds of thousands of ble between El. 1900 m and El. 1690 m and rela-
deaths. The Vajont reservoir slope failure disaster tively weak sandy slate between El. 1900 m and
(Kiersch 1964), as a result a wave over topped the El. 2300 m. The maximum height of excavation
dam by 250 m and swept onto the valley below, at the slope reaches 200~300 m. The faults f5,
with the loss of approximately 2500 lives. Many f8, f42-9, SL44-1 and lamprophyre X are well
slope failures have occurred during the abutment developed, as seen in Figure 1. The deep frac-
excavation process for dams related to hydroelec- tures in the rock are especially disadvantageous
tric projects in China (Huang 2008). For example, to the stability of the left bank slope at the dam
the Jinlongshan landslide at the Ertan reservoir; site, as their strikes are distributed parallel to the
the downstream slope failure at the Lijiaxia arch slope, and the steep slope is already loosened and
dam. The abundance of such socioeconomic significantly splayed. The fault f5, fault f8 lam-
losses justifies the necessary allocation of funds prophyre X are relatively longer and wider being
for slope stability research. A series of super-high poor in property. SL44-1 is a deep tensile crack
arch dams, such as Jinping, Xiaowan, Xiluodu, at the upstream boundary. The fault f42-9 is at
Laxiwa et al, are being built in the alpine-gorge the downstream boundary (bottom slide face),
areas of China (Lin et al. 2011). A question which with lamprophyre X as the internal sliding face.
frequently arises in hydraulic engineering discus- These faults may produce a wedge failure mode,
sions on slope stability is how high and how steep which controls the deformation and stability of
can an excavated rock slope be (Wang et al. 2007, the abutment slope.
Pan & He 1998). How to evaluate the cracking and In this study, the cracking, stability and rein-
stability risk of a major slope after excavation for forcement of the slope at the left hand abutment
abutments has so far been a key dam design and was analysed using the new geomechanism model
foundation treatment issue. Poorly conducted test method and numerical method. The compari-
excavation not only delays construction but also son of the numerical simulations and the experi-
causes secondary rock mass cracking and slope mental observations was carried out.

391

ISGSR2013.indb 391 10/18/2013 9:42:36 AM

 Figure 2. Schematic diagram of slope test assembly.

to include the simulation of weak structure planes,

lamprophyre fault veins and the deep fractures in
the left hand abutment and similarity of bound-
ary constraints. The Jinping slope geomechanical
model was built by laying gypsum briquettes on
to a test bed made of tamped barite powder and
bentonite. The mechanical parameters of the rock
mass, its faults and weak structural planes at the
Jinping site are shown in Table 1. In this study, a
small block masonry technique was employed for
Figure 1. Schematic failure mode and simulated range the manufature of the Jinping slope model. As
of left bank slope (Unit: m). both the material and model tests are performed
in a normal environment, the effect of temperature
change to glue’s strength is not considered.
2 STUDY METHODS AND MODEL

2.1 Experimental method and model 2.2 Numerical method and model

The physical experiments were performed at FEM (Finite Element Method) was used to simu-
the State Key Laboratory of Hydroscience and late the stability of the high and steep left hand
Hydraulic Engineering at Tsinghua University. It bank of the Jinping dam in this paper. Generally,
consists of a rotational structure system, a defor- strength reduction FEM is often applied to analyze
mation monitoring, cracking and data collection the slope stability numerically recent years. But the
system and an analysis system (Zhou 2008). The strength reduction FEM is not suitable to com-
slope model is built on a rotating support system, pare with the physical experiments because it is not
and can be rotated to any angle of θ around the axis easy to operate experiments with reduced mate-
OY by hydraulic lift (as shown in Fig. 2). Rotation rial properties or parameters. In order to compare
of the test platform at a velocity of 0.5°/min. When the numerical modeling method with the physical
the angle reaches 5°, the excavation of the abut- experiments, the same technique is used in numeri-
ment groove in the model and recording of the cal model, just like what had been done in physical
slope deformation of every measuring point is then modeling. That is, rotating the numerical model to
conducted. The testing platform is continually some angle and calculating the stress condition of
rotated at a velocity of 0.5°/min, until the cracks the whole model until slip surfaces occur.
are initiated on the slope, and then coalescence is Also the following equation is utilized to cal-
propagated until ultimate failure occurs. culate the slope stability factor. This traditional
The simulation scale and faults are shown in method can be used to compare with the method
Figure 1. The model range extends 180 m upstream, developed in this paper.
270 m downstream, and is 600 m wide, taking in
left bank. The simulated elevations and heights are Fanti sliding
K= (1)
1600~2250 m and 660 m, respectively, and sufficient Fsliding

392

ISGSR2013.indb 392 10/18/2013 9:42:36 AM

 Table 1. Mechanical parameters of rock mass main faults.

Width E0 (Mpa) f C (Mpa)

Prototype Model Prototype

Name (m) (cm) × 103 Model Prototype Model Prototype Model

II 23∼31 100 1.35 1.35 2 0.007

III1 9.2∼14.6 49 1.07 1.07 1.5 0.005
III2 6.4∼10.2 34 1.02 1.02 0.9 0.003
IV1 2.56∼1.64 9 0.7 0.7 0.6 0.002
IV2 1.4∼2.4 8 0.6 0.6 0.4 0.0013
V 0.37∼0.82 2.7 0.3 0.3 0.02 0
Fresh X 6.5 22 0.8∼1.0 0.9 0.64 0.002
3.0 1.0
Weathered X 3.0 10 0.55∼0.65 0.6 0.45 0.0015
f5 1.0 0.33 0.4 1.4 0.3 0.3 0.02 0
6.0 2.0 0.4 1.4 0.3 0.3 0.02 0
f2 0.5 0.17 0.4 1.4 0.3 0.3 0.02 0
F1 6.0 2.0 0.4 1.4 0.3 0.3 0.02 0
Compressive zone 0.3 0.1 0.4 1.4 0.3 0.3 0.02 0
SL44-1 10–15 4 2 6.7 0.45 0.45 0.1 0.0003
f42-9 0.3–1.0 0.24 0.4 1.4 0.3 0.3 0.02 0

Figure 3. Slope with multiple sliding structural surfaces

and the stress analysis of the ith potential sliding body.

where, Fanti-sliding and Fsliding denote the anti-sliding

force and the sliding force, respectively. For a slope
with multiple structural surfaces (shown in Fig. 3),
the stability factor of the ith potential sliding body
is described by the following equation

(ci Fi + fi Ni ) + Pi
K′ =
Pi ( i i ) Wi sin α i
[ci Fi + fi Wi cos α i + Pi − i ( i −1 − α i ))] + Pi
=
Pi − ( i − i − ) + Wi sin α i
(2) Figure 4. Numerical model of Jinping left bank slope.

393

ISGSR2013.indb 393 10/18/2013 9:42:38 AM

 where αi denotes dip of the ith structural surface.
The numerical model is shown in Figure 4.

3 THE CRACKING RISK ANALYSIS

3.1 Experimental results

In the experiment, the first step was to simulate
excavation for the dam abutment. The excavation
of the dam foundation slot was undertaken at a
model rotation angle of 5 degrees. Cracking pat-
terns and images upstream and downstream of
the dam are shown in Figures 5 and 6. As shown
in Fiures 5 and 6, cracking first appears along
the faults and jointed rock masses above eleva-
tion 1960 m and below elevation 1885 m in those
unreinforced areas upstream. When the rotation
angle reaches 28°, fractures initiate, propagate and

Figure 6. Crack initiation, propagation and slope fail-

ure process (view from downstream surface).

cause the slope to suddenly collapse. The fractures

mainly occur in the trailing edges of reinforced
zones. In the abutment slot and at the mountain
top where there is no reinforcement, sliding occurs
along fault lines f5. Upstream and downstream
slide surfaces in two directions caused by faults
SL44-1 and f42-9 and rock mass were observed
(see Fig. 5b). The detailed slope failure processes
are summarized below.
When the slope angle is 5 degrees, no crack-
ing is seen. Then, during excavation of the abut-
ment slot, no significant cracking occurs, but local
cracks appear in those blocks intersected by faults
f42-9 and SL44-1. When the slope angle is about
18 degree, fractures occur along fault lines and
joints up and downstream of the abutment slot,
where there is no reinforcement above elevation
1960 m and below elevation 1885 m (see Fig. 6a).
These cracks propagate along a structurally weak
plane, and at the boundary between rock masses
of grade IV and grade III2, as shown in Figure 6a.
They propagate downstream along strata aligned
in the opposite direction and structurally weak
planes, as shown in Figure 6b.
When the slope angle is between 25 and 28
degrees, multiple cracks occur in lamprophyre X
in the downstream abutment slot. It then occurs
along fault f5 at upstream abutment slot. Slip then
occurs below EL 1885, i.e. the triangle cut by faults
Figure 5. Schematic diagram of cracking of the up- f42-9, and SL44-1, and ultimately it occurs along
downstream surface of slope model (red lines represent lamprophyre X and rock mass joints. When the
cracking). slope angle exceeds 28°, toppling failure occurs

394

ISGSR2013.indb 394 10/18/2013 9:42:40 AM

 quickly with great release of energy. The model
blocks were crushed, accompanied by a loud fail-
ure noise, as shown in Fig. 6d.

3.2 Numerical results

By numerical modeling, the crack initiation, prop-
agation and slope failure process are analyzed.
Figures 7 and 8 show the slide faces occur gradu-
ally when the numerical model is rotated.
The experimental and numerical results show
that The bending and toppling cracks in the unrein-
forced zone of the slope as well as strong relaxation
and tension cracks were extended simultaneously.
The rock cracking and slope failure controlling
factors could be directly seen in the structurally
weak rock mass. The shear hole and anchor cable
reinforcement system, meets the standard overall
slope safety factor requirements, allowing also for
losses over time to anchorage resisting forces and
to initially applied prestress.

Figure 8. Slope failure modes in different cross sections

by numerical modeling.

Based on the information gained the most likely

failure mode of the left hand slope above the cable
machine platform is the overall slide, mainly caused
by the large blocks lamprophyre X, the faults f42-9
and the deep fissures SL44-1. The test results show
Figure 7. Crack initiation, propagation and slope fail- that the shear hole and anchor cable reinforcement
ure process by 3 dimension numerical modeling. systems enable the overall slope stability safety

395

ISGSR2013.indb 395 10/18/2013 9:42:41 AM

 factor to meet the standard requirements allowing Science Foundation of China (No: 11272178), and
for some anchorage slippage and some loss of pre- Special Foundation for public welfare industry of
stress over time. the Ministry of Water Resources (2010001035).
The field monitoring results indicate that the sta- Special thanks to Prof. Zhou W.Y. and Yang
bility of the abutment has been effectively attained R.Q. for experimental support.
and the adjacent slope is stable due to the reinforce-
ment applied. The slope deformations have settled to
negligible amounts with no further significant degree REFERENCES
of deformation at levels below 1885 m. Further slow
rates of deformation were still observed after com- Chen, Z.Y. 2003. Principles method and program of soil
pletion of the excavation of the topping slope above slope stability analysis. Beijing: China Water Power
1885 m elevation. Nevertheless the rate of deforma- Press.
tion is slowing down appropriately, and converging David, W. 2006. Innovative solutions for slope stability
reinforcement and characterization. Research report of
upon an ultimate steady state. PGA Iowa State University.
Huang, R.Q. 2008. Geodynamical process and stability
control of high rock slope development. Chinese
4 CONCLUSIONS Journal of Rock Mechanics and Engineering 27(8):
1525–1544.
An experimental and numerical cracking, collapse Kiersch, G.A. 1964. Vaiont reservoir disaster. Civil
study of the Jinping Dam left hand steep and high Engineering 34: 32–39.
slope adjacent to the abutment has been success- Lin, P., Wang, R.K., Kang, S.Z., Zhang, H.C. & Zhou,
fully carried out, and the following conclusions W.Y. 2011. Study on key problems of foundation
failure, reinforcement and stability for super high
can be drawn arch dams. Chinese Journal of Rock Mechanics and
1. The controlling failure mode for the Jinping Engineering 30(10): 1945–1958.
high left bank slope is one of overall cata- Liu, X.L., Wang, E.Z., Han, G.F. & Wang, S.Y. 2012.
strophic instability whereby the slope energy is Discontinuous deformation analysis for the slope
stability in Jinping First Stage Hydropower Station,
dissipated totally and the slope destabilizes at Southwestern China. Disaster Advances 5(4):
the ultimate limit state. 1481–1485.
2. Large areas reingforced at the abutment upper Pan, H.Y., He, J.D. & Zhang, L. 1998. Application of
elevations to counter the unloading actions due material strength reserve method to the analysis of
to the excavation of rock laid above the dam stability of high rock slope. Chinese Journal of Sichuan
platform have effectively improved the safety Union University 2(1): 13–18.
factor against sliding, and thus have increased Tan, X.L., Xu, W.Y. & Liang, G.L. 2009. Application of
overall safety factor of the slope. Deformations extenics method to comprehensive safety evaluation
of this slope are converging gradually to an of rock slope. Chinese Journal of Rock Mechanics and
Engineering 28(12): 2503–2509.
approximate steady state. Wang, R.K., Lin, P. & Zhou, W.Y. 2007. Cracking and
3. The study results show that the shear hole and stability problems of high arch dams on complicated
anchor cable reinforcement system, meets the rock foundations. Chinese Journal of Rock Mechanics
standard overall slope safety factor require- and Engineering 26(10): 1951–1958.
ments, allowing also for losses over time to Zheng, Y.Z. & Zhao, S.Y. 2006. Discussion on safety
anchorage resisting forces and to initially factors of slope and landslide engineering design.
applied prestress. Chinese Journal of Rock Mechanics and Engineering
25(9): 1933–1940.
Zhou, W.Y., Lin, P., Yang, Q., Yang, R.Q. & Zhou, Z.
2008. Experimental research on stability of Jinping
ACKNOWLEDGEMENTS high slope with three-dimensional geomechanical
model. Chinese Journal of Rock Mechanics and
This research work was supported by National Engineering 27(5): 893–901.
Basic Research Program of China (973 Program)
Grant No. 2011CB013503, National Natural

396

ISGSR2013.indb 396 10/18/2013 9:42:43 AM

 Geotechnical Safety and Risk IV – Zhang et al. (eds)
© 2014 Taylor & Francis Group, London, ISBN 978-1-138-00163-3

Mechanism and numerical simulation on geological

mechanical model test

X.Q. Luo, J.F. Bi & H. Shen

School of Naval Architecture, Ocean and Civil Engineering, Shanghai Jiaotong University, Shanghai, China

ABSTRACT: Reference to the principle of employing centrifugal force field to simulate gravity field, the
geomechanics magnetic model test is a new test method that simulates gravity field with electromagnetic
force field to study geomechanical engineering problems. Using electromagnetic method generates a mag-
netic field where the magnetic flux density gradient keeps a fixed nonzero value, where the mixture of fer-
romagnetic material and geological mechanics prototype material can get a uniform electromagnetic force.
The FEM software for 3D electromagnetic field numerical simulation is used to study the influence of the
electromagnetic parameters and magnetic factors to the magnetic field intensity, which provides reason-
able design parameters for the production of the experimental machine. The results of simulation indicate
that the geomechanics magnetic model method can effectively compensate for the landslide weight loss
caused by the reduced-scale and nicely reflect the characteristics of the deformation and failure.

1 INTRODUCTION test and permeability model test considering the

role of water medium. In ordinary gravitational
At the beginning of the 20th century, some coun- field, the model test scale is limited, and the stress
tries started on structural model test research, and level at each point is still lower than prototype,
gradually established the similarity theory. In the as well as the soil stress-strain relationship, espe-
1960s, experts, led by Fumagalli (1979), carried on cially nonlinear relationship, in the model cannot
the groundbreaking engineering geology mechani- get true embodiment, so that the reliability of the
cal model test research in Italy structural model test results mainly depends on the similarity of the
test research institute. Ashby (1971) qualitatively similar material and the dependability of the data
studied the dumping slope failure mechanism and acquisition. Domestic and foreign research results
its process by adopting tilt table model technology. show that although the model test technology has
Soon afterwards, in Portugal, the former Soviet gained greater development, due to the hardness
union, France, Germany, Britain, the United on obtaining the similarity material of soft granu-
States, Japan and other countries, scholars also lar media corresponding to similarity theorem, the
developed model test researches (Stewart et al. development of geomechanics model test technol-
1994, Baumgargarter & Stimpson 1979, Bray & ogy has been severely restricted (Fumagalli 1979).
Goodman 1981), which also get rapid and exten- In order to overcome the geomechanics model
sive development in China (Chen 1984, Shen 2001, test for the presence of these defects, in the six-
Luo 2005). Conventional model test is a simulation ties and seventies of the 20th century geotechni-
test which is taken in the 1g gravitational field and cal centrifuge model test method has been widely
reduces the prototype size n times according to the used in the world, carry out a large amount of
geometric similarity. However, in reality it is diffi- studies around the soft soil engineering, and get
cult to convert gravity stress level into 1/n times of fruitful results (Bao 1991, Pu 1996). Geotechnical
prototype by no change of its physical and mechan- centrifuge model test is based on the principle of
ical properties. So the test can only partly meet the simulating gravity field by centrifugal force field.
geometric and mechanics similarity between the The centrifugal force applied to the model can
physical model and prototype, and cannot fully make the volume-weight of the model larger, so
reflect the global deformation and failure char- that each point stress in the model can keep con-
acteristics of prototype. Although conventional sistent with the prototype. This feature results in
model test gained great success in brittle materi- essential difference between the centrifugal model
als (like concrete and rock) simulation test and soil test and the conventional model test. At present
(especially in soft soil) destruction test, there are a centrifuge model test technology can improve
lot of difficulties in soft soil elastic-plastic model the acceleration of the model to 400 g level or

397

ISGSR2013.indb 397 10/18/2013 9:42:43 AM

 higher. Prototype soil by appropriate treatment convenient (Luo 2009a, b, Luo 2011). This idea is
and similar material can be used for experimental the further development of the field similarity the-
research. Because n g simulated gravity field can be ory of the centrifuge model tests.
obtained in the centrifuge model tests, the require- The geomechanics magnetic model test is a new
ments for the mechanical properties of the mate- test method, which not only can be compared with
rial greatly reduced compared to the conventional the conventional geomechanics model test and the
model test, so there is a qualitative leap in reduc- geotechnical centrifuge model test, but also can
ing the difficulty of the test and improving the overcome the inherent shortcomings and defi-
reliability of the test results (Zhang & Dou 1995, ciencies to establish a comprehensive test theory.
Bao & Rao 1998). However, geotechnical centri- There is a broad application prospect in structural
fuge model test also exist some defects. First, the deformation and its stability study of hydraulic
centrifugal inertia force field and gravitational field engineering, civil engineering, traffic engineering,
cannot be completely similar. The centrifugal force marine engineering and other areas.
field is proportional to the radius of inertia, so that
the acceleration, the magnitude and direction, of
each point on the model is different and chang- 2 BASIC PRINCIPLES
ing, which make the centrifugal model test cannot OF GEOMECHANICS MAGNETIC
obtain a uniform force field. So there are some dif- MODEL TEST
ferences with the actual engineering. Secondly, the
generated Coriolis acceleration in the feeding proc- The geomechanics magnetic model test is an entirely
ess also causes centrifugal model error. Finally, the new geomechanics model test method established by
irregular change of the stress state caused by the using electromagnetic force to simulate gravity, and
centrifuge accelerating and braking is also one electromagnetic force field to simulate the gravita-
of the problems existing in centrifuge model test tional field. The test similarity criteria can be derived
method. The process of enhancing and reducing based on the three basic principles of the similarity
the speed of rotation will take some time, which theory (Chen 1979, Chen 1980). Using the dimen-
is also not consistent with the engineering practice sional analysis and equation derivation can obtain
and causes some impact on the reliability of the the similarity criterion of each parameter for:
test results. φ* = ∈* = μ* = 1, p* = σ* = 1, ρ* = 1, k* = ν* = q* = 1,
The two above model test methods have achieved t*= 1, l* = u* = 1, g* = 1/n. In the model, the time t
tremendous development, but still have some short- and geometric dimension l are reduced n times, the
comings. Centrifuge model test with the idea of using gravitational acceleration g expanded n times, while
the “centrifugal force field” to simulate “gravita- the similar ratios of other physical and mechanical
tional field” represents the development direction of parameters (E, μ, c, φ, ∈, σ, k, v, q, p) of the material
the model test. At present, the new patterns of geo- are 1. If we can provide the conditions to expand
mechanics model test research are mainly concen- g n times of the model material, we can carry out
trated in the “field” simulation. Tsinghua University the model test by using prototype material or simi-
used to investigate the drag force produced by the lar material after appropriate processing; thereby
seepage force to increase the bulk density of the soil reduce the difficulty of the model test.
(Ding et al. 1994, Fang et al. 1997, Huang & Wang Magnetic material will be subjected to magnetic
1998). The principle is to use the superposition of force in a non-uniform magnetic field. Let m stand
“gravitational field” and “seepage force field” to for the mass of the magnetic substance, Ms for the
simulate the “gravitational field”, and obtained saturation magnetization intensity, σs for the satu-
some positive results. But this method requires that ration magnetization intensity of unit mass, V for
the object of study must be saturated body, the per- volume, and B for the magnetic flux density of the
meability coefficient must reach a certain degree, magnetic field, then the force for an arbitrary axial
and free boundary must also be plane. However, α ( i.e. x, y, z ) in a non-uniform magnetic field is
the use of “external force field” to simulate “grav-
ity field” provides a direction for the development ∂B ∂B
of the geomechanics model test theory. The core Fα M sV or Fα σ sm (1)
idea of the geomechanics magnetic model test is to ∂α ∂α
use the similarity of the “magnetic field” and the
“gravitational field”, mix the particulate magnetic When the magnetic force, Fα, of m (kg) material
material in the rock-soil body to get n g “gravita- which has been magnetized to Ms in the magnetic
tional field” by the “magnetic field”. If considering field equals to n times of the gravity, we can get
alternating electromagnetic field, we can also get
+n g or –n g simulated gravitational field, making ∂B m 1
Fα = n g ⇔ = ng ⋅ ⋅ (2)
the simulation of dynamic loading easier and more ∂α V Ms

398

ISGSR2013.indb 398 10/18/2013 9:42:43 AM

 Clearly, if required Fα to be a fixed value of
n times the gravity, the magnetic field gradient
required to satisfy the equation (2). That is to say,
the magnetic flux density B should linearly change
along α direction.
As for iron powder, its saturation magnetization
intensity Ms is 1703 × 103 A/m2 under normal tem-
perature, and density is 7800 kg/m3. Substituting
into formula (2), then:

∂B
∂α
(
= n × 0.044 T ⋅ m −1 ) (3)

So if it is wanted to make iron powder be sub-

jected to the magnetic force, n times of gravity,
and the magnetic flux density gradient should be
n × 0.044 T/m.
Theoretically, the magnetic force Fα should be a
linear relationship with ∂B/∂α when the material is
magnetized to saturation intensity. Increasing the Figure 1. Test equipment sketch.
value of the magnetic field gradient, the magnetic
force will increase with it, and the simulated stra-
tum stress level is also higher.
The purpose of loading or overloading can
be achieved by increasing the current to enhance
∂B/∂α (α axial) value, which can give the stabil-
ity evaluation of the slope and dam foundation
structure. As well as, by fixing the ∂B/∂α (α axial)
value of gradient magnetic field test area, we can
research the deformation and failure mechanism
of slope and dam foundation under the action of
water load or excavation load.

3 IMPLEMENTATION OF UNIFORM
GRADIENT MAGNETIC FIELD
FOR GEOMECHANICS MAGNETIC
MODEL TEST

In other to get the uniform magnetic flux density

gradient magnetic field and meet the requirement
of the test space, two electricity solenoids are
employed with radius of 0.5 m and height of 1 m.
The test equipment sketch is shown in Figure 1,
and set the underside elevation 0 for convenient. In Figure 2. Magnetic flux density of test area.
the middle of two solenoids is 1 m height test area.
The coil turns of the two solenoids are same, and the
current is in the same direction. The magnetic field According to the test need, take the lower coil
distribution of the test area can be controlled current 7 times of the upper coil to simulate and
by adjusting the current. Calculated by the finite analyze. Figure 2 depicts the magnetic flux density
element software, when the lower coil current is 2 distribution of one test area longitudinal profile.
times of the upper part, the magnetic flux density of The coordinates are the test area space coordinates
test area appears obvious gradient. When the mag- in m. It shows that the magnetic flux density in the
netic dipole moment produced by the upper coil middle region changes uniformly by gradient of
reaches the level of the ordinary permanent magnet 3 T/m–5 T/m.
and the current of the lower coil is 5 times of the Magnetic flux density and its gradient along
upper one’s, the magnetic flux density gradient of the axis of the cylindrical test area are shown in
the intermediate 1/3 test area reaches 2.5 T/m. Figures 3 and 4, respectively. Figure 4 indicates the

399

ISGSR2013.indb 399 10/18/2013 9:42:44 AM

 measured in mm. The parameters of the test model
will be the same as the slope body in situ.
Considering the uniformity of the magnetic
force on the model, let the central axis of the test
area pass through the undersurface centroid of the
model. Set the model bottom level to 1.4 m, which
can make the model be located in the best test area.
Suppose the volume percentage of iron powder is
33.3%, and the soil volume percentage is 66.7%. Mix
them well. Let μrFe donate the relative permeability
of iron powder, and μrSoil donate the relative perme-
ability of soil. The equivalent relative permeability
of the mixture can be calculated by the equivalent
Figure 3. Magnetic flux density on the axis of the test magnetic resistance. Because of the inappreciable
area. contribution of the μrSoil, the equivalent relative per-
meability can be approximately taken as μrFe/3.
The body force on the mixing-material in the
magnetic field equals the sum of its own gravity
and the magnetic force. According to the volume
percent of iron, the magnitude of the magnetic
force is one third of the situation that the material
is pure iron powder. The weight of the magnetic
model is 776.16 N, which equals the whole volume,
0.044 m3, times the density, 1800 kg/m3. Applying
the magnetic field shown in Figure 2 on the model,
the resultant force can be calculated by using the
volume integral of the body force on the model.
The magnitude of the resultant force is −79472.4 N,
about 102.4 times of gravity. The minus means the
force direction is downward.
Figure 4. The magnetic flux density gradient on the Figure 6 shows the distribution of the body
axis of the test. force. It can be seen the magnetic force on the

magnetic flux density changes greatly at both ends, Table 1. Material parameters.
i.e., the gradient is bigger corresponding to the
Figure 4. In the height of 1.4 m–1.8 m, namely the Name Value Unit
distance from the test area bottom is 0.4 m–0.8 m,
the change of magnetic flux density gradient is Young’s modulus 10e6 Pa
gentle, between 3.1 T/m–3.7 T/m, where can be Poisson’s ratio 0.3 1
taken for the best area for test. Density 1800 kg/m3
Cohesion 42e3 Pa
Angle of internal friction 17 Degree
4 NUMERICAL SIMULATION OF
GEOMECHANICS MAGNETIC MODEL
TEST OF A TYPICAL LANDSLIDE

A three-dimensional slope, on a scale of 1:100,

is used for numerical simulation of geomechan-
ics magnetic model. The geomechanics magnetic
model will be restored to its original size to ana-
lyze the mechanical behavior under the action of
its own gravity by finite element method and the
results will be compared with the geomechanics
magnetic model. According to the similarity the-
ory, the displacement ratio should be 1:100, and
stress and strain ratio should be 1:1.
The material parameters are listed in Table 1 and
the shrunken model is dimensioned in Figure 5, Figure 5. The dimension of slope model.

400

ISGSR2013.indb 400 10/18/2013 9:42:45 AM

 whole model is between 85 times and 115 times
of gravity, as well as the area exceeding 100 times
is located in middle and bottom parts, while the
upper left and right parts are a little smaller, cor-
responding to the distribution of the magnetic flux
density gradient.
Figures 7–9 are displaying the contrast of dis-
placement, stress, and strain, respectively. Through
the contrast of displacement, it can be seen that
the ratio on the corresponding position of the
magnetic model and prototype is 1:100. Likewise,
the ratio of stress and strain is both 1:1. These all
coincide with similarity theorem of model test.
The effective plastic strain, calculated based
on the Drucker-Prager yield criterion, is located
at the slope foot both on the magnetic model
and prototype. The maximum effective plastic
strain of magnetic model is about 2.5 times of the

Figure 8. The contrast of stress.

Figure 6. The distribution of body force.

Figure 9. The contrast of strain.

prototype. The difference between the magnitudes

of the two models can be blamed on the uneven
distribution of the magnetic force and can get rea-
sonable explanation from Figure 6. The magnetic
force at the slope body is more than 100 times of
gravity, while the ratio at the upper left corner is
less than 100. Compared with the prototype, the
Figure 7. The contrast of displacement. slope body of magnetic model is in the condition

401

ISGSR2013.indb 401 10/18/2013 9:42:47 AM

 of overloading while the slope foot is unloading, Bray, J.W., Goodman, R.E., Chen, J.J., Feng, X.B. 1986.
which is beneficial to the development of plastic Bottom friction model theory. Opencast Mining
strain at the slope foot on the magnetic model. Technology 3: 24–31.
Chen, B.C. 1979. Similarity Theory in Model Test.
Tractor.
5 CONCLUSION AND DISCUSSION Chen, X.H. 1984. Structure model test on brittle material.
Beijing: Water Power Press.
Geomechanics magnetic model test exploits the Ding, J.L., Tang, Q.M., Gong, Y.M. 1994. Research on
electromagnetic principle to improve the body force saturated soil foundation load character by using water
of the magnetic material to n times of its gravity. seepage force model. Chinese Journal of Geotechnical
The value of n can be controlled by adjusting the Engineering 16(1): 8–20.
coil current. Therefore, according to the similar- Fumagalli, E. 1979. Statical and geomechanical models.
Translated by JIANG Pengnian, PENG Guanglv.
ity theory, the prototype can be reduced n times to Beijing: Water Power Press.
test. It provides a new convenient and simple test Fang, L., Li, G.X., Huang, F. 1997. A new method of
mode for model test. geotechnical model test-geotechnical model test of
By means of numerical simulation, a magnetic pile foundation water seepage force. Geological Jour-
field with high magnetic flux density gradient is nal of China Universities 3(4): 451–457.
produced by two energizing solenoids. The flux Fumagalli, E. 1968. Model Simulation of Rock Mechan-
density gradient in the middle test area is close to ics Problem. Rock Mechanics in Engineering Practice
be uniform, although the whole magnetic field is (Stagg, K.G., Zienkiewicz, O.C., eds.). London: J.
not ideal. The amplitude error of magnetic force Wiley Bulletin Ismes Nr. 38.
Huang, F., Huang, W.F., Li, G.X., et al. 1998. Study
on the model is less than 15% of the mean value. about the resistances of piles in different loading ways
The finite element numerical simulation results of by hydraulic gradient test. Chinese Journal of Geotech-
the magnetic model and prototype show that their nical Engineering 20(2): 10–14.
mechanical behavior is similar, and the slight dif- Huang, Z.Q. & Wang, S.J. 1998. The application general sit-
ference mainly comes from the non-uniformity of uation of centrifugal model test in China. Chinese Jour-
the magnetic force on the body. Without regard to nal of Rock Mechanics and Engineering 17(2): 199–203.
the change of the material parameters, if the mag- Luo, X.Q., Cheng, S.G., Niu, E.K. 2009a. Research on
netic flux density gradient in the test area keeps aberration correction and application in landslide
spatially homogeneous and temporal stability, the geomechanical model test. Chinese Journal of Rock
Mechanics and Engineering 28(Supp.1): 3082–3088.
numerical simulation results of magnetic model Luo, X.Q., Ge, X.R., Cheng, S.G., et al. 2009b. Study of
and prototype should be highly consistent accord- magnetic material property of geomechanical model
ing to the similarity theory. test in electromagnet field. Chinese Journal of Rock
It is very difficult to get a magnetic field whose Mechanics and Engineering 28(Supp.2): 3801–3807.
magnetic flux density gradient is absolute uniform Luo, X.Q., Liu, D.F., Wu, J., et al. 2005. Model test study
by experimental method. The best way is to deduce on landslide under rainfall and reservoir water fluc-
the solenoid parameters theoretically. How to tuation. Chinese Journal of Rock Mechanics and Engi-
make the magnetic flux density gradient keep uni- neering 24(14): 2476–2483.
form still need further study. Luo, X.Q., Cheng, S.G., Zhang, Z.H., Tang, K. 2011.
Study of similarity theory of geomechanical model
test in electromagnetic field. Rock and Soil Mechanics
REFERENCES 32(4): 457–462.
Pu, J.L. 1996. The application and development tendency
Ashby, J. 1971. Sliding and toppling modes of failure in of geotechnical centrifuge model test. Chinese Journal
model and jointed rock slope [MS Thesis]. Imperial of Geotechnical Engineering 18(5): 92–94.
College Royal School of Mines, London. Shen, T. 2001. Progress of geomechanical model test tech-
Bao, C.G. & Rao, X.B. 1998. Principle of geotechni- nology. Journal of Yangtze River Scientific Research
cal centrifuge model test. Journal of Yangtze River Institute 18(5): 32–36.
Scientific Research Institute 15(2): 2–7. Stewart, D.P., Adhikary, D.P. & Jewell R.J. 1994. Study
Bao, C.G. 1991. The application and progress of geotech- on the stability of model rock slope. In: Proceedings of
nical centrifuge model test in China. Chinese Journal the Int. Conf Centrifuge 94, Singapore: 629–634.
of Geotechnical Engineering 13(6): 92–97. Xia, X.D. 1983. Test technology research on distortion
Baumgargarter, P. & Stimpson, B. 1979. Development model of soil-mechanical system. Journal of agricul-
of a titable base friction frame for kinematics studies tural machinery (4): 10–26.
of caving at various depths. International Journal of Yu, Y.Z., Li, R.J., Li, G.X., et al. 2008. Experimental
Rock Mechanics and Mining Sciences and Geomechan- study on centrifuge model dynamic behavior of slopes
ics Abstracts 1: 265–267. with saturated subgrades during earthquakes. Journal
Bray, J.W. & Goodman, R.E. 1981. The theory of of Tsinghua University 48(9): 1422–1425.
base friction models. International Journal of Rock Zhang, W.M. & Dou, Y. 1995. Development of geotech-
Mechanics and Mining Sciences and Geomechanics nical centrifuge modeling technique. Hydro-Science
Abstracts 18: 453–468. and Engineering (3): 294–301.

402

ISGSR2013.indb 402 10/18/2013 9:42:49 AM

 Geotechnical Safety and Risk IV – Zhang et al. (eds)
© 2014 Taylor & Francis Group, London, ISBN 978-1-138-00163-3

Time-dependent behaviour modeling of geomaterials

using a discrete thermodynamic approach

W. Wang
Key Laboratory of Ministry of Education for Geomechanics and Embankment Engineering, Hohai University,
Jiangsu, Nanjing, P.R. China
Geotechnical Research Institute, Hohai University, Jiangsu, Nanjing, P.R. China

J.F. Shao
Key Laboratory of Ministry of Education for Geomechanics and Embankment Engineering, Hohai University,
Jiangsu, Nanjing, P.R. China
Geotechnical Research Institute, Hohai University, Jiangsu, Nanjing, P.R. China
Laboratory of Mechanics of Lille, UMR 8107 CNRS, Cité Scientifique, France

Q.Z. Zhu & W.Y. Xu

Key Laboratory of Ministry of Education for Geomechanics and Embankment Engineering, Hohai University,
Jiangsu, Nanjing, P.R. China
Geotechnical Research Institute, Hohai University, Jiangsu, Nanjing, P.R. China

ABSTRACT: A discrete thermodynamic approach is presented for the modeling of coupled anisotropic
viscoplastic and damage behaviour in cohesive-frictional geomaterials. The idea is to extend the aniso-
tropic coupled elastoplastic damage model to the anisotropic one using a discrete approach. The visco-
plastic strain is induced by frictional sliding along weakness planes randomly distributed in the elastic
matrix. The evolution of induced damage is controlled by the evolution of weakness planes in connection
with the propagation of microcracks. Perzyna’s viscoplastic theory is applied to formulate the macro-
scopic viscoplastic strain. The modified Coulomb-type plastic viscoplastic potential function and damage
evolution criterion proposed by Mazars are given for each family of weakness planes. The effective elastic
modulus of damaged material is determined by the damage variable associated with each family of weak
sliding planes. The proposed model is applied to a typical geomaterial (e.g. sandstone) for the modeling
of time-dependent behaviour. The comparison between numerical prediction and experimental data is
shown. The validity of the proposed model is verified by the good fitting of strain-time curves obtained
by the multi-step triaxial creep tests for sandstones.

1 INTRODUCTION these microcracks. It is clear from micromechanical

analysis that plastic flow and damage evolution are
In rock engineering, time-dependent behaviour inherently coupled each other (Zhu et al. 2008a).
of rocks should be taken into consideration for Classically, continuous phenomenological mod-
many problems, such as long-term stability analy- els have been developed for modelling of plastic
sis of foundations, rock slopes and underground deformation and induced damage, either sepa-
constructions (e.g. tunnels and caverns), feasibil- rately or in coupled way (Dragon and Mroz 1979,
ity study of nuclear waste disposal, appropriate Ju 1989, Hayakawa & Murakami 1997, Chiarelli
design and performance assessment prediction et al. 2003, Shao et al. 2006). In such models, scalar
of oil and gas storage. Plastic deformation and and tensorial internal variables are used to repre-
induced damage by microcracks are two main sent isotropic and anisotropic spatial distribution
mechanisms of inelastic deformation and progres- of microcracks and plastic hardening variables.
sive time-dependent failure in geomaterials like However, it is not easy to accurately describe actual
concrete and rocks. It is generally accepted that the distribution of damage and plastic hardening state
plastic deformation is mainly related to frictional even with high order tensors. Moreover, the math-
sliding of microcrack surfaces while the damage ematical description becomes very complex when
evolution is directly related to propagation of accounting for anisotropic plasticity and damage

403

ISGSR2013.indb 403 10/18/2013 9:42:49 AM

 coupling and unilateral effects. On the other hand, between numerical simulations and experimental
micromechanical damage models based on frac- data are presented.
ture mechanics and linear homogenization tech-
niques have also been developed for the description
2 PRESENTATION OF THE MODEL
of brittle behaviour of geomaterials (Kachanov
1982, Gambarotta & Lagomarsino1993, Zhu et al.
2.1 Discrete representation of variables
2008a, b). However, most of these models focused
on the modelling of elastic brittle materials. The In the present study, we adopt the discrete thermo-
capability to describe coupled anisotropic plastic dynamic framework proposed by Zhu et al. (2008c)
damage behaviour in quasi brittle materials is still and apply this method to the visco-plastic damage
not well demonstrated. modelling. The assumption of small strains and
Concerning time-dependent modeling, the isothermal conditions are adopted in the present
viscoplastic theory is generally used (Lemaitre & work. The total macroscopic estrain is first decom-
p
Chaboche 1998) and most viscoplastic models for posed into an elastic partvp
E , a plastic part E
geomaterials are based on the overstress concept and a viscoplastic one E :
proposed by Perzyna (1996). Various specific mod-
e vp
els have been proposed for rocks (Cristescu 1986, E = E + Ep E (1)
Shao 1995, Maranini & Yamaguchi 2001). Other
approaches (Shao et al. 2003, Pietruszczak 2004) In our study, for the sake of simplicity, a sim-
have been developed for the description of time- plified specific version of the model is presented
dependent behaviour in rocks. The time-dependent for modelling mechanical behaviour of a typical
deformation is seen as a consequence of progressive quasi-brittle rock. It is assumed that the material
degradation of material due to physical–chemical damage may be considered in an average way and
reactions. as a consequence there is an isotropic distribution
Based on these previous works, we propose of damage in all orientations.
in the present study to develop a new thermody- The total free energy function is then simplified
namics discrete approach for modeling coupled as follows:
visco-plastic damage behaviour in cohesive-
1⎛ vp ⎞
) : ⎛⎜⎝ E − E
frictional geomaterials. The discrete thermody- vp ⎞
E − E − E ⎟ : C(
p p
Ψ= −E ⎟ (2)
namic framework for plastic modeling is firstly 2⎝ ⎠ ⎠
proposed by Zhu et al. (2008c). It is first assumed p vp
+ Ψ +Ψ
that the material contains a random distribution
of Weakness Sliding Planes (WSPs) subject to The state equations are obtained by the standard
frictional sliding and propagation under applied derivation of the total free energy Ψ with respect
stresses. A suitable system of orientations is then
chosen to discretize the continuous distribution
(
to elastic strain tensor E e E − E p Evp : )
of weakness sliding planes. Each orientation rep-
∂Ψ
resents one family of weakness planes. Overall ∑ = ∂Ee = C ( ) : (E − E p − Evp ) (3)
(total) plastic strains are obtained by the deter-
mination of frictional sliding along each weak-
ness plane. Appropriate yield function, plastic 2.2 Discrete form of damaged elastic tensor
potential, plastic hardening law and damage evo-
lution criterion are determined for each family of In order to formulate discrete plastic damage model
weakness planes. The frictional sliding and crack in general loading conditions, we first propose to
propagation are inherently coupled leading to develop a general methodology for projections of
coupled macroscopic plastic damage behaviour. stress tensor and other variables onto each fam-
The effective elastic properties of damaged mate- ily of WSPs. For this purpose, the following two
rial can be deduced as functions of discrete dam- fourth order tensorial operators are introduced,
age variable related to each family of weak planes. N and T , both functions of the unit normal vec-
For the modeling of time-dependent behaviour tor n and with the components, respectively:
of sandstone, the elstoplastic model is extended
to a viscoplastic one. The viscoplastic strain is Nijkl ni n j nk nl
induced by frictional sliding along weakness
1 ⎛ δ n n + δ iil n j nk + δ jjk ni nl ⎞
planes randomly distributed in the elastic matrix. Tijkl = ⎜ ik j l ⎟⎠ (4)
Perzyna’s viscoplastic theory is applied to for- 2 ⎝ + δ jjl ni nk − 4 ni n j nk nl
mulate the macroscopic viscoplastic strain. The
proposed model is applied to a typical quasi brit- Invoking the classic isotropic damage theory,
tle rock material (sandstone), and comparisons only the shear modulus is affected by damage in

404

ISGSR2013.indb 404 10/18/2013 9:42:49 AM

 the case of closed cracks. Let k(d ) and μ(d ) be induced damage is fully responsible for material sof-
the bulk and shear modulus of damaged material, tening due to propagation of various defeats. Further,
respectively, one can write: as for most geomaterials, a non-associated plastic
flow rule is required. In this model, a Coulomb-type
k (d ) = k s, μ (d ) = μ s ( d ) (5) plastic yield function including plastic hardening
and softening is used. The macroscopic stress tensor
The effective elastic stiffness tensor of damaged ∑ is projected on the normal vector to each family
material can be written in the following general of weakness sliding planes using the general projec-
form: tion operators given in (4). The plastic flow in each
weakness sliding plane is controlled by the shear and
C( ) k sJ + 2μ s ( κ ))K
K Cs μ s d K (6) normal stresses applied to this plane. The plastic
yield function f p r is then expressed as:
In (6), C s denotes the initial elastic tensor of
undamaged (sound) material. Two fourth order f pr ( r
t
r
n γ pr
)
d = σ tr + ( − ζ d )α p r σ nr ( )
c ≤0
isotropic tensors J and K verify the relation: (11)
J K = I, with the components:
The parameter c is related to material cohesion.
The function α p r defines the plastic hardening law
I ijkl
1
2
( iik jl il jjk ), d Jijkl
1
= δ iijδ kl
3
(7) and depends on the plastic hardening variable γ .
pr

The hardening variable γ is expressed as:

p r

However, in the case of random distribution of

weakness sliding planes, the damage state is gen-
erally anisotropic in nature due to propagation of
γ pr
∫ (γ r
. γ )
r 1/2
(12)

microcracks in some preferred orientations. For which is seen as the cumulated equivalent plastic
the description of this anisotropic damage, it is distortion in weakness sliding planes. Based on
proposed to replace the isotropic damage variable experimental data obtained on typical geomateri-
d by a damage distribution function ω ( ). In addi- als, the following particular form is proposed:
tion, it is shown that the fourth order tensor T( )

( )
is directly related to the degradation of the shear − b1γ p r
modulus (Walpole 1981). Thus, the term d K can αpr α fp − α fp α op e (13)
be generalized in the following integral form:
Physically, the hardening function defines the
1 current mobilized frictional coefficient along weak-
dK ζ ∫ ω ( n )T( n )dS
4π S ness sliding planes. α op and α fp are respectively
the initial and asymptotic values of the hardening
∫S 2 T( n )dS / 4π 2 K / 5 (8)
function. The parameter b1 controls its kinetics.
In order to define a non-associated plastic flow
It follows, from the case of isotropic damage rule, the following expression is used as plastic
distribution, i.e. ω ( ) d , that ζ = 5 /2 . The effec- potential g :
pr
tive elastic stiffness tensor C can be rewritten into

( )=σ
the form:
gp r r
t
r
n dγ
d, p ,r r
t + ( − ζ d )η p,rσ nr = 0 (14)
1
C(ω ( )) Cs μ sκ ∫ 2 ω ( )T( )dS (9) In which η is a coefficient related to plastic
pr
4π S
volumetric strain for undamaged material. In order
to describe the transition from plastic compress-
with κ 5κ /2 , and in discrete form: ibility to dilatancy observed in geomaterials; η p r is
defined as a function of plastic hardening variable
κ 15 i i γ p r and expressed as follows:
C(ω ( )) Cs μs ∑ω T
15 i =1
= 3k s J
⎛ κ 15 ⎞
μ s I − ∑ ω i Ti ⎟ : K (10)
ηp r (
η fp − η fp ηop e ) − b2γ p r
(15)
⎝ 15 i =1 ⎠
The current value of η is usually called as
pr

current value of plastic dilatancy coefficient.

2.3 Characterization of plastic flow
Therefore, the two parameters ηop and η fp denote
Plastic flow is coupled with induced damage in each respectively the initial and asymptotic value of
family of weakness planes. It is assumed that the dilatancy coefficient.

405

ISGSR2013.indb 405 10/18/2013 9:42:53 AM

 The plastic flow rule in each family of planes instantaneous plastic yield surface and viscoplas-
is defined through plastic potential as in classic tic loading surface are related to the same internal
plasticity: variable, for instance, the equivalent plastic shear
strain. As the viscoplastic flow is delayed com-
∂g p r pared with the plastic deformation, the evolution
d pr
dλ p r
∂∑ of viscoplastic loading surface is then slower than
⎛ s ⎞ that of plastic yield surface. However, the both
= dλ p r ⎜ r ⊗ N r ( )η p r N r ⊗ N r ⎟ , dλ p,r ≥ 0 are described by the same mathematical function.
⎝ ⎠ Thus, the viscoplastic loading surface for each
(16) family of weakness sliding planes is given by:

On the other hand, one obtains the following

relation:
f vp r ( r
t
r
n γ pr
)
d = σ tr + ( − ζ d )α vvp r σ nr ( )
c ≤0
(21)
s
r r r r
d pr
= dβ N ⊗ N + dγ ⊗ N
r
(17) The function α defines the current internal
vp,r

friction coefficient of the viscoplastic loading

The comparison of (16) with (17) gives: surface. Therefore, it controls the evolution of
the viscoplastic flow. The expression of hardening
dγ r =dλ p r r , dβ r dλ p r ( ζ ) η p,r (18) function for viscoplastic flow is given by:

2.4 Characterization of damage evolution

α vp,r (
α fp − α fp α op e ) − b1γ p r
(22)

The damage evolution is physically related to The same function is proposed for the viscoplas-
propagation of microcracks and defeats in various tic potential:
orientations. Generally, the propagation is not uni-
form in space orientation. However, an isotropic
damage is adopted in this simplified version, and
gvp r ( r
t
r
n dγ
d, p,r
)=σ r
t + ( − ζ d )η p,rσ nr = 0 (23)

the essential consequence of damage is material

softening behaviour due to material degradation. Based on Perzyna’s viscoplastic theory (Perzyna,
Further, it is assumed that the damage evolution is 1996), the viscoplastic flow rule is determined by:
mainly controlled by plastic sliding along surfaces n
of microcracks and defeats. Thus, the driving force f vp,rr ∂gvp,r
for damage evolution is taken as the averaged value E vp,r = γ (24)
c ∂σ
of equivalent plastic shear strain γ p r such as:

1 15 where γ, n are the viscosity coefficients.

γ p
γ p,r (19)
15 r=1
2.6 Parameter identification
Inspired by some previous works on damage The proposed model contains 16 parameters: two
modelling of geomaterials, an exponential function elastic constants for initial (undamaged) state
similar to that used by Mazars (1986) for concrete of material, E s and v s ; four parameters involved
is used as the damage criterion: in plastic yield function α mp , α op , b1 a d c ; three
parameters in plastic potential η fp ηop and b2 ;
⎡ −bb3 ⎛ γ p −γ 0 ⎞ ⎤
p
three parameters characterizing the damage evolu-
fd d − d f ⎢1 e ⎝ ⎠
⎥≤0 (20)
⎢⎣ ⎥⎦ tion d f , b3 , and γ 0p ; the parameter κ for describing
the degradation of shear modulus and the param-
eter ζ for the description of damaged-related
The parameter df defines the asymptotic value
material softening, two viscosity parameters γ, n
of damage variable related to residual strength of
characterizing the time-dependent behaviour.
damaged material; γ 0p is the initial threshold of
The initial elastic constants may be determined
damage and b3 controls the kinetics of damage
from the linear part of stress-strain curves before
evolution.
the initiation of plastic yielding and damage
evolution. The parameters α 0p and c are concerned
2.5 Viscoplastic characterization
with the initial yield condition and therefore can be
The viscoplastic loading surface is considered as determined by drawing up the initial plastic yield
the delayed plastic yield surface. The evolutions of surface. Due to the fact that the damage evolution

406

ISGSR2013.indb 406 10/18/2013 9:43:03 AM

 is coupled to plastic flow and affects both plastic Table 1. Parameters of the model for the triaxial creep
yield function and plastic potential, the explicit test of sandstone.
determination of all the parameters involved in
Es c
plastic flow and damage evolution is generally dif-
(MPa) νs α mp α op b1 (MPa) η f
p
ηop b2
ficult. The method used here is based on a best
fitting algorithm using a series of conventional tri- 23900 0.2 0.7 0.1 1000 6 0.3 0.1 200
axial compression tests performed under different
confining pressures. The viscosity parameters are df b3 γ p
κ ζ γ (h )
−1 n
0
determined by the best fitting method using the
triaxial creep tests. 1.0 2000 0.0 1.0 1.0 10−7 0.1

3 NUMERICAL SIMULATION

In this section, the proposed model is applied to

the modelling of time-dependent behaviour of
sandstone. Note that due to the damage evolu-
tion during visco-plastic flow, the material strength
reduces with time. As a consequence, if the applied
stress state is close to the long-term failure surface,
the material failure can be produced by accelerated
creep process. The model can predict three stages
of creep process. An example of simulation of
creep behaviour is shown in Figure 1.
The emphasis here is modelling of time-dependent
deformation of sandstone. A series of creep tests
Figure 2. Simulation of a triaxial creep test with six
under different stress states have been conducted. deviatoric stress (the number on the line indicate the
The purpose is to check the performance of the deviatoric stress).
model in predicting the time-dependent behaviour
of sandstone. The samples were cored from intact
blocks of sandstones located in the underground
cavern of Xiangjiaba hydropower station in the
southwest of China. The off-white fine grain rocks
are mainly composed of feldspar, quartz and rock
debris. The sandstone has an average porosity of
2.64% and dry density of 2.62 g/cm3. The creep tests
were carried out following a multi-step loading pro-
cedure. A detailed description of test method and
test result can be found in Wang et al. (2009). Based
on these tests as well as short-term triaxial compres-
sion tests, the typical values of main parameters
have been identified and shown in Table 1.
Figure 3. Influence of r on creep deformation of
Figure 2 shows the result of simulation of triax- sandstone.
ial creep tests on the sandstone under a confining

pressure of 3 MPa with six different values of

stress deviator. In general, the time-dependent
response of material is well predicted by the pro-
posed model. The creep rate is increasing when the
stress deviator is higher.
In Figures 3 and 4, a sensitivity study of two
main viscoplastic parameters r and n is presented.
Here, we simulate a creep test at the first loading
under the confining pressure of 5 MPa and with
the deviatoric stress of 120 MPa. We can notice
Figure 1. Example of simulation of three stages of the influence of two parameters controlling the
creep behaviour of sandstone. viscoplastic deformation r and n, on the long-

407

ISGSR2013.indb 407 10/18/2013 9:43:09 AM

 Cristescu, N. 1986. Damage and failure of viscoplastic
rock-like materials. Int J Plast 2(2): 189–204.
Dragon, A. & Mroz, Z. 1979. A continuum model for
plastic-brittle behaviour of rock and concrete, Interna-
tional Journal of Engineering Science 17(2): 121–137.
Hayakawa, K. & Murakami, S. 1997. Thermodynamical
modeling of elastic-plastic damage and experimental
validation of damage potential, International Journal
of Damage Mechanics 6(4): 333–363.
Ju, J.W. 1989. On energy based coupled elastoplastic
damage theories: constitutive modeling and compu-
Figure 4. Influence of n on creep deformation of tational aspects, International Journal of Solids and
sandstone. Structures 25(7): 803–833.
Kachanov, M. 1982. A microcrack model of rock ine-
lasticity—Part I: frictional sliding on microcracks;
term response of sandstone. It seems that n has a Part II: propagation of microcraks. Mechanics of
Materials 1: 19–41.
stronger influence on long-term behaviour than r. Lemaitre, J. & Chaboche, J.L. 1998. Mechanics of solid
materials. Cambridge: Cambridge University Press.
Maranini, E. & Yamaguchi, T. 2001. A non-associated
4 CONCLUSION viscoplastic model for the behaviour of granite in
triaxial compression. Mechanics of Materials 33(5):
In this study, we propose a discrete approach for 283–293.
visco-plastic damage modelling of cohesive-fric- Mazars, J. 1986. A description of micro- and macroscale
tional geomaterials. Compared with classic models, damage of concrete structures. International Engineer-
the discrete modelling allows taking into account ing Fracture Mechanics 25(5–6): 729–737.
Perzyna, P. 1996. Fundamental problems in viscoplasticity.
physical mechanisms involved in viscoplastic Advances in Applied Mechanics 9: 243–377.
deformation and damage evolution. For instance, Pietruszczak, S., Lydzba, D., Shao, J.F. 2004. Description
macroscopic plastic strains in geomaterials are of creep in frictional materials in terms of microstruc-
generally generated by frictional sliding along sur- ture evolution. Journal of Engineering Mechanics
faces of microcracks and defeats while damage 130(6): 681–690.
evolution is associated with the propagation of Shao, J.F., Dahou, A., Bederiat, M. 1995. Experimental
these weakness sliding planes. Further, using the and numerical investigations on transient creep of
proposed discrete approach, it is easier to describe porous chalk. Mechanics of Materials 21(2): 147–158.
induced anisotropy in plastic flow. However, a Shao, J.F., Jia, Y. Kondo, D., Chiarelli, A.S. 2006.
A coupled elastoplastic damage model for semi-brittle
specific simplified version is proposed by using materials and extension to unsaturated conditions.
an isotropic damage description. The elastoplastic Mechanics of Mechanics 38(3): 218–232.
damage model can be extended into a viscoplastic Shao, J.F., Zhu, Q.Z., Su, K. 2003. Modeling of creep
model by using the Perzynas viscoplastic theory. in rock materials in terms of material degradation.
Consequently, the time dependent behaviour of Computers and Geotechnics 30(7): 549–555.
sandstone can be described by considering viscous Walpole, L.J. 1981. Elastic behavior of composite mate-
sliding of weakness planes and sub-critical propa- rials: theoretical foundations. Advances in Applied
gation of defeats. A good accord between simula- Mechanics 21: 169–242.
tion and test data verify the validity of the model. Wang, W., Xu, W.Y., Li, L.Q., Jiang, Z.Y. Experimental
study on the time-dependent behaviour of Xiangjiaba
A further application of the model will be carried Sandstone. In Erich Bauer (ed.), Proc. 2nd Interna-
out in the future. tional Conference on Long Term Behaviour of Dams
(LTBD09), 12–13 October 2009. Graz University of
Technology.
ACKNOWLEDGEMENT Zhu, Q.Z., Kondo, D., Shao, J.F. 2008a. Micromechanical
analysis of coupling between anisotropic damage and
This work is supported by National Program friction in quasi brittle materials: role of the homog-
on Key Basic Research Project (973 Program) enization scheme. International Journal of Solids and
(2011CB013504) and National Science Founda- Structures 45(5): 1385–1405.
tion of China (51109069, 11272114). Zhu, Q.Z., Kondo, D., Shao, J.F., Pensee, V. 2008b.
Micromechanical modelling of anisotropic damage in
brittle rocks and application. International Journal of
REFERENCES Rock Mechanics and Mining Sciences 45(4): 467–477.
Zhu, Q.Z., Shao, J.F., Kondo, D. 2008c. A discrete ther-
Chiarelli, A.S., Shao, J.F., Hoteit, N. 2003. Modeling modynamic approach for modeling anisotropic cou-
of elastoplastic damage behavior of a claystone. pled plasticity-damage behaviour in geomaterials.
International Journal of Plasticity 19(1): 23–45. Comptes Rendus Mecanique 336(4): 376–383.

408

ISGSR2013.indb 408 10/18/2013 9:43:12 AM

 Geotechnical Safety and Risk IV – Zhang et al. (eds)
© 2014 Taylor & Francis Group, London, ISBN 978-1-138-00163-3

Study of time-dependent behavior of rocks with polycrystalline

approach

T. Zeng & J.F. Shao

College of Civil and Transportation Engineering, Hohai University, Nanjing, China
University of Lille I, LML, UMR 8107, Cité Scientifique, Villeneuve d’Ascq, France

ABSTRACT: A completely new constitutive model is proposed for describing the viscoplastic behavior
of cohesive geological material—granite, which essentially possesses the morphology of granular texture.
Based on micromechanics consideration, the local inelastic deformation is attributed to the sliding along
specified direction on certain crystallographic (weakness) planes within each granular. The corresponding
macro deformation is obtained with the classical homogenization approach. The special features
of geomaterial, e.g. pressure sensitivity and plastic dilatancy, are taken into account by introducing
Mohr-Coulomb type yield criterion and non-associated plastic potential. Comparisons between model’s
predictions and test data have been carried out to check the validity of proposed model.

1 INTRODUCTION have been proved. One kind is microcrack induced

damage in brittle geomaterials (Pensée 2002, Zhu
In certain practical engineering applications, et al. 2008). The other one is plastic deformation
appropriate constitutive models are necessary to in ductile porous materials (Shen et al. 2012). For
describe mechanical behaviors of various geo- those micromechanical models, the morphology is
materials (soils, rocks and concretes). Generally, generally represented by matrix-inclusion system.
two kinds of models have been proposed to treat Such morphology is acceptable only when certain
this problem, i.e. phenomenological model and phase is dominate and intact. However, when the
micromechanical model. Phenomenological mod- texture of geomaterial is discrete and granular, the
els, generally formulated within the framework of choice of matrix system may be not that obvious.
thermodynamics of irreversible processes, have Therefore, it is necessary to define a more reason-
been largely developed and applied. It usually able morphology.
has strong experiment support and is easy to be Certainly, the matrix-inclusion system will not
implemented into commercial finite element soft- be completely abandoned. Each granular (inclu-
ware. Therefore, it is suitable for structure analysis. sion or grain) is now embedded in a so-called
However, such model does not take into account Homogeneous Equivalent Medium (HEM), which
physical mechanisms involved at pertinent material is an imaginary medium and its properties equal
scales. It usually contains high number of empiri- to the overall response of aggregate. The special
cally based internal state variables, which results in features of the morphology here is that the prop-
the difficulty of identification process for model’s erty of HEM is unknown in advance. It is coupled
parameters. with property of each granular and can be only
In fact, most geomaterials are heterogeneous obtained at the end of calculation through SC
materials at different scales. From microscopi- (Self-Consistent) method (Hill 1965). However, for
cal point of view, they are aggregate with differ- the sake of simplicity and as a first stage of devel-
ent constituents cemented together. The special opment, the KBW model, which is initially pro-
microstructures, for instance, pores and interface, posed by Kröner, Budiansky and Wu (Budiansky
play an important role in material’s macroscopic 1962), is adopted in this paper. It assumes that the
responses. Inelastic deformation and failure proc- elastic interaction applies not only in elastic range
ess of geomaterials are directly related to the evo- but also in plastic range. Although such simplifi-
lution of these microstructures. Therefore, the cation will results in too stiff response when plas-
micromechanical model provides a completely new tic deformation is large, it will greatly reduce the
way to explore the properties of geomaterials. numerical work.
Until now, two kinds of micromechanical The inspirations of my research are from the
model have been proposed and their capabilities widely used polycrystal model for metallic materials.

409

ISGSR2013.indb 409 10/18/2013 9:43:12 AM

 To adopt the polycrystalline model, we made the
following assumptions and simplifications. Each
granular contains many weakness planes. To con-
sider its distribution, we introduce the conception
of crystallography planes in FCC crystal structure.
Moreover, the distribution of granular is assumed
random and therefore, the response of polycrys-
tal is isotropy. The local inelastic deformation is
Figure 2. Schematic representation of one-site self-
attributed to the sliding along specified direction consistent models.
on certain crystallographic (weakness) planes.
However, due to some special features of geoma-
terials, the classical Schmid’s law and associated
plastic potential of each slip system are replaced the local values. The most widely used one is
by a Mohr-Coulomb type criterion and non- the so-called one site self-consistent scheme, as
associated plastic potential. A general hardening Figure 2 shows. The strain or stress in a single crys-
law is adopted to depict the self-hardening and tal is approximately obtained by solving a spheri-
cross-hardening behavior of each slip system. cal single crystal embedded in an infinite deformed
Throughout this paper, lower case letters rep- matrix. HEM (Homogeneous Equivalent Medium)
resents local fields while capital letters are overall is an imaginary matrix, whose represents are iden-
(macroscopic) fields. First order tensor, second tical to the overall response of the polycrystalline
order tensor and forth order tensor are denoted as aggregate.
a , a and A, respectively. The operations between However, it is not easy to determine the prop-
tensors are as follows: a b = ai bi , a : b = aij bij , erties of HEM. Some approaches have been pro-
A : b Aijkl bkl , a ⊗ b = ai b j . posed to avoid this difficulty. The most widely
used KBW model—KBW assumes the elastic
interaction between each single crystal and HEM,
2 KBW MODEL FOR GEOMATERIAL which applies to case when deformation is not too
large.
2.1 Scale decomposition Considering a polycrystalline aggregate sub-
jected to macroscopic stress at remote boundary,
The polycrystalline model in following study the local stress within each single crystal can be
generally decomposed into three different levels: determined with Equation 1
macroscopic (Representative Element Volume
(REV)–aggregate of granular or grains), mes-
oscopic (granular or single crystal), and micro-
σ Σ k( ) (εvin − Evin) − μ( ) (εdin − Edin) (1)
scopic (Crystallographic Slip System, CSS), as
Figure 1 shows. This is the general interaction law of KBW
model, which takes into accounts the influence
from volumetric strain. The parameters k and u,
2.2 Grain/matrix interaction law
respectively, being the bulk and shear modulus of
Imaging a polycrystalline aggregate subjected to each granular. The subscripts v and d represent
a macroscopic stress or strain, the local stress or the volumetric and deviatoric components of the
stain varies not only from crystal to crystal, but corresponding tensor. For the spherical inclusion
also from point to point within each single crystal. embedded in an infinite elastic medium
Many methods have been proposed to determine
3k 6 k + 2μ
c= and d = (2)
3k + 4 μ 5 3k + 4 μ

2.3 Single crystal constitutive relations

To reflect the pressure sensitivity of geomateri-
als, the classical Schmid law for each slip system is
replaced by a Mohr-Coulomb type yield criterion
as Equation 3 shows
Figure 1. Schematic representations of REV, weakness
planes and typical crystallographical plane in FCC unit
cell. The microscopic photograph of granite is from
fα ( )
Rα = σ : mα ( )
μ f H σ nα σ N α (τ α
c Rα )
(Soulié et al. 2007). (3)

410

ISGSR2013.indb 410 10/18/2013 9:43:12 AM

 where When the slip rate (or plastic multiplier) is
determined, the local inelastic deformation is the
m=
2
(
1 α
n ⊗ s α + s α ⊗ nα ) d N α = nα ⊗ nα
summation of all active slip systems within each
granular
(4)
∂F α
ε vp = ∑ λ α (9)
where uf = the frictional coefficient. n = normal α
∂σ
of crystallographic plane. sα = slip direction on
the corresponding plane. In crystal theory, they
are predefined and are expressed by Miller index. 2.5 Polycrystal constitutive relations
σ nα = the normal stress act on each slip system.
τc = initial resistance stress. The introduction of At macroscale, the macroscopic viscoplatic strain
Heviside function H(⋅) indicates that the normal is calculated by an averaging procedure
stress has no effect on the yield function unless the
Ng
normal stress is compressive.
Another characteristic of geomaterials is dila- E vp  vp ∑ fhε vvp,h (10)
h =1
tancy due to plastic flow. To reflect this feature, the
associated plastic potential is modified as a non-
where fh = relative volume fraction. Owing to the
associated one
randomly distribution, each single crystal possesses
the same relative weights, i.e. fh = 1/Ng. Ng = the
( )
F α σ , Rα = σ : mα ( )σ
vF H α
n Nα number of discrete grains (orientations). For
− Rα( bq R )
α α
(5) FCC single crystal, 40 grains are enough to rep-
resent the polycrystal with isotropic behavior. The
macro Cauchy stress can be easily determined by
where vf reflects the roughness of each slip system. Hooke’s law
b reflects the hardening rate. qα and Rα is a couple
of internal state variables describing transgranular
isotropic hardening
Σ = C (E − E )
vp
(11)

qα =
1
b
( − (−bλ )) α
(6) 3 NUMERICAL SIMULATION

( ( ))
3.1 Representation of the model
Rα = Q ∑ hαβ bλ β (7)
The constitutive equations can be combined
where hαβ is the interaction matrix allowing the intro- with any commercial finite element software, e.g.
duction of cross influence of β th slip system on the ABAQUS, without any difficulty. From micro-
hardening of α th slip system. If there is no cross hard- scopic point of view, each integration point is
ening or latent hardening, hαβ is an identity matrix. regarded as a polycrystal, which associates with a
large number of randomly distributed grains. The
stereographic projection of those 40 grains is as
2.4 Determination of slip rate λ α Figure 3 shows.
For time-dependent (viscoplastic) model, the slip
rate is explicitly expressed as the function of σ and 3.2 Parameters calibration
Rα as follows
The predictive capacity of the modified KBW
polycrystalline model for geomaterials is now
⎧ ⎛ 1 ⎞
⎪1 ⎜⎛ f α ⎞p ⎟ α
+ − ⎟ if f > 0
(
λ α σ ,Rα ) ⎪ ⎜⎜ α
= ⎨ η ⎝ τ c + Rα
⎜
1⎟
⎠
1
⎟⎠
⎪ ⎝
⎪0 if f α ≤ 0
⎩
(8)

where η is a viscosity-related material constant,

whose dimension is time, while p is a rate-sensitive
material constant without dimension. Usually, Figure 3. <0 0 1>, <1 1 0> and <1 1 1> pole figures of
these two parameters must be strictly positive. 40 discrete grains with cubic symmetry.

411

ISGSR2013.indb 411 10/18/2013 9:43:16 AM

 checked through the simulation of laboratory tests
on a quasi-brittle rock, Lac du Bonnet granite.
The experimental data of triaxial compressions
and creep tests are taken from (Martin 1997) and
(Lajtai 1987), respectively. These data are used to
determine the model’s parameters and validations.
The detail procedure for parameters identification
can be found in (Zeng 2012). After a back-analysis
procedure, the final parameters for simulation are
listed in Table 1.

3.3 Short term response

For short term response, the strain rates applied
in the simulations are corresponding to labora-
tory tests. With the calibrated parameters given in
Table 1, the final simulations results are given in Figure 5. Comparison between experiment data and
Figures 4 and 5. polycrystal model predictions–triaxial compression with
confining pressure 20 MPa.
3.4 Long term response
The long term responses of brittle rocks are briefly
depicted in dry conditions. Due to the insufficiency
of experiment data, the simulations will be rather
qualitative than quantitative. The comparison is
given in Figure 6, which is obtained by keeping
the specimen strained under Σ 33 = 143 Mpa for
70 days.

Table 1. Parameters for simulation of time-dependent

mechanical behavior.

E η τc Q
(MPa) v uf vf (s) p (MPa) hαβ b (MPa)

68000 0.21 0.4 0.6 109 0.9 35 1 400 40

Figure 6. Comparison between experiment data and

polycrystal model predictions–uniaxial creep test in dry
conditions.

ACKNOWLEDGEMENT

The present work is jointly supported by the China

Scholarship Council (CSC) for the first author
and the Chinese state 973 program with the grand
2011CB013504.

REFERENCES

Budiansky, B. & Wu, T.T. 1962. Theoretical prediction of

plastic strains of polycrystals. Proc. 4th U.S. Congress
on Applied Mechanics 1: 1175–1185.
Figure 4. Comparison between experiment data and Hill, R. 1965. Continuum micro-mechanics of elastoplas-
polycrystal model predictions–triaxial compression with tic polycrystals. Journal of the Mechanics and Physics
confining pressure 10 MPa. of Solids 13: 89–101.

412

ISGSR2013.indb 412 10/18/2013 9:43:19 AM

 Lajtai, E.Z., Schmidtke, R.H., Bielus, L.P. 1987. The Soulié, R., Mérillou, S., Terraz, O., Ghazanfarpour, D.
effect of water on the time-dependent deformation 2007. Modeling and rendering of heterogeneous gran-
and fracture of a granite. International Journal of ular materials: granite application. Computer Graphics
Rock Mechanics and Mining Sciences 24: 247–255. Forum 26: 66–79.
Martin, C.D., Read, R.S., Martino, J.B. 1997. Zeng, T. 2012. Modélisation multi-échelle des
Observations of brittle failure around a circular test comportements plastiques et viscoplastiques des
tunnel. International Journal of Rock Mechanics and géomatériaux polycristallins. Ph.D. thesis. Université
Mining Sciences 34: 1065–1073. Lille 1.
Pensée, V., Kondo, D., Dormieux, L. 2002. Zhu, Q.Z., Kondo, D., Shao, J.F., Pensée, V. 2008.
Micromechanical analysis of anisotropic damage in Micromechanical modeling of anisotropic damage in
brittle materials. Journal of Engineering Mechanics brittle rocks and application. International Journal of
128: 889–897. Rock Mechanics and Mining Sciences 45: 467–477.
Shen, W.Q., Shao, J.F., Kondo, D., Gatmiri, B. 2012.
A micro-macro model for clayey rocks with a plastic
compressible porous matrix. International Journal of
Plasticity 36: 64–85.

413

ISGSR2013.indb 413 10/18/2013 9:43:21 AM

This page intentionally left blank
 9 Reliability-based design and limit state design
in geotechnical engineering

ISGSR2013.indb 415 10/18/2013 9:43:21 AM

This page intentionally left blank
 Geotechnical Safety and Risk IV – Zhang et al. (eds)
© 2014 Taylor & Francis Group, London, ISBN 978-1-138-00163-3

Serviceability limit state design of lime-cement

columns—a reliability-based design approach

N. Bergman, R. Ignat & S. Larsson

Division of Soil and Rock Mechanics, KTH Royal Institute of Technology, Stockholm, Sweden

ABSTRACT: Deep mixing with lime-cement columns is a ground improvement method used to improve
the strength and deformation properties of soft cohesive soils. Due to the complex manufacturing proc-
ess, the variability in the strength and deformation properties is normally high. A rational approach to
include variability in the design process is by introducing Reliability-Based Design (RBD). This paper
presents a reliability-based design approach for Serviceability Limit State (SLS) design of soil improved
by lime-cement columns using the First-Order Reliability Method (FORM). The paper further presents
the impact of uncertainties, reliability indices and area replacement ratios on the relationship between the
characteristic value and the design value with respect to the column modulus of elasticity.

1 INTRODUCTION State (SLS) and presented the relationship between

the area replacement ratio (α) and the reliability
Deep Mixing (DM) using lime-cement columns indices (β). They did not however study the impact
is a ground improvement method that is used to of column characteristic uncertainties on the rela-
improve the deformation and strength proper- tionship between the characteristic value of the col-
ties in soft soils. Because of the complexity of the umn modulus of elasticity (Ek) and the design value
soil-binder mixing process, soils improved by lime- of the column modulus of elasticity (Ed).
cement columns often show a large variability with The present paper presents a RBD approach
respect to their strength and deformation proper- for the SLS design of soils improved by lime-
ties (Larsson 2005). The large variability makes it cement columns. Uncertainties related to inherent
difficult to predict the properties in advance; and it variability, measurement errors, statistical errors
is thus essential to verify the properties after instal- and model and transformation errors are consid-
lation. This is normally done by penetration test ered and put into context with the evaluation of
methods and laboratory tests on samples taken the design value of Ed. Finally, the First-Order
from the columns. Reliability Method (FORM) is used to analyze the
In current design practice, it is assumed that the impact of uncertainties on the SLS design of an
effects of variability and uncertainties on the relia- embankment founded on lime-cement columns.
bility of a mechanical system are taken into consid- The SLS design of a lime-cement-column-
eration by using an adequate total factor of safety. supported embankment is normally based on
A more rational way to deal with uncertainties is by an assumption of Ed. In the present paper, using
including them as parameters in the design model. FORM analyses and considering the assumed Ed,
In fact, Eurocode 7 (Eurocode 7: Geotechnical uncertainties and reliability indices, a correspond-
design—Part 1: General rules 2004) states that the ing value of Ek was evaluated. Present paper also
selection of characteristic values for geotechnical presents and discusses the relationship between
parameters shall take into consideration the vari- Ed and Ek, and gives an illustrative example of
ability of the measured property values. This can be how this relationship can be used as support for
accomplished by using a Reliability-Based Design the design.
(RBD) approach. The need for RBD in DM has
previously been address by Honjo (1982), Kitzume
(2004), Larsson et al. (2005), Liu et al. (2008), Zheng 2 METHODS
et al. (2009), Babu et al. (2011), Al-Naqshabandy
et al. (2012), Al-Naqshabandy & Larsson (2012, This section gives a description of the statistical
2013), Bergman et al. (2013). Among these stud- methods used in this study. It also shows how a sim-
ies, only the study by Zheng et al. (2009) included ple and well established deterministic design model
reliability analyses for the Serviceability Limit can be incorporated in a RBD methodology.

417

ISGSR2013.indb 417 10/18/2013 9:43:21 AM

 2.1 Uncertainties variance, α is a sensitivity factor derived from the
Using a RBD approach, a statistical quantification reliability analysis and β is the reliability index
of the stochastic variables included in the design given in standards and design practices.
model is needed. In the present study, the average For a log-normally distributed variable, the
evaluated column modulus of elasticity ( col ) design value (xd,log) can be described as:
and the average evaluated clay oedometer modu-
xk ,logg
lus ( soil ) are considered and treated as stochas- xd ,log = = e ( μlnx β x)
(3)
tic variables. Bergman et al. (2013) quantified the γm
variability of Ecol , in terms of the total coefficient
2
of variation ( E ,TOT ) evaluated from penetra- where xk,log is the characteristic value, μlnx is the
tion test data (qc), as: mean value of the transformed variable and σlnx is
its standard deviation.
⎛1 ⎞ In present paper, the characteristic value was
VE2,TOT
COV Vqc2 COV
((COV
COV Vm2 qqcc ) ⎜ + Γ 2 ⎟
⎝N ⎠ considered to correspond to an evaluated or
Vm2 qc
COV assumed average value.
+ + COV Vtr2,E (1)
N
2.4 Serviceability limit state design
where COVqc is the evaluated coefficient of vari- For simplicity, a simple and well established model
ation of the penetration test data, COV Vm qc is the was adopted for the serviceability limit state
coefficient of variation associated with random analysis. The total settlement (Semb) of an embank-
measurement noise, COVtr,E is the coefficient of ment founded on normally consolidated clay,
variation associated with transformation model improved by end-bearing lime-cement columns
bias, N is the number of uncorrelated tests and (Fig. 1), can be expressed by an equal strain (Voigt)
Γ 2 is the variance reduction factor. The variance model (Broms 1999, TK Geo 2011):
reduction factor can be described as the reduction
in the standard deviation due to spatial correlation, h j hembb ⋅ γ eemb
and is further described in Bergman et al. (2013). semb = ∑ (4)
Values of COVqc are normally obtained from coll + ( − a ) M soil
a Eco
penetration test data, while COV Vm qc , COVtr,E and
Γ 2 are obtained from experience and literature, where hj is the height of layer j, hemb is the height
thus giving an estimation of COV VE ,TOT (Bergman of the embankment, γemb is the unit weight of
et al. 2013). In the following analysis, a wide range the embankment and a is the area ratio of the
of COV VE ,TOT are studied. lime-cement columns.

2.2 Distributions 2.5 First Order Reliability Methods (FORM)

Soil parameters are often modeled as log-normally Reliability analysis makes it possible to quan-
distributed variables since they are strictly non- tify how close a system is to failure (Baecher &
negative. Al-Naqshabandy et al. (2012) and Christian 2003). In SLS design, failure can be
Bergman et al. (2013) found that qc was either defined as an unacceptable difference between
normally or log-normally distributed. As a con-
sequence, two different analyses were made. One
analysis in which Ecol and Msoil were considered to
be normally distributed parameters, and a second
analysis in which they were considered to be log-
normally distributed.

2.3 Design values

Thoft-Christensen & Baker (1982) defined the design
value (xd) of a normally distributed variable as:
xk
xd = = m + αβ s 2 (2)
γm
where xk is the characteristic value, γm is a partial Figure 1. Embankment founded on end-bearing lime
factor, m is the sample mean value, s 2 is the sample cement columns.

418

ISGSR2013.indb 418 10/18/2013 9:43:21 AM

 observed and allowed performance. To assess the The FORM analyses were performed using the
reliability of a geotechnical structure, a limit state input parameters given in Table 1. In the analy-
function (g(X)) is defined at g(X) = 0. In the present ses, Ek was altered with changing COV VE ,TOT in
study, g(X) was defined as: order to keep β constant at 1.0 and 1.5. The value
of β = 1.5 was provided by Eurocode 0 and β =
g (X ) = δ max − δ ( x1, x2 , , xn ) = 0 (5) 1.0 was adopted for comparative analyses. From
these analyses, the quotient Ek/Ed was evaluated.
This is a relationship that can serve as a basis for
where δmax is the maximum allowed settlement and
the design process, as is shown in the illustrative
δ(X) is the estimated settlement calculated from
example below. Analyses were made assuming
design properties x1, x2, …, xn. A value of G(X) > 0
both normal and log-normal distributions of Ecol
indicates an acceptable difference between allowed
and M soil . Moreover, the following assumptions
and observed performance. Combining Equations
were made:
(2), (4) and (5), the limit state function can be
rewritten as: • The influence of curing time is not considered
• The composite material (soil and lime-cement
column) was assumed to be uniformly strained
g ( Ecol , Msoil ) • Ecol and M soil are assumed to be constant
hemb ⋅ γ eemb
= δ max − ∑ h j ⋅ throughout the soil profile
(
a ⋅ μEcol − α Ecol ⋅ β ⋅ σ Ecol ) • Loads are assumed to be only vertically
( )
+ (1 − a ) ⋅ μM soil − α M soil ⋅ β ⋅ σ M soil
distributed.

(6)
4 RESULTS
where μ E is the mean value of Ecol , α E is the
evaluated sensitivity factor, σ E is the standard
col
col col

The outcome of the FORM analysis, using the

deviation of Ecol , μM is the mean value of M soil ,
col

parameters values presented in Table 1, is illus-

α M soil is the evaluated sensitivity factor and σ M soil is
soil
soil

trated in Figure 2, where Ek/Ed is plotted against

the standard deviation of M soil . The sensitivity fac-
COV VE ,TOT .
tors α E and α M were determined by an iterative
col soil When Ecol and M soil were modeled as nor-
process described by Rackwitz & Fiessler (1978)
mally distributed parameters and with β = 1.0,
and Baecher & Christian (2003).
Ek/Ed ranged from 1.04 to 2.53. For β = 1.5,
Ek/Ed ranged from 1.06 to 9.98. When Ecol and
3 INPUTS AND ASSUMPTIONS M soil were modeled as log-normally distributed
parameters and with β = 1.0, Ek/Ed ranged from
Input parameters used in the reliability analysis are 1.04 to 2.03. For β = 1.5, Ek/Ed ranged from 1.06
summarized in Table 1. to 2.67.

Table 1. Summary of input parameters.

Parameters Value Comments/References

Reliability index (β) 1.0 and 1.5 Adopted value and Eurocode 0, Honjo et al. (2010)
Maximum allowed settlement (δmax) 15 cm Adopted value
Characteristic value of column Varying Parameter is altered with different values of COV
VE ,TOT
modulus of elasticity (Ek) considering acceptable value of β
Characteristic value of clay 500 and 1500 kPa Adopted values. Normally consolidated clay
oedometer modulus ( soil )
Unit weight embankment (γemb) 20 kN/m3 TK Geo (2011). In the present study γemb is considered
and treated deterministically
Height embankment (hemb) 4m Adopted value
COVE,TOT 5–60% (altered Adopted range
in 5% intervals)
COVM,soil 30% Adopted value
Column length (L) 10 m Adopted value
Lime-cement column area 0.1, 0.2 and 0.4 Adopted values
replacement ratio (a)

419

ISGSR2013.indb 419 10/18/2013 9:43:26 AM

 5 DISCUSSION INCLUDING AN
ILLUSTRATIVE EXAMPLE

The use of Figure 2 as a design aid is best presented

with an illustrative example.
If Ecol is assumed to be normally distributed and
Ed is assumed to be 20 MPa, with COV VE ,TOT evalu-
ated to 50%, Ek must exceed 20 * 1.98 = 39.6 MPa
for β = 1.0. For β = 1.5, the minimum value of Ek
is 20 * 3.96 = 79.2 MPa. If Ecol is less than the
required Ek, the designer has several options: (1)
to reduce COV VE ,TOTT by making additional tests or
by decreasing COVtr2,E by performing tests with an
Figure 2. The quotient Ek/Ed plotted against COV
VE ,TOT additional test method (e.g. Ching et al. 2010) (2)
with different values of β. to add columns to the design (3) to wait and per-
form tests on a later occasion when Ek may have
increased or (4) to investigate whether the obtained
β is good enough for the structure and client.
As shown in Figure 3, changes in M soil and a
have only a small impact on Ek/Ed, and in prac-
tice this means that it is almost entirely governed
by COV VE ,TOT . However, it must be noted that
Figures 2 and 3 are valid only for this particular
serviceability limit state model and only within
specified ranges of M soil and a.
Although a value of β = 1.5 is suggested by
Eurocode 0, the authors believe that it may be a
too conservative value for many DM applications.
A value of β = 1.5 corresponds to 6.7% risk of
failure. In the authors’ opinion, for structures where
potential failure is of minor consequence, a value
Figure 3. The quotients Ek/Ed and M soil /M d plotted of β = 1.5 cannot always be economically justified.
against COV
VE ,TOT with different values of M soil and a. For these structures, a wider range of β could be
allowed, where β is decided by balancing the cost
of making higher quality columns and additional
tests against the cost of structure maintenance.
With low COV VE ,TOT and β = 1.5, the difference
between Ek/Ed for the normal and log-normal
cases is small, only 4% at COV VE ,TOT = 20%. 6 CONCLUSION
This difference becomes greater with increas-
ing COV VE ,TOT , the difference being 74% at The RBD methodology is considered to be a
COV VE ,TOT = 50% . For β = 1.0, the difference rational approach to include strength and defor-
between Ek/Ed for the normal and log-normal mation parameter variability in DM design. If it
cases is much smaller, only 1% at COV VE ,TOT = 20% is adapted and presented in the context of its par-
and 11% at COV VE ,TOT = 50% . ticular area of application, the RBD methodol-
Figure 3 shows the impact of M soil on Ek/Ed, ogy can become a useful design aid (cf. Fig. 2) for
when M soil increased from 500 kPa to 1500 kPa. the DM industry. However, it must be noted that
As can be seen in the figure, the impact was low, Figures 2 and 3 are valid only for this particular
only 6% at COV VE ,TOT = 30% and less than 1% serviceability limit state model and only within
at COV VE ,TOT = 50% . The figure also shows the specified ranges of M soil and a.
impact of a on Ek/Ed. At COV VE ,TOT = 30% , the dif- For small values of COV VE ,TOT , the impact of
ference in Ek/Ed when a as increased from 0.1 to different distributions (normal or log-normal) and
0.4 was 6%. In practice, the small impact of M soil β on the quotient Ek/Ed is small, but the differences
and a on Ek/Ed means that Ek/Ed is almost entirely become substantial as COV VE ,TOT increases.
governed by COV VE ,TOT . Figure 3 also shows the In SLS design, Eurocode 0 suggests β = 1.5.
quotient M soil / M d plotted against COV VE ,TOT . However, it is the authors’ belief that a wider
For COV VE ,TOT = 20% , M soil / M d equals 1.4, but it range of β could be allowed for structures whose
approaches 1.0 with increasing COV VE ,TOT . potential failure is of minor consequence. Here β

420

ISGSR2013.indb 420 10/18/2013 9:43:31 AM

 should be decided by balancing the cost of making μ Ecol Mean value of Eccol
higher quality columns and additional tests against μ Esoil Mean value of M soil
the cost of structure maintenance. σlnx Standard deviation of the trans-
formed variable x
σ Ecol Standard deviation of Eccol
7 NOTATION σ M soil Standard deviation of M soil
γemb Unit weight of the embankment
a Lime-cement column area replace- Γ2 Variance reduction factor
ment ratio
hj Height of layer j
hemb Height of the embankment
m Sample mean value ACKNOWLEDGEMENT
qc Cone tip resistance
semb Embankment settlement The authors acknowledge the support of the
s2 Sample variance Development Fund of the Swedish Construction
xd Design value of a normally distrib- Industry, Skanska Sverige AB and the Swedish
uted variable Transport Administration.
xd,log Design value of a log-normally dis-
tributed variable
xk Characteristic value of a normally REFERENCES
distributed variable
xk,log Characteristic value of a log-normally Al-Naqshabandy, M. & Larsson, S. 2013. Effect of uncer-
distributed variable tainties of improved soil shear strength on the reliability
COV
VE ,TOT Total coefficient of variation of Ecol of embankments. ASCE Journal of Geotechnical and
COV
Vm qc Coefficient of variation associated Geoenvironmental Engineering 139(4): 619–632.
Al-Naqshabandy, M. & Larsson, S. 2012. Partial fac-
with random measurement noise
tor design for a highway embankment founded on
COV
VM ,soil Coefficient of variation of M soil lime-cement columns. Proc. ISSMGE—TC 211 Inter-
COVqc Coefficient of variation of penetra- national Symposium on Ground Improvement IS-GI
tion test data Brussels.
COV
Vtr ,E Coefficient of variation associated Al-Naqshabandy, M., Bergman, N. & Larsson, S. 2012.
with transformation model bias Strength variability in lime-cement columns based on
DM Deep mixing CPT data. Ground Improvement 165 (1): 15–30.
Ecol Average column modulus of elasticity Baecher, G.B. & Christian J.T. 2003. Reliability and
evaluated from in-situ or laboratory Statistics in Geotechnical Engineering. England: Wiley
and Sons Ltd.
tests
Babu, G.L., Sivakumar., Srivastava, Amit &
Ed Assumed design value of the column Sivapulliah, P.V. 2011. Reliability analysis of strength
modulus of elasticity of cement treated soils, Georisk: Assessment and
Ek Characteristic value of the column Management of Risk for Engineered Systems and
modulus of elasticity evaluated from Geohazards 5 (3–4): 157–162.
FORM analyses Bergman, N., Al-Naqshabandy, M. & Larsson, S. 2013.
FORM First-order reliability method Strength variability in lime-cement columns evaluated
L Length of lime-cement columns by CPT and KPS. Georisk 7(1): 21–36.
Md Design value of clay oedometer Broms, B. 1999. Keynote lecture: Design of lime, lime/
cement and cement columns. International Conference
modulus
on Dry Mix Methods: Dry Mix Methods for Deep Soil
M soil Characteristic value of clay oedom- Stabilization. Stockholm.
eter modulus evaluated from in-situ Ching, J., Phoon, K. & Chen Y. 2010. Reducing shear
or laboratory tests strength uncertainties in clays by multivariate
N Number of uncorrelated tests correlations. Canadian Geotechnical Journal 47(1):
RBD Reliability-based design 16–33.
SLS Serviceability limit state Eurocode 0: Basis of structural design. EN 1990:2002.
α Sensitivity factor Eurocode 7: Geotechnical Design—Part 1: General rules.
α Ecol Evaluated sensitivity factor of Eccol 2004. PrEN 1997-1:2004:E.
Honjo, Y. 1982. A probabilistic approach to evaluate
α M soil Evaluated sensitivity factor of M soil
shear strength of heterogeneous stabilized ground by
β Reliability index deep mixing method. Soils and Foundations 22 (1):
γm A partial factor 23–38.
δmax Maximum allowed settlement Honjo, Y., Hara, T. & Kieu Le, T.C. 2010. Level III Reli-
μlnx Mean value of the transformed vari- ability Based Design of Examples set by ETC10.
able x Proceedings of 17 SEAGC, Taiwan.

421

ISGSR2013.indb 421 10/18/2013 9:43:40 AM

 Kitazume, M. 2004. Approach to reliability based design Rackwitz, R. & Fiessler, B. 1978. Structural reliability
on deep mixing improved ground. Proceedings of under combined random load sequences. Computers
DFI’s 29th Annual Conference on Deep Foundation, and Structures 9: 489–494.
Vancouver. Thoft-Christensen P. & Baker M.J. 1982. Structural reli-
Larsson, S. 2005a. State of Practice Report ability and its application. Berlin: Springer-Verlag,
Session 6: Execution, monitoring and quality control. Berlin.
Proceedings of the International Conference on Deep TK Geo 2011. Swedish Design Guidelines, Trafikverket,
Mixing, Stockholm. Borlänge (In Swedish).
Larsson, S., Stille, H. & Olsson, L. 2005. On horizontal Zheng, J., Liu, Y. & Xu, Z. 2009. Reliability-based Design
variability in lime-cement columns in deep mixing. Applied to Multi-column Composite Foundations.
Géotechnique 55 (1): 33–44. In Advances in Ground Improvement: Research to
Liu, Y., Zheng, J. & Guo, J. 2008. Statistical evaluation for Practice in the United States and China: Proceedings
strength of pile by deep mixing method. Proceedings of the US-China Workshop on Ground Improvement
of GEDMAR08, Nanjing. Technologies, Orlando, Florida.

422

ISGSR2013.indb 422 10/18/2013 9:43:44 AM

 Geotechnical Safety and Risk IV – Zhang et al. (eds)
© 2014 Taylor & Francis Group, London, ISBN 978-1-138-00163-3

Cost-effective framework for simplified geotechnical reliability-based

design

J.Y. Ching
Department of Civil Engineering, National Taiwan University, Taipei, Taiwan

K.K. Phoon
Department of Civil and Environmental Engineering, National University of Singapore, Singapore

ABSTRACT: Reliability-Based Design (RBD) is known to provide a rational basis for incorporating
uncertainties in the design environment explicitly into geotechnical design. In principle, RBD should
be able to translate site investigation efforts into design savings. However, the constant partial factor
approach prohibits such translation. This study shows that the combination of two recent developments
made by the authors allows such translation. The first development is the multivariate probability distri-
bution of soil parameters. Such a multivariate distribution provides a solid basis that can translate site
investigation efforts into smaller parameter variability, e.g. smaller Coefficient Of Variation (COV). The
second development is called the Quantile Value Method (QVM). The QVM adopts constant quantile val-
ues as the design values, so the resulting partial factors scale with the parameter COV—when parameter
COV is small, partial factors are closer to one. The combination of the above two novel developments is
a cost-effective framework that is able to translate site investigation efforts into smaller parameter COVs
and in turn translate into partial factors that are closer to one.

1 INTRODUCTION RBD methods are used to obtain conservative

design values for soil parameters or resistances.
More economical geotechnical designs can be Although these simplified RBD methods can pro-
achieved by reducing uncertainties through site duce familiar “look and feel” design codes that
investigation, in particular, by reducing uncertain- do not require engineers to perform probabilistic
ties in soil parameters. While it is theoretically cor- calculations, they are unable to achieve the desired
rect that reduction in uncertainties should translate target reliability index under all design scenarios,
to design savings, practical questions such as “how namely, the translation from site investigation to
many meters of reduction in pile length?” and “is final design savings may not be satisfactory.
pile length reduction worth the engineer’s time/ This study shows that the combination of two
money to collect more site information?” cannot recent developments made by the authors allows
be answered in a general theoretical sense. These such translation. The first development is the mul-
questions can only be answered empirically by tivariate probability distribution of soil parameters
applying Reliability-Based Design (RBD) to actual (e.g., Ching et al. 2010, Ching & Phoon 2012b).
design problems where the amount of site informa- Such a multivariate distribution provides a solid
tion can be varied systematically. basis that can translate site investigation efforts
Ching & Phoon (2012a) showed that rigorous into smaller parameter variability, e.g. smaller
RBD (direct probability-based design method) is Coefficient Of Variation (COV). The second devel-
able to link site investigation efforts to final design opment is called the Quantile Value Method (QVM)
savings. They demonstrated that site investigation (Ching & Phoon 2011, Ching & Phoon 2013). The
(as an illustration of “information”) is not only QVM adopts constant quantile values as the design
a cost item but also an investment item, because values, so the resulting partial factors scale with the
reduction of uncertainties through more and/or parameter COV—when parameter COV is small,
better tests can translate directly to design sav- partial factors are closer to one. The combination of
ings through rigorous RBD. However, practicing the above two novel developments is a cost-effective
engineers may not be familiar with the procedures framework that is able to translate site investigation
of rigorous RBD. The partial factors or load and efforts into smaller parameter COVs and in turn
resistance factors associated with these simplified translate into partial factors that are closer to one.

423

ISGSR2013.indb 423 10/18/2013 9:43:44 AM

 2 COST EFFECTIVE FRAMEWORK
FOR RBD

2.1 Multivariate probability distribution

of soils parameters
A database of OCR (overconsolidation ratio),
su (undrained shear strength), qT − σv (net cone
resistance), and N60 (SPT N corrected for energy
efficiency) is compiled in Ching et al. (2010). This
database was used to construct the multivariate
probability distribution of (OCR, su, qT − σv, N60).
Given the multivariate probability distribution,
the Bayesian analysis is taken to derive some use- Figure 1. Probability density functions of resistance Q
ful equations for the (updated) mean and COV for scenarios A (left) and B (right); the vertical dashed
of su given information such as OCR, su, qT − σv, lines indicate the design value Qd with the constant par-
tial factor γQ = 0.6.
and N60. It is concluded that the updated COV
decreases with increasing information. Ching &
Phoon (2012b) presented another clay database
containing multivariate LI (Liquidity Index),
su, sure (remolded undrained shear strength), σ ′p
(preconsolidation stress), and σ ′v (effective ver-
tical stress) data points. This database was used
to construct the multivariate probability dis-
tribution of (LI, su, sure, σ ′p, σ ′v). It is also con-
cluded that the updated COV of su decreases with
increasing information. In summary, a multivari-
ate probability distribution provides a solid basis
that can translate site investigation efforts into
smaller parameter variability, e.g. smaller Coef-
ficient Of Variation (COV).
Figure 2. Probability density functions of resistance Q
for scenarios A (left) and B (right); the vertical dashed lines
indicate Qd with constant probability threshold = 0.02.
2.2 Quantile value method
The first framework (multivariate probability dis-
tribution) alone does not guarantee the link from Ching & Phoon (2011) showed that for prob-
site investigation efforts to final design savings. lems with variable COVs, it may be possible to
To elaborate on this, consider a simple pile design maintain better uniformity with fixed quantiles.
problem involving two lognormal random vari- The strategy of fixed quantiles is formally named
ables, the resistance/capacity Q and the action F. the Quantile Value Method (QVM) in Ching &
Let δQ be the COV of the resistance. Let scenario Phoon (2013) and can be explained in Figure 2
A be a case where very detailed site investigation for the same example. The QVM adopts design
has been conducted (e.g. OCR, qT − σv, and N60 values for resistance Q corresponding to fixed
are simultaneously known). As a result, δQ is small quantiles, e.g. 0.02 quantiles, for both scenarios
and equal to 0.2. Scenario B is a case where the A and B. For the action F, the 1 − 0.02 = 0.98
site investigation is cursory. As a result, δQ is large quantile is taken to be the design value because
and equal to 0.45. Let us consider the strategy it is a destabilizing variable. The number 0.02 is
of implementing a constant partial factor for Q, called the “probability threshold η″ in Ching &
e.g. γQ = 0.6, for both scenarios. This implies that Phoon (2011). For scenario A, the resulting Qd is
Qd = Qk × 0.6 (Qk is the characteristic (mean) value the 0.02 quantile = Qk × 0.65, but for scenario B,
of Q, and Qd is the design value) will be taken for the 0.02 quantile becomes Qk × 0.38. Note that the
the design calculations of both scenarios. For sce- fixed quantile strategy implies different partial fac-
nario A, there is a very small chance that the actual tors should be used for scenarios with different
Q < Qd, but for scenario B, this chance becomes COVs. The above 0.02 probability threshold is just
much larger (Fig. 1). It is quite unlikely that the for illustration. In reality, the probability threshold
same design value of Qd = Qk × 0.6 can maintain a η should be calibrated by the reliability theory
uniform target reliability index over these two dis- developed in Ching & Phoon (2011) to achieve a
parate scenarios. prescribed target reliability index βT.

424

ISGSR2013.indb 424 10/18/2013 9:43:45 AM

 2.3 ETC10 example for pad foundation design is the weight of the foundation. The RBD require-
ment is to enforce the reliability constraint
The pad foundation example adopted in this paper
was developed by ETC10 (Eurocode 2011). The
square pad foundation with width B is embed- ⎡⎣ ( X , B ) < 0 ⎤⎦ ≤ pT = Φ ( − T ) (3)
ded at a depth of 0.8 m. It is designed to support
the following loads: (a) permanent vertical L, and where Φ(⋅) is the cumulative density function of
(b) variable horizontal load Qh and variable verti- the standard normal random variable; pT is the
cal load Qv. All loads are assumed to be (statisti- target failure probability; βT is the target reliability
cally) independent of each other. The soil consists index. The Partial Factor Design (PFD) require-
of the Boulder Clay, with undrained shear strength ment is to enforce
of su and a bulk unit weight of 21.4 kN/m3. The
concrete unit weight γc is 25 kN/m3. The ground ⎛ ⎡ γ Qh rQQLL μL sin (α)⎤ ⎞
water table is assumed to be deeper than B below B 2 ⎜0.6 (π + 2) ⎢1 1− ⎥ γ su μsu + 17.12⎟
⎜⎝ ⎢⎣ B γ su μsu ⎥
2
⎟⎠
the depth of embedment. Hence, it has no influ- ⎦
ence on the bearing capacity of the foundation.
There are four random variables {su, L, Qv, Qh}, ≥ γ L μL + γ Qv rQL μL cos ( ) + 20 × B 2
and their distributions and statistics are listed (4)
in Table 1. For simplicity, the characteristic val-
ues of {su, L, Qv, Qh} are assumed to be equal to where γ ′s are the partial factors.
their respective mean values, i.e. su,k = μsu, Lk = μL,
Qv,k = μQv, and Qh,k = μQh. The ratio between the
2.4 Simplified RBD codes for pad foundation
mean variable load to mean dead load is rQL = μQ/
in clays
μL = (μQv2 + μQh2)0.5/μL, and the angle α of the vari-
able load is defined to be tan−1(μQh/μQv). For the calibration of simplified RBD codes, par-
This study focuses on the Ultimate Limit State tial factors must be calibrated with respect to a
(ULS) requirement. We adopt the equation recom- wide range of design scenarios rather than a spe-
mended in Eurocode 7 [Eq. (D1) in page 157, BS cific design scenario. The range of the design sce-
EN 1997-1:2004 (British Standards Institute 2004)] narios considered in this study are listed in Table 2.
for calculating qu: There are six design parameters: B, μsu, δsu, μL, rQL,
and α. The range for μL is not specified, as it has no
qu ( + ) su ⋅ sc ic + q effect on the calibrated partial factors. The partial
factors for L, Qv, and Qh are chosen to be γL = 1.1,
⎡ Q ⎤
= 0.6 ( + ) ⎢1 + 1 − 2h ⎥ su + 17.12 (1) γQv = 1.5, and γQh = 1.5 to be consistent with the
⎢⎣ B su ⎥⎦ structural codes. As a result, the only partial factor
to be calibrated is γsu. The target reliability index βT
where sc = 1.2 for the square foundation, is chosen to be 3.2.
ic = 0.5[1 + (1 − Qh/B2/su)0.5], and q = total surcharge Two simplified RBD formats are calibrated:
pressure = 0.8 × 21.4 = 17.12 kN/m2. The perform- a. (constant PF) The design value of su is equal
ance function can be written as to the characteristic value of su multiplied by
γsu, and a single γsu is calibrated to cover all
g( X , B ) qu B 2 L − Qv W p (2) design scenarios. This code format is adopted
in most current design codes (e.g. AASHTO
where X denotes the collection of all random varia- and Eurocode 7). A set of m = 360 cases with
bles {su, L, Qv, Qh}, and Wp = γc × B2 × 0.8 = 20 B2 kN (B, μsu, δsu, rQL, α) randomly drawn from the

Table 1. Statistics and characteristic values of the random variables.

Statistics
Characteristic
Variable Description Distribution Mean COV value

su Undrained shear strength Lognormal μsu δ δsu μ μsu

L Permanent vertical load Gaussian1 μL δ δL = 10%1 μL
Qv Variable vertical load Gumbel1 rQLμLcos(α) δQv = 20%1 rQLμLcos(α)
Qh Variable horizontal load Gumbel1 rQLμLsin(α) δQh = 20%1 rQLμLsin(α)
1
According to Sørensen (2002).

425

ISGSR2013.indb 425 10/18/2013 9:43:46 AM

 ranges given in Table 2 are adopted as the cali- Because su is assumed to be lognormally distrib-
bration cases. For the i-th calibration case with uted, γsu can be expressed as:
(Bi, μsu,i, δsu,i, rQL,i, αi), the “actual” reliability
index, denoted by βA,i, depends on the selected suη
value of γsu in Eq. (4). βA,i can be estimated γ su =
su,k
using Monte Carlo simulation. It is impos-
sible to find a perfect γsu so that βA,i = βT for
all m = 360 calibration cases, but a reasonable ⎣
(
= exp ⎡ l 1 + δ s2u ) 1
(η)⎤⎥
⎦
1 + δ s2u (6)

value of γsu can be found by solving the follow-

ing optimization problem: It is critical to note that the resulting partial fac-
tor γsu depends on the COV of su (δsu) of the future
m design case, i.e. γsu depends on the site investigation
min ∑ ⎡⎣ β A,i ( γ su ) 3.2 ⎤⎦
2
(5) efforts of the future design case.
γ su
i=1 Table 3 shows the calibrated γsu for the constant
PF method and the calibrated η for QVM with tar-
b. (QVM) The design value of su is equal to the get reliability index βT = 3.2. Note that the partial
η-quantile of su, and a constant η is calibrated factors γL = 1.1, γQv = 1.5, and γQh = 1.5 are not cali-
to cover to all design scenarios. This is the code brated and are fixed at values suggested in struc-
format associated with QVM (Ching & Phoon ture codes.
2011, 2013). The same 360 cases with random
(B, μsu, δsu, rQL, α) are taken as the calibration
cases. For the i-th calibration case, a probability 3 COMPARISONS AMONG CODE
threshold ηi is calibrated to achieve the target FORMATS FOR A DESIGN
reliability βT. The details for the calibration of IN A VIRTUAL SITE
ηi are presented elsewhere in Ching & Phoon
(2011, 2013). The final η is taken to be the aver- 3.1 A clay virtual site
age of η1, …, η360. With this average η value, the
design value su,d of a future design case is sim- Table 3 was calibrated over a wide range of design
ply its η-quantile, denoted by suη. Equivalently, scenarios, not for a single design case. The purpose
the required partial factor γsu is equal to suη/su,k. of this section is to examine which of the two codes
performs better in linking site investigation efforts
to final design savings for a specific design. In this
Table 2. Ranges of the design parameters selected for
section, site investigation efforts are varied system-
code calibration. atically using the concept of a virtual site developed
for clays by Ching & Phoon (2012a). This clay “vir-
Parameter Description Range tual site” can produce consistent reduction in δsu as
a function of the type and number of field tests.
B Foundation width [2 m, 4 m] The design goal of this design case is to determine
μsu Mean value of su [25 kN/m2, the width B for a pad foundation in this virtual site
200 kN/m2] to fulfill the target reliability index βT = 3.2. The
δsu COV of su [0.1, 0.7] values of (μsu, δsu, μL, rQL, α) for this specific design
L Mean value of L – are summarized below: μL = 1000 kN, rQL = 0.59,
rQL Live to dead load [0.1, 1.0] and α = 33.7°. The values of (μsu, δsu) depend on the
ratio = μQ/μL site investigation efforts. The clay is assumed to be
α Average inclination [0°, 40°]
overconsolidated with OCR uniformly distributed
angle of the variable
tan−1(μQh/μQv) between 5 and 24. Based on the above informa-
tion, important soil parameters can be simulated.

Table 3. Calibrated γsu for the constant PF method and calibrated η for constant QVM.

Calibrated γsu

Constant PF Constant QVM γL γQv γQh

γsu = 0.331 η = 0.0083

γ su ( ( )
δ su × −1
( ) ) 1 + δ su
2
1.1 1.5 1.5

426

ISGSR2013.indb 426 10/18/2013 9:43:47 AM

 The simulation is based on the graphical model Table 4. The information bounds for various precision
developed in Ching et al. (2010) which is deemed to scenarios.
adequately model the correlations among (OCR, su,
N60, qT − σv). The equations used in the simulation Information bounds based on site
investigation
steps summarized below are taken from Eqs. (4), Precision
(10), and (15) in Ching et al. (2010): scenario OCR N60 qT″ (kN/m2)
a. Simulate a MCS sample of OCR from the uni-
P0 Zero precision [1, 50] [0, 100] [200, 6000]
form distribution over the interval [5, 24].
P1 Poor precision [5, 25] [3, 18] [730, 2040]
b. Simulate a ln(su) sample from the following
P2 : [7.5, 16.7] [5, 12] [940, 1580]
equation:
P3 : [8.5, 14.6] [6, 10] [1030, 1450]
ln ( su ) ln (OC ) ln ( ) P4 Excellent [9.5, 13.1] [7, 9] [1100, 1350]
0.. v precision
− 0.874 + 0.237 × Z1 (7)

where 0.237 is the standard deviation of the

transformation uncertainty, and Z1 is stand- purely estimated from general literature appropriate
ard normal. The average vertical effective stress for “clay”. This scenario is the cheapest, because no
σ ′v0 = 2.3 × 21.4 = 49.22 kN/m2. site investigation is conducted. P4 means that suf-
c. Given the ln(su) sample from step (b), simulate a ficiently extensive tests are conducted to narrow the
ln(N60) sample from the following equation: ranges. There are 20 possible scenarios for site inves-
tigation efforts in terms of number of test types and
ln ( N60 ) 1. l ( su ) − 0.403 × ln (
ln v0 ) precision of information measured. For instance,
− 3.845 + 2 0.456 × Z2 (8) the scenario T2-P1 means two test types are avail-
able (N60 and qT − σv) and the information for
where 2 × 0.456 is the transformation uncer- each test type is known with reasonable precision,
tainty, and Z2 is standard normal. N60 = 3∼18 and qT − σv = 730∼2040 kN/m2. Samples
d. Given the ln(su) sample from step (b), simulate a of (OCR, su, N60, qT − σv) have been simulated
ln(qT − σv) sample from the following equation: from the virtual site using the procedure presented
in the previous section. Based on the information
ln (q ″T ) ln ( su ) + 2.54 + 0.34 × Z3 (9) bounds for a given scenario, the conditional sam-
ples of su can be easily obtained. For example, in
the case of T2-P1, samples satisfying N60 = 3∼18
where 0.34 is the transformation uncertainty,
and qT″ = 730∼2040 kN/m2 simultaneously can be
and Z3 is standard normal.
filtered out from the population at large. The su
e. Apply exponential to the ln(su), ln(N60), and
values associated with this filtered set are therefore
ln(qT − σv) samples to get the (su, N60, qT − σv)
the conditional su samples. In this study, the number
samples.
of conditional samples is chosen to be n = 106 for all
Repeat the above steps for n = 106 times to obtain 20 scenarios. Table 5 lists the statistics of the con-
n = 106 samples of (OCR, su, N60, qT − σv). ditional su samples, including conditional mean and
A variety of site investigation efforts are pro- conditional COV. It is clear that μsu and δsu indeed
duced by systematically changing: (1) the number of depend on the site investigation efforts and that δsu
test types and (2) the test precision. Four scenarios decreases with increasing site investigation efforts.
are considered for the number of test types: (T1)
only the range of N60 is known; (T2) the ranges for
3.2 Rigorous RBD for virtual site
N60 and qT − σv are both known; (T3) the ranges for
N60 and OCR are both known; and (T4) the ranges The rigorous RBD can be taken to determine the
for N60, qT − σv, and OCR are all known. Scenario required width B under each scenario that exactly
T1 is considered as the basic case with the least achieves the target reliability βT. This can be easily
effort, while T4 contains the most information in done by MCS—based on the n = 106 conditional
terms of number of test types. Five scenarios (P0 samples of su for each scenario and n = 106 inde-
to P4) are considered for test precision. The ranges pendent sets of samples of (L, Qv, Qh), one can
in Table 4 represent the bounds for OCR, N60, and solve for B such that P[g(X,B)<0] = P[qu × B2 −
qT − σv based on the assumption that more precise L − Qv − Wp < 0] = Φ(−βT). Table 5 shows the required
information on each test type measurement is avail- B for rigorous RBD. It is clear that the required B
able perhaps by increasing the number of tests and reduces from 3.33 m for P0 (zero site investigation
boreholes. For instance, P0 means no site-specific effort) to 2.84 m for T4-P4 (significant effort). Site
tests are conducted and information bounds are investigation efforts can be perfectly linked to final

427

ISGSR2013.indb 427 10/18/2013 9:43:49 AM

 Table 5. Statistics of conditional su samples under various scenarios.

Conditional mean in kPa

(conditional COV) (final design dimension for rigorous RBD)

T1 T2 T3 T4

P0 115.8
(0.35) (3.33 m)
P1 111.3 107.3 111.3 107.3
(0.32) (3.21 m) (0.27) (3.04 m) (0.32) (3.21 m) (0.27) (3.03 m)
P2 111.4 105.7 102.4 100.4
(0.31) (3.19 m) (0.25) (2.94 m) (0.25) (3.02 m) (0.21) (2.87 m)
P3 111.3 105.5 99.9 99.0
(0.31) (3.16 m) (0.24) (2.91 m) (0.24) (2.98 m) (0.20) (2.81 m)
P4 112.0 105.3 99.5 98.1
(0.31) (3.17 m) (0.24) (2.88 m) (0.23) (2.93 m) (0.19) (2.84 m)

Table 6. Final design dimensions (deviation in B to the rigorous RBD)/βA for the
constant PF method.

T1 T2 T3 T4

P0 3.61 m
(0.28 m)/3.72
P1 3.68 m 3.75 m 3.68 m 3.75 m
(0.47 m)/4.14 (0.71 m)/4.75 (0.47 m)/4.10 (0.72 m)/4.75
P2 3.68 m 3.78 m 3.84 m 3.88 m
(0.49 m)/4.16 (0.84 m)/4.75 (0.82 m)/≥4.75* (1.01 m)/≥4.75*
P3 3.68 m 3.78 m 3.88 m 3.90 m
(0.52 m)/4.16 (0.87 m)/≥4.75* (0.90 m)/≥4.75* (1.09 m)/≥4.75*
P4 3.67 m 3.78 m 3.89 m 3.92 m
(0.49 m)/4.16 (0.90 m)/≥4.75* (0.96 m)/≥4.75* (1.08 m)/≥4.75*

Table 7. Final design dimensions (deviation in B to the rigorous RBD)/βA for the
QVM.

T1 T2 T3 T4

P0 3.21 m
(−0.12 m)/2.96
P1 3.14 m 3.02 m 3.14 m 3.01 m
(−0.07 m)/3.05 (−0.02 m)/3.16 (−0.07 m)/3.04 (−0.02 m)/3.16
P2 3.10 m 2.94 m 3.02 m 2.89 m
(−0.08 m)/3.04 (−0.00 m)/3.20 (0.00 m)/3.21 (0.02 m)/3.26
P3 3.10 m 2.92 m 2.99 m 2.84 m
(−0.06 m)/3.08 (0.01 m)/3.22 (0.01 m)/3.21 (0.03 m)/3.28
P4 3.08 m 2.91 m 2.97 m 2.84 m
(−0.09 m)/3.01 (0.03 m)/3.27 (0.03 m)/3.29 (0.00 m)/3.21

3.3 Constant PF
design savings rationally using the rigorous RBD.
However, practicing engineers may not be com- Table 6 shows the required width B based on the
fortable with the rigorous RBD. The design table calibrated partial factor γsu = 0.331 given in Table 3.
in Table 3 developed by the two simplified RBD To find the required B for each information sce-
methods is more likely to be used in practice. nario, one only needs to solve Eq. (4) for B with

428

ISGSR2013.indb 428 10/18/2013 9:43:50 AM

 γsu = 0.331, μL = 1000 kN, rQL = 0.59, and α = 33.7°. eters. However, whether uncertainty reduction due
The mean undrained shear strength, μsu, is extracted to the site investigation efforts can be linked to final
from Table 5 for each information scenario. Note design savings using simplified Reliability-Based
that δsu in Table 5 is not used in the constant PF Design (RBD) methods has not been clarified in
method. The discrepancy between the required B literature quantitatively with a sound theoretical
determined by the constant PF method and the basis. The link can be easily established using the
one determined by the rigorous RBD shown in rigorous RBD method, but practicing engineers are
Table 5 is listed in the parenthesis (⋅). If all discrep- generally not comfortable with the rigorous RBD
ancies are small, the constant PF method is nearly to date. The most widely adopted compromise in
as effective as the rigorous method in linking the existing RBD codes is to use simplified methods
site investigation efforts to design outcomes, and containing reliability calibrated partial factors that
vice versa. Another way of verifying the effective- have the same look and feel as the global factor of
ness is to find the actual reliability index βA for safety method.
each required B, which can be easily calculated by According to the analysis results in this paper,
MCS with sample size n = 106. It is clear that the such a link for simplified RBD methods is possi-
constant PF method is not very effective because ble if: (a) the methods are sufficiently responsive
the discrepancy in B is often large, and βA often to a wide range of geotechnical information, and
significantly deviates from the target value 3.2. (b) the mean and COV of the soil parameters are
properly updated using appropriate correlation
models and Bayesian analysis. It is concluded
3.4 * No failure sample in MCS that the constant partial factor format that is
with n = 106 QVM adopted in many current codes (e.g. AASHTO and
Table 7 shows the required width B based on the Eurocode 7) fails to establish this crucial link. The
calibrated η = 0.0083 shown in Table 3. This prob- design method based on QVM seems to be able to
ability threshold η was also calibrated for a wide link site investigation efforts to final design savings
range of design scenarios shown in Table 2. Note in a more effective way.
that the resulting partial factor is not a constant: for
P0 (δsu = 0.35), γsu = 0.418 is rather small due to the
large COV, and for P4-T4 (δsu = 0.19), γsu = 0.626 is REFERENCES
large due to the small COV. The smaller deviation
in B from the rigorous RBD shows that the QVM British Standards Institute 2004. Eurocode 7:
is more effective than the constant PF method. In Geotechnical Design—Part 1: General Rules. BS EN
fact, the required B shows the same trend as the 1997-1:2004, London.
Ching, J., Phoon, K.K. & Chen, Y.C. 2010. Reducing
rigorous RBD—the largest for P0 and the smallest shear strength uncertainties in clays by multivariate
for T4-P4. As a result, the actual reliability index correlations. Canadian Geotechnical Journal 47(1):
βA for the QVM method is more uniform and 16–33.
closer to the target value 3.2 than the constant PF Ching, J. & Phoon, K.K. 2011. A quantile-based
method. approach for calibrating reliability-based partial fac-
It is evident that the QVM behaves similarly to tors. Structural Safety 33: 275–285.
the rigorous RBD (the required B reduces with Ching, J. & Phoon, K.K. 2012a. Value of geotechnical
increasing site investigation efforts), indicating site investigation in reliability-based design. Advances
that the QVM is effective in linking the site investi- in Structural Engineering 15(11): 1935–1945.
Ching, J. & Phoon, K.K. 2012b. Modeling parameters of
gation efforts to design savings. This is not an easy structured clays as a multivariate normal distribution.
task at all—remember that Table 3 was NOT cali- Canadian Geotechnical Journal 49(5): 522–545.
brated with respect to any specific design case with Ching, J. & Phoon, K.K. 2013. Quantile value method
any particular site investigation effort but were cal- versus design value method for reliability-based code
ibrated with respect to a very wide range of design calibration. Structural Safety 44: 47–58.
scenarios. The constant PF method cannot link Eurocode 2011. Geotechnical design ETC10 Design
the site investigation efforts to design savings at all. Example 2.2. Pad foundation with inclined eccentric
The inability to link to site investigation efforts is a load on boulder. http://www.eurocode7.com/etc10/
severe disadvantage as it nullifies a strong practical Example%202.2/index.html.
Sørensen, J.D. 2002. Calibration of partial safety factors
motivation to adopt RBD in the first place. in Danish Structural Codes. Workshop on Reliability
Based Code Calibration. Swiss Federal Institute of
Technology, ETH Zurich, Switzerland, March 21–22:
4 CONCLUSIONS 2002.

It is well accepted that site investigation efforts may

reduce the uncertainties in the design soil param-

429

ISGSR2013.indb 429 10/18/2013 9:43:50 AM

This page intentionally left blank
 Geotechnical Safety and Risk IV – Zhang et al. (eds)
© 2014 Taylor & Francis Group, London, ISBN 978-1-138-00163-3

Reliability index of pile foundations designed to SANS 1016-5

M. Dithinde
Department of Civil Engineering, University of Botswana, Botswana

J.V. Retief
Department of Civil Engineering, University of Stellenbosch, South Africa

ABSTRACT: In accordance with international trends in development of geotechnical codes, South

Africa has converted to limit state design philosophy through the publication of SANS 10160: 2010.
Within this suite of standards, Part 5 (SANS 10160-5) is devoted to basis of geotechnical design and
actions. Although SANS 10160 set of standards is formulated on the basis of reliability principles, the
resistance partial factors prescribed in SANS 10160-5 were not developed on the basis of calibration stud-
ies but were adopted from the UK National Annex to EN 9197-1. In lieu of the calibration process, the
critical question is what is the reliability index of piles design to SANS 1060-5? The paper uses resistance
statistics from a comprehensive pile load tests database in conjunction with partial resistance factors and
the model factor prescribed in SANS 10160-5 to compute reliability indices (β-values) for various pile
classes designed to this code. It was found that the β values for all pile classes are above the target β of 3.0
for the reference class of structures as specified in SANS 10160-1 indicating that the resistance factors in
SANS10160-5 adopted from BS EN1997-1 are conservative and uneconomical for Southern Africa.

1 INTRODUCTION geology, soil type and conditions, site investigation

practices (extend, methods, standards, equipment
SANS 10160:2010 is essentially a loading code advances). Since these factors generally differ from
and presents the general principles, requirements one country to another, values developed for a spe-
for design of buildings and industrial structures cific country cannot be simply adopted by another
in accordance with the limit state design approach country. The need to calibrate geotechnical resist-
in South Africa. The procedures are deemed to ance factors for different applications utilising local
achieve acceptable levels of structural performance databases was further emphasised in a study to
in terms of safety, functionality and economy. It review developments of limit state design or LRFD
consists of eight self-contained parts, each deal- methods in Canada, Germany, France, Denmark,
ing with a specific action type. Accordingly SANS Norway and Sweden (DiMaggio et al. 1999). For
10160-5 set out the basis of geotechnical design and the same reason, EN1997-1 leaves the calibration
gives guidance on the determination of geotechni- of partial factors to individual countries.
cal actions on buildings and industrial structures. In lieu of the calibration process, the critical
The actions include vertical earth loading, earth question is what is the reliability index of pile
pressure, ground water and free water pressure, foundations designed to SANS 1060-5? To answer
as well as actions caused by ground movement. this question, this paper uses resistance statistics
However, due to the absence of geotechnical limit from a comprehensive pile load tests database for
state code in the country, SANS 10160-5 also pro- Southern Africa (Dithinde et al. 2011) in conjunc-
vides guidance on geotechnical design aspects of tion with partial resistance factors and the model
structures within the scope of SANS 10160 includ- factor prescribed in SANS 10160-5 to compute
ing pile foundations. reliability indices (β-values) for various pile classes
Although SANS 10160-5 have the same basis designed to this code.
and design format as that for structural parts,
in general there is lack of analytical calibration
2 PILE DESIGN TO SANS 10160-5
and verification. The resistance partial factors
for pile foundations have been adopted from BS
2.1 Design approach adopted by SANS 10160-5
EN1997-1. It is important to note that calibra-
tion of partial factors is dependent on local design EN 1997-1 provides three design approaches
practice, experience and environment such as local for design of substructures or structural

431

ISGSR2013.indb 431 10/18/2013 9:43:50 AM

 members involving geotechnical actions. These Table 1. Partial factors for pile foundations.
are: (a) Design Approach 1, (b) Design Approach
2, and (c) Design Approach 3. These three design Parameters Limit states
approaches differ in the way they distribute the
1. Partial factors for actions STR STR-P GEO
partial factors on actions and ground resistances/ Permanent (unfavourable) 1.2 1.35 1.0
properties. The attributes of the three design
Variable (unfavourable) 1.6 1.0 1.3
approaches have been discussed by many commen-
Permanent (favourable) 0.9 0.9 1.0
tators (e.g., Driscoll & Simpson 2001, Orr 2002,
Variable (fafourable) 0 0 0
Orr 2006, Frank 2002, Schuppener & Frank 2006).
2. Partial material factors * **
The selection of a particular design approach is
Shearing resistance (tanφ) 1.0 1.0 1.0 1.25
a matter for national determination. Accordingly
Effective cohesion (c) 1.0 1.0 1.0 1.25
SANS10160-5 has selected Design Approach 1 for
Undrained strength (cu) 1.0 1.0 1.0 1.4
two main reasons (Day & Retief 2009): (i) partial
Weight density (γ) 1.0 1.0 1.0 1.0
factors are applied directly to the source of uncer-
3. Partial resistance factors
tainty (i.e. actions, calculation models, and mate-
Compression piles 1.0 1.0 1.6
rial properties), (ii) the results obtained using this
Tension piles 1.25 1.25 1.7
approach agree reasonably with the current South
African practice based on the overall factor of *Applies for the resistance of piles and anchors.
safety approach. **Applies for unfavourable actions on piles owing to
negative skin friction or lateral loading.
2.2 Pile design methods and procedure
SANS 10160-5 recognises that piles can be designed
using one of the following methods: (i) Full scale ues of the partial factors presented in Table 1 were
pile load tests, (ii) Empirical analysis by directly developed for application in conjunction with the
using standard field tests results (e.g. SPT and correlation actors (ξ). Accordingly these partial fac-
CPT), (iii) Static analysis using engineering prop- tors are not sufficiently large to be used alone in the
erties of the soil as determined from laboratory or analytical approach. Therefore for the static analysis
in-situ field testing and (iv) Dynamic driving resist- method, the partial factors need to be corrected by a
ance (Pile driving formulae and Wave equation). In model factor (γM) larger than 1. In this regard SANS
contrast to EN 1997-1 in which the predominant 10160-5 recommend a model factor of 1.5.
pile design methods are the full scale pile load tests Furthermore, in accordance with Design
and empirical methods, the main design method in Approach 1, two combinations need to be consid-
SANS10160-5 is the static analysis. Accordingly ered as follows:
the analysis presented in this paper is based on the Combination 1 (STR/STR-P)
static analysis. The design resistance (Rd) is:
The static analysis entails directly determining
the characteristic base resistance (qbk) and shaft Rd = Rk/γM γt = Rk/1.5 × 1.0 = Rk/1.5
resistance (qsk) using the static formula. Hence the Combination 2 (GEO)
characteristic resistance (Rc,k) is given by:
Rd = Rk/γM γt = Rk/1.5 × 1.6 = Rk/2.4.
Rc,k = Rb;k + Rs;k (1)

The design resistance (Rd) then given by: 3 RELIABILITY ANALYSIS

Rd = Rk/γMγt (2) 3.1 Performance function and basic design

variables
The design load (Fd) is given by:
In accordance with the algorithm for the compu-
tations of reliability index β, the first step is the
Fd = γG ⋅ Gk + γQ ⋅ Qk (3)
formulation of the performance function. In this
regard the limit state function is given by:
where γM is the model factor, γt is the relevant resist-
ance partial factor from Table 1, γG is the permanent
action partial factor, Gk is the characteristic perma- R D−L=0 (4)
nent action, γQ is the partial factor on variable action
and Qk is the characteristic variable action. where R, D, and L are random variables defined
The partial factors to be used in Eqs. 2–3 are pre- as follows: R = measured resistance; D = measured
sented in Table 1. It is important to note that val- permanent load and; L = measured variable load.

432

ISGSR2013.indb 432 10/18/2013 9:43:51 AM

 Usually the measured load and resistance are setting and pile design practice reported by
presented in terms of their respective predicted Dithinde & Retief (2013) have been adopted and
values and mean model factors as follows: are presented in Table 3. The complete set of
174 cases was further classified in terms of four
R M R Rn ; D = M D Dn ; L M L Ln (5) theoretical principal pile design classes based on
both soil type and installation method. These
where: MR = mean model factor for resistance, fundamental set of classes include: (i) Driven
MD = mean model factor for dead load, ML = mean piles in Non-Cohesive soil (D-NC) with 29 cases,
model factor for live load and the other symbols (ii) Bored pile in Non-Cohesive soil (B-NC) with
are as previously defined. 33 cases; (iii) Driven piles in Cohesive soils (D-C)
Substituting Eq. 5 into Eq. 4, the performance with 59 cases, and (iv) Bored pile in Cohesive soils
function becomes: (B-C) with 53 cases. These principle four data sets
were further combined into various practical pile
M R Rn − M D Dn − M L Ln = 0 (6) design classes considered in design codes such as
SANS 10169-5:2011 and EN 1997-1:2004. The
In Eq. 6, MR, MD, and ML are now the random additional classification schemes include:
variables following specific probability distribu- 1. Classification based on pile installation method
tions while Rn, Dn, and Ln are deterministic values. irrespective of soil type. This is the classifica-
tion adopted in EN 1997-1:2004 and it yields:
87 cases of driven piles (D) and 83 cases of
3.2 Load and resistance statistics
bored piles (B).
and their probability models
2. Classification based on soil type. This classifi-
In reliability index computations for the perform- cation system is supported by the general prac-
ance function in Eq. 6, the statistics (mean, stand- tice where a higher factor of safety is applied
ard deviation and coefficient of variation) as well to pile capacity in clay as compared to sand.
as the type of distribution that best fit the load and This combination results in 58 cases in Non-
resistance are required. With regard to load statis- Cohesive soil (NC) and 112 cases in Cohesive
tics, values assumed in the previous South African soil (C).
loading code (SABS 0160:1988) reported by Kemp 3. All pile cases as a single data set irrespective of
et al. (1987) were adopted and reproduced here as pile installation method and soil type. This is
Table 2. The distribution followed by each load the practical consideration presented in SANS
type is also shown. 10160-5:2011 where a single partial factor is
The resistance statistics based on a pile load test given for all compressive piles. The scheme
database from the Southern African geological yields 174 pile cases (ALL).

Table 2. Load statistics (Kemp et al. 1987).

Mean load/ Coefficient of Type of

Type of load Code nominal load variation distribution

Permanent load SABS 0160 1.05 0.10 Lognormal

Live (office): lifetime max. SABS 0160 0.96 0.25 Type 1

Table 3. Resistance statistics (after Dithinde & Retief 2013).

M n Mean (m) Std. dev (sM) COV Skewness Kurtosis

D-NC 28 1.11 0.36 0.33 0.35 −1.15

B-NC 30 0.98 0.23 0.24 0.14 −0.19
D-C 59 1.17 0.3 0.26 −0.01 −0.74
B-C 53 1.15 0.28 0.25 0.36 0.49
D 87 1.15 0.32 0.28 0.1 −0.95
B 83 1.09 0.28 0.25 0.41 0.47
NC 58 1.04 0.30 0.29 0.55 −0.37
C 112 1.16 0.29 0.25 0.15 −0.29
ALL 170 1.1 0.31 0.28 0.24 −0.75

433

ISGSR2013.indb 433 10/18/2013 9:43:51 AM

 Based on the analysis carried by Dithinde & set on a spread sheet to compute β for the vari-
Retief (2013), a lognormal distribution was adopted ous pile classes. As already mentioned, the calibra-
for resistance. The analysis concluded that tion points were defined by the Ln/Dn ratio. Typical
although at the customary 5% confidence level, the range of Ln/Dn ratios are 0.5–1.5 for concrete struc-
chi-square goodness-of-fit test results indicate that tures and 1–2 for steel structures (Melchers 1999).
both the Normal and Lognormal distributions are Based on this information, a practical range of
valid, the Lognormal distribution has a slight edge, Ln/Dn ratio of 0.5 to 2 was adopted.
particularly towards the lower tail. Reliability index β computations were carried
out using an excel spread sheet developed by
Low & Tang (2007).
3.3 Limit state design equation
To facilitate the computation of reliability index
using the spreadsheet, the limit state design equa- 4 RESULTS AND DISCUSSIONS
tion in addition to the performance function is
required. As already alluded to, limit state design β values were sensitive to the variation in Ln/Dn
approach entails applying partial factors to both ratio and not the variation in the nominal resist-
the actions and resistance. Accordingly, the design ance, implying that the calibration points are only
equation is given by: defined by the Ln/Dn ratio. Accordingly beta val-
ues were calculated for a range of Ln/Dn ratio. The
Rk resulting β values for the GEO limit state as a func-
= γ G Gk γ QQk (7) tion of the Ln/Dn ratio are shown in Figure 1.
γR γM
Further analysis of Figure 1 indicates that:
The shapes of the curves follow a characteristic
where Rk = characteristic predicted pile capacity;
pattern in which for permanent load dominated
γR = partial resistance factor; γM = model factor;
structures (i.e. Ln/Dn < 1), β values increase with
Gk = characteristic permanent action; γG = partial
the increase in the Ln/Dn ratio. Conversely as live
factor for permanent action; Qk = characteristic
loads become significant (i.e. Ln/Dn > 1), β values
variable action; and γQ = partial factor for variable
decreases with the increase in the Ln/Dn ratio. The
action.
scenario is caused by loading and in SANS 10160
For the GEO limit state, the values of γG, γQ, γR,
it has been accounted for by the introduction of
and γM are 1.0, 1.3 and 1.6, and 1.5 respectively.
STR-P (see Table 1) to cater for permanent load
Therefore Eq. 7 becomes:
dominated structures.
The β values vary with pile classes as well as
Rk
= 1.00G
Gk 1 3Qk (8) with individual cases represented by the Ln/Dn
1.6 x1.5 ratios within the same pile class. This suggests that
despite that SANS 10160-5 provides a single resist-
For simplicity the calculation was done in the ance partial factor and model factor for all the pile
load space which entails expressing Ln in terms of classes, it does not produce a consistent level of reli-
Dn. By so doing, it is not necessary to deal specifi- ability across different design situations and even
cally with pile diameter and length as design situa- within the same design situation. This scenario is
tions are now represented by the Ln/Dn ratio. When not surprising as the partial factors prescribed in
Ln is expressed in terms of Dn, Eq. 8 becomes: SANS10160-5 were not determined on the basis
of reliability calibration capturing the distinct soil
Rk ⎛ 1.3Qk ⎞ types for the geologic region of Southern Africa as
= Gk 1 + (9)
2.4 ⎝ Gk ⎟⎠ well as the local pile design and construction expe-
rience base.
Piles in cohesive materials (D-C, B-C, and C)
From Eq. 9, the expressions for Gk and Qk are
depict higher reliability irrespective of installa-
as follows:
tion method compared to piles in non-cohesive
materials (D-NC, B-NC, and NC). The scenario
Rk
Gk = (10) is attributed to the variability exhibited by piles
2.4 ( + Qk Gk ) in non-cohesive materials as demonstrated by
standard deviations or coefficient of variations in
1 ⎛ Rk ⎞
Qk = − Gk ⎟ (11) Table 3. Dithinde & Retief (2013) attributed the
1.3 ⎝ 2.4 ⎠ scenario to the fact that in cohesive materials the
undrained shear strength derived from the SPT
The design equation, the performance function, measurement is directly used in the computation
and the load and resistance statistics were then of pile capacity while in non-cohesive materials,

434

ISGSR2013.indb 434 10/18/2013 9:43:52 AM

 Figure 1. Variation of beta values with Ln/Dn ratio.

the angle of friction obtained from the SPT meas- Nonetheless, the relatively lower β values depicted
urement is not directly used. Instead the key pile by NC are mainly due to the data set from D-NC
design parameters in the form of bearing capac- which has the lowest β. It appears that D-NC sig-
ity factor (Nq), earth pressure coefficient (ks) and nificantly diverges from the rest of the dataset and
pile-soil interface friction (δ) are obtained from the therefore warrants further investigations in future
derived angle of friction on the basis of empirical studies.
correlation and thus introducing some additional A comparison of all Bored (B) versus all Driven
uncertainties. (D) piles irrespective of soil type indicates very lit-
With piles in none-cohesive soils, there is a tle difference in terms of β values. This suggests
significant difference in β values between bored that pile installation method has little influence on
and driven piles. In this regard β values for bored β values. Therefore from reliability perspective, pile
piles are higher than that for driven piles. Again design should be classified on basis of soil only.
this attributed to the higher uncertainty depicted This further implies that resistance partial fac-
by driven piles relative to bored piles presented in tors and model factors should be differentiation
Table 3. Dithinde & Retief (2013) further explains on basis of soil properties. Further differentiation
that this implies that the densification of the soil into pile installation method is just a refinement
surrounding the pile emanating from the pile driv- and elaboration.
ing process is not well captured in the selection of Generally the β values for all pile classes are
the soil design parameters and hence current prac- above the target β of 3.0 for the reference reliability
tice is conservative in selecting design parameters class R2 for which reliability procedures are speci-
for driven piles. fied in SANS 10160. When taking into account the
Comparing β values for all piles in Cohesive redundancy due to group and system effects, the
materials (C) versus all piles in Non-Cohesive (NC), β values will increase appreciably. Therefore for
shows distinctively different level of reliability sug- pile groups (which is the common practice), SANS
gesting the two broad classes based on soil type 10160-5 yields β values that are significantly higher
should be treated separately in calibration studies. than the target β of 3.0 for class R2. Conversely,

435

ISGSR2013.indb 435 10/18/2013 9:43:53 AM

 in exception of piles in cohesive materials β values REFERENCES
for all other pile classes are below target β of 3.8
for the reference class of structures as specified in Day, P.W. & Retief, J.V. 2009. Provision for geotechnical
EN 1990. Even for piles in cohesive materials, the design in SANS10160. Background to SANS 10160.
β values are quite close to that target β in EN1990. SunMedia, Stellenbosch.
DiMaggio, J., Saad, T., Allen, T., et al. 1999. Geotechnical
Perhaps this is an indication that the resistance fac- engineering practice in Canada and Europe. Report
tors in SANS10160-5 adopted from BS EN1997-1 No. FHWA–PL-99-013, Federal Highway Adminstra-
are conservative and uneconomical for Southern tion, Washington, DC.
Africa. Even with Eurocodes, the determination Dithinde, M., Phoon, K.K., De Wet, M., et al. 2011.
of levels of safety, including aspect of durability Characterization of model uncertainty in the static
and economy remains within the competence of pile design formula. Journal of Geotechnical and
individual member states. Therefore it appears that Geoenvironmental Engineering 137(1): 70–85.
in implementation of limit state design, local cali- Dithinde, M. & Retief, J.V. 2013. Pile design practice in
bration studies are inevitable. southern Africa I: Resistance statistics. Journal of the
South African Institution of Civil Engineering 55(1):
60–71.
Driscoll, R. & Simpson, B. 2001. EN 1997 Eurocode 7:
5 CONCLUSIONS Geotechnical design. Proceedings of ICE.
EN 1997: 2004. Eurocode 7: geotechnical design—Part 1:
Using resistance statistics generated from local pile General rules. European Committee for Standardiza-
load tests, load statistics from the previous South tion (CEN), Brussels.
African loading code (SABS 0160), and partial fac- Frank, R. 2002. Development and feature of Eurocode
tors prescribed in SANS10160-5 the paper evalu- 7—Geotechnical design. Proc. of Foundation Design
ated reliability indexes of pile foundations designed Codes and Soil Investigation in View of International
Harmonization and Performance-Based Design,
to SANS10160-5. The key conclusions drawn from Kamakura, Japan.
the analyses and discussions are as follows: Holický M. 2009. Reliability analysis for structural design.
1. SANS10160-5 does not achieve a consistent Stellenbosch: SUNMeDIA Press.
level of reliability as β values vary with pile Kemp, A.R., Milford, R.V. & Laurie, J.P.A. 1987.
Proposal for a comprehensive limit states formulation
classes as well as with individual cases repre- for South African structural codes. The Civil Engineer
sented by the Ln/Dn ratios within the same pile in South Africa 29(9): 351–360.
class. This is attributed to lack of rigorous cali- Low, B.K. & Tang, W.H. 2007. Efficient spreadsheet
bration of partial factors capturing the distinct aligirithm for FORM. Journal of Engineering
soil types for the geologic region of Southern Mechanics 133(12): 1378–1387.
Africa as well as the local pile design and con- Melchers, R.E. 1999. Structural reliability: Analysis and
struction experience base. prediction. Chichester, New York: John Wiley & Sons.
2. The obtained β values are influenced by the Orr, T.L.L. 2002. Selection of characteristic values and
soil type and not pile installation methods. This partial factors in geotechnical designs to Eurocode 7.
Computers and Geotechnics 26(3–4): 263–279.
implies that resistance partial factors and model Orr, T.L.L. 2006. Development and implementation of
factors should be differentiated on basis of soil Eurocode 7. Proc. of the International Symposium on
properties. New Generation Design Code for Geotechnical Engi-
3. β values for all pile classes are above the target neering Practice, Taipei, Taiwan.
β of 3.0 for the reference class of structures as SANS 10160-5:2010. Basis of structural design and
specified in SANS 10160-1. If redundancy due actions for buildings and industrial structures—Part 5:
to group and system effects is accounted for, the Basis for geotechnical design and actions. SABS Stand-
β values will become significantly higher than ards Division, Pretoria.
the target β of 3.0 indicating that the resist- Schuppener, B. & Frank, R. 2006. Eurocode 7 for geo-
technical design—basic principles and implementation
ance factors in SANS10160-5 adopted from BS in the European member states. Proc. of the Interna-
EN1997-1 are conservative and uneconomical tional Symposium on New Generation Design Code for
for Southern Africa. Therefore it appears that Geotechnical Engineering Practice, Taipei, Taiwan.
in implementation of limit state design, local
calibration studies are inevitable.

436

ISGSR2013.indb 436 10/18/2013 9:43:55 AM

 Geotechnical Safety and Risk IV – Zhang et al. (eds)
© 2014 Taylor & Francis Group, London, ISBN 978-1-138-00163-3

Codified reliability-based design of shallow foundations in Shanghai

J.P. Li
Department of Geotechnical Engineering, Tongji University, Shanghai, China

S.N. Liu & S.N. Hou

Shanghai Xiandai Architectural Design Co., Ltd., Shanghai, China

J. Zhang
Department of Geotechnical Engineering, Tongji University, Shanghai, China

ABSTRACT: As part of the worldwide efforts for implementing reliability-based design in geotechnical
engineering, Shanghai also calibrated the resistance factors for design of shallow foundations using
reliability theory. This paper introduces how the resistance factors are determined in Shanghai. A typical
shallow foundation is assumed to be constructed on 142 sites in Shanghai. The mean values of cohesion
and friction angle are determined based on the site-specific data. The Coefficients Of Variation (COV)
of the cohesion and friction angle are determined based on regional experience considering variance
reduction due to spatial correlation of soil properties. The first order reliability method is then used to
determine the partial factors for each site. It is found that if the notional Factor Of Safety (FOS) is 2.5,
the reliability indexes of the 142 foundations are in the range of 2.3–4.2 with a mean of 3.35. If the target
reliability index is 3.35, the partial factor for cohesion for the 142 sites have a mean of 2.76 and a standard
deviation of 0.32, and the partial factor for friction angle has a mean of 1.17 and a standard deviation
of 0.07. It is then recommended that the partial factors for cohesion and friction angle be 2.7 and 1.2,
respectively. The adopted partial factors produce design bearing capacity predictions consistent with local
experience.

1 INTRODUCTION in Shanghai, China using the reliability theory

with a semi-empirical procedure when revising
Although the global Factor Of Safety (FOS) the local foundation design code (SUCCC, 2010).
method has been successfully used for many The objective of this study is to introduce how the
years, its disadvantage is obvious in that it does resistance factors in Shanghai are determined
not explicitly consider the level of uncertainty for design of shallow foundations. This paper
involved in a design. As a result, designs with the is organized as follows. First, the local bearing
same FOS may in fact correspond to different lev- capacity prediction model is introduced. Then,
els of safety. To overcome the disadvantage of the the methodology for partial factors calibration
global FOS method, probabilistic methods can is described. Finally, the calibration results are
be used to model uncertainty explicitly, through interpreted, analyzed, and verified with previous
which the safety in a design can be controlled by experience.
limiting the chance of unsatisfactory perform-
ance to a sufficiently low level. In the past dec-
ades, extensive researches have been conducted 2 BEARING CAPACITY MODEL
to develop partial factors for design of founda-
tions based on the reliability theory (e.g., Barker Experience in Shanghai indicates that Hassan’s
et al. 1991; McVay et al. 2000; Honjo et al. 2002; bearing capacity model (Hassan, 1970) can
AASHTO 2004; Paikowsky et al. 2004; Foye et al. often produce bearing capacity predictions with
2009; Zhang et al. 2009a,b; Zhang & Chu 2009a, reasonable accuracy. Thus, the bearing capacity
2009b). is often calculated based on the Hassan model
As part of the worldwide efforts for imple- adjusted with local field test data as follows (Li &
menting reliability-based design in geotechnical Hou 2009)
engineering, the resistance factors for design of
shallow and deep foundations were also calibrated Ru 0 5 Nγ ζ γ γ B Nq ζ q γ 0 D Nc ζ c c (1)

437

ISGSR2013.indb 437 10/18/2013 9:43:55 AM

 where B = width of the foundation (m); The design equation can then be written as
D = embedded depth of the foundation; follows
c = cohesion of the soil (kPa); γ = effective unit
weight of the soil beneath the foundation (kN/m3); ⎛c ϕ ⎞
γ0 = effective unit weight of the soil above the Ru ⎜ n , n ⎟ − γ LSnnLL D SnD =0 (12)
foundation bottom (kN/m3); ζγ , ζq, ζc = shape ⎝γc γφ ⎠
factors; Nγ , Nq, Nc are bearing capacity factors;
and Ψ = a model correction factor calibrated based where γc, γϕ, γD, and γL = partial factors for c, ϕ,
on results from 38 field tests carried at Shanghai SL, and SD, respectively; and cn, ϕn, SnL, and SnD =
(Li & Hou 2009). The expressions for calculating nominal values of c, ϕ, SL, and SD, respectively.
Ψ, bearing capacity factors and shape factors are For an uncertain variable x, its nominal value xn is
summarized as follows often related to its mean value μx through the use of
a bias factor λx as follows (e.g., Ang & Tang 1984)
⎧0.9 ϕ ≤ 200
ψ =⎨ (2) μx
⎩0.053ϕ 0 166 200 < ϕ ≤ 320 xn = (13)
λx
Nq = 2.0 ~ 3.0 (3)
In Shanghai, the shear strength parameters are
often measured using undrained direct shear tests.
⎡ πt ϕ ⎛ ϕ⎞ ⎤
Nc ⎢e tan2 45° + ⎟ − 1⎥ cot ϕ (4) When the peak values of cohesion and friction
⎣ ⎝ 2⎠ ⎦ angle of the soil measured in the undrained direct
shear tests are reduced by 20%, the predictions
⎡ ⎛ ϕ⎞ ⎤ from the bearing capacity model as introduced
Nγ 1 5 ⎢ eπ t ϕ
tan2 45° + ⎟ − 1⎥ tan ϕ (5)
⎣ ⎝ 2⎠ ⎦ above are most consistent with the measured
bearing capacity for a large number load tests in
ζ γ = 1.0 − 0.4 B / L (6) Shanghai (Li & Hou, 2009). Thus, for the design of
shallow foundations, λc = λϕ = 1.25 are adopted.
ζq ϕ /L (7)
4 METHOD OF CALIBRATION
ζ c = 1.0 + 0.2 B / L (8)
In principle, the partial factors of the foundation
where L = width of the foundation (m) and design can be calculated using the first order relia-
ϕ = friction angle of the soil (°). bility method by comparing the coordinates of the
design point with the nominal values of the ran-
dom variables. One feature in geotechnical design
3 DESIGN EQUATION is that, its load factors are pre-specified, which are
usually different from those obtained directly based
Supposing only live and dead loads are present, the on reliability theory. According to the national
limit state function in the foundation design prob- design code in China (MHURD, 2002; MHURD,
lem can be written as follows 2011), the load factors for design of foundations
should be 1.0, i.e., γL = γD = 1.0. In such cases, the
resistance factors obtained directly using the reli-
Ru SL − SD = 0 (9) ability theory cannot be straightforwardly used in
geotechnical design.
where SL = live load; and SD = dead load. In the In this study, reliability analysis is performed
above equation, the uncertain variables include c, considering only c and ϕ as random variables.
ϕ, SL, and SD. For ease of presentation, we define The values of partial factors of γL and γD are taken
Ru(c, ϕ) as follows as 1.0. Thus, the limit state function in used in
reliability analysis for calibrating resistance fac-
Ru ( c,φ ) 0.5ψ Nγ ζ γ B + ψ Nc ζ c c + N q ζ q γ 0 D tors is as follows
(10) Ru (c,ϕ ) − SnD SnL = 0 (14)

Substituting Eq. (10) into Eq. (9) yields Let c* denote the value of c at the design point.
The partial factor for c can be determined using
Ru (cc φ ) − SL SD = 0 (11) the following relationship

438

ISGSR2013.indb 438 10/18/2013 9:43:56 AM

 cn is in the range of 0.42–0.55, and the COV of the
γc = (15)
c* friction angle is in the range of 0.16–0.25 (Li &
Gao 2001). For calibrating the resistance factors
γϕ can also be determined using the same for regional design, the upper bound values of
procedure. COV are adopted, which is on the conservative
Traditionally, shallow foundations are designed side. Considering the spatial correlation of the soil
based on the global FOS approach. Let K denote properties (Li & Gao, 2001), a variance reduction
the notional FOS adopted in a design as defined factor of 0.27 is applied to cohesion and friction
below angle. Thus, the COV values of the cohesion and
friction angle adopted in the reliability analysis
1 are 0.21 and 0.09, respectively. It is assumed that
SnL SnD = Rnu (16) both cohesion and friction angle are lognormally
K
distributed and statistically independent. Sensi-
tivity analysis indicates that the calibrated partial
Rnu Ru ( cn n ) (17) factors are not sensitive to the value of ρ adopted,
which is consistent with previous findings (Barker
Let ρ denote the ratio of live load to dead load, et al. 1991; McVay et al. 2000; Zhang 2004). In this
study, ρ = 0.2 is adopted.
ρ λL λD (18)

Substituting Eq. (18) into Eqs. (16) and (17), the 5 CALIBRATION RESULTS
mean values of live and dead loads can be calcu-
lated as follows 5.1 Current level of reliability and target
reliability index
Rnu Figure 1 plots the calculated reliability indexes and
SnD = (19)
( + )K mean values of cohesion for the 142 sites. The reli-
ability indexes are in the range of 2.3–4.2 with a
ρRnu mean of 3.35. For comparison, Phoon et al. (2003)
SnL = (20)
( + )K reported that the reliability index of existing drilled
shafts under compression to support transmission
line structures in North America is in the range
Given the values of cn, ϕn, ρ and K, Eqs. (19)
of 2.4–3.6. Figure 2 plots the calculated reliability
and (20) can be used to determine the values of SnD
indexes and mean values of friction angle for the
and SnL based on which the reliability index cor-
142 sites. Comparing Figures 1 and 2, there are less
responding to a specific K can be found out using
scatter in the relationship between reliability index
Eq. (14) as the limit state function. If one knows
and the friction angle, probably because the reli-
the target reliability index, one can adjust the val-
ability indexes of the foundation are more affected
ues of SnD and SnL in Eq. (14) subjected to the con-
by the friction angle.
straints set by Eqs. (19) and (20) until the target
The target reliability index is often determined
reliability index is satisfied, which can then be used
with reference to the reliability index of exist-
to determine the partial factors corresponding to
ing buildings (e.g., Allen 1975; Phoon 2003) or
the target reliability index.
To calibrate the resistance factors for design,
consider a strip foundation with B = 3 m, D = 1 m.
In Shanghai, the typical design FOS is K = 2.5. As
the soil data may vary with site, the foundation is
assumed to rest on 142 different sites in Shanghai
to obtain a set of robust partial factors that can
be generally used in Shanghai. At a site, while the
mean values of soil properties are relatively easier
to determine, the Coefficients Of Variation (COV)
are much hard to estimate based on limited site-
specific data. Thus, the mean values of the cohe-
sion and friction angle take the site-specific values,
and the COV values of the cohesion and friction
angle take the regional values. In Shanghai, the Figure 1. Relationship between cohesion and reliability
COV of cohesion of the clay at shallow depths index for the 142 sites.

439

ISGSR2013.indb 439 10/18/2013 9:44:00 AM

 Figure 2. Relationship between friction angle and Figure 3. Relationship between γc and μc for the 142 sites.
reliability index for the 142 sites.

with reference to experience in other regions. As

the mean reliability index of the foundation at
142 sites is 3.35, 3.35 is taken as the target reliabil-
ity index. Becker (1996) suggested the reliability
index for design of shallow foundations against
bearing capacity failure be 3.5.

5.2 Recommendation of partial factors

Figures 3 and 4 show the relationships between γc
and μc and the relationship between the γϕ and μc for
the 142 sites, respectively. While γc tends to increase
with the cohesion, γϕ tends to decrease with the Figure 4. Relationship between γφ and μc for the 142 sites.
cohesion. This is deemed reasonable, since a larger
cohesion implies increased importance of cohesion
in foundation design, and hence a larger resistance Rn M Bγ B + M Dγ + Mcc (21)
factor should be applied to maintain the same level
of safety. For comparison, Figures 5 and 6 show where MB, MD, and Mc are bearing capacity factors
the relationships between the γc and μϕ and the rela- summarized in Table 2. In Eq. (21), B should be
tionship between γϕ and μϕ, respectively. Comparing reduced to 6 m when the foundation width is larger
Figures 3–6, there are large scatter in the relation- than 6 m, and when the foundation is resting on the
ship between the resistance factors and the cohe- sandy ground, the value of B should be increased
sion, probably because the relationship between to 3 m if the foundation width is smaller than 3 m.
cohesion and reliability index is more uncertain, In a design, it is required that Rn should not be
as shown in Figure 1. It is attempting to suggest smaller than the summation of the nominal live
partial factors based on regression analysis of the and load loads, i.e., SnD + SnL. Compared with the
data in Figures 5 and 6. However, considering the requirement indicated in Eq. (9), it indicates that
tradition of using fixed FOS and fixed partial fac- the bearing capacity predicted using Eq. (10) using
tors in the profession, it is decided that fixed partial soil parameters factored with the calibrated partial
factors be used in design of shallow foundations. factors should be comparable with that produced
Table 1 summarizes the statistics of the reli- using Eq. (20). As a rule of thumb, the values c and
ability indexes, γc and γϕ of the shallow foundation ϕ used in a design are often 70% of their mean val-
when it is constructed at the 142 sites in Shanghai. ues when Eq. (21) is used in Shanghai for shallow
The recommended resistance factors of γc and γφ in foundation design.
design are 2.7 and 1.2, respectively, which are close Let Rns denote the bearing capacity calculated
to their mean values. using according to Shanghai code with cohesion
and friction angle factored by the partial fac-
tors recommended in this study. Let Rnn denote
5.3 Verification of the recommended partial
the bearing capacity calculated according to
factors
the national code with cohesion and friction angle
The national design code suggests the nominal being 0.7 of their mean values. Figure 7 shows the
value of the bearing capacity can be calculated values of Rns and μϕ for the 142 sites. The values
using the following equation (MHURD, 2010): of Rns are in the range of 50–200 kN. As expected,

440

ISGSR2013.indb 440 10/18/2013 9:44:02 AM

 Table 2. Values of bearing capacity factors
recommended in MHURD (2002).

ϕ (°) MB MD Mc

0 0 1 3.14
2 0.03 1.12 3.32
4 0.06 1.25 3.51
6 0.1 1.39 3.71
8 0.14 1.55 3.93
10 0.18 1.73 4.17
12 0.23 1.94 4.42
14 0.29 2.17 4.69
Figure 5. Relationship between γc and μφ for the 142 sites. 16 0.36 2.43 5
18 0.43 2.72 5.31
20 0.51 3.06 5.66
22 0.61 3.44 6.04
24 0.8 3.87 6.45
26 1.1 4.37 6.9
28 1.4 4.93 7.4
30 1.9 5.59 7.95
32 2.6 6.35 8.55
34 3.4 7.21 9.22
36 4.2 8.25 9.97
38 5 9.44 10.8
40 5.8 10.84 11.73

Figure 6. Relationship between γφ and μφ for the 142 sites.

Table 1. Summary of calibration

results of the 142 sites.

β γc γϕ

Mean 3.35 2.76 1.17

COV 0.12 0.32 0.07

Rns tends to increases with μϕ. However, the rela-

tionship between Rns and μϕ is not unique because
Rns is also affected by the cohesion, which varies Figure 7. Relationship between Rns and μϕ for the 142 sites.
from sites to sites. Figure 8 shows the relationship
between the notional FOS (i.e., K) and μϕ for the
142 sites. The notional FOS values of the founda-
tion at the 142 sites are in the range of 2.0–3.0,
which is largely consistent with the local experi-
ence for adopting a K value of 2.5 in shallow
foundation design. Figure 9 shows the relation-
ship between Rns/Rnn and μϕ for the foundation at
each site. When the mean friction angle is smaller
than 20°, the predicted bearing capacity from the
two approaches are largely consistent. When the
friction angle is larger than 20°, the predicted bear-
ing capacity based on Eq. (21) is larger. It is known
that Eq. (21) tends to underestimate the bearing
capacity when the soil has a larger friction angle
(Li & Hou 2009). Figure 8. Relationship between K and μϕ at the 142 sites.

441

ISGSR2013.indb 441 10/18/2013 9:44:03 AM

 Becker, D.E. 1996. Eighteenth Canadian geotechnical
colloquium: limit states design for foundations.
Part II: Development for the national building code
of Canada. Canadian Geotechnical Journal 33(6):
984–1007.
Foye, K.C., Abou-Jaoude, G., Prezzi, M. & Salgado, R.
2009. Resistance factors for use in load and resistance
factor design of driven pipe piles in sands. Journal
of Geotechnical and Geoenvironmental Engineering
135(1): 1–13.
Hassen, J.B. 1970. A revised and extended formula for
bearing capacity. Copenhagen: Danish Geotechnical
Institute.
Honjo, Y., Suzuki, M., Shirato, M. & Fukui, J. 2002.
Figure 9. Relationship between Rns/Rnn and μϕ at the Determination of partial factors for a vertically loaded
142 sites. pile based on reliability analysis. Soils and Foundations
42(5): 91–109.
Li, J.P. & Gao, D.Z. 2001. Study of Variable Behavior
6 SUMMARY AND CONCLUSIONS of Soil Index. China Harbour Engineering, 6: 26–30.
(In Chinese).
The research reported in this paper can be sum- Li, J.P. & Hou, S.N. 2009. Discussion on the formula
marized as follows: of the ultimate bearing capacity of foundation in
Shanghai region. Journal of Geotechnical Investigation
1. The semi-empirical procedure adopted in cali- & surveying 37(4): 6–9. (In Chinese).
brating partial factors for design of shallow McVay, M.C., Birgisson, B., Zhang, L.M., Perez, A. &
foundation is described. Putcha, S. 2000. Load and Resistance Factor
2. For the hypothetical shallow foundation, its Design (LRFD) for driven piles using dynamic
reliability index is in the range of 2.3–4.2 when methods—A Florida perspective. Geotechnical Testing
resting on 142 different sites at Shanghai. Journal 23(1): 55–66.
3. To achieve a target reliability index of 3.35 and Ministry of Housing and Urban-Rural Development
for the 142 sites considered, the partial factor (MHURD). 2002. Code for design of foundations.
Beijing: China Construction Press. (In Chinese).
for cohesion has a mean of 2.76 and a standard Ministry of Housing and Urban-Rural Development
deviation of 0.32, and the partial factors for (MHURD). 2011. Code for design of foundations.
friction angle has a mean of 1.17 and a stand- Beijing: China Construction Press. (In Chinese).
ard deviation of 0.07. Paikowsky, S.G., et al. 2004. Load and resistance factor
4. Adopting 2.7 and 1.7 respectively as partial fac- design (LRFD) for deep foundations. NCHRP Rep.
tors for cohesion and friction angle can produce No. 507. Washington, D.C.: Transportation Research
bearing capacity predictions consistent with Board, National Research Council.
local experience in Shanghai. Phoon, K.K., Kulhawy, F.H. & Grigoriu, M.D. 2003.
Development of a reliability-based design framework
for transmission line structure foundations. Journal
ACKNOWLEDEMENT of Geotechnical and Geoenvironmental Engineering
129(9): 798–806.
Shanghai Urban Construction and Communications
This research was substantially supported by the Commission (SUCCC). 2010. Foundation Design
Natural Science Foundation of China (Project No. Code (DGJ 08-11-2010). Shanghai.
41272288). Zhang, L.M. & Chu, L.F. 2009a. Calibration of meth-
ods for designing large-diameter bored piles: Ultimate
limit state. Soils and Foundations 49(6): 883–896.
REFERENCES Zhang, L.M. & Chu, L.F. 2009b. Calibration of methods
for designing large-diameter bored piles: Serviceability
AASHTO. 2004. LRFD Bridge Design Specifications, limit state. Soils and Foundations 49(6): 897–908.
3rd Ed. Washington, D.C.: American Association of Zhang. J., Zhang, L.M. & Chu, L.F. 2009b. Resistance
State Highway and Transportation Officials. factors for ultimate limit state design of large-diameter
Allen, D.E. 1975. Limit states design—a probabilistic bored piles in Hong Kong. International Journal of
study. Canadian Journal of Civil Engineering 2(1): Engineering under Uncertainty: Hazards, Assessment,
36–49. and Mitigation 1(3–4): 141–152.
Ang, A.H.-S. & Tang, W.H. 1984. Probability concepts in Zhang, J., Zhang, L.M. & Tang, W.H. 2009a. Reliability
engineering planning and design: design, risk and reli- based design of pile foundations considering both
ability, Vol. 2. New York: Wiley. parameter and model uncertainties. Journal of
Barker, R.M., et al. 1991. Manuals for the Design of Bridge GeoEngineering 4(3): 119–127.
Foundations, NCHRP Report 343. Washington, D.C.: Zhang, L.M. 2004. Reliability verification using
Transportation Research Board, National Research proof pile load tests. Journal of Geotechnical and
Council. Geoenvironmental Engineering 130(11): 1203–1213.

442

ISGSR2013.indb 442 10/18/2013 9:44:04 AM

 Geotechnical Safety and Risk IV – Zhang et al. (eds)
© 2014 Taylor & Francis Group, London, ISBN 978-1-138-00163-3

Development and reliability of a pile driving formula for the MnDOT

S.G. Paikowsky
University of Massachusetts Lowell (UML), Lowell, Massachusetts, USA
GeoDynamica, Inc., Newton, Massachusetts, USA

C.M. Marchionda
Nobis Engineering, Inc., Lowell, Massachusetts, USA
Previously UML, MA, USA

S. Amatya
Golder Associates Pty Ltd., Richmond, Victoria, Australia
Previously UML, MA, USA

M.C. Canniff
GeoDynamica, Inc., Newton, Massachusetts, USA

A.S. Budge
Minnesota State University Mankato, Mankato, Minnesota, USA

ABSTRACT: Driven piles are the most common foundation solution used in bridge construction
(Paikowsky et al., 2004). Their safe use requires to reliable verification of their capacity and integrity.
Dynamic analyses of driven piles are methods attempting to obtain the static capacity of a pile, utilizing its
behavior during driving. Dynamic equations (aka pile driving formulas) are the earliest and simplest forms
of dynamic analyses. The development and the examination of such equation tailored for given demands are
presented. After establishing the bridge pile construction practices of MnDOT, a database of driven pile case
histories relevant to these practices was built. The databases were utilized to investigate previous MnDOT
(and other) dynamic formulas and use object oriented programming for linear regression to develop a new
formula that was then calibrated for LRFD methodology and evaluated for its performance. A standalone
control database and a database of dynamic measurements enabled independent evaluation of the for-
mula leading to the final stage in which the equation was adjusted for field application and past experience.
Though developed for MnDOT, its use is universal for driven piles within the provided limitations.

1 BACKGROUND formulas, the evaluation of the uncertainty of their

performance and thereby calculating the resistance
1.1 Overview factor required for a consistent level of safety in
the bridge foundations. The results of the study
Minnesota Department of Transportation
suggested that the formula in use is inaccurate and
(MnDOT) used its own pile driving formula;
inefficient and the development of an alternative
however, its validity and accuracy had never been
formula was, therefore, required.
thoroughly evaluated. With the implementa-
tion of Load Resistance Factor Design (LRFD)
in Minnesota in 2005, and its mandated use by
1.2 MnDOT state of practice
the Federal Highway Administration (FHWA)
in 2007, the resistance factor associated with 1.2.1 Objectives and method of approach
the use of the MnDOT driving formula needed Data relevant to the subsurface conditions and
to be determined. Paikowsky et al. (2009, 2010) practices of substructure design and construc-
addressed this need by establishing the bridge pile tion in Minnesota were collected and summarized.
construction practices of MnDOT and building up This compilation of data ensures the relevance of
a database of driven pile case histories relevant to the pile performance database to the needs of the
these practices. The databases were utilized for the MnDOT. The following steps were used in develop-
investigation of the MnDOT (and other) dynamic ing the state of design and practice: (a) a detailed

443

ISGSR2013.indb 443 10/18/2013 9:44:05 AM

 questionnaire distributed as part of NCHRP
project 12-66 (Paikowsky and Canniff, 2004) and
completed by Mr. Dave Dahlberg of the MnDOT,
(b) review of MnDOT bridge construction man-
ual, (c) review of local construction records of 28
bridges, and (d) interviews of contractors, design-
ers and DOT personnel. Details of the data and its
relevant analyses are presented by Paikowsky et al.
(2009), while the following section provides a short
summary.

1.2.2 Summary of findings

a. Majority of bridge foundations are based on
Closed Ended Pipe (CEP) and H piles.
b. CEP piles range in diameter from 12 to 20”
(0.30 to 0.51 m) and comprise 85% of the driven
piles.
c. Most common CEP piles used are 12” × 0.25”
(0.30 m × 6.4 mm) and 16” × 0.3125”
(0.41 m × 7.9 mm), installed as 40% and 25% of Figure 1. Range of pile capacity based on static load
the total foundation length, respectively (based test (mean +/− 1 S.D.) and MnDOT mean factored design
on 28 bridge projects). loads sorted by H pile type and cross-sectional area.
d. H piles comprise 15% of the driven pile foun-
dations with sizes ranging between 10 × 42 and
14 × 73 (HP 250 × 63 to HP 360 × 109).
e. Typical (average) driven pile length is 66 feet
(20 m) with a load of 186 kips (827 kN). More
specific categorization includes for example
CEP 12” × 0.25” (0.30 m × 6.4 mm) pile length
is 77 feet (23 m) and it carries 155 kips (689 kN)
design (factored) load.
f. Diesel hammers are most commonly used
for driving piles ranging in size from D 19-32
(42.4 kip-ft or 57.5 kN-m) to D 30-32 (75.4 kip-ft
or 102.2 kN-m).
g. Over 90% of the piles are driven beyond
the easy driving resistance zone of 4 bpi
(16 b/10 cm), hence allowing more accurate
capacity evaluation when utilizing the dynamic
methods. Fifty percent (50%) of the piles were
driven to a final penetration of 8 or more bpi
(31 b/10 cm). Figure 2. Distribution of database MnDOT/LT H-piles
by pile area cross-section along with the frequency and
pile type used by MnDOT.
2 DATABASES

2.1 Overview in Figure 1 provides mean static load-test failure

Two new robust databases of H and pipe piles driven load (+/− 1 standard deviation) for each pile type/
and statically load-tested to failure were compiled size category comprising the database, along with
to address the specific needs of Minnesota DOT the number of cases related to that information.
aforementioned pile foundation practices. In addition, the mean LRFD factored (design)
load for the MnDOT for the applicable pile cases
is presented along with the number of piles it is
2.2 MnDOT/LT-H-Piles
based upon. For example, 41 case histories of
Figures 1 and 2 present graphically some of the the database are related to HP12X53. The mean
important features of MnDOT/LT 2008 H Piles failure load of these cases (applying Davisson’s
database along with a comparison of data reflecting failure criterion) was 305 kips +/− 142 kips (1 SD).
the MnDOT foundation practices. The information The mean factored load of this type of pile by the

444

ISGSR2013.indb 444 10/18/2013 9:44:05 AM

 MnDOT is 157 kips based on 337 HP12X53 piles.
These data suggests that the mean safety margin
of the MnDOT is 1.943 +/− 0.904 in comparison
with the database information (not including the
load factor) or the covering of approximately 1
S.D range (lower value of resistance is 163 kips
compared to a load of 157 kips) translates to a tar-
get reliability of β = 1 and a probability of failure
pf = 15.9%. The information in Figure 2 presents
the distribution of the case histories in the data-
base based on the pile sizes in comparison with the
distribution of use of the same pile by MnDOT.
To be relevant, the frequency of use of the H-Piles
by the MnDOT, presented in Figure 2, reflects the
use of the particular pile type out of the H-Piles
only and not out of all driven piles. For example,
41 pile cases of HP12X53 are available in the data-
base (24.7% of all cases), while the MnDOT uses
this pile in 55.5% of the cases where H-Piles are Figure 4. Distribution of database MnDOT/LT pipe
being used. piles by pile area cross-section along with the frequency
and pile type used by MnDOT.

2.3 MnDOT/LT-Pipe-Piles
Figures 3 and 4 present graphically some of the is presented along with the number of piles it is
important features of MnDOT/LT 2008 Pipe Piles based upon. For example, 12 and 65 case histories
database along with a comparison of data reflect- of the database are related to 12.00 and 12.75 inch
ing the MnDOT foundation practices as presented diameter piles, respectively. The mean failure load
earlier in section 1. The information in Figure 3 of these cases was 388 +/− 173 kips and 372 +/−
provides mean failure load (+/− 1 standard devia- 195 kips (1 SD) for the 12 and 12.75 inch diam-
tion) for each pile type/size category (by pipe pile eter piles, respectively. The mean factored load of
diameter) comprising the database, along with the 12 and 12.75 inch piles by the MnDOT is 155
the number of cases related to that information. and 120 kips, based on 1055 and 116 piles, 12.0
In addition, the mean LRFD factored (design) and 12.75 inch diameter pipe piles, respectively.
load for the MnDOT for the applicable pile cases These data suggests that the mean safety margin
of the MnDOT 12 and 12.75 inch diameter piles
is 2.503 +/− 1.116 and 3.100 +/− 1.625 in compari-
son with the database information (not including
the load factor) or the covering of approximately
1.3 S.D range (lower value of resistance is 1.29 and
1.35 standard deviations from the mean, hence
approximately taken as 1.3) translates to a target
reliability of β = 1.3 and a probability of failure
pf = 9.8%. The information in Figure 4 presents the
distribution of the case histories in the database
based on the pile sizes in comparison with the dis-
tribution of use of the same pile by MnDOT. To
be relevant, the frequency of use of the pipe piles
by the MnDOT, presented in Figure 4, reflect-
ing the use of the particular pile type out of the
pipe piles only and not out of all driven piles. For
example, 77 pile cases of 12 and 12.75 inch diam-
eter piles are available in the database (46.1% of
all cases), while the MnDOT uses these diameter
piles in 50.0% of the projects where pipe piles are
being used. The major difficulty of the database
Figure 3. Range of pile capacity based on static load as presented in Figure 4 is evidently related to the
test (mean +/− 1 S.D.) and MnDOT mean factored design 16 inch diameter closed ended pipe piles. This type
loads sorted by pipe pile type and pipe pile diameter. of piles are used by the MnDOT in 22.9% of all

445

ISGSR2013.indb 445 10/18/2013 9:44:06 AM

 projects and 26.9% of all pipe piles but only 10 equation (Gates, 1957) in associating the pile
closed ended and 4 open ended 16 inch diameter capacity to the square of the hammers’ nominal
piles are available at the database. energy, and the simplicity in using logarithm of the
blow count, a linear regression analysis of the data
was performed looking for the best fit parameters
3 DEVELOPMENT OF A NEW MNDOT to the anticipated formulation. This process is out-
DYNAMIC EQUATION lined in the following section.

3.1 Plan of action 3.3 Method of approach

a. Developing an independent equation for the S-Plus is a commercial advanced statistics pack-
needs of the MnDOT using object oriented age sold by Insightful Corporation of Seattle,
programming. Washington. It features object oriented programming
b. Examine the new MnDOT dynamic equations capabilities and advanced analytical algorithms. To
by the following steps: (1) evaluate the ulti- develop the New MnDOT Dynamic Equation, a
mate pile static capacity of all tested piles using linear regression was performed using the S-Plus
Davisson’s failure criterion (Davisson, 1972), program. Static capacity and Gates parameters
(2) evaluate the pile capacity of the database (which are square root of hammer energy and log
case histories using the developed new equa- of 10 times the blow count) were provided as input
tion, (3) evaluate the bias of the method as the parameters into an S-Plus worksheet. Linear regres-
ratio between measured (stage ‘a’) to calculated sion was then performed for each of the eight (8)
(stage ‘b’) capacity, and (4) examine the statisti- different examined cases. The coefficient for the New
cal parameters of the bias. MnDOT Dynamic Equation and the coefficient of
c. Conduct an in-depth evaluation to the new determination, r2 of the proposed relationship were
MnDOT equation by examining subsets of var- calculated, as well as Cook’s Distance graph, ena-
ious conditions. bling to identify the data outliers.
d. Examine the new MnDOT distribution func- Cook’s Distance is a commonly used estimate
tions fit to LRFD calibration. of the influence of a data point when doing least
e. Develop the resistance factors associated with squares regression. Cook’s distance measures the
the different conditions, using both FOSM and effect of deleting a given observation. Data points
MC simulation methods. with large residuals (outliers) and/or high lever-
f. Examine the recommended resistance factors age may distort the outcome and accuracy of a
in comparison to the performance of the tra- regression. Points with a Cook’s distance of 1 or
ditional MnDOT and other dynamic equations. more are considered to merit closer examination in
This examination includes the use of a control the analysis. Cook’s distance is a measurement of
databases the data of which not included in the influence of the ith data point on all the other
the development of the equation and the use data points. In other words, it tells how much influ-
of dynamic measurements data from MnDOT ence the ith case has upon the model. The formula
projects. to find Cook’s distance, Di, is, (Cook, 1979):
g. Adjust the equation for the practice of field
∑ (Yj )
observations by MnDOT field inspectors. 2
Yj i
h. Examine the equation for use with timber and Di = (1)
concrete driven piles. p MSE
M
i. Develop final recommendations now imple-
mented by the MnDOT. where Yj is the predicted (fitted) value of the ith
observation; Yj (i ) is the predicted value of the jth
observation using a new regression equation found
3.2 Principle
by deleting the ith case; p is the number of parameters
A regression analysis can provide parameters that in the model; and MSE is the Mean Square Error.
connect the major factors affecting the pile capac- Using the F distribution to compare with Cook’s
ity (e.g. energy, driving resistance, etc.) and allow distance, the influence that the ith data point has
the development of a dynamic equation. A lim- on the model can be found. Values in the F distri-
ited attempt was made in that direction, but most bution table can be used to express the percentage
obtained equations had no engineering “feel” (or of influence the ith data point has. A percentage of
“logic”) to them and are constructed of arbitrary 50% or more would indicate a large influence on
terms and parameters. the model. The larger the error term implies that
A different approach was then taken. the Di is also larger which means it has a greater
Recognizing the unique success of the Gates influence on the model.

446

ISGSR2013.indb 446 10/18/2013 9:44:08 AM

 The new MnDOT equation was developed in c. The final equation recommended as the new
two stages described in the following sections. The MnDOT dynamic equation for the most general
most generic form is described in section 3.4 and is case (all hammers, all conditions) is therefore:
investigated in section 3.5. A more specific form is
described in section 3.6 and is investigated in sec- Ru 35 Eh * log (10 * N )
tion 3.7. The final adoption of the equation to field (2)
application is described in section 5.
where Ru = predicted pile capacity in kips,
Eh = rated hammer energy kips ⋅ ft, and N = Blows
3.4 The general new MnDOT dynamic equation Per Inch (PBI) at the End of Driving (EOD).
development d. Based on the data presented in Table 2, the rec-
Table 1 presents a summary of the results, obtained ommended preliminary resistance factors for
by applying the analysis to all cases and EOD cases equation 13 are φ = 0.55 for H-piles and φ = 0.35
only for H and pipe piles suggesting the following: for pipe piles.
a. For both pile types under all data selection
criteria, the recommended coefficient varied
3.5 Investigation of the new general MnDOT
between 34.5 and 37.1.
dynamic equation
b. All regressions resulted with a coefficient of
determination greater than 0.85. As a lower Initial examination of the uncertainty of the proposed
coefficient means a more conservative evalua- new equation and the associated resistance factors
tion and the scatter of the pipe piles predictions is summarized in Table 2. Figures 5 and 6 present
is higher than that for the H-piles, it is reason- the scatter of the new equation in the form of static
able to use one coefficient, 35. (measured) capacity vs. predicted capacity for H and
pipe piles, respectively. The obtained results suggest
a consistent higher performance of the equation for
Table 1. Summary of S-PLUS linear regression analysis H piles (efficiency factor of about 53% to 54%) and
results for the new general MnDOT dynamic equation.
a resistance factor of 0.55. The results also suggest
Pile # of Searched Coeff. of the highest performance of the equation for pipe
type Condition cases coeff.1 determ. r2 piles (36%) with a recommended resistance factor
of 0.40. Further in-depth investigations of the new
H All cases 135 35.814 0.880 equation are presented by Paikowsky et al. (2009).
H All excluding 132 35.170 0.896 Section 3.6 follows the more restrictive MnDOT pile
Cook’s outliers driving conditions in examining the applicability of
H EOD only 125 35.660 0.896 the equation or a variation of it.
H EOD excluding 123 34.550 0.914
Cook outlier’s
Pipe All cases 128 35.866 0.861 3.6 The development of a detailed new MnDOT
Pipe All excluding 125 34.875 0.877 dynamic equation
Cook’s outliers
Pipe EOD only 102 37.142 0.851
The approach and method of analysis presented in
Pipe EOD excluding 99 35.866 0.868 sections 3.1 to 3.5 were used for searching an optional
Cook outlier’s new MnDOT equation that would (if possible) bet-
ter fit the specific conditions than the general case
1
Searched coefficient for the equation Ru = Coeff. of equation 2 presented earlier. Table 3 presents a
Eh log(10 N ). summary of the results obtained by applying the

Table 2. Statistical parameters and LRFD calibration for the developed equation (2) for H piles and pipe piles, EOD
condition only.

Resistance factor φ
Mean bias β = 2.33, pf = 1%, φ/λ
# of measured/ Stand. Coeff. Best fit line Coeff. of redundant Efficiency
Pile cases calculated dev. of var. equation determination factor
type (n) (mλ) (σλ) (COVλ) (least square) (r2) FOSM MC3 Recom. (%)

H 125 1.0163 0.3599 0.3542 Ru = 0.880 * Rs 0.896 0.495 0.542 0.55 54.1
Pipe 99 1.1089 0.5955 0.5370 Ru = 0.805 * Rs 0.849 0.364 0.385 0.35 31.6

Note: 1MC—Monte Carlo simulation for 10,000 simulations.

447

ISGSR2013.indb 447 10/18/2013 9:44:08 AM

 Figure 5. Measured static capacity vs. new general Figure 6. Measured static capacity vs. new MnDOT
MnDOT dynamic equation prediction for 125 EOD cases. dynamic equation prediction for 99 EOD cases.

Table 3. Dynamic equation predictions for H piles and pipe piles EOD condition only.

Pile # of Searched Coeff. of

type Condition cases coeff.1 determ. r2

H EOD only 125 35.637 0.896

H EOD only, Excl. Cook’s outliers 122 34.151 0.925
H EOD, diesel hammer, B.C. ≥ 4 BPI 39 33.527 0.907
H EOD, diesel hammer, B.C. ≥ 4 BPI, Excl. Cook’s outliers 38 32.126 0.935
H EOD, diesel hammer, MnDOT energy2, B.C. ≥ 4 BPI 13 34.401 0.870
H EOD, diesel hammer, MnDOT energy2, B.C. ≥ 4 BPI, 12 31.181 0.924
Excl. Cook’s outliers
Pipe EOD only 99 36.746 0.850
Pipe EOD only, Excl. Cook’s outliers 97 35.839 0.859
Pipe EOD, diesel hammer, B.C. ≥ 4 BPI 41 30.532 0.918
Pipe EOD, diesel hammer, B.C. ≥ 4 BPI, Excl. Cook’s outliers 38 29.983 0.946
Pipe EOD, diesel hammer, MnDOT energy2, B.C. ≥ 4 BPI 16 33.294 0.974
Pipe EOD, diesel hammer, MnDOT energy2, B.C. ≥ 4 BPI, 14 33.146 0.989
excluding Cook’s outliers

Notes: 1Searched coefficient for the equation Ru = Coeff. Eh log(10 N ), 2MnDOT energy range contains
hammers with rated energies between 42.4 and 75.4 k-ft.

analysis under the various sub-categorizations to good performance of the proposed format and
the EOD cases only. The obtained results summa- obtained coefficients.
rized in Table 3 suggest the following: c. When restricting the EOD data to diesel ham-
mers only and a blow count of equal or greater
a. For both pile types under all EOD data selec- to 4 BPI (with or without the outliers) the rec-
tion criteria (with or without the outliers), the ommended coefficients are 32.1 to 33.5 for the
recommended coefficient varied between 34.2 H piles and 30.0 to 30.5 for the pipe piles. Both
to 36.7 reaffirming the coefficient of 35 recom- subsets contain significant number of cases
mended for the general equation as appeared in (38 H piles and 38 pipe piles when eliminating
equation 2. outliers).
b. All regressions resulted with a coefficient of d. When further restricting the conditions
determination greater than 0.85 suggesting described in (c) above by looking at the energy

448

ISGSR2013.indb 448 10/18/2013 9:44:09 AM

 range of the diesel hammers typically used in

factor (%)

Notes: Calculated capacity using MnDOT new dynamic equation Ru = Coeff. Eh log(10 N ), Rs is the static capacity of the pile examined by Davisson’s failure criterion,
Efficiency
MnDOT practice, the subsets decrease to 13/12
H pile cases and 16/14 pipe pile cases, with and

54.1
50.6
55.8
62.2

48.2

57.0
φ/λ
without outliers, respectively. These are mar-
ginal size sets that result with coefficients vary-

Recom
β = 2.33, pf = 1%, redundant
ing between 31.2 to 34.4 for H piles and 33.1 to

0.55
0.60
0.60
0.65

0.55

0.60
33.3 for pipe piles.
e. Close examination of the most restrictive sub-

Resistance factor φ
sets described in (d) above (i.e. 13 H piles and

0.542
0.632
0.628
0.674

0.553

0.620
MC3
16 pipe piles before removing the outliers) show
that in both subsets a relatively (to the subset
size) large group of cases are of different piles

FOSM
of the same size tested at the same site (e.g. 7

0.495
0.578
0.566
0.598

0.508

0.558
out of the 16 pipe piles are 14” diameter piles
from Deer Island project in Massachusetts and
6 of the H piles are 12 × 53 from site no. 37

Coeff. of
determ.
in Canada). As such, the data are too biased

0.896
0.896
0.907
0.935

0.870

0.924
as not only the set is marginal in size, but

(r2)
about 50% of the cases are related to the same
project. The statistics and coefficient obtained Statistical parameters and resistance factors of the new detailed MnDOT dynamic equation for H piles.

Ru = 0.888 Rs
Ru = 0.754 Rs
Ru = 0.812 Rs
Ru = 0.873 Rs

Ru = 0.759 Rs

Ru = 0.889 Rs
from that subset should, therefore, cautiously

(least sq.)
be applied.
line eq.
Best fit
f. As a lower coefficient means a more conserva-
tive capacity evaluation, the above discussion
and the observations presented in (c) should
serve as the guideline for the new MnDOT
(COVλ)
of var.

0.3542
0.3542
0.3176
0.2762

0.3968

0.3135
dynamic equation that suits better to MnDOT
Coeff.

pile driving practice of diesel hammers and a

Blow Count (BC) ≥ 4 BPI.
g. The equation recommended as the new MnDOT
0.3599
0.4199
0.3417
0.2888

0.4531

0.3297
Stand.

dynamic equation for the specific practice (diesel

dev.
(σλ)

hammers) is therefore:

Ru 30 Eh * log(10 * N ) (3)
Mean bias
measured/
calculated

1.0163
1.1856
1.0760
1.0458

1.1419

1.0518
(mλ)

h. The associated recommended preliminary

resistance factors for equation 3 are φ = 0.60
for H-piles and φ = 0.45 for pipe piles. Further
MnDOT energy, C ≥ 4 BPI

details and discussion of these recommenda-

BC ≥ 4 BPI w/o outliers

BC ≥ 4 BPI w/o outliers

MC—Monte Carlo Simulation for 10,000 simulations.

tions are presented in Paikowsky et al. (2009)

EOD, diesel, BC ≥ 4 BPI

and section 3.7.

MnDOT energy,

3.7 Investigation of the detailed new MnDOT

EOD, diesel,

EOD, diesel,
EOD, diesel,

dynamic equation
EOD only
EOD only
Condition

Table 4 presents the statistical details of the new

MnDOT dynamic equation (equation 3) applied to
H piles. For the most generic case of EOD with all
piles, the statistics of both equations (2 and 3) are
Coeff.

presented in Table 4. This is done so to examine

35
30
30
30

30

30

the applicability of using the detailed equation (3)

under all driving conditions.
cases

The presented information suggests that the

# of

125

39
38

13

12
(n)

performance of equation (3) is consistent and

reliable for all H piles driven with diesel hammers
Table 4.

regardless of the energy range. The use of equa-

Case
no.

tion (3) for all type of hammers at EOD and all

1

2
3

449

ISGSR2013.indb 449 10/18/2013 9:44:11 AM

 driving resistances naturally would provide a safer

Notes: Calculated capacity using MnDOT new dynamic equation Ru = Coeff. Eh log(10 N ), Rs is the static capacity of the pile examined by Davisson’s failure criterion,
factor (%)
Efficiency
evaluation compared to that of equation (2) that
was developed for that situation specifically. The

31.6
34.8
42.0
49.6

45.3

81.3
φ/λ
greater mean bias obtained when using equation
(3) allows, therefore, to select a consistent resist-
ance factor of φ = 0.60 to be used for all the cases
Recom
β = 2.33, pf = 1%, redundant

0.35
0.45
0.40
0.45

0.85

0.90
when applying equation (3). This conclusion was
further examined and reaffirmed against an inde-
Resistance factor φ

pendent control database.

0.385
0.450
0.427
0.492

0.905

1.012
MC3

Table 5 presents the statistical details for the new

MnDOT dynamic equation (equation 3) applied to
pipe piles. For the most generic case of EOD with
FOSM

all piles, the statistics of both equations (2 and 3)

0.364
0.424
0.396
0.449

0.766

0.825
are presented in Table 5. This is done to examine
the applicability of using the detailed equation (3)
under all driving conditions.
Coeff. of
determ.

The presented information suggests that equation

0.849
0.850
0.918
0.946

0.974

0.988

(2) provides accurate predictions for all cases (mean

(r2)

about 1.0), however, due to the larger scatter associ-

Statistical parameters and resistance factors of the new detailed MnDOT dynamic equation for pipe piles.

ated with the capacity prediction of pipe piles, the

Ru = 0.694 Rs
Ru = 0.902 Rs
Ru = 0.946 Rs

Ru = 0.878 Rs

Ru = 0.895 Rs
Ru = 0.805 Rs

coefficient of variation is typically higher than that

(least sq.)

for the H piles, and hence, the associated resistance

line eq.
Best fit

factors are lower. Exception to that are the cases

of the most restrictive subsets, matching closely the
MnDOT practice by the hammers energy range in
addition to diesel hammers and BC ≥ 4 BPI. These
(COVλ)
of var.

0.5955
0.5370
0.4284
0.3472

0.1818

0.1151

subsets result with an under-prediction and, hence,

Coeff.

a bias greater than 1.0 along with low coefficients

of variation, resulting with very high resistance
factors. The reasons for that behavior were previ-
0.5955
0.6947
0.4078
0.3149

0.2051

0.1273
Stand.

ously discussed, as the small subset is biased due to

dev.
(σλ)

large number of piles from the same site, the use of

these parameters is, therefore, not safe. A consistent
Mean bias

resistance factor of 0.45 could be used when apply-

measured/
calculated

ing equation (3) for all pipe pile cases.

1.1089
1.2937
0.9519
0.9071

1.1284

1.1065
(mλ)

4 EVALUATION OF THE DEVELOPED

EQUATION
MnDOT energy, BC ≥ 4 BPI

MC—Monte Carlo Simulation for 10,000 simulations.

BC ≥ 4 BPI w/o outliers

BC ≥ 4 BPI w/o outliers

4.1 Overview
EOD, diesel, BC ≥ 4 BPI

Following the above described process and

MnDOT energy,

development, three additional stages took plac:

(a) evaluation of equation (3) via a control data-
EOD, diesel,
EOD, diesel,

EOD, diesel,

base, (b) evaluation of equations (2) and (3) via a

EOD only
EOD only
Condition

MnDOT database of dynamic measurements, and

(c) adjustment of the equation to field practices,
local procedures and past experience.

4.2 Re-evaluation of equation (3)

Coeff.

and recommended resistance factors

35
30
30
30

30

30

An independent control database was assembled

cases

beyond the one described in section 2 and not being

# of

(n)

99

41
38

16

14

part of the data originally used to develop equa-

Table 5.

tions (2) and (3). The control database included

24 H piles of which 20 cases related to EOD with
Case
no.

blow counts greater or equal to 4 bpi (BC ≥ 4 bpi).

1

2
3

450

ISGSR2013.indb 450 10/18/2013 9:44:11 AM

 Equations (2) and (3) were examined to the general
driving conditions and appropriate resistance fac-
tors were developed as summarized below.
Figure 7 summarizes the resistance factors devel-
oped for H piles using the original and the control
databases. Figure 7 presents the findings leading
to the conclusion that the appropriate resistance
factors to be used for H piles under MnDOT prac-
tices (and all conditions as well) would be φ = 0.60
assuming redundant pile use.
Figure 8 summarizes the findings regarding
resistance factors developments for pipe piles with
the use of relevant databases. The small datasets
associated with the best match to the MnDOT
practices has several sets of identical piles from
a small number of sites and hence result with a
reduced variability (i.e. COV) and increased resist- Figure 8. Developed and recommended resistance fac-
ance factor. Figure 8 expresses this trend showing tors as a function of pipe piles’ database and its subsets
a consistent increase in the resistance factor with for existing and proposed MnDOT dynamic equations.
the decreased number of cases in the database (or
more accurately, with an increased alliance of the
database with MnDOT practices). mended to be used with the new MnDOT dynamic
In summary, the recommendations of the resist- equation (equation 3) for all H and Pipe Piles,
ance factors for the new dynamic equation proposed driven by Diesel hammer to EOD BC ≥ 4 BPI
to be used by the MnDOT (equation 3) was inves- being φ = 0.60 for H piles and 0.45 for pipe piles.
tigated leading to the conclusion that the appropri-
ate resistance factors to be used for pipe piles under
4.3 Evaluation of the developed equation via
MnDOT practices (and all conditions as well) would
dynamic measurements database
be φ = 0.45 assuming redundant pile use.
The difference in the behavior of the two pile 4.3.1 Overview
types is evident. While H piles are predominantly MnDOT had not accumulated its own database
small displacement piles, closed-ended pipe piles of load tested piles as the process was not part of
are large displacement piles and would be, there- the common practice. As such, the above described
fore, more sensitive to soil inertia effects expressed equation was developed by tailoring a generic data-
via blow count and hammer type and energy. As base to the practices of MnDOT.
a result, it is unwise to rely on the smaller subsets A local database containing dynamic measure-
that provide resistance factors of φ = 0.80 to 0.90. ments and signal matching analyses was devel-
A unique resistance factor is therefore recom- oped by the MnDOT. The database was compiled
and provided by Messrs. Ben Borree and Derrick
Dasenbrock of the MnDOT Foundations unit.
The database consists of accumulated PDA data
from various projects with supplemented DOT
data (stroke, blow count, etc.), and was entered
into formatted spreadsheets provided by UML.
The compiled database contains 126 pipe-pile
cases including hammer type and rated energies
that match for the most part the MnDOT practice
as previously established. 95 cases included signal
matching analyses (CAPWAP) and sufficient data
to evaluate the new equation. The dataset did
not include any static load test information and
50% of it was assessed to relate to bridges within
one square mile area of Minneapolis/St. Paul
(Rowekamp, 2011). The evaluation of the data was
aimed at comparing the performance of the signal
Figure 7. Developed and recommended resistance fac- matching analyses to various dynamic equations
tors as a function of H piles’ database and its subsets for including the above described newly developed
existing and proposed MnDOT dynamic equations. MnDOT equation.

451

ISGSR2013.indb 451 10/18/2013 9:44:12 AM

 4.3.2 Comparison results statistical parameters provided in Table 6 present
The comparison between the signal matching comparisons between the signal matching analyses
analyses (CAPWAP) and the various selected and equations (2) and (3), respectively.
dynamic equations is described in detail by
Paikowsky et al. (2013). Figures 9 and 10, and the 4.3.3 Conclusions
Noting that the analyses are not compared to a
benchmark (i.e. SLT) capacity, the analyses presented
lead to the following observations and conclusions:
a. The CAPWAP values to which the new equa-
tions are compared most likely represents a
typical conservatism of CAPWAP at EOD.
Previous analyses of static load test ultimate
capacity over CAPWAP capacity prediction
(377 cases of EOD and BOR) resulted with a
ratio of 1.368 (Paikowsky et al., 2004).
b. The proposed new MnDOT equation (3) desig-
nated for Diesel hammers, EOD and BC > 4 bpi
(coeff = 30) performed the best of all examined
dynamic formulas with a mean ratio of 1.0
compared to CAPWAP.
Figure 9. CAPWAP vs. new MnDOT formula c. The general proposed new MnDOT equation
(coefficient 30) all CIP piles. (2) designed for all hammers, all piles, all cases
(coeff = 35) provided the second best ratio com-
pared to CAPWAP. A bit on the unsafe side but
probably considering CAPWAP conservatism,
very close to what one could expect in a SLT.

5 EQUATION ADJUSTMENT
AND FINAL FORMULATION

5.1 Overview
The equations presented above were found to
best predict pile capacity when examined with
separate data not related to its development.
Several additional developments were required in
order to finalize the equation format to be used
by the MnDOT. These developments include the
Figure 10. CAPWAP vs. new MnDOT formula following:
(coefficient 35) all CIP piles.

Table 6. Summary of statistical analysis and best fit line correlation.

Statistics
Figure No. of Coeff. of
Relations no Category cases n Mean S.D. COV Best fit line determ. (r2)

CAPWAP/ 10 All piles 95 0.8260 0.1632 0.1976 Ru = 1.178 * Rc 0.964

proposed EOID 40 0.8438 0.1591 0.1886 – –
MnDOT BOR 55 0.8131 0.1664 0.2047 – –
(35) En EOID & BC ≥ 4 BPI 33 0.8314 0.1533 0.1843 – –
CAPWAP/ 9 All piles 95 0.9637 0.1904 0.1976 Ru = 1.009 * Rc 0.964
proposed EOID 40 0.9844 0.1857 0.1886 – –
MnDOT BOR 55 0.9486 0.1941 0.2047 – –
(30) En EOID & BC ≥ 4 BPI 33 0.9700 0.1788 0.1843 – –

Notes: Ru is the calculated capacity using each of the dynamic formulae. Rc is the Capacity determined by
CAPWAP.

452

ISGSR2013.indb 452 10/18/2013 9:44:13 AM

 a. Modify hammer energy to rely on ram’s stroke reason for the difference. Overall, it was found
measurements by field inspectors. that under normal driving conditions, the recom-
b. Scrutinize the database for specific case histo- mended new equation results with equal or better
ries to be examined as to their applicability to economic outcome. However, under unrestricted
MnDOT practices (e.g. pile size, soil type, driv- high driving resistance, it was found that past
ing conditions, etc.). practice resulted with an economic advantage.
c. Design, construction, and cost ramifications This, however, was a result of using a dynamic
when moving from past practice to the use of equation that had an almost linear increase in
the new equations. capacity with increase in blow count, result-
d. Examine the equation applicability for analyz- ing with unsafe high capacity prediction under
ing the driving of concrete and timber piles. refusal.
e. Configure the equation’s format to resemble
existing field practices.
5.5 Equation examination for concrete
and timber driven piles
The detailed procedures of the above develop-
ments are beyond the scope of the present paper Expansion of the work previously described
and are presented in details by Paikowsky et al. included the assembly of two additional data-
(2013). The following sections provide a brief bases for timber and Prestressed Precast Concrete
description of the work done tracking the above (PPC) piles. These databases included 137 PPC
list and concluding with the equations adopted for and 28 timber cases for which static load tests
current use by the MnDOT. were carried out. The equations were examined
for their performance and for statistical param-
eters to allow LRFD resistance factors to be
5.2 Field observations
developed.
While equations (2) and (3) were developed using
nominal hammer energy, the actual developed energy
5.6 Equation format
(in particular for diesel hammers) depends on the
resistance to penetration and the fuel-pump setting. Final stage of development included restructuring
As such, inspector’s field observations, albeit difficult of the equations to be compatible with MnDOT
and of limited accuracy, can greatly help in assess- field practices, e.g. using ton rather than kips in
ing the actual hammer’s energy. This value is typically capacity evaluation, weight of ram in pounds and
75% of the nominal energy, but can vary significantly the use of penetration per blow.
especially during easy driving. Examination of those
values and the adjustment of the equations to Eh = Wr
5.7 Final formulation
⋅ h, i.e. hammer energy equal to the ram’s weight times
the stroke, was performed including statistical data The following final formulation known as MPF12
based on existing observations and limiting values to (Minnesota Pile Formula 2012) was adopted for
prevent unreasonable values. use:

5.3 Database outliers of common MnDOT W H ⎛ 10 ⎞

Rn = 20 × log ⎜ ⎟ (4)
practice 1, 000 ⎝ s⎠
As the databases described in section 2 do not
include case histories from Minnesota, different where Rn = nominal resistance (tons), H = stroke
cases (of high and low biases) were examined by (height of fall) (ft), W = weight of ram (lbs),
the MnDOT research panel and unrelated cases s = set (pile permanent displacement per blow)
were excluded. These cases included, for example, (inch). The value of the energy (W ⋅ H) used in
very short piles, piles of easy driving, piles designed the dynamic formula shall not exceed 85% of the
for loads out of range, etc. manufacturer’s maximum rated energy for the
hammer used considering the settings used during
driving. Equation (4) is to be used with the follow-
5.4 Ramification of change in practice ing Resistance Factors (RF) in order to obtain the
Data for a large number of piles observed dur- factored resistance:
ing bridge construction were assembled and the
previous practice was compared to the proposed Rr = φ ⋅ Rn (5)
practice. The change was examined; i.e. the
acceptability of a pile under previous practice vs. – for pipe and concrete piles, φ = 0.50, 2 < BC ≤ 15
under the proposed new practice, and then the – for H piles, φ = 0.60, 2 < BC ≤ 15 BPI

453

ISGSR2013.indb 453 10/18/2013 9:44:14 AM

 and for timber piles: Previous students of the Geotechnical
Engineering Research Laboratory at the University
WH 10 of Massachusetts Lowell are acknowledged for
Rn = 10 log (6) their contribution to the aforementioned databases
1000 S and various studies, namely: John J. McDonell,
John E. Regan, Kirk Stenersen, Colin O’Hearn,
where φ = 0.60. Jorge Fuentes and Christopher Jones.
See http://www.dot.state.mn.us/bridge/docs
down.html.
Although the equation was developed for ham-
mers with Eh ≤ 165 kip-ft (224 kN-m), its use should REFERENCES
be applicable for hammers with higher energies but
this was not verified directly in the study. Cook, R.D. 1979. Influential Observations in Linear
Regression. Journal of the American Statistical Asso-
ciation 74, 169–174.
Davisson, M. 1972. High Capacity Piles. In Proceedings,
ACKNOWLEDGEMENTS Soil Mechanics Lecture Series on Innovations in Foun-
dation Construction, ASCE, Illinois Sect., Chicago,
The presented research was supported by Minne- pp. 81–112.
sota Department of Transportation (MnDOT) via Gates, M. 1957. Empirical formula for predicting pile
a grant to Minnesota State University at Mankato capacity, Journal of the Soil Mechanics and Founda-
and is presented in detail by Paikowsky et al. (2009, tions Division, ASCE 27(3): 65–66.
Paikowsky, S.G. with contributions by Birgission G.,
2013). The Technical Advisory Panel (TAP) is
McVay M., Nguyen T., Kuo C., Baecher G., Ayyub B.,
acknowledged for its support, interest, and com- Stenersen K., O’Mally K., Chernauskas L., and O’Neill
ments. In particular we would like to mention, M. 2004. Load and Resistance Factor Design (LRFD)
Mr. Paul Rowekamp and Mssrs. Richard Lamb, for Deep Foundations, NCHRP Report 507, Transpor-
Gary Person, Dan Mattison and Derrick Dasen- tation Research Board, National Research Council,
brock of the Foundations Unit, and Mssrs. Paul Washington, DC., pp. 134 (not including Appen-
Pilarsky, Dustin Thomas, Dave Dahlberg, Kevin dices), http://onlinepubs.trb.org/onlinepubs/nchrp/
Western, Bruce Iwen, and Paul Kivisto of the nchrp_rpt_507.pdf.
Bridge Office. Paikowsky, S.G., and Canniff, M.C. 2004. AASHTO
LRFD Specifications for Serviceability in the Design
The research presented in this manuscript makes
of Bridge Foundations, Appendix A: Questionnaire.
use of a large database specifically developed for Interim Report Appendix A submitted for the
MnDOT purposes. This database makes use of research project NCHRP 12-66 to the National
data originally developed for a Federal Highway Academies, Geosciences Testing & Research, Inc., N.
Administration (FHWA) study by Paikowsky et al. Chelmsford, MA.
(1994) followed by an updated database denoted Paikowsky, S.G., Canniff, M.C., Robertson, S.O., and
as PD/LT 2000 presented by Paikowsky and Sten- Budge, A.S. 2013. Load and Resistance Factor Design
ersen (2000), which was also used for the LRFD (LRFD) Pile Driving Project—Phase II Study.
development for deep foundations (presented in Final report submitted to Minnesota State Univer-
sity Mankato, University of Massachusetts Lowell,
NCHRP Report 507, see Paikowsky et al., 2004).
Lowell, MA.
The contributors for those databases are acknowl- Paikowsky, S.G., Marchionda, C.M., O’Hearn, C.M.,
edged for their support as detailed in the refer- Canniff, M.C., and Budge, A.S. 2009. MnDOT
enced publications. Mssrs. Carl Ealy and Albert Research Project: Developing a Resistance Factor for
DiMillio of the FHWA were constructive in sup- MnDOT’s Pile Driving Formula. Final Report submit-
port of the original research studies and facilitated ted to Minnesota State University Mankato, October
data gathering via FHWA sources. Significant 29, pp. 209, website: http://www.lrrb.org/pdf/200937.
additional data were added to those databases, pdf.
most of which were provided by six states: Illinois, Paikowsky, S.G., Marchionda, C.M., O’Hearn, C.M.,
Canniff, M.C., Budge, A.S., Dasenbrock, D.,
Iowa, Tennessee, Connecticut, West Virginia,
Person, G. and Dahlberg, D. 2010. The reliability of
and Missouri. The data obtained from Mr. Leo MnDOT pile driving formula—new equation and
Fontaine of the Connecticut DOT was extremely resistance factors. Proc. University of Minnesota
valuable to enlarge the MnDOT databases to the 58th Annual Geotechnical Eng. Conf., University of
robust level presented in this study. In addition, the Minnesota, St. Paul, Feb. 26, pp. 79–99.
data provided by Ms. Betty Bennet of the Ontario Rowekamp, P. 2011. Email communication Monday,
ministry of Transportation were invaluable for April 11, 2011 to A. Budge and S. Paikowsky.
developing the timber piles database.

454

ISGSR2013.indb 454 10/18/2013 9:44:14 AM

 Geotechnical Safety and Risk IV – Zhang et al. (eds)
© 2014 Taylor & Francis Group, London, ISBN 978-1-138-00163-3

On the validation of reliability and partial safety factors for axially

loaded piles in dense sand

K.A. Schmoor & M. Achmus

Institute for Geotechnical Engineering, Leibniz University of Hannover, Germany

ABSTRACT: By designing mainly axially loaded offshore foundation piles in North Sea conditions,
two commonly used design methods, the β-method and the ICP-method, were used. Thereby the required
minimum embedded pile length may vary a lot. The actual safety which is achieved with a certain pile
length and the corresponding prescribed partial safety factors is not known. For general assumed pile and
soil conditions it could be determined that a reliability index between β = 3.5 and β = 4.1 is established.
Further it could be shown that the prescribed partial safety factors should be decreased for the β-method,
where an increase for the ICP-method should be considered. By performing reliability based design also a
more accurate determination of the required pile length for a specified safety is possible.

1 INTRODUCTION 2 DESIGN METHODS

One of the main reasons for the increase of The tension bearing capacity of mainly axially
renewable energy in Germany over the last years loaded piles consists basically of the mobilized
is the continued extension of wind energy. For friction between the pile outer shaft area and
the next years additionally several offshore wind the surrounding soil. Additionally if an open
farms are planned to be built in the North Sea. ended pile is used two different condition states
Since many projects at the German North Sea sites “plugged” or “unplugged” have to be considered.
are going to be realized outside of the 12 sea mile In the unplugged case also the friction resistance
border, water depths greater than 30 m are often between the inner pile shaft area and the inner soil
faced. For such water depths jacket and tripod can be taken into account. By assuming a plugged
supporting structures with mainly axially loaded condition the effective weight of the inner soil plug
foundation piles are mostly used. can be added to the bearing capacity.
Thereby the pile resistance can be calculated by Generally it can be observed that in almost all
applying the well known β-method (also known as design cases the tension limit state is the control-
API-method) recommended in the guideline of the ling one with regard to the required pile length.
American Petroleum Institute (API 2007). In the Therefore only tensile capacity is considered
course of publishing errata and supplements for here. The pile resistance for the unplugged or
the existing guideline, new simplified CPT-based plugged condition can be computed by applying
design methods were introduced following the Equation 1:
recommendation of the corresponding research
groups, like the ICP-method introduced by Jardine Rt Ao ∫ ft ( z ) z Ai ∫ ft ( z ) dz G′
G p] (1)
et al. (2005).
For many design cases the ICP-method leads to
considerably differing resistances compared with where Ao = outer pile shaft area; ft(z) = skin fric-
the API-method. This affects also the deterministic tion for tension loading; Ai = inner pile shaft area;
design, where this deviation leads to a strong vari- and G′p = effective weight of the inner soil plug.
ation of the required embedded pile length. For Following the API-method the skin friction of
practical applications it is of interest which method offshore foundation piles in sand for tension con-
is more reliable for determining the pile length and ditions can be computed by Equation 2. Thereby
which method should be used, respectively? In the friction is basically determined by multiplying
addition it is also unknown which safety level— the effective vertical stress with a β-value. Further
measured by the reliability index β—is finally a limitation of the skin friction is included. Both,
achieved by applying one of these methods with the shaft friction factor as also the limiting fric-
the corresponding partial safety factors in a deter- tion value, only depend on the relative density of
ministic design. the soil and may vary with depth. The additional

455

ISGSR2013.indb 455 10/18/2013 9:44:15 AM

 factor of 2/3 is not recommended by the API, but
is regularly used in practice since it is prescribed
by certification companies like the Germanischer
Lloyd (GL 2010). It is not clearly indicated whether
the reduction factor should also be applied to the
limit value. However, here the approach is used in
that way.

ft ( z ) = 2 3 β ′v ≤ 2 3 ft ,max (2)

where β = shaft friction factor from Table 1;

σ′v = effective vertical stress; and ft,max = limit skin
friction from Table 1.
To determine the relative density from the
Figure 1. Considered pile system.
results of a CPT test, the approach proposed
by Jamiolkowski et al. (API 2007) according to
Equation 3 is recommended by the API. 3 PILE SYSTEM UNDER
CONSIDERATION
1 ⎛ qc ⎞
Dr = ln ⎜ 0.51 ⎟
(3)
2.93 ⎝ 205 p′ m ⎠ For the performed study typical site conditions for
the North Sea and typical ranges for the foundation
pile properties were assumed.
where qc = measured cone tip resistance in kPa; and
Since the subsoil in the German North Sea
p′m = effective mean in-situ soil stress in kPa.
mostly consists of dense sands with only limited
The ICP-method stated in API (2007) assumes
intermediate cohesive layers, two idealized CPT
a plugged condition for the determination of the
profiles for homogeneous dense (Dr = 0.75) and
tension pile resistance. Also no additional resist-
very dense (Dr = 0.93) sands were considered. The
ance by the inner soil plug is taken into account.
effective unit weight was assumed to be 10 kN/m3,
The skin friction for non-cohesive soils according
since this value represents a reasonable estimate for
to the ICP-method can be determined by applying
both soil densities.
Equation 4. Therefore the trend of the skin fric-
The choice of pile properties depends on the
tion is mainly influenced by the measured cone tip
type of foundation, the water depth and the sub-
resistance.
soil condition at the desired location. In general
0.1 it can be said that pile slenderness ratios (embed-
⎛σ ′ ⎞ ded length to diameter) between L/D = 10 and
ft z ) = 0 016 qc⎜ v ⎟
⎝ pa ⎠ L/D = 40 are used. Thereby the pile outer diameter
−0.4 is varying between D = 1 m and D = 3 m, where
⎛ ⎡L z ⎤⎞ the pile embedded length is commonly chosen to
× Ar 0.2
⎜mmax ⎢ ;ν ⎥⎟ tan δ cv (4)
⎝ ⎣ Do ⎦⎠ be between L = 20 m and L = 60 m. For the con-
sidered range of the load in this study a diameter
where pa = atmospheric pressure = 100 kPa; of D = 2 m was chosen. The regular pile wall thick-
Ar = effective area ratio Ar = 1 − (Di/Do)2; Di = ness which is commonly used can approximately be
pile inner diameter; Do = pile outer diameter; determined by t = D/40.
L = embedded pile length; ν = dimensionless param- A characteristic load of Vk = 9 MN for a pile
eter ν = 4(Ar)0.5; and δcv = interface friction angle. foundation with a supporting structure for a 5
MW windmill in the German North Sea with a
water depth of 30 m was considered. Therefore
typical 50-year extreme environmental values were
Table 1. Design parameters for the API-method (API assumed (Achmus et al. 2007). Figure 1 shows the
2007). considered boundary conditions.
β ft,max
Relative density (−) (kPa)
4 DETERMINISTIC DESIGN
Medium dense (Dr = 0.36–0.65) 0.37 81
Dense (Dr = 0.65–0.85) 0.46 96 The partial safety factors for the determinis-
Very dense (Dr = 0.85–1.00) 0.56 115 tic design of an offshore pile foundation in
Germany are defined by the Eurocode 7 with the

456

ISGSR2013.indb 456 10/18/2013 9:44:15 AM

 applying only one single random variable. In case
of a pile where properties along the vertical axis
are varying much more, a random field could be
applied to get a more reliable subsoil model.
The soil model used for the performed study
mainly bases on a 1-D autocorrelated field for the
cone tip resistance. Therefore the two CPT pro-
files for constant densities were specified as trend
functions and a random component was added to
represent uncertainties from inherent variability
wqc and measurement errors eqc.
To account for the genesis of the soil addition-
ally an autocorrelation was taken into account. This
implies that the properties of two points which are
close together are more related as the properties
of two points which are far away from each other.
Figure 2. Deterministic resistances for the API and ICP This relation can be expressed by an autocorrela-
method with increasing pile length. tion function. Equation 6 shows the applied single
exponential function. Thereby the autocorrelation
length describes how strong the correlated field is.
national supplementary code DIN 1054. Here Since only a 1-D field was simulated the cor-
a 50-year extreme event is studied. Thereby the relation structure was computed by applying the
corresponding partial safety factor for the load is Cholesky technique (Fenton & Griffiths 2008).
defined as γL = 1.35. For the tension resistance of
a pile a partial safety factor of γR = 1.5 according ⎛τ ⎞
−
to DIN 1054 is applied. Together a product of ⎝θ ⎠
ρ (τ ) = e (6)
partial safety factors—indicating the global factor
of safety—of 2.03 is applied.
For the assumed CPT profiles with their cor- where τ = distance between two points; and
responding relative densities the two introduced θ = autocorrelation length.
design methods were evaluated. Figure 2 shows the Figure 3 elucidates one realization of the simu-
development of the characteristic resistance with lated CPT profile with the presumed trend func-
increasing pile length. For very dense soil condi- tion for the very dense soil condition.
tions a strong deviation between the methods must
be noticed. 5.2 Model error
By assuming the characteristic load and par-
tial safety factors a characteristic resistance of A model error can be determined by comparing
Rk = 18.23 MN is required to fulfill the determinis- measured properties with calculated ones. In this
tic design proof according to Equation 5. The cor- way a model error is an indicator how reliable a
responding pile lengths for the different methods calculation method is.
and densities can be obtained from Figure 2. Also
Table 4 summarizes the deterministically deter-
mined pile lengths.

Vk Lγ R Rk (5)

where Vk = characteristic load; γL = partial safety

factor for the loading; γR = partial safety factor for
the resistance; and Rk = characteristic resistance.

5 RELIABILITY BASED APPROACH

5.1 Stochastic subsoil model

Since soil properties are subject to spatial varia-
tion, a certain volume has to be considered. For
most geo-technical applications which affect a local
soil area, soil properties could be approximated by Figure 3. One realization of a simulated CPT profile.

457

ISGSR2013.indb 457 10/18/2013 9:44:17 AM

 Table 2. Model error (Rm/Rc) for ICP and API-method. Table 3. Parameters used for all Monte-Carlo
simulations.
μ σ COV
Method (−) (−) (−) Basic variables Notation μ σ Type*

ICP-method, all pile tests 1.01 0.28 0.28 Unit weight γ′ in kN/m3 10 1 N
API-method, all pile tests 1.15 0.58 0.50 Inherent wqc in MN/m2 0 6 N
ICP-method, only related tests 1.16 0.19 0.16 variability qc
API-method, only related tests 1.26 0.14 0.11 Measurement eqc 1 0.15 N
error qc
Transformation eϕ in ° 0 2.8 N
error ϕ
Jardine et al. (2005) obtained a model error for Model mAPI 1.26 0.14 N
both methods where the following pile tests were error API
considered: Closed-ended and open-ended piles, Model mICP 1.16 0.19 N
tension and compression test, steel and concrete error ICP
piles and loose to very dense soil state. Pile diameter D in m 2 – D
Altogether 81 pile tests were evaluated. The sta- Pile wall t in m 0.05 – D
tistical values are shown in Table 2. Therein the thickness
model error is defined as the ratio of the measured Pile length L in m 20–60 – D
to the calculated resistance. Axial loading V in MN 5.4 1.89 G
Achmus & Müller (2010) compared measured *N = Normal distributed; D = deterministic value; and
with calculated results for the API and the ICP- G = Gumbel distributed.
method only with pile tests which are closely
related to the boundary conditions in the North
Sea (see section 3). By this it can be seen how relia- The vertical autocorrelation length for the cone
ble these methods are to a special boundary sector. tip resistance is indicated to be between 0.1 m and
In detail the following pile tests were considered: 2.2 m. Hence, the value of θ = 0.6 m was chosen for
Open-ended piles, tension test, steel piles, dense to all simulations.
very dense soil state and slenderness ratios between The buoyant unit weight γ ′, the measurement
10 and 40. error for the cone tip resistance eqc and the transfor-
By taking into account only pile tests within mation error for the internal friction angle eϕ were
a range of 95% confidence interval, only 6 pile modeled as uncorrelated and normally distributed
test for the ICP-method and only 4 pile tests for random fields with typical values for the mean and
the API-method could be found. The mean and standard deviation. The stochastic formulation for
standard deviations for these related tests are sum- the load was chosen according the recommenda-
marized in Table 2 and should be seen as rough tion by Holicky et al. (2007) for a 50-year extreme
approximations of the real values. environmental event.
As it can be seen the overall model error for All applied variables with their corresponding
the ICP-method is smaller in mean and standard stochastic moments are summarized in Table 3.
deviation compared to the overall model error for
the API-method. By comparing the model error
6 OBSERVED RESULTS
only for the related tests, it can be said that in aver-
age both methods underpredict the resistance. The
6.1 Impact of uncertainties
standard deviation seems also to be closer related
to each other as in the overall case. For the computed data sets first the COVs of the
resistance distributions are evaluated. Addition-
ally the corresponding parts which arise from
5.3 Stochastic variables and simulation
the model error and soil uncertainties were esti-
For each performed simulation with a certain pile mated, as depicted in Figure 4 for very dense soil
length 6 million realizations within a plain Monte- condition.
Carlo Simulation (MCS) were computed. For both methods the total variation of the
The inherent variability of the cone tip resist- resistance is much lower than of the load. For the
ance wqc was assumed to be constant with depth. API-method almost 11% and for the ICP-method
By applying a standard deviation of 6 MN/m2 about 17% were obtained. Further the model error
the COV is varying between 0.75–0.20 for dense is almost exclusively responsible for the variation of
and 0.40–0.10 for very dense soil conditions. the resistance, especially for the API-method. Also
According to Phoon & Kulhawy (1999) the COVs it can be noticed that the deviation arising from
were thereby in a typical range of 0.81 and 0.10. soil uncertainties is decreasing with embedded pile

458

ISGSR2013.indb 458 10/18/2013 9:44:19 AM

 Figure 4. COV for the resistance and load with
increasing pile length.

length, since an averaging effect becomes more

dominant.

6.2 Safety
The safety of a system can be described in terms of
a reliability index β corresponding to Equation 7.
In the following, β is also termed safety.
Figure 5. Evaluated safety with increasing pile length.
In the Eurocode 0 three reliability classes with
the corresponding safeties were proposed. Further
it is remarked that the partial safety factors are
construed for the second reliability class and there- Table 4. Required deterministic pile lengths with
fore a safety of β = 3.8 should be established in a corresponding reliability index.
design system. However, offshore pile foundations
may also be associated to the first reliability class Dr = 0.75 Dr = 0.93
which demands a safety of β = 3.3. Design method API ICP API ICP

β = Θ −1
−
( − ) (7) Deterministic length 51.85 45.91 45.19 30.24
Reliability index 4.1 3.5 4.0 3.5

where Θ-1 = inverse of the cumulative standard

deviation; and pf = failure probability.
For the computed pile lengths the derived reli-
6.3 Partial safety factors
ability index is shown in Figure 5. Table 4 also
presents the corresponding safeties for the deter- Figure 6 elucidates the sensitivity values depending
ministic required pile lengths. on safety. It can be seen that the sensitivity for the
Therefore it can be generally said that by apply- load is much higher than for the resistance. Generally
ing the usual partial safety factors a safety of the values are getting more uniform for higher reli-
β = 3.5 according to the ICP-method and about ability values and more dense soil conditions.
β = 4.1 according to the API-method is achieved. The corresponding partial safety factors which
Nevertheless, it should also be noticed that the should be used for a given safety are shown in
API-method indeed yields a longer embedded pile Figure 7. For the API-method a resistance par-
length. tial safety factor lower than one was computed.
For a certain embedded pile length or safety This uncommon result is achieved due to applying
it could be generally assumed that the real estab- model uncertainty in combination with nonuni-
lished safety or required pile length is between the form COVs. However, the product of both partial
computed values corresponding to these methods, safety factors affects the safety within the deter-
respectively. ministic design proof. In contrast both partial

459

ISGSR2013.indb 459 10/18/2013 9:44:19 AM

 Table 5. Product of partial safety factors for a safety
according to the minimum of both methods.

Dr = 0.75 Dr = 0.93
Design
method API ICP API ICP

γL γR
β = 3.3 1.59 (1.52)* 1.86 1.55 2.60 (1.87)*
β = 3.8 1.96 (1.82)* 2.29 1.84 3.12 (2.31)*

*Corresponding partial safety factors product by not

taking into account the minimum of both methods.

Table 6. Comparison of the separation of required pile

lengths for different designs.
Figure 6. Load and resistance sensitivity values with
increasing safety. ΔL ΔL
(Dr = 0.75) (Dr = 0.93)

For deterministic design 5.94 14.95

For reliability β = 3.3 1.46 8.61
For reliability β = 3.8 3.12 8.77

deterministic values, generally a decrease for the

API-method and an increase for the ICP-method
should be applied. Single partial safety factors as
also the product can be estimated for a desired
safety from Figure 7.

6.4 Deterministic design vs. reliability based

design
By performing reliability based design the same
safety can be aspired as also more information,
like a model error, can be taken into account.
This leads to a more robust determination of the
required pile length. Therefore Table 6 summarizes
the separation according to deterministic design
and two reliability indexes. As it can be seen the
gap between the obtained pile lengths via reliabil-
ity based design is decreased about 61% for dense
and 41% for very dense soil conditions.

Figure 7. Partial safety factors required to achieve 7 CONCLUSIONS

a certain safety.
It could be determined that the use of partial safety
factors according to DIN 1054 for mainly axially
safety factors for the ICP-method are increasing loaded foundation piles of offshore wind turbines
continuously, where for very dense sand condition leads in case of the API-method to a safety of
the partial safety factors tend to be more uniform, β = 4.1, where in case of the ICP-method a safety
like also the sensitivity values. of β = 3.5 was detected. Based on that, it generally
Table 5 shows the products of partial safety fac- can be assumed that the real safety is in the range
tors which should be used within a deterministic of these values.
design to achieve the given safeties according to Also it was found that the variation for the
the minimum of both methods. Comparing to the resistance arises mainly from the model error and

460

ISGSR2013.indb 460 10/18/2013 9:44:21 AM

 that the soil uncertainty is of much lower effect DIN1054 2010. Baugrund-Sicherheitsnachweise im
than it would be expected. Erd und Grundbau, Berlin: Deutsches Institut für
According to the computed results the partial Normung.
safety factors should be decreased for the Eurocode 0, 2010. DIN EN 1990: Eurocode: Grundlagen
der Tragwerksplanung, Berlin: Deutsches Institut für
conservative API-method, where an increase for Normung.
the ICP-method is recommended. Eurocode 7, 2009. DIN EN 1997: Eurocode 7: Entwurf,
A more specific determination of the pile length Berechnung und Bemessung in der Geotechnik—Teil
by applying reliability based design is possible, 1: Allgemeine Regeln, Berlin: Deutsches Institut für
since more information is evaluated. Normung.
Fenton, G.A. & Griffiths, D.V. 2008. Risk assessment in
geotechnical engineering. Hoboken: Wiley & Sons.
REFERENCES GL G.L. Rules and Guidelines, IV Industrial Services,
Guideline for the Certification of Offshore Wind
Achmus, M., Abdel-Rahman, K. & Tom Wörden, F. Turbines. Germanischer Lloyd Wind Energie GmbH,
2007. Geotechnical design of piles supporting foun- Hamburg/Germany, Edition 2005.
dation structures for offshore wind energy converters. Holicky, M., Markova, J. & Gulvanessian, H. 2007. Code
Proceedings of the 17th (2007) International Offshore Calibration Allowing for Reliability Differentiation
and Polar Engineering Conference (ISOPE 2007): Held and Production Quality. In Kanda, Takada & Furuat
in Lisbon, Portugal, July 1–6, 2007. Lisbon: ISOPE. (eds), Application of Statistics and Probability in Civil
Achmus, M. & Müller, M. 2010. Evaluation of pile capac- Engineering: Proceedings of the 10th International
ity approaches with respect to piles for wind energy Conference, London. Taylor & Francis.
foundations in the North Sea, In S. Gourvenec, D.J. Jardine, R., Chow, F.C., Overy, R. & Standing, J. 2005.
White (eds), Frontiers in Offshore Geotechnics Two, ICP design methods for driven piles in sands and clays.
CRC Press: Balkema London: Thomas Telford.
API 2007. Errata and Supplement 3—API Recommended Phoon, K.K. & Kulhawy, F.H. 1999. Characterization
Practice 2A-WSD, Recommended Practice of geotechnical variability. Canadian Geotechnical
for Planning, Designing, Constructing Fixed Offshore Journal 36(4): 612–624.
Platforms—Working Stress Design.

461

ISGSR2013.indb 461 10/18/2013 9:44:22 AM

This page intentionally left blank
 Geotechnical Safety and Risk IV – Zhang et al. (eds)
© 2014 Taylor & Francis Group, London, ISBN 978-1-138-00163-3

Reliability-based evaluation of vertical bearing capacity

of piles using FORM and MCS

A. Teixeira & A. Gomes Correia

University of Minho—C-TAC, Guimarães, Portugal

A.A. Henriques
University of Porto—LABEST, Porto, Portugal

ABSTRACT: This paper aims to show reliability concepts for axial pile foundations’ design and safety
aspects, revealing that reliability analysis do not imply a significant increased effort when compared to
the traditional design methodologies. In fact, calculations using reliability concepts provide very useful
information, such as the uncertainty that mostly influence the probability of failure. Therefore, the most
used reliability methods, FORM and MCS, are here presented and used to evaluate the safety of the bear-
ing capacity of an axial bored pile considering the geotechnical uncertainties and different load values.
For the application example presented, the physical uncertainties of actions, the inherent soil variability
and model error (pile’s resistance) were determined by experimental in situ tests (SPT) or collected from
bibliography.

1 INTRODUCTION Therefore, a simple reliability-based approach is

presented, using both FORM and MCS to evalu-
The new regulation codes and social concerns, such ate the safety of the bearing capacity of a pile
as the sustainability or the economy, are imposing concerning different load values. Finally, an appli-
geotechnical engineers to increase their ability to cation example of a vertically loaded pile founda-
deal with uncertainty in a direct and more rational tion, a bored pile (reinforced concrete) installed in
way. Designs should be economic, sustainable and a residual soil of granite, illustrates the applicability
reliable at the same time. Concerning the reliability of the proposed reliability-based approach and the
part of design, different approaches can be done, consistency between the FORM and MCS results.
but the most important thing to keep in mind is
the simplicity of the tools and applied methods.
2 RELIABILITY
To prevent loss and intuitive understanding of the
problem it is necessary to fully understand what is
Reliability Analyses (RA) have the intention of
being done, regarding data collection task, uncer-
assessing the probability of a particular behavior
tainties characterisation and the reliability calcula-
by introducing randomness of the variables.
tion itself.
Because in geotechnical practice these tools are
not as present as they are in the structural engi- 2.1 RA levels
neering, this paper aims to show simple reliability
Usually, the RA levels considered are five—Table 1:
concepts for axial pile foundations’ design and
safety aspects. It is also shown that calculations – RA level zero: deterministic methods, the
using reliability concepts provide very useful tools Random Variables (RV) involved are taken
to model the uncertainties and to quantify and as deterministic and uncertainties taken into
give information about the ones that mostly influ- account by a global Safety Factor (SF);
ence the behaviour under study. The most used – RA level I: semi-probabilistic methods, where
and simple reliability methods are FORM (First deterministic formulas are applied and the repre-
Order Reliability Method) and MCS (Monte Carlo sentative values of RV (statistically determined)
Simulations). These methods denote different reli- are multiplied by partial SF (calibrated by RA
ability levels. FORM is an approximate method, level II or III);
while MCS is a pure probabilistic method with – RA level II: approximate probabilistic methods,
higher accuracy, and it is used as reference method. where the RV are characterised by their statistical

463

ISGSR2013.indb 463 10/18/2013 9:44:22 AM

 Table 1. Levels of reliability.

Zero I II III IV

Geotechnical parameters
Calculation method (deterministic)
Design parameter (statistical basis)
Variability of parameters:
– mean and SD
– PDF
Costs/consequences
Type of analysis: Global SF Partial SF e.g. FORM e.g. MCS Risk

parameters—mean and Standard Deviation in calculations (Phoon & Honjo 2005, Phoon
(SD) or coefficient of variation (COV = SD/ 2008, Zhang et al. 2011).
mean); A full account of the reliability methods devel-
– RA level III: full probabilistic analyses, it takes opment and evolution can be found in Manohar &
into account all the probabilistic characteristics Gupta (2005) and Baecher & Christian (2003).
of the RV—mean, SD and Probability Distribu-
tion (PDF) and when the problem is complex
simulations methods are used; 2.3 RA-based safety evaluation
– RA level IV: risk analyses, where the conse- In order to carry the reliability-based safety evalu-
quences of failure are also taken into account and ation, the following procedure is used for both
the risk is used as a measure of the reliability. FORM- and MCS-based RA:
– definition of the significant failure modes and
formulation of their functions → g(Xi):
2.2 RA methods
One of the first attempts to solve this kind of M R−E g(X i ) (1)
problems involving probabilities and statistics,
with non-normal variables and complex integral
where M is the safety margin, R denotes the
calculations, was the First Order Second Moment
resistance, E denotes the action, g is the perform-
(FOSM) proposed by Cornell (Cornell 1969) that
ance function, and Xi are the random variables;
first introduced the concept of reliability index.
– identification of the random and deterministic
FOSM ignores the shape of the PDF, using only
variables → Xi;
the mean and SD of the RV, and the calculation
– description and characterisation of the RV,
model is linearized using Taylor’s expansion.
namely the statistical parameters—mean, SD,
Next, other authors (e.g. Hasofer & Lind 1974)
COV and PDF type—as well as identification of
continued to develop the method, and FORM
the dependencies among them (if existent);
was proposed, solving the invariance problem
– selection of the target reliability index → βT and
of FOSM and being nowadays one of the basic
correspondent probability of failure → pf.
tools for RA level II. These are called first and
second order methods and are approximation The target depends on many factors, such as
techniques. the type of structure, the social tolerance for non-
For more accurate calculations (RA level III) compliance, among others. Its selection can be
the simulation methods are used, these methods based on previous similar construction projects
can be applied to RV with non-normal distribu- that met predefined requirements or on recom-
tion and complex performance functions (e.g. mendations in design codes—βT for ultimate limit
requiring nonlinear functions or finite element state design should be 3.3 to 4.3 (CEN 2002a)
methods). RA level III use all the statistical infor- or 1.3 to 4.3 (ISO 2394, 1998). These values cor-
mation about the RV and the most known and respond to pf of approximately 10−1 and 10−5, as
method of reference is the MCS. To MCS it is shown in Figure 1.
possible to add reduction of variance techniques, For safety evaluation of the bearing capacity,
such as Stratified sampling, Latin hypercube or the problem main characteristics are fixed, such
Importance sampling, with the goal of reducing as dimensions and materials, and probabilities of
calculation time, by selecting only points near fail- failure (reliabilities) are calculated for different
ure, or selecting specific groups to be represented load values.

464

ISGSR2013.indb 464 10/18/2013 9:44:22 AM

 general steps previously described, RA using MCS
is accomplished as follows:
– based on the desired β select the number of sim-
ulations → n;
– generate n values for each RV based on the varia-
bility information (mean, SD or COV and PDF)
by applying existing correlations;
– compute the value of the performance function
for each generation;
– determine pf as the sum of the simulations that
fail (g(Xi) < 0) divided by n.
The total number of simulations n must be
chosen carefully. Its stability should be studied by
analysing the fluctuation of the final results of dif-
ferent n (Teixeira 2012).
Figure 1. Probability of failure vs. reliability index.
2.6 Uncertainties
Uncertainties are associated with the RV. Mostly,
RV are continuous variables and uncertainties
characterised by their statistical moments, mean
and SD or COV, and PDF. Very commonly this
PDF is assumed to be Normal or Lognormal. The
uncertainties are usually divided as follows:
– physical uncertainties, are the inherent uncer-
tainties of the material;
– modelling uncertainties, come from the theoreti-
cal approaches (transformations and models);
Figure 2. Graphical representation of FORM. – statistical uncertainties, include the uncertainty
associated with the finite size and fluctuations in
the samples (always present);
2.4 FORM – human errors (usually not considered in geotech-
nical RA, however its influence can be reduced
RA level II using FORM is based on successive
by an adequate quality control during design
linear approximations to a nonlinear performance
and construction);
function (Fig. 2) (Henriques et al. 1999). Following
– and spatial variability.
the previously described steps, a RA using FORM
is accomplished as follows: Gathering all RV and uncertainties may be a
consuming task, because there is a great number
– transforming all RV into standard normalised
of variables, but only the variability of the most
RV → Z∼N(0,1);
important and influent ones are worth consider-
– rewriting the performance function with nor-
ing (Baecher & Christian 2003, Teixeira et al.
malised RV → g(Zi);
2012). For that purpose, sensitivity analyses can
– selecting the design point → Z*, that is, the one
help evaluate the relative influence of the uncer-
closest to the origin in the normalised space;
tainty associated with each RV on the final result.
– evaluating β as the distance between the origin
However, the knowledge of the important and
and the design point Z*.
influent uncertainties is sometimes limited. Data
This method includes sensitivity factors (α), from the specific site in study may not be available
determined as shown in Figure 2b (for the case of or may not be sufficient to estimate variability,
two RV, E and R), that help to evaluate the influ- in this cases, uncertainty can be characterized by
ence of each RV in the result. COV observed in similar sites. Kulhawy & Mayne
(1990), Phoon & Kulhawy (1999a, 1999b) and
Uzielli et al. (2007) did a literature review for the
2.5 MCS
COV of inherent variability, scale of fluctuation
Simulation methods are RA level III. They can (soil spatial correlation), and COV of measurement
be applied to more complex RV and performance error. Also, typical values of COV for soil proper-
characteristics Once again, after following the ties and in situ test results have been compiled and

465

ISGSR2013.indb 465 10/18/2013 9:44:23 AM

 6 m in length and 0.6 m in diameter. The ultimate
capacity of the pile, measured under static loading
to failure, was 1350 kN.
Since the case study is an experimental pile
designed according to the available resources and
sponsorships, the actions values were not available.
Therefore, the permanent and variable loads range
(considered equal in magnitude) was determined
based on the prediction of the vertical bearing
capacity and using partial safety factors proposed
by Eurocodes (CEN 2002b, 2007).

3.2 Bearing capacity

Figure 3. Methodology and uncertainties considered in
Throughout the world, the SPT is the most com-
RA (adapted from Honjo et al. 2010).
mon method of soil investigation and often is the
only available source of information for pile design.
Also, the major specifications pertaining to piles
reported by Phoon et al. (1995), Jones et al. (2002),
have also adopted pile bearing capacity estimation
and more recently by Phoon (2008). Geotechnical
formulas based on the N value obtained from the
variability is so great that sometimes none of
SPT. As such, the vertical bearing capacity of the
these references can be applied. Finally, some RV
pile is evaluated by an SPT based empirical method
in geotechnical problems, vary continuously over
from the Specifications of Highway Bridges of
a space or time. These variables are referred as
Japan (JRA 2001, Honjo et al. 2002, Teixeira et al.
random fields, and autocorrelation between vari-
2012). This method was selected because it has
ables in space or time can and should be consid-
statistical information about the model error asso-
ered (Vanmarcke 1977). The spatial variability
ciated with the bearing capacity values predicted
is normally ignored because of the difficulties in
(Okahara et al. 1991).
practical application. But when introduced, spatial
Considering the result of the static load test,
correlation of variables can reduce the variances.
and also considering that the permanent and vari-
It could be said that it is a conservative choice, but
able loads were equivalent in magnitude, E (total
it is technically incorrect to perform probabilistic
Gk + Qk) for the RA-based safety evaluation it was
analyses without considering spatial correlation
assumed to be between 400 and 1400 kN.
(Lacasse & Nadim 1996, Kulhawy & Phoon 1996,
Honjo et al. 2007). In Figure 3 are depicted the
uncertainties that will be taken into account in this 3.3 Performance function
paper, for pile foundations. The basic formula for the performance function
presented in Equation 1, is transformed into
Equation 2 when using an empirical method
3 APPLICATION EXAMPLE for the evaluation of the pile vertical bearing
capacity.
3.1 Case study
The case study pertains to an experimental site in M (R + R )
tip
ti
ip side
d (G + Q )
the north of Portugal (Viana da Fonseca & Santos
2008). Residual soil from granite is found at this = ( ×Q
t tip f Fside) (
d − G × Gk Q )
Qk (2)
site. The site is characterised geologically by an
upper layer of heterogeneous residual (saprolitic) where M is the safety margin, Rtip the tip resist-
granite soil of varying thickness, overlying a rela- ance of the pile, Rside the side resistance of the
tively weathered granite in contact with high-grade pile, G is the permanent action, Q is the variable
metamorphic rocks. Bedrock is found at a depth action, δ are the factors to take into account the
of approximately 20 m, and the ground water line uncertainties (δt for model error uncertainty on
is found at a depth of approximately 10 m. An tip resistance, δf for model error uncertainty in
extensive in situ and laboratory investigation was side resistance, δG for permanent actions uncer-
conducted, but because the SPT is one of the most tainties and δQ for variable actions uncertainties),
commonly used in situ tests for geotechnical design Qtip is the predicted tip resistance, Fside is the pre-
and soil characterisation, SPT results were used for dicted side resistance, Gk the permanent charac-
the calculations performed. The pile considered is teristic actions and Qk the variable characteristic
a reinforced concrete bored pile (id: E9) that is actions.

466

ISGSR2013.indb 466 10/18/2013 9:44:24 AM

 3.4 Uncertainties
The uncertainty sources considered taking into
account the Equation 2 were:
– the modelling uncertainty (or model error) in the
evaluation of resistance by an empirical method,
both side and tip components, δf and δt;
– the inherent soil variability (spatial and estima-
tion error) considered through the number of
blows (N) of SPT or other soil test, both side
and tip components, NSPT,side and NSPT,tip;
– and the physical uncertainties of actions (per-
manent and variable), δG and δQ.
Note that all uncertainties were considered as
independent and the values for each uncertainty
are presented in Table 2.
As referred before, uncertainties for model error
were gathered from Okahara et al. (1991), while
for actions, the studies made by JRA (2001) and
Holicky et al. (2007) were used. The SPT trend
was fitted to the SPT data at the site, presented in
Figure 4. The soil variability is therefore, taken into
account with this information (trend and residuals)
and by a normal distribution for the residuals. Figure 4. SPT data from application example.
A graphical method (probability plot, also known
as Q-Q plot) was used to assess this approxima-
tion. The reduction of SD of the soil uncertainty/
variability was based on Vanmarcke’s principle
of autocorrelation (Vanmarke 1977, Honjo &
Setiawan 2007, Honjo et al. 2011).

3.5 Results
Recall that the safety evaluation approach assumes
a fixed length and diameter of the pile while it cal-
culates the probabilities for different load values.
Safety evaluation approach was carried out for the
following load values E = Gk + Qk = [400; 600; 700;
750; 800; 850; 900; 1000; 1200; 1400] kN. FORM

Table 2. Uncertainties values for application example.

Uncertainties Mean SD PDF type

Model
– tip 1.12 0.706 Lognormal Figure 5. Reliability-based safety evaluation of vertical
– side 1.07 0.492 Lognormal bearing capacity of application example.
Soil
– NSPT,tip 10.26.1.91 z 4.6* Normal
– NSPT,side 4.6** Normal and MCS results are presented in Figure 5. MCS
Actions†
using n1 = 200,000 and n2 = 1,000,000 are presented
– permanent 1.0 0.10 Normal
to see that the number of simulations stabilizes
– variable 0.6 0.21 Gumbel
from n > 200,000. Figure 5 also depicts a light line
*Will be reduced taking into account the influence zone on marking the hypothetical/considered design load
the pile tip (3 × diameter) as averaging over the thickness. (total of 800 kN) and the load test result (1350 kN).
**Will be reduced taking into account the length of the Moreover, Figure 6 presents the obtained values
pile as averaging over the thickness. †Considered Gk = Qk. for α factors (FORM).

467

ISGSR2013.indb 467 10/18/2013 9:44:25 AM

 ACKNOWLEDGEMENTS

The authors wish to thank to “Fundação para

a Ciência e a Tecnologia” (FCT) for the finan-
cial support under the strategic project PEst-OE/
ECI/UI4047/2011 and the doctoral grant SFRH/
BD/45689/2008. Also, authors wish to thank the
contributions of Professor Yusuke Honjo and
Doctor Kieu Le Thuy Chung for their contribu-
tions and ideas.

REFERENCES
Figure 6. FORM sensitivity factors for the application
example.
Baecher, G.B., & Christian, J.T. 2003. Reliability and
statistics in geotechnical engineering. Chichester, UK:
John Wiley & Sons.
CEN. 2002a. European Committee for Standardization.
It is possible to conclude that:
Eurocode 0: Basis of Structural Design. EN 1990.
– for MCS with n1 and n2 the results are CEN. 2002b. European Committee for Standardization.
consistent; Eurocode 1: Actions on Structures. EN 1991.
– FORM results give very acceptable approxima- CEN. 2007. European Committee for Standardization.
Eurocode 7: Geotechnical Design. EN 1997.
tions to MCS results;
Cornell, A.C. 1969. A probability-based structural code.
– the results also show a clear relationship (expo- Journal American Concrete Institute 66(12): 974–985.
nential type) between the reliability and the Hasofer, A.M. & Lind, N.C. 1974. Exact and invariant
load; second-moment code format. Journal of the Engineer-
– if one considers the hypothetical action (800 kN) ing Mechanics Division, ASCE 100(EM1): 111–121.
it is possible that it does not meet the require- Henriques, A.A., Calheiros, F. & Figueiras, J.A. 1999.
ments, that usually recommend a reliability Probabilistic modelling of nonlinear behaviour of
index between 2.5 and 4.0; concrete structures. In Schueller, & Kafka (Ed.), Proc.
– and it is also possible to see that the uncertainty ESREL, ‘99 (pp. 495–500) Munich, Germany: A. A.
Balkema.
that has more influence in the reliability is the
Holicky, M., Markova, J. & Gulvanessian, H. 2007. Code
model error, being the other uncertainties con- calibration allowing for reliability differentiation and
siderably less important. production quality. Proc.10th International Confer-
ence on Applications of Statistics in Civil Engineering,
Tokyo, Japan. Taylor & Francis.
Honjo, Y. & Setiawan, B. 2007. On conditional estima-
4 CONCLUSIONS tion accuracy of spatial average of soil proprieties
and sample size. Proc.10th International Conference on
This paper presented a simple methodology for a Applications of Statistics in Civil Engineering, Tokyo,
reliability-based safety evaluation of an axial pile Japan. Taylor & Francis.
foundation, based on FORM and MCS. These are Honjo, Y., Suzuki, M., Shirato, M. & Fukui, J. 2002.
believed to be easy methodologies and following Determination of partial factors for a vertically loaded
the trend of reliability-based assessments. They pile based on reliability analysis. Soils and Foundations
can help support the design of pile foundations 42(5): 91–109.
Honjo, Y., Jliati, M.N. & Ishino, J. 2007. Effects of spa-
and also try to eliminate the possible confusions
tial variability and statistical estimation error in pre-
and difficulties that traditional reliability method- diction of settlement of shallow foundation. Proc.10th
ologies used in structures can cause to geotechni- International Conference on Applications of Statistics
cal designers in practice. This paper also presented in Civil Engineering, Tokyo, Japan. Taylor & Francis.
an application example, for which calculations Honjo, Y., Hara, T. & Kieu Le, T.C. 2010. Level III
were repeated for different loads. One of the main reliability based design of examples set by ETC10.
results is the relationship between probability of Proc. 2nd International Workshop on the Evaluation of
failure and the load value and the high influence Eurocode 7, Pavia, Italy.
of the model error in these results. Furthermore, ISO. 1998. International Organization for
Standardization. ISO 2394: General principles on reli-
the actions uncertainties were demonstrated to
ability for structures.
have not an important role in the reliability. This Jones, A.L., Kramer, S.L. & Arduino, P. 2002.
type of reliability-based assessments allows a more Estimation of uncertainty in geotechnical properties
rational way to deal with uncertainties of a prob- for performance-based earthquake engineering. Pacific
lem, instead of just introducing safety factors into Earthquake Engineering Research Center, University
the calculation. of Washington, USA.

468

ISGSR2013.indb 468 10/18/2013 9:44:26 AM

 JRA. 2001. Japan Road Association. Specifications for Phoon, K.K. & Honjo, Y. 2005. Geotechnical reliability
Highway Bridges. analyses: towards development of some user-friendly
Kulhawy, F.H. & Mayne, P.W. 1990. Manual on estimat- tools. Proc. 16th ICSMGE: 2845–2848, Osaka, Japan.
ing soil properties for foundation design. Electric Power Phoon, K.K. 2008. Reliability-based design in geotech-
Research Institute, Cornell University, USA. nical engineering: computations and applications.
Kulhawy, F.H. & Phoon, K.K. 1996. Engineering judg- London and New York: Taylor & Francis.
ment in the evolution from deterministic to reliability- Teixeira, A., Honjo, Y., Gomes Correia, A. &
based foundation design. Uncertainty in the Geologic Henriques, A.A. 2012. Sensitivity analysis of verti-
Environment: From Theory to Practice, Geotechnical cally loaded pile reliability. Soils and Foundations
Special Publication, ASCE, 58: 29–48. 52(6): 1118–1129.
Lacasse, S. & Nadim, F. 1996. Uncertainties in charac- Teixeira, A. 2012. Reliability and cost models of axial
terising soil properties. Proc. UNCERTAINTY’96, pile foundations. PhD Thesis. Dep. Civil Engineering,
Madison, USA. ASCE. University of Minho, Portugal.
Manohar, C.S. & Gupta, S. 2005. Modeling and evalua- Uzielli, M., Lacasse, S., Nadim, F. & Phoon, K.K. 2007.
tion of structural reliability: current status and future Soil variability analysis for geotechnical practice.
directions. In K.S. Jagadish. & R.N (eds). Iyengar, Proc. 2nd Int. Workshop on Characterization and Eng.
Recent Advances in Structural Engineering: 90–187. Properties of Natural Soils. 3. Singapore: Taylor &
Indian Institute of Science, Bangalore, India. Francis.
Okahara, M., Takagi, S., Nakatani, S. & Kimura, Y. Vanmarcke, E.H. 1977. Probabilistic modeling of soil
1991. A study on bearing capacity of as single pile profiles. Journal of the Geotechnical Engineering
and design method of cylinder shaped foundations. Division, ASCE 103(GT11): 1227–1246.
Technical Memorandum of PWRL. Viana da Fonseca, A. & Santos, J.A. 2008. International
Phoon, K.K., Kulhawy, F.H. & Grigoriu, M.D. 1995. Prediction Event. Behaviour of CFA, Driven and Bored.
Reliability-based design of foundations for transmis- ISC’2 Experimental Site. FEUP, Porto and IST,
sion line structures. Electric Power Research Institute, Lisboa.
Cornell University, USA. Zhang, J., Zhang, L.M. & Tang, W.H. 2011. Kriging
Phoon, K.K. & Kulhawy, F.H. 1999a. Characterization numerical models for geotechnical reliability analysis.
of geotechnical variability. Canadian Geotechnical Soils and Foundations 51(6): 1169–1177.
Journal 36(4): 612–624.
Phoon, K.K. & Kulhawy, F.H. 1999b. Evaluation of geo-
technical property variability. Canadian Geotechnical
Journal 36(4), 625–639.

469

ISGSR2013.indb 469 10/18/2013 9:44:27 AM

This page intentionally left blank
 Geotechnical Safety and Risk IV – Zhang et al. (eds)
© 2014 Taylor & Francis Group, London, ISBN 978-1-138-00163-3

Reliability-based design for the serviceability state design

of an excavation with cross walls in clays

S.H. Wu & J. Ching

Department of Civil Engineering, National Taiwan University, Taipei, Taiwan

C.Y. Ou
Department of Construction Engineering, National Taiwan University of Science and Technology, Taipei, Taiwan

ABSTRACT: Cross walls are commonly used for constraining the excavation-induced deformations as
well as the subsequent building damages in Taiwan. Based on the real case histories, a regression model
that is capable of predicting maximum wall displacements for excavations with or without cross walls is
proposed. With the consideration of uncertainties in the proposed model, a simplified Reliability-Based
Design (RBD) method in the form of partial factors for the Serviceability Limit State (SLS) design of
excavations is developed. The partial factors are calibrated by rigorous reliability theory and are verified
to produce design outcomes that meet the target reliability index. The use of cross walls which is extremely
effective in achieving the target reliability level is highlighted.

1 INTRODUCTION is applicable to excavations with or without cross

walls. They concluded that the use of cross walls is
Large deformations induced by excavations would fairly effective in reducing wall displacements in a
cause serious damage to adjacent buildings, especially qualitative way, but the effectiveness of cross walls
for excavations in soft clay. In Taiwan, cross walls are was not verified in a quantitative way in that study,
known to be extremely effective for constraining the i.e. how the reliability index would increase after the
excavation-induced deformations in soft clay (Lin use of cross walls? To address this quantitative ques-
2010, Ou et al. 2011). A cross wall is a wall connect- tion, quantifications of uncertainties are required.
ing opposite diaphragm walls, as shown in Figure 1. For instance, the regression model developed in
Cross walls are constructed prior to excavation and Wu et al. (2013) is not 100% accurate, i.e. there are
behave as the lateral support with large axial stiffness. transformation and parameter uncertainties. These
An important feature of cross walls is that it devel- uncertainties must be taken into account for the
ops wall displacement behavior similar to the corner purpose of reliability-based design.
effect of diaphragm walls (Lin 2010, Ou et al. 2011). In this paper, the uncertainties associated with
Wu et al. (2013) developed a regression model this model will be quantified and a simplified
for predicting the maximum wall displacement that Reliability-Based Design (RBD) methodology
based on partial factors will be developed for the
SLS design in excavations—these partial factors
are calibrated by reliability theory rather than
specified by judgments and experiences. A design
example is illustrated for the application. The
effectiveness of cross walls in achieving the target
reliability level will be demonstrated in the design
example.

1.1 Review of the regression model

Based on the 11 cases with cross walls (all are
cast in-situ prior to excavation) and another
11 cases without cross walls in Taipei, Wu et al.
(2013) developed a regression model for predict-
Figure 1. Schematic diagram of cross walls. ing maximum wall displacement δm* (in mm) for

471

ISGSR2013.indb 471 10/18/2013 9:44:27 AM

 the A-A section in Figure 1 for braced excavations and B are treated as deterministic parameters, but
in soft clay: the values of su/σv′ and K are treated as random
variables. Furthermore, Equation 1 is not exact,
ln ⎡⎣δ m* ( mm ) ⎤⎦ ≈ a0 + a1B 2 ln
l (He ) and the random model error is characterized by an
error term ε in the following:
(
l s σ
+ a3ln ′
) + a ln (L )
4 a5 ln (K )
+ a6 ⎡⎣ l ( H e )⎤⎦ + a7 ⎡⎣ ln (L′′ ) ⎤⎦ + a8 ⎡⎣ ln (K ) ⎤⎦
2 2 2
ln ⎡⎣δ m* ( mm ) ⎤⎦ = a0 + a1B 2 ln
l (He )
+ a9 (s u
′
v ) ( ′) + a10 ln (L ) ln ( )
′
(1) l s σ
+ a3ln ( ) + a ln (L ) a ln (K)
′
4
′
5

+ a ⎡⎣ l ( H )⎤⎦ + a ⎡ ln (L )⎤ + a ⎡⎣ ln (K ) ⎤⎦
2 2 2
′
where L′, B, and He (in meters) are respectively 6 e ⎣ ⎦ 7 8

the spacing between the two cross walls, excava- + a ln ( s σ ) ln (L ) + a l (L ) l ( ) + ε

9 u
′
v
′
10
′
(3)
tion width, and excavation depth (see Fig. 1 for
the definition of L′ and B); su/σv′ is the normal- where ε is the model error. The regression model
ized undrained shear strength based on isotropic error ε is modeled as a zero-mean normal ran-
Consolidated Undrained Compression (CIUC) test; dom variable. The STD of ε is determined to be
K (in kN/m) is the system stiffness; coefficients are roughly 0.33 based on the 22 cases given in Wu
a0 = 11.1908, a1 = −0.0048, a2 = −0.0168, a3 = 1.5855, et al. (2013).
a4 = −0.5071, a5 = −1.1914, a6 = 0.2354, a7 = −0.0691, As mentioned earlier, in real application the
a8 = 0.0390, a9 = −0.8365, and a10 = 0.0196. Note that value of su/σv′ may not be exactly known due to
Equation 1 is applicable to cases with or without measurement and transformation errors. Only the
cross walls. For cases with cross walls (e.g. Fig. 1), if measured (or nominal) value of su/σv′, denoted by
the maximum wall displacement at the A-A section (su/σv′)m, may be known. The relation to the meas-
in Figure 1 is of concern, L′ is the spacing between ured value is assumed to be
the two cross walls, whereas for cases without cross
walls (i.e. the hatched cross walls in Figure 1 do not
exist), L′ is the spacing between the two diaphragm ( )l ⎡ su σ v′
ln su σ v′ = ln
⎣ ( ) ⎤+e
m⎦
1 (4)
walls. The system stiffness K can be modeled as
an in-series system consisting of the following
where the error e1 is modeled as a zero mean nor-
two stiffnesses: (a) the flexural stiffness of the dia-
mal distribution with Standard Deviation (STD)
phragm wall (k2) and (b) the stiffness of the parallel
in the range of 0.1–0.5 [the STD of e1 is roughly
sub-system consisting of the axial stiffness of the
the same as the Coefficient of Variation (COV)
cross walls (k1) and the axial stiffness of the lateral
of su/σv′; the range of 0.1–0.5 is consistent to the
supports (struts) (k3):
low to medium variability summarized in Phoon &
Kulhawy (2008)].
k2 × ( k1 + k3 ) The value of K is also treated as a random
K= (2)
k1 + k2 + k3 variable. It depends on the values of the more fun-
damental input parameters, such as the Young’s
For cases without cross walls, k1 should be cal- moduli of the concrete (Ec) for the diaphragm wall,
culated as the axial stiffness of the diaphragm walls cross walls, and concrete floor slabs as well as the
in dark and grey (see Fig. 1). The details for the Young’s modulus for the steel struts (Est) [see Wu
calculations of K can be found in Wu et al. (2013). et al. (2013) for the detailed formulas for K]. These
Based on the results in Wu et al. (2013), the concrete Young’s moduli (Ec) depend on the con-
use of cross walls was found to be fairly effective crete compressive strengths (fc):
in reducing wall displacements. Nonetheless, it is
not clear how the use of cross walls can effectively ln ( Ec ) lln ⎡⎣ 4700 fc ⎤⎦ + e2 (5)
increase the reliability index and how to design the
spacing L′ to meet the target reliability level. This
where e2 is the transformation error, modeled as a
is the main objective of this study.
zero mean normal random variable with STD in
the range of 0.1∼0.3. The transformation errors
of the Ec′s (namely, e2′s) for the diaphragm walls,
2 MONTE CARLO SAMPLES OF δM*
cross walls, and concrete floor slabs are considered
to be independent. Finally, the fc in the above equa-
The reliability-based design method adopted in
tion is related to the measured value (fc)m by
this paper requires the simulations of Monte Carlo
samples of δm*. As seen in Equation 1, δm* depends
on su/σv′, K, He, L′, and B. Among them, He, L′, ln ( fc ) l ⎡⎣( fc )m ⎤⎦ + e3
ln (6)

472

ISGSR2013.indb 472 10/18/2013 9:44:28 AM

 where e3 denotes the measurement error, modeled the technical details. The technical details of QVM
as zero mean normal random variables with STD are given in Ching & Phoon (2011, 2013).
in the range of 0.1∼0.2. Again, the measurement The main idea of QVM is to reduce stabilizing
errors of the fc’s (namely, e3’s) for the diaphragm random variables to their η quantiles (η is small)
walls, cross walls, and concrete floor slabs are and the increase destabilizing ones to their 1 − η
considered to be independent. The struts Young’s quantiles. For Equation 3, stabilizing random vari-
modulus (Est) for the steel struts is related to its ables include su/σv′ and K, and destabilizing ran-
measured value by dom variables include ε. Taking K as an example,
its η quantile, denoted by Kη, satisfies
ln ( Est ) l ( Est )m + e4
ln (7)
(
P K < Kη = η ) (8)
where e4 denotes the measurement error, modeled
as a zero mean normal random variable with STD
For example, if K is a normally distrib-
in the range of 0.1∼0.3.
uted random variable, the 5% quantile,
In summary, the system stiffness K can be simu-
K0.05 = μK(1 – 1.645δK), where μK and δK are the
lated as a function of {(fc)m, (Est)m, e2, e3, e4} and the
mean and COV of K, respectively. This definition
su/σv′ can be simulated as a function of {(su/σv′)m,
is appropriate for stabilizing factors, because the
e1}. One also needs to know the dimension param-
design value is typically a conservative value less
eters {B, L′, He} and needs to simulate a sample of
than the mean. For destabilizing factor such as ε,
ε. A Monte Carlo simulation (MCS) sample of δm*
the (1 − η) quantile is applied:
can be then computed by Equation 3. Repeating
the above steps to yield N samples of δm*, and the
failure probability (Pf) of exceeding a prescribed P ε( ε1 η ) 1− η or (
P ε > ε1−η
1 )
=η (9)
limiting displacement δL can be estimated to be
Nf/N, where Nf is the number of samples that sat- The probability of ε larger than ε1−η is 100η%.
isfy δm* > δL, and the reliability index β = −Φ−1(Pf), If ε is normally distributed, the (100 − 5)% = 95%
where Φ−1(.) is the inverse cumulative density func- quantile of ε is ε0.95 = με (1 + 1.645δε), where με
tion of standard normal distribution. Note that and δε are the mean and COV of ε, respectively,
the above calculation of Pf or β can be conducted which is a conservative value larger than the mean.
at any excavation stage—one needs to replace (He, Therefore, the RBD in the form of QVM can be
K) by the depth and stiffness of the current excava- achieved by first calculating a conservative design
tion stage. value for ln(δm*). This can be done by inserting the
η quantiles of su/σv′ and K and the 1 − η quantile
3 SIMPLIFIED RELIABILITY-BASED of ε into the right hand side of Equation 3:
DESIGN
ln ⎡⎣δ m,
*
d ( mm ) ⎦ = a0 + a1B
⎤ 2 ln
l (He )
Although cross walls are effective in increasing
reliability level, it is not clear at this point how to + a3ln (⎣ η⎥
⎦
)
l ⎡ s σ ′ ⎤ + a4 ln L′ + a5 ln ( ) ( )
design a real case (e.g. select the cross wall spac-
+ a6 ⎡⎣ l ( )⎤⎦ + a7 ⎡⎣ ln ( ) ⎤⎦
2 2
ing L′) in order to meet the target reliability level
because L′ is highly related the amount of wall dis-
( ) (
⎤ + a9 ln ⎡ su σ v′ ⎤ ln L′ ) ( )
2
+ a8 ⎡⎣ l ⎦
placement (Wu et al. 2013). This is the main pur- ⎣⎢ ⎥
η⎦
pose of this section.
In this study, a simplified RBD design equation ( ) ( )+ε
+ a10 l L l ′
η 1 η (10)
in the form of partial factors will be developed.
The First Order Reliability Method (FORM) where δm,d* is the design value for δm*. If the design
(Ang & Tang 1984) was commonly used for cali- value δm,d*, is less than the limiting value δL, the
brating partial factors in major design codes (e.g. design is satisfactory, and vice versa. Apparently,
AASHTO codes & Eurocode), however, based a design dimension that satisfies δm,d* ≤ δL is a con-
on the recent study given by Ching & Phoon (2011), servative design, because the random variables
they indicated that the partial factors calibrated by su/σv′, K, and ε are taken conservative quantile val-
the quantile approach is more sensible and robust ues in order to determine δm,d*.
than FORM. This approach is formally called the Equation 10 may not yet be convenient for
Quantile Value Method (QVM) in Ching & Phoon practical engineers because it requires the knowl-
(2013). Therefore, QVM is used in this study for cal- edge for the definition of quantiles. Practical
ibrating the partial factors. What follows will only engineers are familiar with the deterministic RBD
present the basic idea of QVM without mentioning design equation in the form of partial factors.

473

ISGSR2013.indb 473 10/18/2013 9:44:30 AM

 Equation 10 can be readily converted to the form
of partial factors:

δ m,d ( mm )
*

⎛ a0 + a1B + a2 l ( H e ) ⎞
⎜
( )
l ⎡ γ su × su σ v ⎤ + a4 ln L ⎟
⎜ + a3ln ⎣
′
m⎦
′
( )
⎟
⎜ ⎟
⎜ + 5l ( ) + a6 ⎡⎣ l ( )⎤⎦ ⎟
2

= exp ⎜ 2⎟ × γ
( )
2
⎜ + a7 ⎡⎣ l L ⎤⎦ + a8 ⎡⎣ l ( )⎤⎦ ⎟ δ
′

⎜ ⎟
⎜ ⎣ ( m
)
l ⎡ γ su × su σ v′ ⎤ ln L′
⎜ + 9 ln
⎦ ( ) ⎟
⎟
⎜⎝ + 10 ln ( )
l L ln (
′
) ⎟⎠
(11)

where γ′s are the partial factors: γδ = exp[ε1−η]

is for the model uncertainty; γsu = (su/σv′)η/ Figure 2. Relations between η and βT for BU and TD
(su/σv′)m = exp[(e1)η]; γK = Kη/Km, where Km is the methods.
nominal K evaluated based on {(Est)m, (fc)m}.
Because ε and e1 are zero-mean normal random
variables with STDs equal to σε and σ1, γδ and γsu
have the following analytical forms:

(
γ δ = exp ε −η ) = exp ⎡⎣σ ε × Φ 1 (1 − η)⎤⎦ (12)
γ su = exp ⎡(e1 )η ⎤ = p ⎡⎣σ1 × Φ −1 ( η) ⎤⎦
⎣ ⎦

However, γK does not have the analytical form, as K

is not a random variable of well-known type, hence
Kη does not have an analytical form. However,
Kη can be obtained by MCS with a large sample
size—n samples of K can be obtained from MCS,
and Kη ≈ the 100η% sample quantile of K, which
is simply the n × 100η% ordered statistics (sorted
from small to large K samples) of the n samples of
K. Once Kη is obtained, γK is simply Kη/Km.
Based on the theory developed in Ching &
Phoon (2011), the η − βT relations can be calibrated.
Figure 2 shows the calibrated η − βT relation. It is
found that the calibrated η − βT relation is fairly Figure 3. Relations between target reliability index (βT)
unique with respect to various construction and partial factors (γi).
methods (bottom-up or top-down methods), vari-
ous design dimensions (B ranging from 15 m to
85 m; L′ ranging from 10 m to 136 m; He rang- βT − γ′s relations. Note that these calibrated rela-
ing from 1.5 m to 35 m), various typical values tions do not depend on the choice of δL. It is note-
of {(fc)m, (Est)m, (su/σv′)m}, and various choices of worthy mentioned that five sets of βT − γsu relations
STDs of {e1, e2, e3, e4}. The only exception is when are calibrated (see Fig. 3) for five different STD’s of
7.5 m < L′ < 10 m, the calibrated η − βT relation e1 (0.1, 0.2, …, 0.5). The practical engineers should
somewhat changes (see Fig. 2). The lower bound choose the appropriate βT − γsu relation based on the
7.5 m is a possible practical bound for L′ for the STD of e1. The practical engineers can check the
BU construction method. required partial factors {γδ, γsu, γK} from this figure
Given the prescribed βT, the required η can and implement these factors in Equation 11 to
be checked from Figure 2, and Equation 12 can obtain δm,d*. If δm,d* ≤ δL, the design should roughly
be subsequently used to derive the required par- satisfy β ≥ βT. Verification (not shown herein) indi-
tial factors γδ and γsu, while γK can be determined cates that these partial factors can indeed achieve
by MCS. Figure 3 shows the resulting calibrated the target βT fairly accurately—the actual reliability

474

ISGSR2013.indb 474 10/18/2013 9:44:32 AM

 indices of the designs based on these partial factors of the cross walls should not be computed based on
are reasonably close to the target value. the initial height of the cross walls (Hw – 3.0 = 28 m)
but should be computed based on the height at
the final stage the height (Hw – final He = 14.4 m)
4 DESIGN EXAMPLE because the cross wall portion above the excavation
depth will be dismantled.
A design example is taken to illustrate the use of It is desirable to determine the spacing of the
the partial factor charts (Fig. 3) and the effective- cross walls (L′) to fulfill δm,d* ≤ δL = 60 mm at
ness of cross walls. Suppose a rectangular excava- the final excavation stage. The introduction of the
tion area in Figure 1 (B = 35 m, the total length of cross walls will reduce L′ from 80 m to a smaller
the excavation area = 80 m) to be constructed by value. This reduction in L′ will affect k1, k2, and
the bottom-up method with the following param- k3 in Equation 2 and consequently K. Moreover,
eters: final He = 16.6 m, and (su/σv′)m = 0.325 with it will also change the L′ value input to Equation
σ1 = 0.2 (COV = 0.2). The total depth and thick- 11. The parameters {He, B, (su/σv′)m} will not
ness of the diaphragm wall (Hw) are 31.0 m and change, and the required partial factors {γδ, γsu,
0.8 m, respectively. There are five levels of steel γK} also remain the same. As a result, the process
struts with 4 m horizontal spacing. The nominal of determining L′ is iterative. It is found that the
compressive strength of concrete (fc)m = 27.5 MPa. spacing L′ corresponding to δm,d* = δL = 60 mm
The nominal Young’s modulus of the steel struts is about 20 m. This indicates that the spacing L′
(Est)m = 2.1 × 108 kPa. The design target is to ensure is required to be less than 20 m in order to meet
the maximum wall displacement at the A-A section the design target of δL = 60 mm with βT = 1.5. A
is less than δL = 60 mm at the final excavation stage simple MCS shows that the actual reliability index
with target reliability index βT = 1.5. for this design spacing is 1.47, fairly close to the
Let us first consider the design without cross target value of 1.5. The simplest way to fulfill this
walls (the hatched walls in Fig. 1 are absent). In this L′ ≤ 20 m requirement is to divide the site into four
case, B = 35 m and L′ = 80 m. Recall that the input subzones with same geometry, i.e. L′ can be taken
parameters for Equation 11 are {He, B, L′, (su/σv′) to be 80/4 = 20 m.
m, Km, γδ, γsu, γK}. The nominal K (Km) is computed
to be 2.44 × 104 kPa based on {(fc)m, (Est)m} and
4.1 Effectiveness of cross walls for RBD
problem dimensions (thickness of the walls, cross
sections of the struts, total depth of the diaphragm The aforementioned design example only illus-
wall, etc.). Because there is no cross wall, k1 in trated a case where δL = 60 mm and βT = 1.5. In
Equation 2 is the axial stiffness of the diaphragm this section, sensitivity analysis is taken to dem-
walls in dark and grey. The partial factors {γδ, γsu, onstrate the RBD for other choices of δL and βT.
γK} can be checked from Figure 3 for βT = 1.5—they
are 1.415, 0.812 and 0.805, respectively.
Once all input parameters {He, B, L′, (su/σv′)m,
Km, γδ, γsu, γK} are obtained, the design maximum
wall displacement (δm,d*) at the A-A section is com-
puted to be 196.6 mm by the right hand side of
Equation 11. It should be noted that the δm,d* cal-
culated above is a conservative value, since (su/σv′)m
and Km are multiplied by its corresponding reduc-
tion factors, i.e. γsu = 0.812, and γK = 0.805, and
the resulting δm* is multiplied by an amplification
factor, i.e. γδ = 1.415, that accounts for the model
error. Without such conservatism (γsu = γK = γδ = 1),
δm,d* is 84 mm. The design value δm,d* = 196.6 mm is
greater than δL = 60 mm, indicating that the serv-
iceability is not satisfied for the requirement of
βT = 1.5 at the final excavation stage. As a result,
cross walls are needed.
To reduce the wall displacement at the A-A
section, cross walls (the hatched wall in Fig. 1)
are introduced. Assume that the cross walls are to
be constructed from 3 m below ground surface to
the bottom of the diaphragm wall. Note that when
computing k1 in Equation 2, the cross section area Figure 4. Design example illustration.

475

ISGSR2013.indb 475 10/18/2013 9:44:34 AM

 Figure 4(a) shows how the required L′ changes by either the Bottom-Up (BU) or Top-Down (TD)
with respect to δL values and βT requirement for the method using steel struts and concrete floor slabs
A-A section, while Figure 4(b) shows the relation as lateral supports.
between L′ and K. Figure 4(c) is developed based Finally, a design example is used to illustrate
on Figure 4(a), (b). The red dashed line in Figure 4 for the use of the partial factor charts developed
corresponds to the design outcome demonstrated in this study and to show the effectiveness of cross
previously for βT = 1.5 and δL = 60 mm. It is clear walls. The design outcomes show that reducing the
from Figure 4(a) that reducing L′ is fairly effective cross wall spacing L′ is very effective to the increase
for increasing βT. For instance, for a fixed δL = 60 of the reliability index.
mm, reducing L′ from 20 m to 15 m leads to an
increase of βT from 1.5 to 2.1 (Pf from 0.067 to
0.018). This is because reducing L′ is very effective REFERENCES
for increasing K, as shown in Figure 4(b). Sensi-
ble trends are observed in Figure 4(a): for a fixed AASHTO. 1997. LRFD highway bridge design
βT, smaller δL results in smaller L′ and larger K, specifications. AASHTO, Washington, D.C.
because the reduction in L′ leads to the reduction Ang, A.H.S. & Tang, W.H. 1984. Probability Concepts
of wall displacement. in Engineering Planning and Design, Vol. II—Decision,
Risk and Reliability. New York: Wiley.
CEN (European Committee for Standardization) 2001.
Eurocode 7 Part 1: Geotechnical Design: General
5 CONCLUSIONS Rules. CEN, Brussels.
Ching, J. & Phoon, K.K. 2011. A quantile-based
Based on the regression equation proposed by approach for calibrating reliability-based partial
Wu et al. (2013), this study further developed a factors. Structural Safety 33(4–5): 275–285.
simplified RBD method for the serviceability Ching, J. & Phoon, K.K. 2013. Quantile value method
limit state design of excavations. Easy-to-imple- versus design value method for reliability-based code
ment charts for the calibrated partial factors are calibration. Structural Safety (in press).
produced. These partial factors are rigorously Lin, Y.L. 2010. The Effect of Cross Walls on the Movements
of Excavations in Clay. Ph.D. Dissertation, National
calibrated by the Quantile Value Method (QVM)
Taiwan University of Science and Technology, Taipei,
recently developed by Ching & Phoon (2011, Taiwan.
2013). Strictly speaking, the proposed method Ou, C.Y., Hsieh, P.G. & Lin, Y.L. 2011. Performance of
should be only applicable to Taipei cases that are excavations with cross walls. Journal of Geotechnical
mostly dominated by soft clays, because the regres- and Geoenvironmental Engineering ASCE 137(1):
sion model was calibrated based on these cases. 94–104.
The Taipei soft clays are with Liquidity Index Phoon, K.K. & Kulhawy, F.H. 2008. Serviceability limit
(LI) = 0.5∼1.0, Plasticity Index (PI) = 7∼20, and su/ state reliability-based design. In Phoon, K.K. (ed.),
σv′ (CIUC) = 0.3∼0.37. Nonetheless, the proposed Reliability-based Design in Geotechnical Engineering:
Computations and Applications: 344–384. London:
method may be still applicable to non-Taipei cases
Taylor & Francis.
that are dominated by soft clays with properties Wu, S.H., Ching, J. & Ou, C.Y. 2013. Predicting wall
similar to the Taipei soft clays. Finally, the pro- displacements for excavations with cross walls in soft
posed method should only be applied to cases that clay. Journal of Geotechnical and Geoenvironmental
are retained by diaphragm walls and constructed Engineering ASCE 139(6): 914–927.

476

ISGSR2013.indb 476 10/18/2013 9:44:35 AM

 10 Risk assessment and management in geotechnical engineering
and infrastructural projects

ISGSR2013.indb 477 10/18/2013 9:44:35 AM

This page intentionally left blank
 Geotechnical Safety and Risk IV – Zhang et al. (eds)
© 2014 Taylor & Francis Group, London, ISBN 978-1-138-00163-3

Indicator Kriging for locating risk zones: An application to buildings

at risk in the Barcelonnette Basin

S.A. Arnaouti, S. Fotopoulou, K. Pitilakis & Th. Chatzigogos

Civil Engineering Department, Aristotle University of Thessaloniki, Thessaloniki, Greece

A. Puissant
Laboratoire Image, Ville, Environnement, CNRS and University of Strasbourg, Strasbourg, France

J.-P. Malet
Laboratoire Image, Ville, Environnement, CNRS and University of Strasbourg, Strasbourg, France
Institut de Physique du Globe de Strasbourg, CNRS and University of Strasbourg, Strasbourg, France

ABSTRACT: In this study, landslide risk to buildings is assessed through an index-based, semi-
quantitative approach by integrating information obtained from multiple sources: landslide inventory and
susceptibility maps; information concerning the exposed buildings and their vulnerability. Once the risk
value of the individual buildings have been determined, a methodology based on the geostatistical tech-
nique of Indicator Kriging (IK) is proposed in order to cluster the exposed buildings into zones of equal
risk. Each risk zone consists of the areas where the probability of exceedance of the risk zone value—that
is produced through geostatistical analysis—is greater than a given threshold. The final risk zone map is
produced by the synthesis of each risk zone. The proposed methodology is applied to the Barcelonnette
Basin, where areas of low, moderate and high landslide risk zones are determined. The risk zone map
produced objectively can provide useful guidelines for risk management design.

1 INTRODUCTION Two studies, those of Maquaire et al. (2004)

and Ferlisi & Pisciotta (2007), cluster the exposed
An efficient landslide risk assessment, which allows buildings to units of equal vulnerability. The first
integration of the hazard analysis, information on authors present a semi-automatic procedure using
the elements at risk and their vulnerability, is always Hierarchical Ascending Classification to aggregate
required for the sustainable development of moun- the buildings at risk based on their geometrical
tainous areas and the safety of the citizens. Index- and contextual characteristics. Their methodology
oriented, semi-quantitative approaches have been is compared with an expert’s vulnerability zona-
widely used at medium scales (1:50,000–1:10,000) tion and the two results are in accordance for the
for mapping landslide risk due to their flexibil- majority of the buildings. Also, Ferlisi & Pisciotta
ity and their ability to combine a vast amount of (2007) homogenized the buildings at the Lazio
diverse information into a simpler, more usable and Abruzzo regions in Central Italy according to
form (e.g., Puissant et al. 2006, Puissant et al. their use in order to reduce the time for data col-
2013). lection in a working scale of 1:25,000. Each homo-
Even though landslide hazard and susceptibil- geneous unit consists of buildings with specific
ity zoning is an everyday practice in landslide risk use (residential, commercial and industrial) whose
assessment, this is not the case for spatially non- distance between them does not exceed 100 m. In
continuous parameters, such as vulnerability or terms of loss index, Kappos et al. (2009) estimated
risk, whose spatial extent is constrained within the risk in urban blocks; buildings that belong to the
the area of the exposed elements. However, at the same urban block form a cluster and the loss index
same time, van Westen et al. (2005) points out the of this cluster is the average of the loss indexes of
need of integrated risk information for emergency each building weighted upon the built area.
management purposes, which is the main purpose In this context, the present work first provides an
of risk analysis in meso-scales (1:50,000–1:10,000). index-based, semi-quantitative approach to assess
In the same spirit, Puissant et al. (2006) note that landslide risk of individual buildings at medium
at these scales the final goal of risk analysis is the scales. Then, a statistical methodology is proposed
determination of the most sensitive areas. based on indicator kriging to cluster buildings in

479

ISGSR2013.indb 479 10/18/2013 9:44:35 AM

 units of equal risk. Especially in meso-scales, this distance and its determination is based on the n
risk clustering is the spatial information necessary known values of the parameter under study, here
to the local authorities to easily establish efficient the indicator function I. By physical intuition, γ
management strategies since it highlights the areas is expected to increase as h increases, since close
where the high-risk buildings are concentrated and measurements tend to show similar results while
thus immediate action at those areas should follow measurements that lie far apart exhibit greater
the landslide event. Further, the risk zonation can discrepancies.
trace the sensitive areas where a more detailed risk
analysis should be conducted. The proposed meth-
odology is implemented in the Barcelonnette Basin 3 METHODOLOGY
(South East France), an area highly susceptible to
landslide activity. 3.1 Landslide risk assessment of buildings
The process of risk estimation integrates the haz-
2 GEOSTATISTICAL FRAMEWORK ard analysis with the elements at risk and their
vulnerability in the form of the generic equation
Geostatistics is an interpolation method that allows (Varnes, 1984):
the estimation of a regionalized variable R(x) at an
unsampled location. The Indicator Kriging (IK) Risk = Hazard × Vulnerability × Elements at risk
approach presented by Journel (1983) applies the (5)
geostatistical principles to the indicator function
I(x). For a specific threshold Ro, the indicator This simple equation allows identifying sepa-
function is defined as: rately the principal factors contributing to landslide
risk. These include the probability of occurrence of
⎧1, if R ( x ) < Ro a hazardous landslide of a given intensity within
I ( x; Ro ) = ⎨ (1) a given period of time (i.e. hazard), the degree of
⎩0, if R ( x ) ≥ Ro loss to an element or set of elements exposed to a
landslide of given type and intensity (i.e. vulner-
By assuming that R(x) has second order station- ability) and the valued assets at risk (i.e. elements
ary indicators, the IK predictor I(xo, Ro) at location at risk).
xo, that is actually an estimate of the probability Within the framework of this study, a semi-
P(R(xo) < Ro), can be determined by following the quantitative procedure fully implemented in a GIS
ordinary kriging equations (Cressie, 1993): environment is suggested to assess landslide risk.
The study is focused herein on the risk assessment
 n of buildings but the procedure is quite flexible and
I ( xo ; Ro ) ∑ λi I (xi ; Ro ) (2) can be easily adjusted to different assets.
i =1 The proposed framework includes four main
steps that are shortly outlined below:
where I(xi; Ro) is the known value of the indicator
function at the xi location given the Ro threshold, n – The collection of the susceptibility and hazard
is the number of the I known values in the field and information based on the characteristics of the
λi is the weight of this value. landslide inventory.
According to the kriging theory, the estimator Î – The identification of the buildings at risk and
should be unbiased and of minimum square error. of the main factors influencing their vulner-
Based on those two criteria, the equations to esti- ability (e.g. construction material, number of
mate weights λi are respectively: floors, state of maintenance etc.). The value of
the exposed buildings that is associated to their
n function and thus to their importance to the
∑ λi = 1 (3) local community should also be specified.
i =1 – The vulnerability assessment of the build-
n ings by weighting the different contributing
∑ λi ( )
γ Ro xij + m (Ro ) γ Ro ( xio ) , j 1 n (4) factors using engineering judgment based on
the framework proposed by Papathoma et al.
i =1
(2007). A different score is assigned to each
where xij the distance between locations xi and xj, category of the given factors and the total vul-
m(Ro) a Lagrange multiplier and γRo(x) the vari- nerability value is calculated for each building
ogram that corresponds to the Ro threshold. by means of a Weighted Linear Combination
The variogram describes the difference between Method (e.g., Papathoma et al. 2007, Mousavi
measurements of distance h in respect to that et al. 2011).

480

ISGSR2013.indb 480 10/18/2013 9:44:35 AM

 – The risk estimation integrates information on of exceedance Q(x; Ro,k) regarding the Ro threshold
the hazard, the vulnerability and the value of can be produced according to the equation:
the element (see Eq. 5). Both hazard and vul-
nerability values are given in standardized terms. Q( x; Ro,k ) I ( x; Ro,k ) (6)
Landslide risk can be finally assessed either
qualitatively or semi-quantitatively expressed on To produce the kth risk zone from the prob-
a scale from 0 to 1.0. ability Q(x; Ro,k) map, a probability threshold Qb
The proposed procedure, although it involves is required—common for all risk zones—that will
some degree of subjectivity, is generally compatible define the boundary of the risk zone. Indeed, areas
with the availability and quality of data taking also of the probability exceedance map that exhibit
into account the scale of the analysis (1:10,000). probability Q(x; Ro,k) greater than Qb form the kth
In addition, this index-based semi-quantitative risk zone.
approach allows treating on a hierarchical basis the Finally, the risk zone map is produced by super-
level of risk supporting, in this way, effective risk position of each risk zone.
management and decision making processes.

4 IMPLEMENTATION TO THE
3.2 Indicator Kriging for locating risk zones BARCELONNETTE AREA
Indicator Kriging can be used to cluster the
exposed elements in groups of equal risk. The 4.1 The study area
whole procedure that can be totally implemented The test site (Fig. 1) is located on the north-facing
in a GIS environment is summarized in the fol- hillslope of Barcelonnette Basin extending over an
lowing steps. area of about 100 km2 and it is representative of
Once the risk value, either quantitative or quali- the climatic, lithological, geomorphological and
tative, of the buildings has been determined, the landcover conditions common to several regions
risk R(x) at location x (x∈D, D the study area) of the South French Alps. Barcelonnette Basin is
can be considered as a regionalized variable whose a highly landslide prone area characterized by a
value is equal to the building’s risk value inside the large variety of slope movements (e.g., Maquaire
building’s area and zero otherwise. et al. 2003, Malet et al. 2005, Thiery et al. 2007),
The number of the desired risk zones nk form-
ing the risk zone map has to be selected, along
with the risk value Ro,k of each zone. For the kth
risk zone, the indicator function I(x; Ro,k) and the
variogram γk that corresponds to its spatial varia-
tion is determined. Then, by applying the kriging
equations at different locations xo of the field, an
estimate of the indicator function I(xo; Ro,k) at the
whole study area can be calculated. At this point
it should be noted that in order to provide moni-
tonicity, a single variogram—that of the median
of thresholds—can be used for the whole analysis
(Cressie 1993) instead of estimating different var-
iograms for each risk zone. Also, the kriging esti-
mations equal the input data values. To produce
more smooth risk zones that can include exposed
buildings of different risk value this property of
kriging should be relaxed. Thus, a nugget effect
ne, i.e. an offset at the origin, fully attributed
to measurement errors should be added to the
variogram.
According to the geostatistical theory, the indi-
cator function I(xo; Ro,k) is an estimate of the prob-
ability of the R(x) not to exceed the threshold Ro at
that point. However, the risk zones should contain
all the areas where R(x) is equal or greater than Figure 1. (a) Typical landscape of the north-facing
Ro, so, it is the complementary event that is actu- hillslope of the Barcelonnette Basin and (b) shaded relief
ally of interest. Thus, a map with the probability map and distribution of landslides (Thiery et al. 2007).

481

ISGSR2013.indb 481 10/18/2013 9:44:36 AM

 some of which directly affect the local community Table 1. Factors contributing to building vulnerability
causing considerable physical damage and eco- to landslides, the score of each category and their rel-
nomic losses. evant weightings.

Factor Categories Score Weighting

4.2 Landslide risk assessment of buildings
Construction Masonry 0.6 4
The proposed framework for assessing land- material RC 0.3
slide risk is implemented to the buildings of the Steel 0.3
Barcelonnette Basin that have been repeatedly Wood 0.8
affected by different landslide hazards. Informa- Mixed 0.7
tion on the susceptibility classes for the active Number of 1 0.5 1
landslides type of the North-facing slope and floors 2–3 0.3
the corresponding landslide susceptibility map at ≥4 0.1
1:10,000 scale has been taken from previous pub- Outdatedness Good 0.1 4
lished work (Thiery et al. 2007)—see Figure 2. It Average 0.3
should be noted that information on the magni- Bad 0.6
tude-frequency relationships of the potentially Destroyed 0.9
damaging landslides of different types was not Age Before 1900 0.9 3
made available and thus, only a preliminary hazard 1900–1950 0.7
analysis based on its spatial component (i.e. sus- 1950–1970 0.5
ceptibility) was possible. In addition, a database of 1970–1990 0.3
the exposed buildings in Barcelonnette area includ- 1990–2000 0.1
ing their structural characteristics and their urban After 2000 0
function (Puissant et al. 2006) has been provided in
GIS format for the purpose of this study.
Table 1 presents the description of the various Table 2. Function and the corresponding
factors considered for the vulnerability assessment value of the building.
of the buildings in Barcelonnette Basin as well as
their relevant scores defined by expert knowledge Value of the
and judgment. Different weightings are assigned Function building (E)
to each of the factors on the basis of their relative
importance in the vulnerability assessment. Table 2 Residential 1
presents the various functions of the exposed Commercial 1.1
buildings and the corresponding values assigned Industrial and craft 1
Agricultural 1
for each function.
Leisure and sportive 1.1
Finally, Figure 2 illustrates the estimated risk of
Hotel 1.1
each individual building in the Barcelonnette Basin
Place of religious worship 1.2
in qualitative terms (low, moderate, high). It is seen
Private service 1.2
that the higher risk buildings are generally concen-
Public service 1.3
trated in highly susceptible areas. Education 1.3
Civil service 1.3
Urgency service 1.3

4.3 Indicator Kriging for characterising risk zones

Three risk zones (nk = 3), that of low, intermediate
and high risk, were selected for the present analy-
sis, in accordance with the qualitative terms used in
buildings’ risk assessment.
The indicator function changes sharply near the
buildings’ limits since non-null risk values at the
building’s area exist adjacent to null values outside
of it. To take account of the non-smoothness of
the I function, the exponential model was adopted
for the variogram (Kitanidis 1997). The variogram
Figure 2. Landslide risk to individual buildings in the parameters selected for the median risk threshold
Barcelonnette Basin. (moderate risk) were sill = 0.10 and range = 70 m.

482

ISGSR2013.indb 482 10/18/2013 9:44:38 AM

 Figure 3. Landslide risk zone map in Barcelonnette Basin.

To validate the selected parameters, a 10% of the occur. So, an emergency plan should include extra
input data was randomly selected and excluded caution for this area in case of a landslide event.
from the analysis. The selected exponential vari-
ogram was used with the remaining data to predict
the indicator function at the excluded locations. 5 CONCLUSIONS
The very small value of the mean absolute error
(0.03) verifies the efficiency of the adopted vari- The present paper provides an index-based,
ogram and, in general, the legitimacy of the use semi-quantitative approach for spatial analysis
of geostatistics for the determination of the risk of landslide risk at medium scales by combining
zones. Also, to allow for a more smooth risk zone, information on hazard analysis with the character-
a nugget effect ne equal to 50% of the sill parameter istics and function of the elements at risk as well as
is artificially added to the selected variogram. The their vulnerability. The proposed approach is quite
same variogram is used for all three risk zones. general in concept and depends primarily on data
To finally determine the risk zones, a value of Qb availability and the expertise of the user.
equal to 0.2 was selected. Once the risk value of each individual element
The risk zones are illustrated in Figure 3. No at risk (building in our case) is assessed, this study
high risk zone exists because the buildings that presents a geostatistical framework based on indi-
exhibit high risk are few and spatially scattered. cator kriging equations, which offers an automatic
Six moderate risk zones—the larger one at the procedure to objectively cluster the elements at risk
North-West side of the Barcelonnette Basin cov- in zones of equal risk. Although the procedure has
ers an area of about 445,000 m2—and multiple been applied to landslide hazard, it can be eas-
low risk zones are formed. The risk zone map ily extended to other hazards as well. The whole
highlights the prone areas on the basin. Indeed, in procedure can by fully implemented in an ArcGIS
the Barcelonnette Basin, the South-West moder- environment.
ate risk zone encloses a vast area where the most The resulting risk zone map offers an inte-
severe and densely-spaced damages are expected to grated risk representation necessary for efficient

483

ISGSR2013.indb 483 10/18/2013 9:44:39 AM

 emergency strategies since it highlights the most Maquaire, O., Thiery, Y., Malet, J.-P., Weber, C.,
sensitive areas where high risk elements are Puissant, A. & Wania, A. 2004. Current practices and
concentrated. assessment tools of landslide vulnerability in moun-
tainous basins—identification of exposed elements
with a semi-automatic procedure. In Lacerda, Ehrlich,
Fontoura and Sayão (eds), Landslides: Evaluation and
ACKNOWLEDGMENT Stabilization; Proc. of the IX intern. symp. on Land-
slides, Rio de Janeiro, 2004. Taylor & Francis Group,
A special acknowledgement is due to Mr. Bouchounas London.
and the MDS Marathon Data Systems Co. for their Mousavi, M.S., Omidvar, B., Ghazban, F. & Feyzi, R.
technical support on ArcGIS. 2011. Quantitative risk analysis for earthquake-
induced landslides - Emamzadeh Ali, Iran. Engineering
Geology 122(3–4): 191–203.
REFERENCES Papathoma, K., Neuhauser, B., Ratzinger, K.,
Wenzel, H. & Dominey-Howes, D. 2007. Elements at
risk as a framework for assessing the vulnerability of
Cressie, N.A.C. 1993. Statistics for spatial data. N.Y.:
communities to landslides. Natural Hazards and Earth
John Wiley and Sons, Inc.
System Sciences 7: 765–779.
Ferlisi, S. & Pisciotta, G. 2007. A preliminary study of
Puissant, A., Malet, J.-P. & Maquaire, O. 2006. Mapping
landslide induced property damages towards conse-
landslide consequences in mountain areas: a tenta-
quence analysis. In Schuster, Schuster, Turner (eds),
tive approach with a semi-quantitative procedure. In
First North American Landslide Conference; Proc., Vail
Weber and Gancarski (eds), Proc. intern. conf. on
(Colorado), 3–9 June 2007. AEG Publication n.23.
Spatial Analysis and Geomatics SAGEO, Strasbourg,
Kappos, A.J., Stylianidis, K.C., Sextos, A.G.,
2006.
Papanikolaou, V.K., Panagopoulos, G., Kouris, L. &
Puissant, A., Van Den Eeckhaut, M., Malet, J.-P. &
Goutzika, E. 2009. Seismic risk scenarios for the
Maquaire, 0, 2013. Landslide consequence analysis: a
building stock in Grevena. 16th Hellenic conf. on Con-
region-scale indicator-based methodology. Landslides
crete, Paphos, Cyprus, 2009. Paper no. 111108.
(in press).
Journel, A.G. 1983. Nonparametric estimation of spatial
Thiery, Y., Malet, J.P., Sterlacchini, S., Puissant, A. &
distribution. Mathematical Geology 15(3): 445–468.
Maquaire, O. 2007. Landslide susceptibility assess-
Kitanidis, P.K. 1997. Introduction to geostatistics: appli-
ment by bivariate methods at large scales: application
cations to hydrology. Cambridge: Press Syndicate of
to a complex mountainous environment. Geomorphol-
the University of Cambridge.
ogy 92: 38–59.
Malet, J.-P., Van Asch, T.W.J., van Beek, R. & Maquaire, O.
van Westen, C.J., van Asch, T.W.J. & Soeters, R. 2005.
2005. Forecasting the behaviour of complex land-
Landslide hazard and risk zonation—why is it still
slides with a spatially distributed hydrological model.
so difficult? Bulletin of Engineering Geology and the
Natural Hazards and Earth System Sciences 5: 71–85.
Environment 65(2): 167–184.
Maquaire, O., Malet, J.-P., Remaître, A., Locat, J.,
Varnes, D.J. 1984. Landslides hazard zonation: A review
Klotz, S. & Guillon, J. 2003. Instability conditions of
of principles and practice. UNESCO, Paris.
marly hillslopes: towards landsliding or gullying? The
case of the Barcelonnette Basin, South East France.
Engineering Geology 70: 109–130.

484

ISGSR2013.indb 484 10/18/2013 9:44:41 AM

 Geotechnical Safety and Risk IV – Zhang et al. (eds)
© 2014 Taylor & Francis Group, London, ISBN 978-1-138-00163-3

Development of a probabilistic model for the prediction of building

damage due to tunneling induced settlements

Carles Camós & Climent Molins

Universitat Politècnica de Catalunya, Barcelona, Spain

Olga Špačková & Daniel Straub

Engineering Risk Analysis Group, Technische Universität München, Munich, Germany

ABSTRACT: Tunnel construction can cause deformations of the surrounding ground, which endanger
buildings and other structures located in the vicinity of the tunnel. The prediction of these deformations
and the damages to the buildings is difficult due to limited knowledge of the geotechnical conditions and
uncertainty in the response of the structures to the settlements. This paper presents a probabilistic model
for prediction of the damage to buildings due to tunneling, which combines the volume loss method with
the equivalent beam model. It furthermore proposes a probability-based method for determination of the
limiting value of settlement that is used for control purposes during the tunnel construction. Updating of
the limiting settlement with measurements gathered during the construction is described. The proposed
methodology is applied to a masonry building affected by the construction of the L9 metro line tunnel
in Barcelona.

1 INTRODUCTION methods such as the equivalent beam method

(Burland & Wroth, 1974; Boscardin & Cording,
The prediction of damages to buildings caused by 1989), which is widely used in tunnel engineer-
underground constructions such as tunnels entails ing. This method determines the maximum tensile
uncertainty due to our limited knowledge of the strain in the building by modeling it as a linear
geotechnical conditions and the response of the elastic beam subjected to a given deflection ratio.
structures subjected to differential settlements. Pre- This strain value is then compared with limiting
diction of damages is important as a basis for the strain values, which define different categories of
design, the selection of the construction technol- damage according to the severity of affection. An
ogy and for setting allowable limits on settlements. iteration process is performed in order to assess the
These limiting values of settlement are then used in limiting value of settlement that leads to damages
the construction phase for control purposes: if the below an acceptable level.
measured settlement exceeds the limiting values, This paper proposes a computationally efficient
the construction is stopped or additional safety probabilistic model for the estimation of building
measures must be taken. damage due to tunneling, combining the volume
At present, settlement profiles and resulting loss method for approximation of the subsidence
damages in buildings are commonly modeled trough and the equivalent beam method for mod-
deterministically. Settlement profiles are typically eling the response of the building (Sec. 2). The
predicted by means of 2D Finite Element (FE) parameters of the volume loss method are usually
models combining the soil, the tunnel and the selected based on expert judgment. The uncertainty
foundations of a given building. Alternatively, the connected to the choice of these parameters is typ-
volume loss method (Peck, 1969; Attewell et al., ically high. The proposed methodology allows tak-
1986) can be used for approximation of the subsid- ing into account these uncertainties as well as the
ence trough. The volume loss method is an empiri- uncertainty in the building response.
cal approach to determining the settlement profile; The paper further proposes a novel methodol-
this empirical approach is computationally more ogy for the determination of the limiting settle-
efficient than 2D FE simulation and it is fully suf- ment value on a probabilistic basis (Sec. 3). The
ficient for many engineering applications. Once the limiting settlement is here defined as a settlement,
settlement profile is calculated, it is possible to pre- for which the probability of damage to the build-
dict damages in buildings by means of empirical ing is acceptably low. Two approaches for setting

485

ISGSR2013.indb 485 10/18/2013 9:44:41 AM

 this limiting value are proposed: (1) a simple mode of deformation. Therefore, extreme modes
approximate approach using a plot of the results of bending and shear are analyzed separately. The
of the probabilistic analysis, (2) an advanced extreme fiber strains in bending and shear are
approach based on reliability updating (Straub, given by the following equations:
2011). Additionally, a procedure for updating the
limiting values with observations gathered during ⎛ E⎞
the tunnel construction is described. ε br f2 VL K Eε br (ε b ε h ) Eε br (3)
⎝ G⎠
The proposed methodology is applied to a case
study of masonry buildings affected by the con- ⎛ E⎞
ε dr = f3 VL , K , ⎟ ⋅ Eε dr
struction of the L9 metro line in Barcelona. ⎝ G⎠
⎡ ε2 ⎛ E ⎞
2 ⎤ (4)
⎛ E ⎞
= ⎢ε h 1 − ⎟ + h ⎜ ⎟ + ε dm 2 ⎥
ax ⋅ Eε ddr
⎢ ⎝ 4G ⎠ 16 ⎝ G ⎠ ⎥
2 PROBABILISTIC MODEL OF BUILDING ⎣ ⎦
DAMAGE DUE TO TUNNELING
where E G is the ratio between the Young and
The shape of the settlement profile in a plane, shear moduli of the building material, which is
which is close to perpendicular to the tunnel axis, modeled as a RV, Eε br d Eε ddr represent the model
can be modeled by means of a Gaussian curve errors and εh is the horizontal strain at the base of
(Peck, 1969). The settlement at the distance y from the beam, which is obtained as the derivative of the
the tunnel axis then equals: horizontal displacements u:

⎛ y2 ⎞ s( y ) ⋅ y
s( y ) = Smax exp ⎜ − 2 ⎟ (1) u( y ) = (5)
⎝ 2i cos θ ⎠ z0
du( y )
where i is the location of the inflection point (hor- ε h ( y) = (6)
izontal distance from tunnel axis), θ is the angle dy
between the modeled plane and the perpendicular
plane and Smax is the maximum settlement in the The model errors Eε br d Eε ddr are considered
center of the Gaussian curve, i.e. above the tunnel as multiplicative RVs with mean value equal to 1.
axis. Smax can be calculated as: They result from the assumption of linear elastic-
ity, the position of the neutral axis and the omis-
sion of the presence of openings.
VL d 2
Smax f1(VL , K ) = (2) Maximum bending (εbmax) and shear (εdmax)
3.192 ⋅ K z0 strains in the equivalent beam are calculated as:

where d and z0 are the diameter and depth of the Δ

tunnel, respectively, VL is the expected volume L
ground loss (i.e., the ratio between the area of the ε bmax = (7)
⎛ L 3I E ⎞
settlement trough and the cross-section area of the +
⎝ 12t 2aLH G ⎠
tunnel), which is dependent on the tunneling tech-
nology, and K is a shape parameter of the curve Δ
which depends on the type of soil. The product K⋅ L
z0 determines the location of the inflection point i ε dmax = (8)
⎛ HL L2 G ⎞
of the Gaussian curve with respect to tunnel cen- ⎜1 + 18I E ⎟
terline. VL and K are modeled as random variables ⎝ ⎠
(RVs). The model error is considered as described
later in Eq. (12). where H is the beam height, I is the inertia per unit
Knowledge of the shape of the settlement trough length, t is the assumed position of the neutral axis
allows determining the deflection ratios Δ/L that and a is the location of the fiber where strains are
are affecting the building, where L is the distance calculated.
between two reference points and Δ is the relative The calculation of Eqs. (3)–(8) is performed
deflection between these two points. separately for the zone of the building undergoing
The response of the building is modeled using sagging deflection (upwards concavity) and for the
the equivalent beam method, which represents the zone undergoing hogging deflection (downwards
building by means of a weightless linear elastic rec- concavity). The errors of the equivalent beam model
sag sag hog hog
tangular beam. The aim is to calculate the tensile in sagging, Eε br Eε ddr , and hogging, Eε br Eε ddr ,
strains in the beam for a given deflected shape. The are assumed to be independent. In case of sagging
distribution of strains in the beam depends on the deflection, the neutral axis is assumed to be at middle

486

ISGSR2013.indb 486 10/18/2013 9:44:41 AM

 Table 1. Classification of damage (Burland et al., 1977).

Limiting tensile
Category of damage Normal degree of severity Typical damage strain (εlim)(%)

0 Negligible Hairline cracks less than 0.1 mm 0–0.050

1 Very slight Fine cracks up to 1 mm 0.050–0.075
2 Slight Cracks easily filled up to 5 mm 0.075–0.150
3 Moderate Cracks from 5 to 15 mm 0.0150–0.300
4 Severe Extensive repair work. Cracks from 15 > 0.300
to 25 mm
5 Very severe Partial or complete rebuilding. Cracks
> 25 mm.

height (t = H/2). In case of hogging deflection, the Reliability Method (FORM)) were used, separate
LSFs for ε ε br ε ε dr ε ε br , ε ε dr should be defined
sag sag hog hog
neutral axis is assumed to be at the top fiber (t = H).
Strains are calculated in the most critical fiber from and the failure event should be described as a series
the position of the neutral axis, so that a = t in both system.
cases. The damage on the buildings is determined
depending on the maximum strain εmax:
3 DETERMINATION OF LIMITING
SETTLEMENT
ε max = max ⎡⎣ε εsabrg ε εsag
dr
ε εhog
bbr
, ε εhog ⎤
dr ⎦
(9)
The measured maximal settlement above the crown
of the tunnel Sm equals:
where ε εsag
br
ε εhog
are the maximum bending strains
br
in sagging and hogging respectively,
p both obtained Sm Smax + E f Em = Smax EE (12)
using Eq. (3), and ε ε dr ε ε dr are the maximum shear
sag hog

strains in sagging and hogging respectively, both

obtained using Eq. (4). where Smax is the maximal settlement calculated
Based on εmax, one can estimate the size of the using Eq. (2), Ef is the model error representing
cracks in the building. The approach of Burland the deviation of the real settlement from the ide-
et al. (1977) is used in this paper for classification alized Gaussian shape described by Eqs. (1) and
of the damage magnitudes as shown in Table 1. (2), Em is the error of measurement on site, which
The different damage categories can be used for reflects imprecision of the instruments, human
the definition of system failure Fε lim . Failure occurs errors, effect of temperature changes, etc., and
if the maximum strain (εmax) obtained from Eq. (9) EE = Ef + Em.
exceeds a given limiting tensile strain value εlim for The goal is to find the limiting value of settle-
a target category of damage according to Table 1. ment slim from the following condition:
For example, if cracks with a width larger than
0.1 mm are considered inacceptable, the limiting ( ε llim | Sm = Slim ) pT (13)
strain defining the failure is εlim = 0.05%.
The Limit State Function (LSF) is then pT is the required (target) safety level. A measured
defined as settlement Sm > slim thus implies an unacceptably
high probability of failure Fεlim and would trigger
g ( ) = ε lilim − ε m (10) further actions.
max
In the following, the value of slim will be deter-
mined using two different approaches. In Sec. 3.1,
where X is the vector of variables that are consid-
an approximate approach based on engineering
ered to be random. The LSF determines the failure
judgment is utilized. In Sec. 3.2, the exact value
domain ΩF = {g(x) ≤ 0}. The probability of fail-
of slim will be determined using a reliability-based
ure then equals the probability of X taking a value
approach. Finally, Sec. 3.3 describes the updating
within the failure domain:
of the limiting settlement based on observations
gathered during the tunnel construction.
Pr( ε llim ) = Pr ( F ) (11)
3.1 Approximate approach
Note that this definition of LSF is suitable when
applying sampling methods for the computation An approximate estimate of the limiting settle-
of probabilities. If methods such as First-Order ment slim can be determined based on evaluation

487

ISGSR2013.indb 487 10/18/2013 9:44:45 AM

 of the probabilistic model described in Sec. 2 using 3.3 Updating with the measurements gathered
Monte Carlo (MC) simulation. For each sample during construction
of the input variables X, the settlement trough is
After the construction starts, N measurements of
evaluated according to Eqs. (1) and (2). Based on
the settlements s = (s1, s2, … sN)) are obtained along
the estimated settlement, the maximal tensile strain
the tunnel. These measurements can be used for
in the building is calculated using Eqs. (3) to (8).
updating the probabilistic model and the value of
For simplification, the error terms are disre-
the limiting settlement. The measurements are car-
garded and only VL, K and E/G are considered as
ried out in the same quasi homogeneous geotech-
RVs. Eq. (12) then reduces to Sm = Smax. The limit-
nical section of the tunnel, where also the analyzed
ing settlement is approximately determined from
building is located.
a scatter plot of the maximum tensile stains εmax
The uncertain geotechnical conditions in this
against the maximal settlement Smax as is shown
quasi-homogeneous section, characterized by
later in Figure 4. The value is determined visually
volume loss VL and shape parameter K, are now
from the plot using engineering judgment.
described as stationary spatial stochastic processes
with constant autocorrelation functions RK(l) = ρK
3.2 Reliability-based approach
and RVL(l) = ρV where l is the distance between
L
The conditional probability of Eq. (13) can be two locations within the section. In other words, K
determined by means of Bayesian updating tech- has the same marginal distribution at any location
niques with equality type information as proposed within the section and the values of K at any two
in Straub (2011) and applied to geotechnical safety locations are correlated with correlation coefficient
in Papaioannou and Straub (2012). With this ρK, independent of the distance between them. The
approach, all the model and measurement errors same holds for VL. This simple correlation model
are included. was selected based on a preliminary analysis of
First, the likelihood of VL and K for given meas- data from a constructed tunnel; its validity should
ured settlement sm is calculated: be tested in the future based on a more detailed
analysis. The new measurements at locations 1, …,
L(vL , k sm = sm | VL vL , K = k ) N can be expressed by separate likelihood func-
(14)
= fE sm f1 vL , k )) tions L1, …, LN following Eq. (14). For each likeli-
hood function Li, one can find the corresponding
where fE is the Probability Density Function (PDF) observation domain Ωi defined by means of a LSF
of the error EE in Eq. (12). Following Straub hi(vL,i, ki, ui) as described in Eq. (15). Here, vL,i and
(2011), this likelihood function can be expressed ki are the realizations of the random processes K
by a LSF: and VL at the location of measurement i.
To update the limiting value of settlement for
h(vL , k u ) = u Φ −1[cL(vL , k )] ≤ 0 (15) Sm conditional on the existing measurements s, the
failure probability conditional on Sm and on s is
computed (compare with Eq. (16)):
where u is the realization of a standard Normal
RV, Φ−1 is the inverse standard normal CDF and
c = σ EE ⋅ 2π is a scaling constant chosen to ensure Pr ( ε llim m m ∩ )
Pr ( ε llim | Sm = Sm , )=
that cL(vL, k) ≤ 1 for all vL, K. This LSF defines the Pr ( m m ∩ )
observation domain Ω0 = {h(x, u) ≤ 0} in a space Pr ([ x, u, u1, N] ∩ ΩO ∩ 1 ∩ ∩ ΩN )
= F
that contains the original RVs X = ( L , K E G ) Pr ([ x, u, u1, N] O ∩ Ω1 ∩… ∩ ΩN )
and the standard Normal variable U. With this
approach, the conditional probability of failure (17)
Fεlim for a given observed settlement sm is computed
as: Analogous to the procedure in Sec. 3.2, this con-
ditional probability is evaluated for different values
of sm. The updated limiting settlement value slim *
Pr ( ε llim m m)
Pr ( ε llim | Sm = Sm ) = ensuring Eq. (13) is found iteratively.
Pr ( m )
(16)
Pr ([ x, ] ∈ΩΩ0 ∩ ΩF )
=
Pr ([ x, ] ∈Ω
Ω0 ) 4 CASE STUDY

This probability can be evaluated using a MC The proposed method is applied to a case study
simulation for different values of sm. The limiting of the L9 metro line construction in Barcelona.
settlement value slim ensuring Eq. (13) is then found The damage produced by the tunnel construction
iteratively. to a complex of masonry buildings from the late

488

ISGSR2013.indb 488 10/18/2013 9:44:48 AM

 1920’s located in the Bon Pastor area is studied. the c.o.v. is supposed to be 0.4. VL is modeled by a
An equivalent beam analysis of the buildings was lognormal distribution.
already performed in Camós et al. (2012), show- A value equal to 2.5 is typically assumed for the
ing the validity of this model. The location of the ratio E G of masonry buildings. Uncertainty is
building and the tunnel is shown in Figure 1. also present in this parameter due to the variety
of orthotropic materials composing a building, yet
this uncertainty is relatively small. Therefore, it is
4.1 Model parameters
here modeled by a Beta distribution defined on the
The tunnel diameter d in the studied section is interval 2.4 to 2.6. The measurement error Em and
12 m, the depth of the tunnel is z0 = 23 m. The the model error Ef are represented with normal dis-
length of the building complex is L = 46 m, the tribution with zero mean and standard deviations
angle between the building wall and the plane per- 0.5 mm. The multiplicative model errors of the
pendicular to the tunnel axis is θ = 26º, the build- equivalent beam model Eεsag br
Eεsddrag Eεhog
bbr
, Eεhog
ddr
, are
ing height is H = 3 m and thus, the inertia per unit described by lognormal distributions with mean
length of the cross-section of the building is equal equal to 1.
to I = 2.25 m4/m. The parameter t equals 1.5 m in
the sagging zone and 3 m in the hogging zone and
4.2 Results of the probabilistic analysis
a = t for both zones.
The probabilistic model is summarized in The results of the MC simulation of the model
Table 2. The shape parameter of the settlement described in Sec. 2 are presented here. They show
profile K usually varies from 0.2 to 0.3 for granular the influence of the different uncertain parameters
soils to 0.4 to 0.5 for stiff clays to values as high as on the assessment of maximum strain εmax calcu-
0.7 for soft silty clays (Burland, 2008). The ground lated following Eq. (9) and on the associated dam-
in the analyzed tunnel section is formed by typi- age category as defined in Table 1.
cal alluvial soil with coarse sand, limes and a small
quantity of gravel. K is likely to be in the interval
from 0.2 to 0.4, which is thus assumed to be a 90% Table 2. Random parameters of the model.
confidence interval. The mean is assumed to be
0.3 and coefficient of variation (c.o.v.) is assumed Parameter [units] Distribution Mean St.dev.
equal to 0.2. K is non-negative and the lognor- K [-] Lognormal 0.3 0.06
mal distribution is thus an appropriate model for (−1.22, 0.20)
this RV. VL [%] Lognormal 0.4 0.16
Experience from tunneling constructions in (−0.99, 0.39)
similar conditions (TYPSA, 2003) shows that Beta (2,2, [2.4, 2.5 0.045
E
the expected interval of volume loss VL is in the [-] 2.6])
range 0.1% to 0.6%. Nevertheless, the uncertainty G
on these values is high due to many unpredictable Em, Ef [mm] Normal 0.0 0.50
factors that influence ground losses (unexpected (0.0, 0.5)
geological units, technical problems of the TBM, Lognormal 1.0 0.05
Eεsag Eεsddrag Eεhog , Eεhog [-] (0.0, 0.05)
human errors, etc.). The interval of 0.1–0.6% is br bbr ddr

thus assumed to be a 90% confidence interval and

Figure 1. Location of buildings and tunnel track.

489

ISGSR2013.indb 489 10/18/2013 9:44:51 AM

 Figure 2 shows the influence of the volume loss 4.4 Exact reliability-based determination of the
VL on the maximum strain εmax. A positive linear limit value of settlement
correlation is observed; higher values of volume
The reliability based approach shown in Sec. 3.2 is
loss are likely to lead to more severe damages on
used to find the limiting settlement slim that satisfies
the building. Figure 3 displays the influence of
Eq. (13) for p = 0.05 and εlim = 0.05% (the failure
the shape parameter K on εmax. Higher values of
event is defined in accordance with the previous
K produce flatter settlement troughs, which cause
Sec. 4.3.). Figure 5 displays the conditional prob-
smaller tensile strains in the building and thus
ability of failure for different values of measured
lead to milder damages. The relationship is clearly
settlement sm from 20 to 30 mm (denoted as prior
nonlinear.
estimate). The limiting settlement is determined as
The a-priori probability of the building damage
slim = 23 mm.
being in category 0, which corresponds to negligi-
ble damages, is 0.6. The probability of only aes-
thetical damages, corresponding to categories 0–2, 4.5 Results of updating with observations from
is 0.95. monitoring instruments
The prior estimate of the limiting settlement
4.3 Approximate determination of the limiting described is now updated with the measure-
settlement value ments gathered during the construction process,
following the procedure described in Sec. 3.3.
The limiting settlement slim is determined using the
Two measurement of the settlement in the
approximative approach described in Sec. 3.1. Only
same quasi-homogeneous section are utilized:
a negligible damage (category 0) is acceptable, as is
s1 = 14 mm, s2 = 19 mm. Correlation coefficients
usual in tunneling construction. More severe dam-
of the underlying normal distributions of shape
ages to buildings are considered as a failure, there-
parameter and volume loss are estimated by expert
fore the limiting tensile strain is set to εlim = 0.05%.
judgment: The shape of the settlement trough
Figure 4 shows the scatter plot of settlement Smax
(described by parameter K) is dependent on the
and maximum strain εmax obtained from the MC
geotechnical conditions. A high correlation is there-
simulation. An approximate value of the limiting
fore assumed within a geologically homogeneous
settlement is determined slim = 22 mm.
section and ρK = 0.7. On the contrary, the volume

Figure 2. Scatter plot of volume loss VL and max. Figure 4. Scatter plot of settlement Smax and max.
strain εmax. strain εmax.

Figure 3. Scatter plot of parameter K and max. strain Figure 5. Conditional probability of failure for different
εmax. values of measured settlement, Pr ( ε 0. % | sm ).
llim

490

ISGSR2013.indb 490 10/18/2013 9:44:53 AM

 loss VL is strongly influenced by the construction these parameters were thus not taken into account.
process and it is typically highly variable within Finally, the value of the limiting settlement was
one homogeneous section. It is therefore assumed updated with observations gathered during the
to be uncorrelated and ρVL = 0 . The updated con- construction. The updated limiting settlement
ditional probabilities of failure for different values is 27 mm and is thus higher that the prior value
of the settlement measured at the vicinity of the determined during the design phase. The increase
building, sm, are depicted in Figure 5. The updated of the the limit is possible thanks to the reduc-
*
value of limiting settlement is slim = 27 mm . tion of uncertainty after including the additional
measurements.

5 CONCLUDING REMARKS
REFERENCES
The paper presented a computationally efficient
model for probabilistic prediction of building dam- Attewell, P.B., Yeates, J. and Selby, A.R. (1986) Soil move-
age due to tunnelling that is applicable in engineer- ments induced by tunneling and their effects on pipelines
ing practice (Sec. 2). Further, a novel method for and structures. Blackie Academic & Professional.
determining the limiting settlement was presented Boscardin, M.D. and Cording, E.J. (1989) Building
response to excavation-induced settlement. Jour of
(Sec. 3), which is a more systematic and traceable Geo. Eng. 115(1–21).
reliability-based approach with an explicit ration- Burland, J.B. (2008) The assessment of the risk of dam-
ale than the deterministic methodology typically age to buildings due to tunnelling and excavations. J.T.
used in practice. Additionally, the reliability-based Payma Cotas: Mov. de edificios inducidos por excava-
approach allows to incorporate measurements made ciones. p. 3.
during the construction. The proposed procedure Burland, J.B., Broms, B. and De Mello, V.F.B. (1977)
was demonstrated on a case study of a tunnel con- Behaviour of foundations and structures. Proc. 9th
struction in Barcelona (Sec. 4). First, the influence Int. Conf. on Soil Mech. and Found. Eng., 2, 495–546.
of the uncertainty in the model parameters (volume Burland, J.B. and Wroth, C.P. (1974) Settlement of build-
ings and associated damage. London: Pentech Press.
loss, shape parameter of the settlement through, Camós, C., Molins, C. and Arnau, O. (2012) A case
Young and shear moduli of the building material) study of damage on masonry buildings produced by
on the estimated damage was presented. Second, tunneling induced settlements. Int. Jour. of Architec-
the value of the limiting settlement was determined tural Heritage.
with the approximate approach as 22 mm. Third, Papaioannou, I. and Straub, D. (2012) Reliability updat-
the limiting settlement was determined more pre- ing in geotechnical engineering including spatial vari-
cisely using an advance realiability-based approach ability of soil. Computers and Geotechnics 42: 44–51.
as 23 mm. Both approaches provide similar values Peck, R.B. (1969) Deep excavations and tunneling in soft
and the simpler method appears to be satisfactory ground. SOA Report, 7th Int. Conf. SM&FE.
Straub, D. (2011) Reliability updating with equality infor-
for practical applications. Both of these values are mation. Probabilistic Engineering Mech. 26: 254–258.
more strict than the value that was used in the real TYPSA. (2003) Análisis de los movimientos del terreno
case, where a settlement of up to 24 mm was con- producidos por la excavación mecánica del túnel de
sidered to be safe. The reason for this difference is la L9 en la zona de Santa Coloma de Gramanet. Pro-
the fact that in the real case, the uncertainties in the jecte de construcció de la Línia 9 de metro de Barcelona.
ground parameters and building parameters were Tram 4t Bon Pastor—Can Zam. Infraestructura i Esta-
not considered and some unfavorable values of cions. Annex 8.

491

ISGSR2013.indb 491 10/18/2013 9:44:55 AM

This page intentionally left blank
 Geotechnical Safety and Risk IV – Zhang et al. (eds)
© 2014 Taylor & Francis Group, London, ISBN 978-1-138-00163-3

A three-level framework for multi-risk assessment

Z.Q. Liu & F. Nadim

Norwegian Geotechnical Institute/International Centre for Geohazards, Oslo, Norway

ABSTRACT: Many regions of the world are exposed to and affected by several natural hazards.
Implementation of effective risk management strategies in these areas requires that all relevant threats
are assessed and considered. Compared to single-risk analysis, the examination of multiple risks poses a
range of additional challenges due to the different characteristics of hazards. This paper summarizes pre-
vious research on multi-risk assessment and proposes a new three-level framework for multi-risk assess-
ment that could account for the possible interactions among the threats. The framework is developed as
part of the EU FP7 Collaborative Research Project MATRIX. The first level is a simple flow chart that
guides the user in whether a multi-hazard, multi-risk approach is required for the problem at hand. The
second level is a simplified, semi-quantitative approach to explore if a detailed assessment is needed. The
third level is a detailed quantitative multi-risk analysis based on Bayesian networks. The key components
of this framework, such as assessment of cascading hazards, time-dependent vulnerability estimation,
and the choice of the required level of sophistication are addressed in the paper. The multi-risk assess-
ment procedure outlined in the paper integrates the results of risk posed by each threat, cascade effect,
and appropriate consideration of uncertainties, to provide a rational estimate of multiple risks. Simple
examples that demonstrate the application of the method are presented in the paper.

1 INTRODUCTION optimum alternative among those options avail-

able without doing a detailed, rigorous multi-risk
Many regions of the world are exposed to and analysis. Therefore, the framework recommended
affected by several types of natural hazard. The in this study is based on a multi-level approach
assessment and mitigation of the risk posed by where the decision-maker and/or the risk analyst
multiple natural and man-made threats at a given will not need to use a more sophisticated model
location requires a multi-risk analysis approach than what is required for the problem at hand, or
that could account for the possible interac- what would be reasonable to use given the available
tions among the threats, including possible cas- information.
cade events. Performing quantitative multi-risk
analysis using the methodologies available today
presents many challenges (e.g., Kappes et al. 2012, 2 THE RECOMMENDED THREE-LEVEL
Marzocchi et al. 2012). The risks associated with FRAMEWORK FOR MULTI-RISK
different types of natural hazards such as volcanic ASSESSMENT
eruptions, landslides, floods, and earthquakes are
often estimated using different procedures and the The recommended multi-risk assessment frame-
produced results are not comparable. Furthermore, work is a multi-level process which assumes that
the events themselves could be highly correlated the end-user (decision-maker or risk analyst)
(e.g., floods and debris flows could be triggered has identified the relevant threats and has car-
by an extreme storm event), or one type of threat ried out an assessment of the risk(s) (at the
could be the result of another (e.g., a massive land- level of sophistication required for the problem
slide that is triggered by an earthquake, so-called at hand) associated with each single hazard(s).
cascade effect). Figure 1 shows the general steps of the multi-risk
It is obvious that a mathematically rigorous assessment framework. The overall multi-risk
approach to multi-risk assessment that addresses assessment process comprises the following
all the challenges named above, as well as the uncer- stages: (1) risk assessment for single hazards,
tainties in all steps of the analysis, will be compli- (2) Level 1: qualitative multi-risk analysis,
cated and require resources and expertise. On the (3) Level 2: semi-quantitative multi-risk
other hand, in many situations the decision-maker analysis, and (4) Level 3: quantitative multi-risk
in charge of risk management can identify the analysis. The details are described below.

493

ISGSR2013.indb 493 10/18/2013 9:44:56 AM

 3 LEVEL 1 ANALYSIS

The Level 1 analysis comprises a flow chart type

list of questions that guides the end-user as to
whether or not a multi-type assessment approach,
which explicitly accounts for cascading hazards
and dynamic vulnerability within the context of
conjoint or successive hazards, is required. Each
question will be supplied with an exhaustive list
of answers that the user should choose from. This
process is shown schematically in Figure 3.
The flow chart will include, for example, these
Figure 1. Schematic view of the steps followed in the questions:
proposed multi-risk assessment framework.
• What is the purpose of the risk assessment
exercise? (answers: identifying the most criti-
cal risk scenarios and choosing the optimal risk
mitigation measures, assessing the adequacy of
resources and level of preparedness for post-
event response, etc.).
• Which natural threats are relevant for your area
of interest? (answers: earthquake, landslide,
volcanic eruption, tsunamis, wildfire, winter
storm, storm surge and coastal flood, fluvial
flood, snow avalanche, other perils, etc.).
• (If the user has chosen only one natural hazard
Figure 2. Stages of risk assessment for single hazard. from the list) How likely is it that the dominant
natural threat could happen more than once dur-
ing the time window of concern with an inten-
sity that will cause significant loss? (answers:
In the first step, it is assumed that the risk
very likely, likely, unlikely, very unlikely, virtually
assessment for the single hazard(s) roughly follows
impossible).
the classical approach that is comprised of the fol-
lowing stages (Fig. 2): Note: At this stage, if the user has chosen only
one natural hazard from the list and chooses very
• Definition of space/time assessment window
unlikely or virtually impossible as the answer to
(target area, time window) and the risk metric
the above question, then there is no need to go any
quantifying the expected losses.
further and a more detailed multi-risk assessment
• Threat(s) identification (e.g., earthquake, vol-
is irrelevant.
cano, landslide, etc.).
• Single hazard assessment (e.g., rate of occur-
rence, pathway, intensity measure, etc.).
• Assessment of the vulnerability of the elements
at risk (e.g., people, buildings, etc.).
• Assessment of the consequences in terms of the
chosen metric (e.g., loss of life, economic losses,
environmental degradation, etc.).
Once the results of the single-hazard risk
assessment(s) are available, the user embarks on a
three-level process, which becomes more detailed
and rigorous as the user moves from one level to
the next. The user moves to a higher level analy-
sis only if the problem at hand requires a more
accurate risk estimate and, equally important,
if the data needed for doing the more detailed
analysis are available. The selection of which of
these three levels is to be used depends on the
outcome of the preliminary risk assessment for Figure 3. The steps involved in the Level 1 multi-risk
single hazard(s). analysis.

494

ISGSR2013.indb 494 10/18/2013 9:44:56 AM

 • Cascading events: Could a hazard trigger results strongly suggest that a multi-type assess-
another hazard in your list (example: an earth- ment is required, then the end user moves on to
quake triggering a landslide, landslide debris Level 2 to make a first-pass assessment of the
blocking a river and causing flooding when the effects of dynamic hazard and time-dependent
landslide dam breaks, earthquake causing col- vulnerability (see Fig. 4). If cascading events are
lapse of flood defence structures and leading to potentially a concern, the user goes directly to the
flood, etc.)? (answers: yes or no). Level 3 analysis.
• Conjoint events: Could several hazards in your
list occur simultaneously because they are
4 LEVEL 2 ANALYSIS
caused by the same external factors (example:
earthquakes and volcanic eruptions are both
In Level 2 analysis, the interactions among hazards
caused by tectonic processes, winter storms and
and dynamic vulnerability are assessed approxi-
storm surges, fluvial floods and debris flows
mately using semi-quantitative methods. The
caused by extreme precipitation events, ..)?
steps involved in the Level 2 analysis are shown in
(answers: yes or no).
Figure 4.
Note: The user will be provided with some To consider hazard interactions and time-
guidance and examples for questions such as the dependent vulnerability, the suggested method in
previous two, which may not be straightforward. multi-risk Level 2 analyses is a matrix approach
based on system theory. This kind of matrix has
• Dynamic vulnerability: Could the occurrence of
been used in various fields, including environmen-
one of the hazards in your list significantly influ-
tal issues (Simeoni et al. 1999, de Pippo et al. 2008),
ence the vulnerability of some of the elements at
rock engineering (Hudson 1992) and natural hazard
risk to another event of the same type or to other
assessment (Kappes et al. 2010). The assumption
hazards (example: a building partially damaged
of this approach consists of the comprehension
by an earthquake has a higher vulnerability to
and description of the relationships among agents
the next earthquake or to floods and landslides,
and processes in the evolution of system.
ash fall from a volcanic eruption on roof tops
Figure 4 shows an example to explain this
will increase the mass and hence may increase
approach. Firstly, a matrix is developed by means
the seismic vulnerability of the building, etc.) ?
of the choice of a couple of hazards, considered
(answers: yes or no).
as the basic components of the system (Fig. 5a). It
• Dynamic hazard: Could the occurrence of one
will be followed by a clockwise scheme of interac-
of the hazards in your list significantly influence
tion (Fig. 5b), with the description of the mutual
the occurrence probability of other hazards
influence between different hazards (Fig. 5c). More
(example: a strong earthquake could weaken
specifically, each element of the row, which crosses
the soil in a slope and increase the probability
one of the hazards in the mean diagonal, shows the
of landslide during extreme precipitation events,
influence of this hazard on the system, thus indi-
etc.)? (answers: yes or no).
cating the cause of the phenomena; whereas each
Additional questions may, of course, be added, element of the column, which crosses the same
depending upon the situation at hand. If the Level 1 hazard analysed, shows the influence of the system

Figure 5. Matrix approach for the identification of the

Figure 4. The steps involved in the Level 2 multi-risk interactions between hazards in Level 2 analysis (Modified
analysis. after de Simeoni et al. 1999 and Kappes et al. 2010).

495

ISGSR2013.indb 495 10/18/2013 9:44:56 AM

 on this hazard, thus focusing on the effect of the the threshold hazard interaction index calculated
phenomena. After the descriptions contained in by Equation 1 is HI = 18 (50% of 36), while the
the matrix, they are assigned numerical codes vary- total value of causes and effects is 16, hence we do
ing between 0 (No interaction) and 3 (Strong inter- not need to do Level 3 analysis.
action) with intervals of 1, as a function of their
degree of the interaction intensity (Fig. 5d, 5e).
5 LEVEL 3 ANALYSIS
Once all the hazards in the matrix are filled, it is
possible to verify the degree of the impact of each
In Level 3 analysis, the interactions among hazards
hazard on the others and the effect from other
and dynamic vulnerability are assessed quantita-
hazards. In order to avoid the excessive weighting
tively with as high accuracy as the available data
of a single hazard, the sum of the codes for the row
allow.
and the column is considered. Table 1 shows the
A new quantitative multi-risk assessment
coding result for each hazard.
model based on Bayesian networks (BaNMuR)
It can be seen that slides are the dominant haz-
is introduced to both estimate the probability
ard caused by other hazards because they have the
of a triggering/cascade effect and to model the
maximum number of causes. On the other hand,
time-dependent vulnerability of a system exposed
river floods are the hazards most sensitive to the
to multi-hazard. The flexible structure and the
influence of other hazards, with the maximum
unique modelling techniques offered by Bayesian
number of effects.
networks make it possible to analyze cascade
In the scoring system above, the maximum pos-
effects through a probabilistic framework.
sible value of each off-diagonal cell in Figure 5a
Furthermore, the interactions between hazards
is 3. Therefore the maximum possible value for
and the uncertainties involved may be captured
the total sum of each row is 3 ⋅ (n − 1), where n is
using a Bayesian network. The uncertainties
the number of hazards. Likewise, the maximum
in each hazard/vulnerability and their inter-
possible value for the total sum of each column is
relationships are represented with probabilities.
3 ⋅ (n − 1). This means that the maximum possible
The prior (conditional) probabilities can be
value for the total sum of causes and effects is:
updated with information of specific cases by
Bayes’ theorem. Therefore, the uncertainties
HI,max = 2 ⋅ 3 ⋅ n ⋅ (n − 1) = 6 ⋅ n ⋅ (n − 1) (1) would become smaller and the updated multi-risk
where n is the number of hazards and HI is the haz- results would become more reliable based on the
ard interaction index. new information. In particular, this methodology
Therefore, the maximum possible value for the is well suited for treating uncertainties associated
hazard interaction index is HI,max = 6 ⋅ 3 ⋅ 2 = 36 for with hidden geodynamic variables, which are not
the example considered in Table 1. Given the uncer- directly observable from the Earth’s surface (e.g.,
tainties and possible excessive or moderate weight- model uncertainty in causal relationships between
ing of single hazards, a threshold hazard interaction unobservable volcanic processes and surface man-
index HI equal to 50% of HI,max is recommended for ifestations or monitoring data).
considering a more detailed Level 3 analysis. If the The probabilities of hazardous events are
hazard interaction index is less than this threshold, updated on the basis of any new information
Level 3 analysis is not recommended because the gathered. This framework shows how the updat-
additional accuracy gained by the detailed analyses ing of probabilities due to the interaction of haz-
is most likely within the uncertainty bounds of the ards helps to update the vulnerability and total risk
simplified multi-risk estimates. Otherwise, Level 3 quantitatively and how mitigation measures influ-
analysis is recommended. In the example above, ence the multi-risk consequences.
A conceptual Bayesian network multi-risk model
may be built as shown in Figure 6. To determine
the whole risk from several threats, the network
Table 1. Coding of each hazard in the system. takes into account possible hazards and vulner-
ability interactions. This would include the events:
Causes Effects Causes +
Number Hazard (rows) (columns) effects 1. Independent but threatening the same elements
at risk with or without chronological coinci-
1 Slides 4 1 5 dence (the column marked in deep orange color
2 Debris 2 3 5 in Fig. 6);
flows
2. Dependent on one another or caused by the
3 River 2 4 6
floods
same triggering event or hazard; this is mainly
Total 8 8 16 the case of ‘cascading events’ (the column
marked in green color in Fig. 6).

496

ISGSR2013.indb 496 10/18/2013 9:44:57 AM

 losses and should be estimated with an accept-
able degree of credibility in order to determine the
potential losses that are dependent upon the per-
formance of elements at risk subjected to various
hazard excitations. Fragility curves represent the
cumulative distribution of damage, which specify
the continuous probability that the indicated dam-
age-state has been reached or exceeded, and could
provide graphical information on the distribution
of damage.
The Limit State (LS) probability for a structure
exposed to a single hazard can be expressed in
terms of discrete random variables as follows:
∞
Pf ∑ P ⎡⎣LS I ⎦P [ I = i ]
i ⎤⎤P
i =0
∞
= ∑ P ⎡⎣D > C I = i ⎦⎤⎤PP [I i ] (2)
Figure 6. Bayesian network for quantitative multi-risk x =0
assessment.
where I is the intensity measure of the hazard
and LS (limit state) is the condition in which
the load demand D due to the hazard is greater
than the capacity C. The conditional probability
P[ LS | I = i ] is the probability of reaching LS at a
given hazard intensity level, I = i. The term P[I = i]
is the marginal hazard probability. For continuous
random variables, Eq. (2) can be expressed as
i =∞
Pf ∫i = 0 Fr (i )gI (i ) di
d (3)

where Fr ( i ) is the fragility function in the form of

a cumulative distribution function and gI ( i ) the
hazard function in the form of a probability den-
sity function.
Figure 7. Possible scenarios of multi-hazard interac- In the case of a structure subjected to a multi-
tion as considered in the MATRIX project. hazard situation involving additive load effects
(e.g., earthquake + landslide), the convolution con-
cept must be expanded. This multi-hazard form is
This network consists of two main sub-networks calculated as
for (1) multi-hazard and (2) time dependent vul-
nerability, as detailed in the following sections. ∞ ∞ ∞
Pf ∑ ∑ ∑ P ⎡⎣LS I1 i1 ∩ I 2 i2 ∩ ... ∩ I n in ⎤⎦
5.1 Multi-hazard analyses i1 =0 i2 =0 in =0

A number of possible scenarios of single haz- × P [ I1 = i1 ∩ 2 2 ∩ ∩ In in ]

ards and cascade events have been identified
for the MATRIX (EU FP7 Project: New Multi- (4)
HAzard and MulTi-RIsK Assessment MethodS Eq. (3) can also be expressed in terms of con-
for Europe) case studies (Garcia-Aristizabal et al. tinuous random variables as
2012). A Bayesian network may therefore be built
as shown in Figure 7 to describe the interactions
i1 =∞ i2 =∞ in =∞ Fr (ii1 i , in )gI1 (i ) gI 2 (i )
between hazards. It is obvious that one hazardous Pf = ∫ ∫i = 0 …∫i = 0 …ggI (in ) dii1dii2 … diin
event could trigger other hazardous events. i1 = 0 2 n (5)
n

5.2 Time-dependent vulnerability assessment

An example of the seismic fragilities, includ-
Predicting the damage of elements at risk (e.g., ing the combination of probabilistic debris flow
buildings) is critical for the evaluation of economic load (additional seismic weight), are presented

497

ISGSR2013.indb 497 10/18/2013 9:44:57 AM

 ACKNOWLEDGEMENT

The research leading to these results has received

funding from the European Community’s Seventh
Framework Programme [FP7/2007-2013] under
Grant Agreement n 265138 New Multi-Hazard
and Multi-Risk Assessment Methods for Europe
(MATRIX).

REFERENCES

de Pippo, T., Donadio, C., Pennetta, M., Petrosino, C.,

Terlizzi, F. & Valente, A. 2008. Costal hazard assess-
ment and mapping in Northern Campania, Italy.
Geomorphology 97: 451–466.
Figure 8. Fragility surface for a scenario involving a Garcia-Aristizabal, A. & Marzocchi, W. (with: Woo G.,
seismic event and debris flow for a low rise, low code RC Reveillere A., Douglas J., Le Cozannet G., Rego F.,
building. Colaco C., Fleming K., Pittore M., Tyagunov S.,
Vorogushyn S., Nadim F., Vangelsten B.V., and ter
Horst W.) 2012. Review of existing procedures for
multi-hazard assessment, Deliverable D3.1. New
over three-dimensions, where the x-axis is the peak
methodologies for multi-hazard and multi-risk assess-
ground acceleration, the y-axis the deposition ment methods for Europe (MATRIX), contract No.
height, and the z-axis the fragility (see Fig. 8). 265138.
Hudson, J.A. 1992. Rock engineering system. Ellis
Horwood Ltd., Chichester.
6 CONSLUSIONS Kappes, M.S., Keiler, M., Glade, T. 2010. From single-
to multi-hazard risk analyses: a concept address-
Quantification of all the natural and anthropo- ing emerging challenges. In Malet, J.-P., Glade, T. &
genic risks that can affect an area of interest is a Casagli, N. (Eds.), Mountain Risks: Bringing Science
to Society. Proceedings of the International Conference,
basic factor for the development of a sustainable
Florence. CERG Editions, Strasbourg: 351–356.
environment, land-use planning, and risk miti- Kappes, M.S., Keiler, M., von Elverfeld, K. & Glade, T.
gation strategies. In this study, we put forward a 2012. Challenges of analysing multi-hazard risk:
consistent framework for multi-risk assessment. a review, Natural Hazards 64(2): 1925–1938.
The developed procedure consists of three levels: Marzocchi, W., Garcia-Aristizabal, A., Gasparini, P.,
(1) Level 1: qualitative analysis, (2) Level 2: semi- Mastellone, M.L. & Di Ruocco, A. 2012. Basic prin-
quantitative analysis, and (3) Level 3: quantitative ciples of multi-risk assessment: a case study in Italy.
analysis. In this way, multi-risk assessment can be Natural Hazards 62(2): 551–573.
performed step by step. At the same time, the inter- Simeoni, U., Calderoni, G., Tessari, U., Mazzini, E. 1999.
A new application of system theory to foredunes
actions among different threats are considered in a
intervention strategies. Journal of Coastal Research
systematic structure, and the uncertainties in the 15(2): 457–470.
different stages can be accounted for in the recom-
mended framework.

498

ISGSR2013.indb 498 10/18/2013 9:44:59 AM

 Geotechnical Safety and Risk IV – Zhang et al. (eds)
© 2014 Taylor & Francis Group, London, ISBN 978-1-138-00163-3

Rockfall risk management based on survey data of real slopes

S. Moriguchi
Tohoku University, Sendai, Miyagi Prefecture, Japan

Y. Otake, M. Iwata, Y. Honjo, A. Takagi, F. Kurauchi, T. Hara, K. Sawada & A. Yashima

Gifu University, Gifu, Gifu Prefecture, Japan

N. Asano
Chubu University, Kasugai, Aichi Prefecture, Japan

ABSTRACT: This paper presents a framework of risk management of rockfall using real survey data.
Hida area located in Gifu prefecture in Japan is selected as a target area. In the evaluation process of
rockfall provability, relative probability is calculated based on results of the survey data using the logistic
regression analysis. Then, absolute probability is calculated using history data of occurrence of rockfall
in the target area. In the evaluation process of the economic loss, road closure induced by rockfall is con-
sidered. Direct and indirect losses are taken into consideration. A trial calculation with different strate-
gies were conducted based on calculated risk, and then it was summarized the obtained results are useful
information for rockfall risk management.

1 GENERAL INTRODUCTION 2 SETTING CONDITIONS

Rockfall is one of the serious slope disasters. Once 2.1 Target area
rockfall take place, it has potential to cause dam-
Hida area, northern part of Gifu prefecture in
age on infrastructures and loss of human lives.
Japan (Fig. 1), is selected as a target area. In the
Japanese governments, researchers, and engineers
area, mountain areas are widely distributed, and
have been making effort to minimize the damages.
There are however still huge numbers of danger-
ous slopes that has high risk of rockfall in Japan.
Furthermore it is becoming difficult to construct
new countermeasures due to the budget cut of
public works, and maintenance cost of existing
countermeasures is becoming larger year by
year. It is therefore quite important to develop a
method which realizes a strategic decision mak-
ing of the maintenance and new construction of
countermeasures.
In order to overcome the problem mentioned
above, this study presents a framework of risk
management of rockfall. Risk of rockfall is cal-
culated based on real survey data. In this paper,
first, setting conditions, such as definition of
risk, study area, and assumptions are explained,
then calculation procedures of risk of rockfall
are explained. This paper finally shows a result
of a trial calculation is shown. Based on the
result, effectiveness of the proposed framework
is discussed. Figure 1. Study area.

499

ISGSR2013.indb 499 10/18/2013 9:45:00 AM

 there are a lot of slopes that has high risk of the occurrence of rockfall with the occurrence of
rockfall. The reason why the area is selected is road closure. Therefore, the probability of occur-
that detailed survey data of slopes and accident rence of rockfall can be treated as the probability
data of rockfall has been accumulated by local of the road closure.
government. The probability is calculated based on a survey
data. An intensive slope survey was carried out in
1996. A total of 3023 slopes were investigated by
2.2 Definition of risk and risk event
skillful engineers in target area. The slopes were
Risk is defined as the product of the economic loss separated into two failure types, rockfall and rock
and the probability of risk event as follows, failure. In this study, rock failure is treated as a
kind of rockfall. The slopes are also classified into
R P∑ D (1) three levels, namely Measures Required (MR),
Observation (OB) and No Measures (NM). In
where R is risk, D is the economic loss and P is the addition, information of each slope, such as geo-
probability of risk event. Road closure induced by logical characteristics and geometric configura-
rockfall is defined as the risk event. tion, were accumulated.
The flow of the evaluation process of rockfall
probability is shown in Figure 2. Based on the
2.3 Assumptions survey data, relative rockfall probability is firstly
Following assumptions are used in this study. calculated using the logistic regression analysis.
Then the absolute probability is calculated by
a. Risk event occur independently. Thus, con- calibrating the relative probability using the
dimental occurrence of risk event is not accident data of rockfall. The accident data has
considered. also been accumulated in target area. Number
b. Effects of disturbances such as earthquakes of slopes and rockfall accidents are summa-
and rainfalls are not considered directly. This rized in Table 1. As shown in the table, the tar-
means these effects are not considered directly get area includes three regions, Gero, Takayama
in calculations of the risk. However, because and Furukawa. Different models are built for
the effects should be included in rockfall acci- each region in consideration of the rationality.
dent data, a part of the effects is reflected Histograms of annual absolute rockfall probabil-
indirectly. ity are shown in Figures 3–5. It is found the prob-
c. Degradation of countermeasures is not taken ability of Furukawa region is higher than other
into consideration. Therefore, once a counter- two regions.
measure constructed, it is assumed that risk of
slope is zero for eternity.
d. Effect of size of rockfall is not considered. It is
therefore assumed that the road closure always
occurs when rockfall takes place.
There are two reasons why the assumptions are
used in this study. First reason is shortage of survey
data. Large amount of survey data are accumu-
lated, but it is difficult to build a database because
of partial lack of data. Second reason is to simplify
the calculation conditions. Although it is better to
consider real condition, in this case, it becomes dif-
ficult to understand calculation results. Because
the objective of this study is to build a frame-
work of risk management of rockfall, we tried to
build the framework under the assumptions. It is
however important to remove the assumptions in
future study.

3 PROBABILITY OF RISK EVENT

As mentioned in subsection 2.2, the road closure

induced by rockfall is defined as risk event. In addi- Figure 2. A flow of evaluation process of rockfall
tion, as mentioned in subsection 2.3, we identify probability.

500

ISGSR2013.indb 500 10/18/2013 9:45:00 AM

 Table 1. Number of slopes and rockfall accidents.

Numbers of slopes Number of rockfall accidents

Region Evaluated level Rockfall Rock failure Investigated Uninvestigated

Gero MR 186 51 5 7
OB 399 68 12
NM 204 10 0
Takayama MR 275 109 7 17
OB 517 380 7
NM 133 36 0
Furukawa MR 277 36 11 16
OB 126 20 5
NM 186 11 3

Figure 3. Histogram of rockfall probability (Gero). Figure 5. Histogram of rockfall probability (Gero).

Figure 6. Economic losses.

restoration cost (D2), and the indirect loss includes

the circumvention loss (D3) and the emergency
medical service loss (D4). Thus equation 1 can be
Figure 4. Histogram of rockfall probability (Gero). described using the economic losses as below,
R = P (D D +D D ) (2)
4 ECONOMIC LOSS
Detailed information of each economic loss is
4.1 Items of economic loss explained in following subsections.
As shown in Figure 6, the economic losses consid-
4.2 Economic loss from traffic accidents
ered in this study are separated into two categories;
direct loss and indirect loss. The direct loss includes The economic loss from traffic accidents means the
the economic loss from traffic accidents (D1) and the loss related death and injury induced by rockfall.

501

ISGSR2013.indb 501 10/18/2013 9:45:01 AM

 In other words, the loss is calculated based on a sup- travel distance. These are calculated using a using a
position that drivers or passengers are injured or user equilibrium traffic assignment.
killed due to the rockfall directly. Although much of
discussions have been made about the definition of C α(t − t ) β(l − l ) (4)
risk in terms of human life, but at least in Japan,
confronting at various natural disasters, the value where α is unite amounts of the talue of time,
of human life is generally considered explicitly to β is unite amounts of the travel cost, t is travel time
quantify the benefit of infrastructure investment. and l is travel distance. Subscripts 0 and 1 indicate
Public Works Research Institute (PWRI) (2004) states before and after road closure, respectively.
proposed a concept of the economic loss from The values of α and β are obtained from a manual
traffic accidents. Figure 7 shows an image of the published by The Ministry of Land, Infrastruc-
concept. The value of the economic loss from ture, Transport and Tourism (2008).
traffic accidents is evaluated based on trapezoid
distribution. Height of the trapezoid means the
value of the human life. The value of the human 4.5 Economic medical service loss
life (242 million Japanese Yen) is originated from a The Emergency medical service loss relates to death of
report published by Japanese government (2007). emergency patients induced by road closure. Fatality
rate of emergency patients is strongly depends on
4.3 Restoration cost travel time of ambulances. Hashimoto et al. (2002)
reported a relation between the fatality rate and the
The restoration cost is evaluated from history data travel time. In their study, brain hemorrhage, sub-
of rockfall accidents. 115 cases of restoration cost arachnoid hemorrhage, acute myocardial infarction,
data are used and the averaged cost (410,000 JPY) acute cardiac failure, pneumonia, cardiopulmonary
is employed as the restoration cost in this study. arrest, and brain infarction are employed as target
Although it is better to evaluate the cost depending diseases. The diseases and the model proposed by
on size of rockfall and level of importance of road, Hashimoto et al. (2002) are used in this study. Onset
these effect is not considered in this study, thus we probability of each disease and averaged travel time
used the averaged value for all cases. are obtained using a medical data accumulated in
hospitals and fire departments in the target area.
These information are shown in Tables 2 and 3.
4.4 Circumvention loss
The circumvention loss is calculated by following
equation. Table 2. Onset probabilities of diseases in study area.

D3 CT (3) Hida Takayama Gero

Disease (×10−4/day) (×10−4/day) (×10−4/day)
where C and T are amount of change of the con-
Brain hemorrhage 6.2 10.5 5.4
sumer surplus and number of days of road closure,
Subarachnoid 4.7 12.4 1.9
respectively. Number of days of road closure is hemorrhage
assumed as 1 day. Although this is also big assump- Acute myocardial 6.2 14.3 8.2
tion, but the number of days is strongly depends infarction
on degree of damage of road and size of rockfall Acute cardiac 5.5 11.4 6.0
and level of importance of road, thus we assumed failure
the number of days in consideration of past tends. Pneumonia 9.9 0.0 9.8
C is calculated from increments of travel time and Cardiopulmonary 7.3 40.9 4.9
arrest
Brain infarction 14.6 33.3 11.2

Table 3. Averaged time required for transportation to

hospital.

Region (number of data) Avaraged time required

Hida (979) 28 min. 18 sec.

Takayama (346) 32 min. 29 sec.
Figure 7. Image of economic loss from traffic accidents Gero (1311) 24 min. 00 sec.
of slope (PWRI, 2004).

502

ISGSR2013.indb 502 10/18/2013 9:45:02 AM

 5 RISK MANAGEMENT

5.1 Definition of B/C

In this section, results of evaluated risk are shown.
Because the cost-benefit ratio (B/C) is frequently
used, the definition of B/C is described here.
Generally, B/C is defined as follows,
ΔR
( /C ) = (5)
C
where ΔR is a decrease in risk and C is cost of coun-
termeasure. In this study, it is assumed the risk of
slope is zero after taking countermeasure. Because
degradation of countermeasures is not considered
in this study, we employed this assumption. But in Figure 9. Evaluated economic loss.
the future study, it is important to check the effect
of this assumption. Based on the assumption,
amount of risk change equal to risk as follows, Figure 10 shows values of risk of top 100 slopes.
Values of B/C of each slope are also described in the
R
( /C ) = (6) figure. As we can understand from the figure, high
C risk slopes don’t always have high values of B/C.
This indicates that it is necessary to make decision of
The cost of countermeasure C is obtained from taking countermeasures using both risk and B/C.
the database of the survey data, because informa-
tion of assumed cost is recorded in the data. 5.3 A trial calculation of risk management
A trial calculation was carried out to validate the
5.2 Results of evaluated risk effectiveness of calculated results in this study.
Figures 8 and 9 show histograms of risk and B/C. Following three different strategies of taking coun-
As mentioned before, the slopes are categorized termeasures were considered in this validation.
into 3 levels, required (MR), Observation (OB) and Plan A
No Measures (NM). Total of 1706 slopes that are Information of risk and B/C obtained in this study
grouped in MR and OB are described. As shown in are not used. Generally, decision making of tak-
Figure 9, values of B/C are over 1.0 in 288 slopes. ing countermeasure should strongly depend on the
It is therefore considered from the view point of results of the survey and the traffic volume. Thus
the efficiency that there are still a lot of slopes that priority of taking countermeasure is made based
should be taken countermeasures. on levels of survey results (MR, OB, NM) and
traffic volume in this plan.
Plan B
Results obtained in this study are efficiently used
in this plan. Priority of taking countermeasure is
determined based on B/C evaluated in this study.
Plan C
Results obtained in this study are used in this plan.
Priority of taking countermeasure is determined
based on amount of risk evaluated in this study.
Figure 11 shows a relation between residual risk
and amount of investment. Results of each plan
mentioned above are shown in the figure. As we
can see in the Figure, plans B and C can remove the
residual risk much more effectively than plan A.
This indicates evaluated risk and B/C in this study
are effective information for decision making of
taking countermeasure. Here, let us discuss about
Figure 8. Evaluated economic loss. the result of plans B. Plan B can remove risk most

503

ISGSR2013.indb 503 10/18/2013 9:45:03 AM

 Figure 10. Risk top 100 and B/C.

Figure 11. A relation between residual risk and investment.

effectively, but the strategy of plan B is not best such as fairness, are not considered in proposed
solution of risk management. The reason is that framework. In addition, as mentioned in this
the strategy depends on only economic efficiency, paper, some big assumptions are still included. It
and other viewpoints, such as fairness, are not is therefore necessary to develop the framework in
considered. It is therefore important to introduce the future works.
another indexes in future study.
REFERENCES
6 CONCLUSION Cabinet office (Japanese Government), A report of eco-
nomic analysis of damage of traffic accident, 2007. (in
This paper presents a framework of risk manage- Japanese)
ment of rockfall risk using real survey data. Road Hashimoto, T., Kurihara, M., Inoue, K., Iwasaki, Y. &
closure is defined as risk event, and the provability Fujimoto, A. 2002. Relationship between the carry-
of occurrence and economic losses are calculated. ing time and the survival rate in emergency medical
Finally, the trial calculation with different strategies service. Journal of Japanese Society for Emergency
of taking countermeasure was shown. According Medicine 5: 285–292. (in Japanese)
Public Works Research Institute (PWRI), A manual of
to the result of the trial calculation, it could be risk analysis and risk management (draft version),
summarized the obtained results are useful infor- Records of PWRI, No. 3926, 2004. (in Japanese)
mation for risk management of rockfall. The Ministry of Land, Infrastructure, Transport and
Although effectiveness of the proposed frame- Tourism, A manual of cost-benefit analysis, 2008. (in
work was validated, some important viewpoints, Japanese)

504

ISGSR2013.indb 504 10/18/2013 9:45:04 AM

 Geotechnical Safety and Risk IV – Zhang et al. (eds)
© 2014 Taylor & Francis Group, London, ISBN 978-1-138-00163-3

Implementing geo risk management in a client organization

M.Th. van Staveren

VSRM, Breda, The Netherlands

P.P.T. Litjens & J.J. Heerema

Rijkswaterstaat, Ministry of Infrastructure and the Environment, The Netherlands

ABSTRACT: Rijkswaterstaat, the largest public client organization for infrastructural projects in
The Netherlands, is responsible for the design, construction, management, and maintenance of the main
civil infrastructure in The Netherlands. Their key success factor is realizing infrastructure projects, not
only within time, budget, safety and quality standards, but also with a minimum of hindrance for the
public during construction. For these reasons, Rijkswaterstaat initiated the Dutch Geo-Impuls programme
for reducing geotechnical failure. Consequently, this public organization embraced Geotechnical Risk
Management (GeoRM) as the leading geotechnical working method and is serious about implementing
GeoRM in its own organization. This paper presents the selected approach for the GeoRM implementation
process, which includes embedding GeoRM in existing formal working procedures and empowering
geotechnical engineers and managers to adopt and apply GeoRM in their day to day activities. The main
hurdles encountered and solutions applied are discussed. Finally, the paper draws the main conclusions
for implementing GeoRM in public client organizations in the construction industry.

1 INTRODUCTION of dealing with geotechnical risks that is executed

in all project phases, in order to achieve project
Rijkswaterstaat is the executive organisation objectives effectively and cost-efficiently. GeoRM is
that manages and develops the main national a new name for the GeoQ process for geotechnical
infrastructure facilities on behalf of the Dutch risk management (van Staveren 2006). The process
Minister and State Secretary for Infrastructure and of geotechnical risk management is similar to the
the Environment. Rijkswaterstaat is responsible for process of project risk management and involves
the design, construction, management, and mainte- the same sequence of steps. Therefore, GeoRM fits
nance of the main infrastructure in The Netherlands. well in any sort of project risk management. The
This makes Rijkswaterstaat the largest public cli- difference is that GeoRM is a more specific and
ent organization for infrastructural projects in in-depth approach of project risk management,
The Netherlands. Realizing these projects, not only for giving geotechnical risk the attention it requires
within time, budget, safety and quality standards, in all phases of engineering and construction
but also with a minimum of hindrance for the pub- projects.
lic during construction, is a key success factor. As initiator of the Geo-Impuls program,
For these reasons, Rijkswaterstaat is the ini- Rijkswaterstaat is serious about implementing
tiator of the Dutch Geo-Impuls programme for GeoRM in its own organization. This paper
reducing geotechnical failure. This joint industry presents the selected approach for the GeoRM
programme aims to strengthen the geotechnical implementation process and its results to date.
community by substantially reducing geotechnical First, the GeoRM implementation objectives of
failures in all types of construction projects (Cools the public organization will be presented. Next,
2011). Over 200 geotechnical engineers and man- the selected GeoRM implementation approach is
agers from some 40 Dutch organizations, including introduced, by the three dimensions of method,
client organizations, contractors, engineering firms, organization and people. Then each dimension
and knowledge institutes, work closely together in is explored in more detail, including the main
this programme that runs from 2009 up to 2015. hurdles that were encountered and the actions
The Geo-Impuls participants embraced Geo- that were executed. Finally, the paper draws the
technical Risk Management (GeoRM) as the lead- main conclusions for implementing GeoRM in
ing geotechnical working method. GeoRM is an public client organizations in the engineering and
explicitly structured and well-communicated way construction industry.

505

ISGSR2013.indb 505 10/18/2013 9:45:05 AM

 2 GeoRM IMPLEMENTATION
OBJECTIVES

Before starting any GeoRM implementation activi-

ties, an organization should establish a set of clear
implementation objectives. These objectives provide
the rationale for the entire implementation process.
The main GeoRM implementation objective for
Rijkswaterstaat is to embed GeoRM and its sup-
porting tools in all relevant working processes and
developments within the entire organization. Well-
implemented GeoRM should contribute to the
organizational objectives of realizing infrastructure
projects within time, budget, safety and quality Figure 1. The three GeoRM implementation dimensions.
standards, with a minimum of hindrance for the
public during construction. The GeoRM imple-
mentation is initiated by the Rijkswaterstaat top
management and executed by a working group of are inter-related and their organization-specific
geotechnical engineers within Rijkswaterstaat. There balance will allow effective implementation of
are two types of members of these working group. GeoRM in an organization. This can be a client
The first type of members are engineers who are organization, but also a project organization,
actually working in projects, by supporting and facil- or an organization of a contractor or engineer-
itating geotechnical activities within these projects. ing company. Figure 1 presents these GeoRM
Examples of these activities are geotechnical reviews implementation dimensions.
of the scope of site investigations, specification of Given the previously mentioned implementation
geotechnical boundary conditions, and reviews of objectives, an implementation plan has been derived.
geotechnical reference designs. Specific GeoRM The members of the GeoRM working group have
implementation objectives for these engineers are been interviewed in order to acknowledge their
being able to apply GeoRM activities in the most views, expectations, and concerns regarding the
complex projects of Rijkswaterstaat, and being able GeoRM implementation, which will have an effect
to delegate GeoRM activities to other profession- on their working routine. Moreover, the actual
als within Rijkswaterstaat, as well as to subcontract organizational conditions for the GeoRM imple-
GeoRM activities to engineering firms. mentation have been identified and discussed with
The second type of members of the working the working group members. All these activities
group responsible for GeoRM implementation are resulted in personalized implementation objectives
engineers who are not directly working in projects. of the working group members, as presented in the
Specific GeoRM implementation objectives for previous section.
these engineers are expanding the motivation and Based on the organizational and personal
commitment to routinely applying the GeoRM objectives, as well as the identified organizational
working method, by using their networks. Also, conditions for the GeoRM implementation,
these engineers monitor and steer the execution a series of activities have been defined for
of GeoRM related topics within research devel- the implementation dimensions of method,
opment programmes within Rijkswaterstaat, with organization and people. At the moment of
knowledge institutes and universities, as well as writing of this paper, these activities are being
with similar organizations abroad, including the executed and the first results are gained. These
US Corps of Engineers in the United States. activities and results are presented in the next three
Finally, all members of the geotechnical working sections, including any hurdles encountered and
group should participate in at least one working the solutions applied to overcome these hurdles.
group of the Geo-Impuls development program.

4 THE GeoRM METHOD DIMENSION

3 GeoRM IMPLEMENTATION
APPROACH The GeoRM method dimension includes the
procedure and the supporting tools for applying it
The applied GeoRM implementation approach in an effective and cost-efficient way. The GeoRM
acknowledges three implementation dimensions of procedure has been summarized in a thin guide-
(1) method, (2) organization and (3) people (van line and is similar to the GeoQ approach for
Staveren 2009). These implementation dimensions geotechnical risk management (van Staveren 2006).

506

ISGSR2013.indb 506 10/18/2013 9:45:05 AM

 the geotechnical design for a deep foundation. To
be effective and cost-efficient, all organizational
members need to agree on their formal tasks and
responsibilities. This creates some sort of formal
organizational structure. Such a structure may
stimulate and support, or frustrate and even
obstruct the implementation of geotechnical risk
management.
Figure 2. The GeoRM structure in six cyclic steps. In addition to this structure each organization
has also some sort of informal way of working,
behaving and talking: an organizational culture.
Van Staveren (2013) provides an summary in However, an organizational culture is not tangible.
English of this Dutch guideline. The GeoRM It is “the way we are working here”, with the
process fits well in the ISO risk management shared convictions of the people working within
process (ISO 2009) and also in the RISMAN proc- an organization (Cameron & Quinn 1998). The
ess for project risk management (van Well-Stam organizational culture will to an considerable
et al. 2004). This latter process is widely applied in degree determine whether the formalized part of
The Netherlands. geotechnical risk management will be turned into
Figure 2 presents the structure of the GeoRM action, or just stay on paper.
procedure, with six cyclic risk management steps. Due to the widely accepted importance of
Two types of tools can be distinguished that the concepts of organizational structure and
proved to be useful in several phases of infrastructure culture in the organization sciences, the GeoRM
projects: (1) conventional geotechnical tools that dimension organization has been be divided into
can be applied in a risk-driven way and (2) specific the organizational structure and the organizational
GeoRM tools. culture.
Examples of the first type of conventional
tools are geological, geohydrological, and
5.1 GeoRM and organizational structure
geotechnical maps and databases, geotechnical
data reports, geotechnical modelling software for For becoming embedded, it is important that the
providing sensitivity analysis, risk-driven site and GeoRM method, its processes and supporting
laboratory investigations, risk-driven geotechnical tools, fits well within the formal work procedures
monitoring, and, not to forget, specialist geotechni- of the organization. Otherwise GeoRM would be
cal knowledge and judgement of well-experienced considered as “unofficial” and therefore not becom-
geotechnical experts. ing applied in a routine way of working within all
Examples of specific GeoRM tools are of course projects. Within Rijkswaterstaat a formal project
geotechnical risk checklists and risk registers. Other management procedure for realizing infrastruc-
tools are well-structured processes for geotechnical ture projects is used. This procedure distinguishes
risk identification, risk classification and risk three main project phases: (1) the definition phase,
remediation sessions, that can be supported by (2) the development phase and (3) the realisation
specific software in so-called electronic board phase. In each of these phases a number of formal
rooms for dealing effectively with different risk documents have to be developed. Specific reviews
perceptions (van Staveren 2006). Another useful are applied to allow the project to proceed to the
instrument is the GeoRiskScan (Bles et al. 2009) next phase.
for appraising the GeoRM quality. Finally, a well- Specifically for the formal project management
defined process for the communication of geotech- procedure, a formal and organization-specific
nical risk towards stakeholders of projects proved GeoRM guideline has been developed by a number
to be valuable. of members of Rijkswaterstaat’s geotechnical
group. This guideline presents the core elements
of GeoRM, by providing definitions of geotech-
5 THE GeoRM ORGANIZATION nical risk and its management, a description of
DIMENSION the GeoRM process, and the presentation of the
GeoRM principles. In addition, the guideline
An organization can be defined as a stable and defines for each of the three project phases which
structured social system, where individuals work GeoRM activities need to be performed and
together to achieve one or more joint goals reported in which formal document. As a result
(Rogers 2003). Organizations include different of these activities these documents will benefit, in
types of sub-organizations, such as units, projects order to fulfil the requirements of the formal project
and teams. An example is a design team realizing management procedure of Rijkswaterstaat.

507

ISGSR2013.indb 507 10/18/2013 9:45:05 AM

 For example, during the project definition phase, aspect, for a considerable number of years Rijks-
the first four GeoRM process steps of sampling waterstaat transforms from a rather technically-
project information, identifying geotechnical risk, oriented to a more project management-oriented
classifying geotechnical risk, and remediation geo- organization. The well-structured GeoRM proc-
technical risk can be applied in order to support ess suits well in the organizational culture of risk-
the costs estimations of the project with a required driven project management and control. Over the
degree of accuracy. Tools applied are for instance years, many technically-oriented managers devel-
geotechnical risk checklists for the specific type of oped towards more process-oriented managers,
project (a tunnel, a road, a railway, and so on) and who increasingly recognize and appraise the proc-
databases with geological, geotechnical and geo- ess structure of GeoRM. In addition, project risk
hydrological information in the neighbourhood management is a key topic in Rijkswaterstaat and
of the project location. The resulting GeoRM GeoRM is in fact the risk-driven and geotechnical
product is a cost assessment of geotechnical risk specialism of project risk management.
remediation. This product can become part of the Nevertheless, on-going attention is needed to
cost estimation report of the formal project man- communicate the benefits of applying GeoRM in
agement procedure. Rijkswaterstaat projects to project managers, tech-
During the project development phase, Rijks- nical managers and contract managers, because it
waterstaat has to define and procure for instance changes the way of dealing with geotechnical engi-
location specific ground information that is needed neering in a project from rather implicit and reac-
for reference designs and contracting. The risk- tive towards more explicit and pro-active ways of
driven definition and execution of a site investiga- working. For this reason, best practice flyers have
tion is a GeoRM tool that is for instance useful in been provided. These flyers summarize the GeoRM
this phase. By using the formal GeoRM guideline, process in only one brief text block, make the link
Rijkswaterstaat is able to subcontract the execu- with the formal project management approach and
tion of risk-driven site investigations with a scope give examples of GeoRM steps taken, GeoRM
that satisfies the geotechnical risk profile of the tools applied, and the results for the projects in
project. A tunnel project adjacent to an existing terms of cost savings, avoided delays, improved
tunnel in poor soil conditions has typically a higher quality and higher satisfaction of stakeholders in
risk profile than the same tunnel in isolation in bet- the projects.
ter soil conditions. Consequently, the risk-driven Furthermore, specific flyers have been developed
scope of the site investigation for the first tunnel for the technical managers, contract managers and
would be more extensive than that for the second project teams. These flyers summarize the proven
tunnel. Quite obvious for a geotechnical engineer, GeoRM benefits for these specific roles within the
however, by applying the GeoRM approach the project. An example of these benefits for the tech-
need for more or less site investigations becomes nical manager are the execution of additional risk-
also clear for less geotechnically educated project driven site investigations in a motorway expansion
managers and decision makers. project in The Netherlands, which proved the fea-
Finally, during the project realization phase for sibility of embankments in a conventional way and
example contractual management of geotechnical saved a lot of money. An example of GeoRM ben-
risk becomes important for controlling the budget efits for the contract manager was the application
and planning of the project. The Rijkswaterstaat of a geotechnical baseline report in a large bridge
specific GeoRM guideline provides clear sugges- construction, which settled the occurrence of dif-
tions on how to deal with the contractual alloca- fering site investigation effectively and efficiently,
tion of geotechnical risks, for instance by using the without delaying the project. The entire project
concept of the geotechnical baseline report (Essex team benefitted for instance from the results of
1997). a GeoRiskScan (Bles et al. 2009) in a project for
broadening a canal, which revealed in time a gap
in essential yet lacking ground information which
5.2 GeoRM and organizational culture
has been filled by hiring specific geotechnical
In addition to the formal part of GeoRM imple- expertise.
mentation, by developing the organization-specific
GeoRM guideline that is part of the formal project
management procedure, also the informal part of 6 THE GeoRM PEOPLE DIMENSION
the organizational culture needs attention. This is
a combination of adapting GeoRM to the exist- The third dimension for GeoRM implementation
ing organizational culture and trying to change is about the people who need to have the motiva-
the existing organizational culture in favour of tion and competencies to apply geotechnical risk
routinely applying GeoRM. Regarding the latter management in their work processes. Obviously,

508

ISGSR2013.indb 508 10/18/2013 9:45:05 AM

 the people dimension and cultural part of the or her GeoRM hurdle in his or her project. The
organization dimension are closely coupled, as other members reflect on the problem and provide
organizational culture is developed by interactions possible solutions from their own practice by
between individuals. The people dimension has a brainstorming. It is in fact a sort of action learn-
focus in the individual, and in particular on the ing or learning by doing and reflection, in combi-
motivation and competencies to apply GeoRM. nation with training on the job, according to the
For this reason, in the current phase of geotechni- concepts of Argyris & Schön (1978).
cal risk management implementation at Rijkswa- In the next meeting, some two months later,
terstaat the people dimension concerns especially the experiences with the solution of the previous
the members of the geotechnical working group. meeting is discussed and also a GeoRM hurdle
In order to identify their view on GeoRM, as of another team member is considered. Also, the
well as the hurdles they encounter when trying to overall progress of the GeoRM implementation
apply GeoRM in their project, six interviews were within the entire Rijkswaterstaat organization is
held with members of the geotechnical group. discussed in each session.
Questions included the expectation and results of
GeoRM, as well as hurdles and solutions for rou-
tinely applying GeoRM. There revealed wide con- 7 CONCLUSIONS
sensus about the GeoRM objective of supporting
the realization of successful projects by effectively, This paper presents the selected approach for the
cost-efficiently and continuously managing geo- GeoRM implementation process within a large
technical risk. In particular explicit communica- Dutch public client organization, Rijkswaterstaat.
tion about geotechnical risk was considered new This implementation process involves embedding
and valuable. GeoRM in existing formal working procedures, as
However, also a few hurdles arose. First of all, well as empowering geotechnical engineers during
how to put the GeoRM process in the practice inter-vision sessions to adopt and apply GeoRM
of projects appeared rather difficult for some of in their day to day activities.
the engineers. Also, the selection and allocation Rijkswaterstaat is the largest public client
of GeoRM tasks between the geotechnical engi- organization for infrastructural projects in The
neers of Rijkswaterstaat acting as client of pub- Netherlands, with a responsibility for the design,
lic projects, engineering firms performing design construction, management, and maintenance of
and contractual services, and contractors realizing the main civil infrastructure in The Netherlands.
projects, appeared to be not always clear. As such, Rijkswaterstaat is front-running with
In general, engineers of engineering firms seem implementing GeoRM in all phases of its project
to hesitate to apply GeoRM, perhaps because of a management activities in order to contribute to suc-
lack of awareness of the need for it and a lack of cessful project results in terms of budget, planning,
knowledge about what Rijkswaterstaat is expecting quality, safety, and user-friendliness. The GeoRM
from them. Moreover, within the organization of strategy of this client organization is expected
Rijkswaterstaat some of the engineers encountered to result into a boost of GeoRM applications by
difficulties to communicate the need and benefits engineering firms and contractors, who work for
of GeoRM with non-technical project managers, Rijkswaterstaat. In addition, other important pub-
contract managers, and other project team lic client organizations in The Netherlands, such as
members. the municipalities of large cities and the organiza-
For the latter hurdle, development of earlier tions that are responsible for flood protection, are
introduced best practice flyers and specific fly- already following Rijkswaterstaat in the application
ers for the project managers and teams have been of GeoRM in their most complicated projects.
developed. Regarding outsourcing of GeoRM It proved to be essential to formalize GeoRM
activities to engineering firms, guidance is in devel- in the existing project management procedures, in
opment. A procedure how to outsource risk-driven order to be taken seriously and applied within the
site investigations within the GeoRM process for projects of Rijkswaterstaat. For this reason, the
road projects to engineering firms has already been development of the organization-specific GeoRM
developed and applied successfully. guideline, which entirely fits in the formal project
Regarding the hurdle of putting the GeoRM management procedures, proved to be a neces-
process in the practice of projects, in combination sary step. Developing this guideline resulted in
with outsourcing GeoRM tasks to engineering clarifying the required GeoRM steps, the targeted
firms and contractors, so-called inter-vision meet- GeoRM results and the available GeoRM tools in
ings are held every two months. During these meet- the subsequent project phases.
ings of about one day each, one of the members In addition to this organization-specific
of the geotechnical working group presents his implementation of the GeoRM method in the

509

ISGSR2013.indb 509 10/18/2013 9:45:05 AM

 organizational structure, it was also necessary to Cameron, K.S. & Quinn, R.E. 1998. Diagnosing
pay attention to the organizational culture and the and Changing Organisational Culture. Boston:
people dimension. Writing GeoRM best practice Addison-Wesley.
flyers, as well as specific flyers for project manag- Cools, P.M.C.B.M. 2011. The geo-impuls programme
reduc-ing geotechnical failure in The Netherlands.
ers and project teams, proved to be valuable for In N. Vogt et al. (eds.), Geotechnical Safety and
revealing the GeoRM benefits in an explicit way Risk; Proc. 3rd Int. Symp., Munich, 2–3 June 2011.
to non-geotechnical professionals. Technical man- Karlsruhe: Bundesanstalt fur Wasserbau.
agers, contract managers, and project teams have Essex, R.J. 1997. Geotechnical baseline reports for
now specific examples of why and how GeoRM underground construction. Danvers: American Society
contributes to their own tasks and responsibilities of Civil Engineers.
for realizing successful projects, within the inher- ISO, 2009. International organization for standardization
ent limitation of budgets that are funded by public 31000:2009 risk management: principles and guidelines.
money. In conclusion, the combination of paying Geneva: ISO.
Rogers, E.M. 2003. Diffusion of innovations (5th edition).
attention to methods, organizational structure, New York: Free Press.
organizational culture, and people works, when van Staveren, M.Th. 2006. Uncertainty and ground
implementing geotechnical risk management in a conditions: a risk management approach. Oxford:
large public client organization. Elsevier.
van Staveren, M.Th. 2009. Risk, innovation & change:
design propositions for implementing risk management in
ACKNOWLEDGEMENTS organizations. Köln: Lambert Academic Publishing.
van Staveren, M.Th. 2013. Geotechnics on the move:
The authors are grateful to all members of the geo- guidance for a risk-driven way of working. Georisk
(accepted).
technical working group within Rijkswaterstaat, as van Well-Stam, D., Lindenaar, F., van Kinderen, S. &
well as to all project managers, technical managers, van den Bunt, B.P. 2004. Project risk management:
contract managers and other managers who dare an essential tool for managing and controlling projects.
to apply the still rather new concept of GeoRM in London: Kogan Page.
their day to day practices.

REFERENCES

Argyris, C. & Schön, D. 1978. Organizational learning:

a theory of action perspective. Boston: Addison-Wesley.
Bles, T.J., van Staveren, M.Th., Litjens, P.P.T. & Cools,
P.M.C.B.M. 2009. Geo risk scan: getting grips on geo-
technical risks. In Y. Honjo et al. (eds.), Geotechnical
Risk and Safety; Proc. 2nd Int. Symp., Gifu, 11–12
June 2009. London: Taylor and Francis.

510

ISGSR2013.indb 510 10/18/2013 9:45:06 AM

 Geotechnical Safety and Risk IV – Zhang et al. (eds)
© 2014 Taylor & Francis Group, London, ISBN 978-1-138-00163-3

Implementing geo risk management in the construction industry

M.Th. van Staveren

VSRM, Breda, The Netherlands

P.P.T. Litjens & P.M.C.B.M. Cools

Rijkswaterstaat, Ministry of Infrastructure and the Environment, The Netherlands

ABSTRACT: Geotechnical Risk Management (GeoRM) aims to contribute to successful infrastruc-

tural and other types of construction projects by effectively and cost-efficiently dealing with inherent
uncertain ground conditions. GeoRM has been adopted as the leading geotechnical working method by
over 200 participants of 40 organizations that participate in the Geo-Impuls geotechnical development
program. This is a Dutch industry wide program that aims to strengthen the geotechnical community
by substantially reducing geotechnical failures in all types of construction projects. Routinely applying
GeoRM in projects requires embedding GeoRM in organizations. This paper presents some of the efforts
the Dutch geotechnical community is currently making for implementing geotechnical risk management.
The main results of three industry wide work-shops about organizational key conditions for routinely
applying GeoRM are presented. Representatives of client organizations, engineering firms, and contrac-
tors assessed the presence of these key conditions and identified 36 pragmatic actions for strengthening
their organizational GeoRM conditions.

1 INTRODUCTION The Geo-Impuls participants embraced geo-

technical risk management (GeoRM) as the
1.1 Why geo-risk management? leading geotechnical working method. The two
key objectives of the Geo-Impuls program are
Applying Geo Risk Management (GeoRM) in
(1) a completed GeoRM toolbox, including the
construction projects becomes more and more a
process, tools, guidelines, and best-practices, and
key success factor. It is because GeoRM aims to
(2) in total 100 projects that demonstrate appro-
contribute to successful infrastructural and other
priate GeoRM application by 2015. This requires
types of construction projects.
embedding GeoRM principles and practices in
Successful projects are by definition completed
organizations. Therefore, the implementation of
within budget and planning, as well as according
GeoRM in (project) organizations is a key issue in
to pre-set safety and quality standards. In addition,
the Geo-Impuls program. The quality of GeoRM
these projects fulfill the expectations and needs of
application in these projects will be validated by an
its end users and other stakeholders. However, still
update of the GeoRiskScan (Bles et al. 2009).
too often unexpected ground conditions or behav-
There are three central development themes within
ior reduce project success or even creates project
the Geo-Impuls program: (1) contracts for dealing
failure.
with legal issues, (2) geo-engineering for dealing with
technical issues, and (3) project communication for
paying attention to differences in risk perception by
1.2 Geo-Impuls development program
the human factor. For each theme, the by the Dutch
Geo-Impuls is a Dutch industry-wide geotechnical geotechnical community considered most important
development program. It aims to strengthen the topics are worked out in 12 Working Groups.
geotechnical community by substantially reduc- Examples of these working groups are risk-
ing geotechnical failures in all types of construc- based soil investigation planning (contracts theme),
tion projects. Over 200 geotechnical engineers guidance for the application of the Observational
and managers from some 40 Dutch organizations, Method (geo-engineering theme), and a protocol
including client organizations, contractors, engi- for communicating geotechnical risk to commu-
neering firms, and knowledge institutes, work nities (project communication theme). Moreover,
closely together in this program that runs from international knowledge exchange, education, and
2009 up to 2015 (Cools 2011). training are part of the Geo-Impuls program.

511

ISGSR2013.indb 511 10/18/2013 9:45:06 AM

 1.3 Paper set up
This paper presents some of the efforts the Dutch
geotechnical community is currently making for
implementing geotechnical risk management in
their organizations.
First, the concept of GeoRM will be briefly pre-
sented, including its so-called Geo-Principles. The
core of this paper consists of presenting the execu- Figure 1. The GeoRM process in six cyclic steps.
tion and results of three industry-wide GeoRM
implementation workshops. These have been exe-
cuted for the three main players of client organiza-
2.2 GeoRM principles
tions, engineering firms, and contractors. In these
sessions, the presence of four organizational key Basically, there are two main routes of doing risk
conditions for routinely applying GeoRM in the dif- management: rule-based and principle-based. The
ferent types organizations were identified, classified, rule-based approach is using rules and results into
and thoroughly discussed. This resulted in 36 prag- one best way for managing risk. This is not appro-
matic actions for strengthening these four organiza- priate for GeoRM, because of the large diversity
tional key conditions for effective GeoRM. of engineering and construction projects in com-
Finally, the paper will draw the main conclu- plexity, size, location, and ground conditions. For
sions and remaining challenges for implement- this reason the Geo-Impuls takes the principle-
ing GeoRM in organizations in the construction based route for allowing fit-for-purpose geotechni-
industry. cal risk management. Based on the eleven generic
risk management principles from ISO31000 (2009),
eight specific geotechnical or GeoRM principles
2 GeoRM CONCEPT & PRINCIPLES have been defined by a number representative of
Geo-Impuls participants.
2.1 GeoRM concept By definition, principles are abstract and need
a translation into activities. Such a translation can
Before explaining the GeoRM concept, it may
be done for specific projects and even for specific
be helpful to define the term geotechnical risk.
project phases, again in a fit-for-purpose way. By
According to ISO31000 (2009), a risk is the effect
this approach GeoRM is a mean for realizing
of uncertainty on realizing objectives. Similarly,
project objectives, and not an end in itself. Table 1
a geotechnical risk can be defined as the effect of
presents the eight GeoRM interrelated principles
geotechnical uncertainty on realizing objectives,
and a number of examples of related actions in the
such as settlements or horizontal deformations
format of a simple checklist.
within pre-set limits. Geotechnical uncertainty may
result from randomness, fuzziness, incompleteness,
or simply incorrect geotechnical information (van
3 GeoRM WORKSHOPS
Staveren 2006). A geotechnical risk has a probabil-
ity of occurrence, one or more geotechnical causes,
3.1 Why GeoRM implementation workshops?
and usually a number of different effects when hap-
pening, such as damage, cost overruns and delay. Managing risk is difficult. Applying risk man-
GeoRM is an explicitly structured and well- agement is more difficult. Implementing risk
communicated way of dealing with geotechnical management in organizations is the most difficult
risk. It should be executed in all project phases, (van Staveren 2009b). The Geo-Impuls program
in order to achieve project objectives effectively provides a lot of guidelines and tools for execut-
and cost-efficiently. GeoRM is a new name for the ing GeoRM. However, routine application of these
GeoQ process for geotechnical risk management guidelines and tools, in other words its implemen-
(van Staveren 2006). The process of geotechnical tation, requires serious attention. Experiences, as
risk management is similar to the process of project well as research, teaches that this is not a sponta-
risk management and involves the same sequence neous process in many (project) organizations (van
of steps. Therefore, GeoRM fits well in any sort Staveren 2009b). For this reason, the implemen-
of project risk management. The difference is that tation of GeoRM within the organizations and
GeoRM is a more detailed and in-depth approach projects of the Geo-Impuls participants started
of project risk management, for giving geotechni- already during the program, parallel to the devel-
cal risk the attention it requires in all phases of opment of guidelines and tools. The Geo-Impuls
engineering and construction projects. Figure 1 implementation target is to apply GeoRM in 100
presents the six cyclic GeoRM steps. Dutch projects by 2015.

512

ISGSR2013.indb 512 10/18/2013 9:45:06 AM

 Table 1. GeoRM principles in a checklist format.

Done

No. GeoRM principles Examples of GeoRM actions Yes No

1 Generate and protect value Make all geotechnical risks in each project phase explicit,
including risk effects and risk remediation measures
2 Participate in decision making Make a geotechnical risk file from the start of the project
in all project phases and use it for decision making
3 Make geotechnical uncertainty Include geotechnical sensitivity analyses with margins
explicit in project reports
4 Work systematically, Include GeoRM explicitly in the project planning and
structured, and in time reserve resources for it
5 Use all available information Work from a general level to a detailed level, from using
geological maps to geotechnical monitoring
6 Work transparently together Indicate and communicate any dependencies of geotechnics
with all stakeholders with other disciplines in the project
7 Include the role of the human Make differences in organizational culture of all involved
factor project parties visible and feasible
8 Use experiences and lessons Use all available and relevant project evaluations, risk
for continuous improvement checklists, and experiences from professionals

A first series of three workshops focussed on 3.3 Workshop execution

the application of GeoRM in three large Dutch
The three workshops were held in the LEF Future
tunnelling projects in different phases: one in the
Centre of Rijkswaterstaat. This is a special accommo-
development phase, one in the design phase and
dation for working with groups that want to change
one in the construction phase. These workshops
their routines. Implementing GeoRM in (project)
resulted in fine-tuning of the Geo-Impuls activities
organizations requires some sort of organizational
within the themes contracts, geo-engineering and
and individual change, because dealing with inherent
project communication. The next year, a second
geotechnical uncertainties becomes well-structured
series of workshops were executed with a focus
and well-communicated in an explicit way.
on the implementation of GeoRM in the partici-
The interactive workshop program of one morn-
pating client organizations, engineering firms and
ing or afternoon each consisted of three parts. The
knowledge institutes, as well as contractors. This
first part consisted of an introduction, in which the
second series of workshops are presented in more
participants indicated their own degree of GeoRM
detail in this paper.
implementation. Also, the core definitions of geo-
technical risk and its management were shared, to
3.2 Workshop objective and participants work on a shared language of risk management.
The second part was a quick audit, in which the
The objective of the three workshops was to
participants assessed the degree of presence of
identify what should be done in the organizations
four key conditions for GeoRM implementation
of the Geo-Impuls participants in order to shape
in their organizations. During the third part of the
the right organizational conditions for imple-
workshop, the participants defined and discussed
menting the GeoRM way of working in their
their most appropriate individual actions that were
organizations.
based on the assessed degree of presence of the key
The participants consisted of geotechnical
conditions. In other words, what should they do by
engineers and (project) managers who are work-
themselves, within their (project)organizations, in
ing for client organizations, engineering firms
order to realize quick wins with regards to GeoRM
and knowledge institutes, as well as contractors.
implementation?
In total over 40 persons of more than 20 differ-
ent organizations participated in the three work-
shops. In order to learn as much as possible from
4 GeoRM WORKSHOP RESULTS
organizations with more or less the same role in
projects, these workshops were held separately
4.1 GeoRM positioning
for the client organizations, the engineering firms
together with the knowledge institutes, and the During the first part of the workshop the partici-
contractors. pants should indicated their own degree of GeoRM

513

ISGSR2013.indb 513 10/18/2013 9:45:06 AM

 implementation. They were literally asked to take positive side of uncertainty, being opportunities
their position on a line of one to ten. The number with favourable effects on project results.
one position indicated no GeoRM implementation
at all, while the ten number ten position indicated
4.2 GeoRM conditions
perfectly implemented GeoRM within the activi-
ties of the particular workshop participant. During the second part of the workshop the partic-
This exercise provided a wide array of dif- ipants completed a quick audit on the presence of
ferent GeoRM implementation positions. The four key conditions for GeoRM implementation in
participants of the client organizations took posi- their organizations. These key conditions, which are
tions ranging from 2,5 to 8. The positions of the derived from van Staveren (2009b), include (1) an
participants of engineering firms and knowledge uniform and shared understanding of GeoRM
institutes varied between 3,5 and 8,5, while the within the entire (project) organization, (2) an
positions of the contractors ranged from 4 to 8. inter-disciplinary application of GeoRM (such
It became paramount that the degree of GeoRM as involving contracts and planning departments
implementation is still highly dependent on the in GeoRM), (3) a formally embedded GeoRM in
persons involved. Even within the same organi- the working procedures of the organization, and
zation, different persons took sometimes entirely (4) a GeoRM cooperation with external parties,
different positions with regard to the degree of including project stakeholders, subcontractors,
GeoRM implementation. This situation indicated suppliers, and so on. The results of the audit on
that GeoRM was not yet implemented in the entire the presence of these four key conditions are pre-
(project)organization, but at best in a number of sented in Table 2.
parts of the organization. From Table 2 a few remarkable observations
During the discussion that followed a number can be made. Key condition number 1, a uniform
of points of attention were raised for implement- and shared GeoRM understanding, seems present
ing GeoRM more widely in organizations. These for some 25% to 75% in all of the three types of
included making the benefits of applying GeoRM organization.
clear (expressed in terms of quality, positive finan- With regard to key condition number 2 about
cial effects, stronger professional reputation), the inter-disciplinary application of GeoRM, there
developing a shared risk management language, is considerable variation within the client organiza-
rewarding the application of GeoRM, relating tions and between the contractors and the clients,
risks to expectations, communicating geotechnical engineering firms and knowledge institutes. The
risks effectively to clients and their clients (who do contractors showing almost no inter-disciplinary
not always want to hear about risk), persisting as GeoRM application (assessment of 0% to 25%)
geotechnical engineers in making geotechnical risk explained the score by stating that in their organi-
and their potential effects explicit, acknowledging zations there is no cooperation between geotechni-
the inherent subjective role of the human factor in cal engineering and other disciplines with regard to
risk assessments, and finally considering also the explicitly applied geotechnical risk management.

Table 2. Results of quick scan on GeoRM key conditions.

Assessment of degree of presence

GeoRM key conditions 0–25% 25–50% 50–75% 75–100%

1. There is uniform & shared GeoRM understanding Cl Cl

EK EK
Co Co
2. There is interdisciplinary GeoRM application Cl Cl Cl Co
Co EK EK
3. GeoRM is formally embedded in working procedures Cl EK Co
EK Co
4. GeoRM is cooperated with external party involvement EK Cl Cl
Co EK
Co
Legend:
Cl: Client Organisation
EK: Engineering forms & Knowledge institutes
Co: Contractors.

514

ISGSR2013.indb 514 10/18/2013 9:45:06 AM

 The contractors have already a reasonable degree increasing the degree of presence of the GeoRM
of formally embedded GeoRM in their work pro- key conditions.
cedures, as expressed by their assessment of key
condition number 3. The client organizations not
4.3 GeoRM actions
yet started with it, similar to the engineering firms
and knowledge institutes. During the third and final part of the workshops
Finally, regarding the cooperation of GeoRM each participant identified at least one individual
with external parties, the client organisations are and pragmatic action for increasing the degree of
front-running, as do part of the engineering firms, presence of one of the GeoRM key conditions.
knowledge institutes and contractors. In addition, Obviously, key conditions with a relatively low
some of these latter organizations nearly did not degree of presence should be prioritized.
cooperate at all with external parties, when apply- The participants wrote their action(s) on a post-
ing GeoRM. card, which they received three weeks later as a
In conclusion, a number of the participating kind reminder of taking action. In total 36 prag-
organisations, either clients, engineering firms, matic actions were identified and discussed. These
knowledge institutes or contractors, have already actions are summarized, slightly adapted when
some of the GeoRM key conditions present to a required for clarification purposes, and presented
considerable degree. Nevertheless, for all of the in Table 3 up to and including Table 6. In total 17
participating organizations there is still some, or actions for establishing or strengthening key condi-
even a lot, of opportunity to increase the level of tion number 1 about the uniform shared GeoRM
presence of key conditions, which will support the understanding are presented in Table 3.
GeoRM implementation in their organizations. In total 5 actions for establishing or strengthening
Therefore, the next task during the workshop was key condition number 2 about the inter-disciplinary
to identify and discuss appropriate actions for application of GeoRM are presented in Table 4.

Table 3. Actions for establishing key condition number 1.

No. Action

1 Apply GeoRM in a uniform way.

2 Explicitly mentioning geotechnical risks during project management meetings.
3 Include geotechnical risk in project presentations.
4 Communicate geotechnical risk in the terminology of lawyers and non-technical decisions makers.
5 Share GeoRM knowledge and experiences in the entire organization.
6 Take a strong position as a geotechnical engineer in the project team in order to take care of GeoRM.
7 Make project team members and colleagues aware of the GeoRM need by organizing internal workshops.
8 Establish checklists with geotechnical risks.
9 Integrate GeoRM in the education program of structural engineers.
10 Integrate GeoRM in internal project management courses.
11 Stimulate GeoRM thinking and acting within your department.
12 Put GeoRM on the agenda of strategy meetings.
13 Invite external GeoRM experts for sharing their knowledge in the organisation.
14 Develop GeoRM best practices that show the GeoRM successes and use these best practices.
15 Draw attention to effective attitude and behaviour when applying GeoRM.
16 Explain differences in applying GeoRM for different types and scopes of projects.
17 Stimulate learning about GeoRM by reading GeoRM books and papers.

Table 4. Actions for establishing key condition number 2.

No. Action

1 Give specific attention to geotechnical risks during project risk management sessions.
2 Organize internal and interdisciplinary GeoRM workshops for a number of projects.
3 Demonstrate to the internal organization how to apply interdisciplinary GeoRM, including its benefits.
4 Make geotechnical risks, including its inter-disciplinary remediation actions, explicit in the project risk registers.
5 Provide internal checklists for geotechnical risks that demonstrate their inter-disciplinary character.

515

ISGSR2013.indb 515 10/18/2013 9:45:06 AM

 In total 10 actions for establishing or strengthen- 4.4 Examples of GeoRM implementation
ing key condition number 3 about formally embed-
As initiator of the GeoImpuls program,
ding GeoRM in work procedures are presented in
Rijkswaterstaat is very serious about implement-
Table 5.
ing GeoRM in its own organization (van Staveren
Finally, in total 4 actions for establishing or
et al. 2013). Parallel to embedding GeoRM in
strengthening key condition number 4 about the
existing work procedures, Rijkswaterstaat used
coordinating GeoRM with the external project envi-
several tools within different types of projects. This
ronment and stakeholders are presented in Table 6.
resulted in successful examples of GeoRM imple-
Together, all these individual actions demon-
mentation, as presented in Table 7.
strate how geotechnical engineers and (project)
Table 7 shows also the main GeoRM key condi-
managers are able to give their own appropriate
tions that were realized within these projects. Using
attention to implementing GeoRM in their organi-
the GeoRM tools proved to be helpful for the geo-
sations. On the one hand, participants representing
technical engineer to communicate the impact of
client organizations, as well as engineering firms
subsoil risk in terms of money and time effects, as
and knowledge institutes, dedicated their actions
well as reputation consequences.
in particular to developing an uniform and shared
Therefore, geotechnical risks became a specific
GeoRM understanding and to formally embed-
topic of interest for the project management team
ding GeoRM in work procedures. On the other
members. This resulted in dedicated risk con-
hand, the contractors focussed their actions on
trol activities, such as gathering missing subsoil
increasing internal and external cooperation dur-
information linked to geotechnical risks, specific
ing the application of GeoRM.

Table 5. Actions for establishing key condition number 3.

No. Action

1 Apply internal lobby activities in the organization for getting GeoRM embedded in existing processes.
2 Include geotechnical engineering with GeoRM in primary process descriptions.
3 Provide a clear internal protocol for applying GeoRM.
4 Include the GeoRM process in the formal quality management protocol.
5 Formally embed GeoRM in the geotechnical department.
6 Consider together with the geotechnical consultancy team how to formalize GeoRM in projects.
7 Formally apply the relevant GeoRM tools in projects.
8 Design and apply formal GeoRM reporting structures.
9 Organize formal sessions about the GeoRM embedment in the organization.
10 Develop a formal scan for deriving the top 3 geotechnical risks in each project.

Table 6. Actions for establishing key condition number 4.

No. Action

1 Make non-technical decision makers enthusiastic by presenting the specific GeoRM benefits for them.
2 Identify geotechnical risks with the entire project environment and stakeholders at the start of the project.
3 Integrate geotechnical risks systematically and explicitly in contract between project parties.
4 Show to the project environment and stakeholders the need for and specific benefits of GeoRM.

Table 7. Examples of GeoRM implementation in projects.

Type of project Type of GeoRM tool applied GeoRM key condition

Waterway GeoRiskScan 1, 4
Third Party Expertise
Motorway Risk-based soil investigation 1, 3
Bridge construction Geotechnical baseline report 1, 2
Tunnelling GeoRM Communication protocol 1, 4

516

ISGSR2013.indb 516 10/18/2013 9:45:06 AM

 contractual statements for managing these risks, and contractors. These organizations all have their
early requests of necessary licences for groundwa- stake in realizing successful construction projects,
ter extraction, and strategies for communicating against reasonable costs.
geotechnical complexity to local residents who live
near the project.
The projects mentioned in Table 6 are now ACKNOWLEDGEMENTS
used as Best Practices within Rijkswaterstaat. In
this way, these projects help in further formalis- The authors would like thank all participants of
ing and applying GeoRM within this major client the Geo-Impuls workshops for sharing their pre-
organization. cious insights and experiences with GeoRM.

5 CONCLUSIONS
REFERENCES
The three Dutch and industry-wide workshops, as
Bles, T.J., van Staveren, M.Th., Litjens, P.P.T. &
presented in this paper, demonstrate that the status Cools, P.M.C.B.M. 2009. Geo risk scan: getting
of GeoRM implementation, as well as the ways to grips on geotechnical risks. In Y. Honjo et al. (eds.),
increase the degree of GeoRM implementation, is Geotechnical Risk and Safety; Proc. 2nd Int. Symp.,
rather organization specific. Between and within Gifu, 11–12 June 2009. London: Taylor and Francis.
organizations, quite different degrees of GeoRM Cools, P.M.C.B.M. 2011. The Geo-impuls programme
implementation emerged by applying the quick reducing geotechnical failure in The Netherlands. In
scan on the degree GeoRM key conditions pres- N. Vogt et al. (eds.), Geotechnical Safety and Risk; Proc.
ence. Obviously, similar types of workshop can 3rd Int. Symp., Munich, 2–3 June 2011. Karlsruhe:
be easily organized in other countries with simi- Bundesanstalt fur Wasserbau.
ISO 2009. International organization for standardization
lar programs for reducing geotechnical failure, or 31000:2009 risk management—principles and guide-
within individual companies for strengthening their lines. Geneva: ISO.
GeoRM implementation capabilities and effects. van Staveren, M.Th. 2006. Uncertainty and ground
The results and examples in this paper demon- conditions: a risk management approach. Elsevier,
strate that a lot of pragmatic actions for embed- Oxford.
ding GeoRM are readily available. Therefore, it is van Staveren, M.Th. 2009a. Extending to geotechnical
simply up the responsible engineers and managers risk management. Georisk 3(3): 174–183.
to apply the quick scan for assessing the degree of van Staveren, M.Th. 2009b. Risk, innovation & change:
presence of GeoRM key conditions in their organ- design propositions for implementing risk manage-
ment in organizations. Köln: Lambert Academic
izations, and to select and apply suitable actions Publishing.
for strengthening any yet underdeveloped key van Staveren, M.Th., Litjens, P.P.T. & Heerema, J.J.
conditions. By doing so, considerable steps can be 2013. Implementing geo risk management in a client
made with implementing GeoRM in client organi- organization. In Geotechnical Risk and Safety; Proc. 4th
zations, engineering firms, knowledge institutes, Int. Symp., Hong Kong, 2–4 December 2013 (in press).

517

ISGSR2013.indb 517 10/18/2013 9:45:06 AM

This page intentionally left blank
 Geotechnical Safety and Risk IV – Zhang et al. (eds)
© 2014 Taylor & Francis Group, London, ISBN 978-1-138-00163-3

The buildup and assessment of environmental indices

for storm induced disaster prone areas

M.-H. Wu
Department of Civil and Environmental Engineering, National University of Kaohsiung, Kaohsiung, Taiwan

J.P. Wang
Department of Civil and Environmental Engineering, Hong Kong University of Science and Technology,
Hong Kong, China

H.-R. Liao & Y.-R. Chen

Department of Civil and Environmental Engineering, National University of Kaohsiung, Kaohsiung, Taiwan

ABSTRACT: In recent years, global warming has resulted in serious extreme climate events. Natural
disasters occur more frequently world widely, and the corresponding impacts are more and more
serious, for example of the tragedy of Hsiaolin Village, Kaohsiung due to the impact of Typhoon
Morakot in 2009. In this research, the study areas are primarily those who were suffering from typhoon
storm-related disasters and are characterized with high disaster potential. A series of field reconnais-
sance works were conducted in flood and slopeland disaster-prone areas in Kaohsiung city in order
to analyze the causative factors of the hazards. There are ninety eight communities and ninety nine
villages under investigation for flood and slopeland disaster potential analyses. With the environmen-
tal index buildup and assessment, the districts of Shanlin, Namaxia, Meinong, Maolin, Taoyuan and
Liugui are categorized as high slopeland disaster-prone districts, and the causative factors are exist-
ence of serious landslides and poor vegetated bare slopes. The districts of Yongan, Mituo, Gangshan,
Qiaotou, Nanzi, Dashe and Renwu are classified as high flood disaster-prone districts, and the causes
are abundant rainfall and poor drainage system. The results obtained from this study will contribute
to keep abreast of high disaster-prone areas and to promote the corresponding disaster prevention and
preparedness tasks.

1 INTRODUCTION (ISDR 2004). But vulnerability research, if it is to

contribute to wider debates on resilience and adap-
Natural hazards are threatening events, capable of tation faces significant challenges, in measurement,
producing damage to the physical and social space in handling perceptions of risk, and in governance
in which they take place not only at the moment (Adger 2006). The diversity and apparent lack of
of their occurrence, but in the long-term, due to convergence over time are, in many ways, a reflec-
their associated consequences. When these conse- tion of the divergent objectives of the research and
quences have an impact on society and/or infra- the phenomena being explained (Janssen 2006).
structure, they become natural disasters (Irasema A comprehensive theory of vulnerability to global
2002). In average, there are four typhoons per year change therefore needs to account for a range of
around the region of Taiwan. They usually bring risks, thresholds and institutional responses and
a huge amount of rainfall that occasionally causes resources, given that vulnerability will manifest
natural hazards such as floods, landslides, and itself differently at different scales (Kasperson &
debris flows. In recent years, due to extreme climate Kasperson 2001).
condition caused by global warming, there come The City of Kaohsiung in Taiwan suffered seri-
more frequent natural disasters with aggravat- ous typhoon induced disasters in 2009 and 2010,
ing impacts around the world. Disaster reduction including Typhoon Morakot in 2009 causing a
becomes an important issue. Disaster reduction massive landslide burying a village and killing
strategies include an assessment of vulnerability most villagers as many as 474 people. As a result,
and risk, as well as assessment of institutional more attention is placed on the environmental
capacities describing an operational framework vulnerability of 38 individual districts of the city.

519

ISGSR2013.indb 519 10/18/2013 9:45:06 AM

 Table 1. Survey form of flooding prone sites.

Figure 1. Flow chart of study procedure.

The assessment of vulnerability essentially includes

the characteristics of critical facilities, social and
economic infrastructure, the use of effective early
warning systems, and the application of scientific,
technical, and other skilled abilities (UNDP 2004).
To develop environmental index assessment, we
carry out an investigation for environmental char-
acteristics in this research. Vulnerability has two
elements: exposure and susceptibility (ADPC
2006). Exposure is the nature and degree to which
a system experiences environmental or socio-
political stress (Adger 2006). The characteristics of
these stresses include their magnitude, frequency,
duration and areal extent of the hazard (Burton
1993). Susceptibility takes into account those
social, economic, political, psychological and envi-
ronmental variables that work together in produc-
ing susceptible impacts amongst people within the
same exposure (DFID 2004).
This study investigated the hazard prone areas
including those with frequent disaster history or
those with high disaster potential in the near future, Table 2. Survey form of slopeland disaster prone sites.
in order to understand the spatial distribution and
causative factors of the storm induced disasters.
A series of reconnaissance works were conducted
in the sites with typhoon induced flooding and
slopeland disaster history. The information was
collected and integrated to build up the disaster
database, so as to facilitate the pre-warning analy-
sis and interpretation of emergency response oper-
ations during typhoon events. Moreover, not only
qualitative but also quantitative analyses were car-
ried out the assessment of environmental indices.
The study’s procedure is shown in Figure 1. With
the analyses, it is more efficient to establish the
assessment method and to master the risks of
villages for the use of foundation of emergency
response strategies.

2 FLOODING & SLOPELAND DISASTER

PRONE AREA INVESTIGATIONS

The flooding and slopeland disaster-prone area

investigations were mainly conducted with field
reconnaissance and literature review. As for flood-
ing prone site investigations, the team went to the
field and surveyed flood attributes at the site includ-
ing elevations, flood heights, and the factors such
as if the site is close to water ways or reservoirs,

520

ISGSR2013.indb 520 10/18/2013 9:45:06 AM

 and if there are draining facilities around the site. each cluster or village would be calculated by sum-
Slopeland disaster attribute includes geological marized the total factors (points) applied to them,
conditions such as if the site on the dip-slope or namely the environmental score. Through the
seismic fault zone, and geographical conditions analysis of each cluster’s (village’s) environmental
such as if the site is on the stream terrace and the score, the levels as well as the corresponding envi-
slope gradient. Moreover, some observed charac- ronmental indices are established for flooding and
teristics such as slope seepage, crack of retaining slopeland disaster prone clusters or villages. The
wall, and poor vegetation were reported. In this levels included high vulnerability, medium vulner-
study, 99 villages of the 30 districts and 96 clusters ability and insensitivity.
within 7 districts of Kaohsiung City were selected In terms of district vulnerability, the district
for site investigations of flooding and slopeland hazard environmental score is obtained by the fol-
prone areas, respectively. The surveying forms are lowing equation:
shown in Tables 1 and 2.
District environment score
3 DATA CLASSIFICATION NumberrHVV × 4 + NumberrMVV
AND STATISTICAL ANALYSIS × 3 + NumberrIV × 1 (1)
=
NumberrRVD
3.1 Data classification
In this study, the environmental assessment indices where HVV = high vulnerability villages, MVV =
(factors) are divided into categories of exposure, medium vulnerability villages, IV = insensitive vil-
environmental sensitivity and disaster prevention lages and RVD = reconnaissance villages in the
measures, as shown in Figure 2. district. Through the calculation of the equation,
the levels as well as the corresponding environ-
mental indices are established for flooding and
3.2 Statistical analysis
slopeland disaster prone districts. The classifica-
According to the aforementioned classification, tion of district vulnerability is as follows: ≤3 points
each factor is given up to 1 point, and the score for for medium vulnerability and >3 points for high
vulnerability.

4 RECONNAISSANCE AND ANALYSIS

RESULT

In recent five years, Typhoon Morakot in 2009

and Typhoon Fanapi 2010 are two principal events
inducing serious flooding and slopeland disasters.
Typhoon Morakot, which coupled with Southwest
Monsoon, in total brought a rainfall of 2792 mm
in four days, about 150% of the annual rainfall in
Kaohsiung. The disasters resulted in serious land-
slides and debris flows in mountainous districts,
killing 523 people, 17 people missing (Ministry of
the Interior); Typhoon Fanapi brought a rainfall
of 321 mm in Kaohsiung within 3 hours, resulting
in serious flooding in plain areas, and railway or
highway traffic tie-up.
This study investigated the areas with disaster his-
tory within the two principal typhoon events, along
with referencing to the scope of “2010 Kaohsiung
City Preservation Plan of Flooding Hazard Poten-
tial Areas” and “2010 Investigation and Analysis of
Vulnerability Factors for Kaohsiung City Disaster-
Prone Area”. The sites under investigation and
analysis are shown in Figure 3.
This study categorized the environmental indi-
ces (factors) into three groups, which are exposure,
Figure 2. Classification of environmental indices. environmental sensitivity and disaster prevention

521

ISGSR2013.indb 521 10/18/2013 9:45:07 AM

 Figure 5. The variation of environmental characteris-
tics of flood hazard areas.

Figure 3. Distribution of reconnaissance sites.

Figure 6. Statistical chart of slopeland disaster induc-

ing factors.

Figure 4. Statistical chart of flood inducing factors.

measures. According to the reconnaissance report,

each site would be given a score from summariz-
ing the points of its environmental factors. Based
on analysis of the scores, the environmental indi-
ces for regional (village and district levels) were
determined.

4.1 Flooding reconnaissance result

Figure 7. The variation of environmental characteris-
The analysis shows that the main reason for flood- tics of slopeland hazard areas.
ing disaster in Kaohsiung should be the exces-
sive rainfall, followed by poor drainage and river
embankment overflow, as shown in Figure 4.
4.2 Slopeland reconnaissance result
The survey also found that among the past
flooding events, the primary environmental char- Among the 96 clusters under reconnaissance in this
acteristics of flooding areas are near rivers, drain- study, 71% of those have landslides nearby, and
age systems or coast (75%), followed by lowland 64% have suffered debris flow disasters. The statis-
or ground subsidence area (70%), as shown in tical analysis of disaster inducing factors is shown
Figure 5. in Figure 6. Poor vegetation and cluster close to

522

ISGSR2013.indb 522 10/18/2013 9:45:08 AM

 dangerous slope are the main features. Besides,

score
Total
the study also reported that all of the mountain-

8
9
8
8
ous clusters have no substitute access roads, which

potential
means once the main roads are interrupted, the

Flood
clusters would be isolated.

zone
The statistical analysis for geographical features

1
1
1
1
flood prone of slope hazard prone areas showed that 81% of the
sites have more than one sensitive geological char-
Approved

acteristic, for example of dip slope, rockfall, debris

slides and rock slides, and 53% of the sites nearby
area

potential debris flow creeks, shown in Figure 7.

1
1
1
1
prevention
plan area

4.3 Flooding indices assessment result

Flood

In recent five years, the flooding disasters which

1
1
0
1

resulted in live and property loss in the villages were

primarily in Morakot and Fanapi typhoon events
Others

in 2009 and 2010, respectively. Observed from the

analysis, the minimum score of these villages with
1
1
1
1

flooding depth below 0.5 meters is 4.0, and the

conservancy
remediation

mean is 6.8. Therefore, the village flooding disaster

or water
Lack of

facility

Table 4. District flooding vulnerability assessment.

0
1
1
0

District
Poor drainage,

environmental
District score Vulnerability
pumping
dalay

Yong An 4 High
Mi Tuo 4
1
1
1
1

Gang Shan 4
overflow,
silting or
blocking

Hu Nei 4
Stream

Yan Chao 4
Lu Jhu 4
0
0
0
0

A Lian 4
Flooding

Da She 4
history

Ren Wu 4
Cian Jin 4
1
1
1
1

Cian Jhen 4
San Min 4
subsidence
or ground
Lowland

Ling Ya 4
Mei Nong 4
Village flooding indices grading form.

Zih Guan 4
1
1
1
1

Nan Zih 3.9

seawater intrusion,

Ciao Tou 3.8

Excessive rainfall,

reservoir spillway

Da Shu 3.8
Cie Ding 3.7
Niao Song 3.5
Tian Liao 3.3
Da Liao 3.3
1
1
1
1

Fong Shan 3.3

Ci Shan 3.2
Sin Gang

Jiugang
Wei Sin

Gu Shan 3 Medium
Village

Ci Jin 3
Tadi

Liou Guei 3
Sin Sing 3
Table 3.

District

Zuo Ying 2.8

Mituo
An
Yong

Yan Cheng 2.5

523

ISGSR2013.indb 523 10/18/2013 9:45:10 AM

 Table 6. District slopeland vulnerability

score
Total

6.2

7.2
assessment.

10
District

subsitute
environmental

access
road
District score Rank
No

1.0
1.0
1.0
debris flow Na Ma Sia 4 High
Mei Nong 4
dip slope
Potential

Mao Lin 4
creek,

Tao Yuan 3.8

1.0
1.0
1.0
Shan Lin 3.6
Jia Sian 3.1
vegetation

Liou Guei 3.1

Poor

0.0
1.0
1.0
Building foundation
subsidence, retaining

prone classification is determined as follows: less

than 3.0, 3.1∼6.9 and 7.0∼10 represent insensitiv-
dangerous slope
wall crack, near

ity, medium vulnerability, and high vulnerability,

respectively. The grading method is sorting the fac-
tors into 10 items, and allocating 1 point to each
item, as shown in Table 3. The village vulnerability
1.0
1.0
0.0

assessment was completed. Moreover, in terms of

district vulnerability, through the Equation (1), it
surface

was found that 84% of the 30 districts with flood-

Slope

crack

ing history in Kaohsiung were categorized as high

0.0
1.0
1.0

vulnerability, and 16% of them were classified as

medium vulnerability, as shown in Table 4.
seepage
Slope

0.0
1.0
0.0

4.4 Slopeland indices assessment result

Disaster
history

After analysis, it is observed that the minimum

score of the reconnaissance clusters is 3.7, and
1.0
1.0
1.0

the mean score is 7.0. Therefore, the village slope-

land disaster prone classification is determined as
Debris slide,

follows: less than 3.6, 3.7∼6.9 and 7.0∼10 repre-

rock slide,
rockfall

sent insensitivity, medium vulnerability, and high

vulnerability, respectively. The grading method is
1.0
1.0
1.0

grouping the factors into 10 items, and allocat-

ing 1 point to each item, as shown in Table 5. The
stream terrace

village vulnerability assessment was completed.

Slopeland,

Moreover, in terms of district vulnerability,

Village slopeland indices grading form.

or valley

through the Equation (1), it was found that all of

the 7 districts with slopeland disaster history in
1.0
1.0
1.0

Kaohsiung were categorized as high vulnerability,

as shown in Table 6.
households
No. of

0.2
1.0
0.2

5 CONCLUSIONS
Guanglin

In recent five years, Typhoon Morakot in 2009

Village

and Typhoon Fanapi 2010 are two principal events

Maya
Ji Lai

inducing serious flooding and slopeland disasters

in Kaohsiung City. In this study, a series of recon-
Na Ma Sia

naissance works were conducted on the investiga-

Mei Nong
Shan Lin
Table 5.

tion of 99 villages with flooding disaster history

District

and 96 clusters with slopeland disaster history in

Kaohsiung within five years. From the observation

524

ISGSR2013.indb 524 10/18/2013 9:45:10 AM

 of the result, the primary inducing factors for REFERENCES
flooding disaster are excessive rainfall (97%),
poor drainage system (60%) and stream overflow Adger, W.N. 2006. Vulnerability. Global Environmental
(34%). As far as geographical characteristics are Change 16: 268–28.
concerned, 75% of flooding villages are located in ADPC (Asian Disaster Preparedness Center), 2006.
lowland areas, and 75% of them are close to river. Critical guidelines: community-based disaster risk
management. Bankok: Asian Disaster Preparedness
The environmental indices assessment shows that Center.
among the investigated villages, 68% of them are Burton, I., Kates, R.W. & White, G.F. 1993. The environ-
classified as high vulnerability and 28% as medium ment as hazard (second edition). New York: Guilford.
vulnerability. With regard to the 30 districts with DFID (Department for International Development),
flooding disaster history, 80% of them are catego- 2004. Disaster risk reduction: a development concern.
rized as high vulnerability and 20% as medium Norwich: Overseas Development Group, School of
vulnerability. The study also reports that all of Development Studies, University of East Anglia.
the 96 clusters under slopeland hazard investiga- Irasema, A.A. 2002. Geomorphology, natural hazards,
tion have no substitute access road, which would vulnerability and prevention of natural disasters in
developing countries. Geomorphology 47: 107–124.
encounter threat of isolation in slopeland disasters. ISDR (International Strategy for Disaster Reduction),
Among them, 66% are located close to poor- 2004. Living with risk: a global review of disaster reduc-
vegetated slopeland, and all of the seven districts tion initiatives. Geneva.
are classified with high vulnerability. Janssen, M.A., Schoon, M.L., Ke, W. & Börner, K. 2006.
The results obtained from this study will pro- Scholarly networks on resilience, vulnerability and
vide the reference of high disaster-prone areas in adaptation within the human dimensions of global
Kaohsiung City, and contribute to disaster preven- environmental change. Global Environmental Change
tion and preparedness tasks. 16(3): 240–252.
Kasperson, J.X. & Kasperson, R.E. 2001. Global environ-
mental risk. London: Earthscan and UNU Press.
Mohammed, S.I., Mohammad S.H.S. & Shamim M.H.
ACKNOWLEDGEMENTS 2013. Disaster risk index: How far should it take
account of local attributes? International Journal of
This study was possible through the “2012 Kaoh- Disaster Risk Reduction 3: 76–87.
siung City Disaster-Prone Areas Environmen- UNDP (United Nations Development Programme),
tal Reconnaissance & Assessment” sponsored 2004. Reducing disaster risk: a challenge for develop-
by National Science and Technology Center for ment, a global report. New York.
Disaster Reduction (NCDR). Special thanks to
Dr. Chih—Hsin Chang in NCDR.

525

ISGSR2013.indb 525 10/18/2013 9:45:10 AM

This page intentionally left blank
 11 Geotechnical applications and case studies

ISGSR2013.indb 527 10/18/2013 9:45:10 AM

This page intentionally left blank
 Geotechnical Safety and Risk IV – Zhang et al. (eds)
© 2014 Taylor & Francis Group, London, ISBN 978-1-138-00163-3

Ultrasound as a new approach for the prediction of collapsible soils

K. Abbeche
Department of Civil Engineering, Laboratory LARHYA, University of Batna, Batna, Algeria

M.S. Laouar Mohamed Salah

Department of Civil Engineering, University of Tebessa, Tebessa, Algeria

ABSTRACT: Collapsible soils types are found in natural unsaturated soils, which are characterized by
a metastable structure, and undergo an abrupt collapse when they are flooded (with or without loading),
which causes important damages. Based on the structural composition of these soils, samples made up of
various proportions of sand and fine particles were reconstituted. The first phase of work consists in the
experimental determination of the geotechnical characteristics of these samples. It is thereafter proposed
a test program based primarily on the use of the cone penetrometer and the ultrasonic apparatus as the
approaches for the prediction of the factors which influence collapse. The results obtained clearly show
the influence of certain parameters such as; water content and the energy of compaction on the collapse
potential and the ultrasonic speed.

Keywords: collapsible soils, collapse potential, cone penetrometer, ultrasonic speed

1 INTRODUCTION The collapse of Cheria in Eastern of Algeria

which occurred in 2009 constitutes a good example.
Collapsible soils are metastable soils of loose open In the latter, a great collapse was recorded, in
structure, unsaturated nature, being in deposits which tens of constructions were inserted of more
form. In the dry state, a natural cementing between than two meters and half in the ground. While
the grains confers an important inter-granular waiting to achieve measurements of the technical
connection and can support very high loads. expertise, the preliminary report charges this catas-
However, saturation, even without additional load- trophe to a movement in the ground water. A geo-
ing, causes the disintegration of the connections technical study made by Batna (2000), within the
giving a denser structure followed by a sudden scope of the realization of a natural gas station in
collapse of the soil particles. Among the causes of Hassi Messaoud shows that the site is composed of
saturation there is the raise of the ground water, two layers of collapsible nature, the adopted solu-
water infiltration and canalization leaks. Because tion is substituting the first layer and taking meas-
of the important collapse potentials and serious ures to avoid the infiltration of water to the second
consequences which can occur in the construc- layer. In addition, degradations of several resi-
tion, this type of soil is considered unstable where dence buildings which took place in Biskra (2002)
foundations sit. These soils are primarily local- are due to water infiltrations. A building of three
ized in the arid and semi arid region. They relate floors with Xining, Qinghai, was destroyed beyond
to a significant number of countries in particular repair because of collapse Qian & Lin (1988). This
those of the northern hemisphere located between problem occurred because the loess beneath the
the 30th and 55th parallels as well as countries of foundations underwent a structural collapse when
South America (Abelev 1988). The cycles of pro- it was flooded. Experimental and theoretical stud-
longed dryness which occurred these last years ies aiming to understanding the great number of
on several occasions and in several regions of the uncertainties implied in the phenomenon of col-
world modify the parameters governing the behav- lapse are currently undertaken. The literature
ior of the soil and give rise to new collapsible soils revealed that the majority of research was devoted
zones. The following are considered as collapsible to the collapse mechanisms and the identification
soils: the alluvial and eolian deposits, flows mud, methods, of treatment and prediction.
residual grounds, volcanoes rejections, loesses, and Dudley (1970) qualitatively described that the
embankments slightly compacted or compacted in collapse phenomenon for cemented structures does
the dry slope of the compaction curve. not depend on dampness, it occurs only when the

529

ISGSR2013.indb 529 10/18/2013 9:45:10 AM

 cementing connections are broken by mechanical density is higher than 0.65% and the moisture con-
constraints. In revenge, if the ground is a mixture tent is close to the optimum of Proctor there is no
of grains and fine particles which induce important risk of collapse (Abbeche et al. 2005). In spite of
connections due to suction or cementing, dampness having a great range of ultrasonic equipment and
lead to the cancellation of suction what decreases a large use of this process in various fields, the lit-
cohesion and supports collapse. This result was con- erature reveals that, except geotechnical marine and
firmed by Cui & Magnan (2000), Morgenstern & De some applications, little attention was granted to
Matos (1975) and Ganeshan (1982), who affirmed this technique regarding soil mechanics problems.
that the cause of collapse lies in the low water This experimental work presents the results of three
contents. Booth (1975) and Ting (1979) explained series of tests. In addition to the compression tests,
that collapse depends on the initial dry density, the a series of tests using the cone penetrometer and
void ratio and the degree of saturation. Markin the first time of the original experimental, curves
(1969) propose an interval of degree of saturation of the non-destructive tests with the ultrasounds
between 60% and 65% beyond which collapse does are put in parallels, in the objective to propose a
not appear any more. The same result is confirmed predicting method of the collapsible soils based on
by Ganeshan (1982), Booth (1977) and Lawton ultrasonic tests.
(1989), who observed that for a given dry density
the overload which causes collapse is inverse pro-
portion to the natural moisture content of the soil. 2 CHARACTERISTICS OF MATERIALS
The destruction of the capillary forces can explain
the sudden collapse by flooding the ground. The The tests were carried out on six reconstructed soils
suctions developed in the clay connections can be made up of sands and of kaolin in various pro-
different from those developed between the silt portions for which the application of the various
grains. Up to our days, there is no means to meas- criteria of collapse, reported by Ayadat & Bellili
ure these differences (Fredlund & Rahardjo 1993). (1995), shows that those are collapsible.
The examination of the macroscopic and micro- Two types of sands lesser than 2 mm of diameter
scopic aspects of the sudden collapse is recom- are used for the soils reconstruction; sand of Dunes
mended (Feda 1994). Abrupt collapse occurs when of Oum Ali region and sand of stream extracted
the dry density and the initial moisture content are from Melag stream of El Aouinet region washed
low (Cui et al. 1999; Cui et al. 1998). If the relative and dried at 105 °C during 24 hours. In view of the

Table 1. Characteristics of materials.

Materials Characteristics

Sand of Sand equivalent: 73.26%

dunes Grain size distribution ranged between 0.08 and 2 mm with 1.36% of particles < 80 μm
Coefficient of uniformity: 3.91
Coefficient of curvature: 1.33
Sand of Sand equivalent: 68.69%
stream Grain size distribution ranged between 0.08 and 2 mm with 3.01% of particles < 80 μm
Coefficient of uniformity: 2.19
Coefficient of curvature: 0.94
Kaolin % <2 μm 43%
Liquid limit: 65.83%
Plastic limit: 39.64%
Specific density of grains Gs 2.42
Reconstructed Label S1 S2 S3 S4 S5 S6
soils % Kaolin 15 35 50 20 30 40
% Sands of dunes 85 65 50 – – –
% Sands of stream – – – 80 70 60
GS 2.65 2.59 2.46 2.62 2.56 2.48
ωL% 16.47 26.63 35.37 18.47 28.97 33.42
ωP% 11.03 5.37 20.87 11.95 14.77 19.03
γdmax(g/cm3) 2.04 1.95 1.84 1.95 1.82 1.75
ωopt% 8.62 9.43 13.88 12.82 14.67 117.82
% <2 μm 4.91 1.73 16.74 7.03 9.84 14.12

530

ISGSR2013.indb 530 10/18/2013 9:45:10 AM

 small percentage of fine particles that they contain, Table 2. Characteristics of consistency of soils.
these two types of sands are used for the concretes
making. The kaolin used (<80 μm) is extracted from Soil W0 AC IP IL IC IW
of Hamame Debagh Mine of Guelma region it has
Soil 1 2 1.11 5.44 −1.66 2.66 0.37
white color and is used generally in the manufacture
4 −1.29 2.29 0.73
of the fine porcelain, pottery and ceramic products.
6 −0.92 1.92 1.10
The soils S1, S2 and S3 are reconstructed with sands
8 −0.56 1.56 1.47
of Dunes and kaolin, while the soils S4, S5 and S6
are reconstructed with sands of stream and kaolin. Soil 2 2 0.96 11.26 −1.19 2.19 0.18
The geotechnical characteristics of sands, kaolin 4 −1.01 2.01 0.35
and reconstructed soils are presented in Table 1. 6 −0.83 1.83 0.53
The gradation curves of the reconstructed soils are 8 −0.65 1.65 0.71
presented in Figures 1 and 2. Soil 3 2 0.87 14.5 −1.30 2.30 0.14
4 −1.16 2.16 0.27
6 −1.02 2.02 0.41
2.1 Characteristics of consistency of the soils 8 −0.89 1.89 0.55
The literature revealed that a soil is expected to Soil 4 2 0.93 6.52 −1.53 2.53 0.31
collapse if at least, one of the following criteria is 4 −1.22 2.22 0.61
checked (Ayadat & Ouali 1999): AC < 1, IL < 0, IP ≤ 20, 6 −0.91 1.91 0.92
IC > 1, IW ≤ 1. 8 −0.61 1.61 1.23
The results presented in Tables 2, show that Soil 5 2 1.44 14.2 −0.90 1.90 0.14
these soils are expected to collapse and that the 4 −0.76 1.76 0.28
6 −0.62 1.62 0.42
8 −0.48 1.48 0.56
Soil 6 2 1.02 14.39 −1.18 2.18 0.14
4 −1.04 2.04 0.28
6 −0.90 1.90 0.42
8 −0.77 1.77 0.55

characteristics of consistency of the reconstructed

soils depend basically on the initial moisture content.

3 USED MATERIALS

3.1 Oedometric tests

The odometer whose essential elements are:
Figure 1. Grains size distribution curves (Soils 1, 2 and 3).
– Oedometer Mold of 50.4 mm diameter and
20 mm height.
– Frame of consolidation of lever arm of 1/10.
– Set of weights.
The compaction tamper Figure 3 which is con-
ceived especially at the laboratory for the compac-
tion of the soil in the oedometer ring. Entirely
manufactured of steel, it is consists of:
– A base of 48.42 mm diameter and 3 mm thick-
ness attached to a column of guidance of
280 mm length, through which a piston slips.
A sliding stopper along the rod makes it possible
to adjust drop height of the hammer.
– A hammer in circular shape of dish of 84.422 mm
diameter and 8.40 mm thickness. Its weight is
de121g, having a centered drilling of 8.45 mm
Figure 2. Grains size distribution curves (Soils 4, 5 and 6). diameter.

531

ISGSR2013.indb 531 10/18/2013 9:45:10 AM

 – A calibration bar.
– A set of two transducers of 54 kHz with cables,
acting differently as transmitter or receiver.
– Paste pot of contact.

4 TEST PROGRAM

Three series of principal tests were carried out on

six reconstructed soils; Table 3 illustrates the pro-
gram of these tests.
Figure 3. Compaction tamper.

5 TEST PROCEDURE AND REALIZATION

5.1 Oedometric tests

The test soil consists of sand and kaolin accord-
ing to the proportions mentioned above. The soil is
brought to the required moisture content by addi-
tion of distilled water; the set soil-water must be
well homogenized in a porcelain mortar. The mix
of soil is then poured in the mold of the odometer
then compacted using the compaction tamper.
The compaction of the soil consists in dropping
the hammer which slides through the rod of the
Figure 4. Cone penetrometer. Tamper from a height H = 15 cm, which will
strike the dish that transmits the shock to the
specimen. To make it perfectly plane, the higher face
of the sample must be leveled using a rigid blade.
The compression tests with the oedometer are
made according to Jennings & Knight (1975) pro-
cedure which consists to succeeding application
of the following loads: 25, 50, 100, and 200 kPa.
Then, proceeding to the flood of the sample and
recording the new settlement value, afterward
increasing the loading up to 400 kPa. During the
test the readings of settlement are noted at 15s, 30s,
1 min, 2 min, 5 min, 10 min and 24 h.
Figure 5. Ultrasonic analyzer.

5.2 Penetration tests

3.2 Penetration tests
They are achieved with a cone penetrometer pro-
The manual cone penetrometer. Figure 4 is made vided with a metal dish. The soils reconstruction,
up of: the mix filling and the compaction in the dish are
carried out in the same manner as for the com-
– A stainless steel cone, 30 degrees of opening
pression tests. The cone, with its rod, is placed in
surmounted by a rod. The weight of the mobile
contact with the upper face of the soil sample. The
system is of 80 g.
penetration of the cone in the soil is measured with
– A dial gauge of 36 mm/0.01 mm.
the dial gauge. The Δh penetrations of the cone are
– A metal dish with a diameter of 53 mm and
carried forward relating to the selected parameters.
36.4 mm height. Its mass is about 56.2 g.

5.3 Ultrasonic tests

3.3 Ultrasonic tests
This series of tests starts with the calibration of
The equipment Figure 5 includes:
the analyzer, by measuring the speed transmission
– An analyzer for velocity measurement of the of the wave through the calibration bar. The veloc-
ultrasonic waves. ity of an ultrasonic wave train, which crosses soil

532

ISGSR2013.indb 532 10/18/2013 9:45:11 AM

 Table 3. Tests program.

Test type Selected parameters Test # Observation

Odometric Moisture contents: 2%, 4%, 6% and 8%. 96 Realized according

tests Compaction degrees: 10, 25, 40 and 60 blows. to Jennings and
Knight procedure.
Penetration Moisture contents: 2%, 4%, 6%, 8%, 10%, 12% and 14%. 168 Realized with the
tests Compaction degrees: 10, 25, 40 and 60 blows. cone penetrometer.
Ultrasonic Moisture contents: 2%, 4%, 6% and 8%. 96 Led to the ultrasonic
tests Compaction degrees: 10, 25, 40 and 60 blows. analyzer.

specimens, produced in the oedometric mold, is Table 4. Classification.

measured, according to the procedure of the com-
pression tests. To ensure a good transmission of CP Degree of problem
the waves in the body of the specimen and before
0% to 1% No risk
adjusting the system of measurement, thin layers 1% to 5% Moderate trouble
of contact grease to the two faces of the transduc- 5% to 10% Trouble
ers (transmitter and receiver) are applied. 10% to 20% Severe trouble
On the screen of the analyzer are represented, >20% Very severe trouble
the time or the transit speed of the wave, that
according to the configuration of the analyzer.

6 TEST RESULTS AND INTERPRETATION

The results of this experimental work are presented

in two parts; the first concerns the standard tests
(Tables 1 and 2) (characteristics of materials and
characteristics of consistency), necessary to the
geotechnical identification of used materials and
the reconstructed soils. The second is distributed
as follows.

6.1 Oedometric tests

6.1.1 Depiction of the collapse of the soil
The variation of the moisture contents and ener-
gies of compaction are made in the purpose to
check whether these soils have the properties of Figure 6. Typical oedometric curve of a collapsing soil
collapsible soils. The variation of moisture content (Knight and Jennings, 1975).
and energies of compaction allow controlling the
collapse potential. The curves obtained are simi-
lar to that of Jennings and Knight (1975). • Soil S3: from 0.83% to 8.92%
The Collapse Potential CP (%) is calculated by • Soil S4: from 0.66% to 7.61%
the relation: • Soil S5: from 0.74% to 7.84%
• Soil S6: from 0.77% to 7.90%.
Δec
CP = × 100% (1) According to the classification suggested by
1 + e0 Jennings and Knight (1975), (Table 4), these results
correspond to the headings going from “No risk”
where: to “Troubles”.
Δec = e1 (200 kPa) − e2 (200 kPa, flooded)
e0: Initial void ratio 6.1.2 Influence of the water content
and the energy of compaction
The results of these tests show that the collapse
The high collapse potentials are noted for low
potential CP varies for;
initial moisture contents. For a given initial water
• Soil S1: from 0.52% to 7.54% contents the collapse potential is decreasing with
• Soil S2: from 0.59% to 8.34% the increase in the energy of compaction.

533

ISGSR2013.indb 533 10/18/2013 9:45:13 AM

 Figure 7. Oedometric curve Soil 1 (E = 10 blows).

Figure 10. Variation of penetration with moisture (Soil 1).

6.2 Penetration tests

6.2.1 Interpretation of the penetrations
versus the initial moisture content ω0
The moisture contents between ω = 2% and ω = 8%
do not give a clear idea on the behavior of the stud-
ied soils; thus the increase in the water contents is
increased up to 14%.
The curves obtained are divided into two slopes
Figure 10. In the first, collapse decreases gradually
with the increase in the moisture content until a
lower limit when the moisture content approaches
Figure 8. Variation of collapse potential with number the Proctor optimum. In the second slope it is noted
of blows (ω = 6%). an opposite behavior in which collapse increases
with the growth of the moisture content.
Considering its speed and its convenience com-
pared to the Proctor test, it can be more practi-
cal for the compaction projects of the collapsible
soils to use the test of the cone penetrometer for
the determination of the limit penetration and the
corresponding moisture content which divide the
penetration curve into two slopes, the first is dry
and the second is wet. This is analogue to Proctor
test that the optimum separates also the curves into
two slopes, dry and wet. A similar performance is
noted for all tested soils. One can deduce on col-
lapsible soils, that there is an opposite relationship
Figure 9. Variation of collapse potential with moisture between the penetration test and the Proctor test,
content (E = 25 blows). the first being used to determine the limit penetra-
tion and the second the maximum dry density.

The decrease of collapse is more obvious that 6.2.2 Interpretation of the penetrations
the moisture content increases. In the same condi- versus ωopt/ω0
tions of compactness and moisture content of the Collapsible soils are characterized by the condition
soil containing the greatest percentage of kaolin ωopt/ω0 > 1, (Holtz & Hilf 1961). Analysis of
exhibit greatest collapse potential. the penetration curves versus the ratio ωopt/ω0,
These results agree with those of Lawton et al. Figure 11 confirm the existence of two distinct
(1989) and Ayadat et al. (1998) and confirm the behaviors and separated by the line ωopt/ω0 = 1.
observations of Abbeche et al. (2007). One can On the left of this line, the penetration knows a
conclude that the reconstructed soils at the labora- gradual decrease for then growing in a roughly
tory hold a similar behavior to those met in situ, regular way as one moves away from the limit sepa-
therefore suitable for the series of tests suggested. rating the collapsible soils (ωopt/ω0 >1) of the non

534

ISGSR2013.indb 534 10/18/2013 9:45:14 AM

 Figure 11. Variation of penetration with ωopt/ω0 (Soil 1).

Figure 13. Variation of ultrasonic speed with number

of blows (Soil 4).

Figure 12. Variation of penetration with γd/γs (Soil 5).

collapsible soils. This limit corresponds to the limit

penetration indicated by Plim.

6.2.3 Interpretation of the penetrations

versus γD/γS Figure 14. Variation of ultrasonic speed with moisture
The same statement is visualized in the representa- content (Soil 1).
tion of the penetration against the ratio of density
γd/γs, Figure 12. We notice a gradual decrease of the
penetration up to a limit value corresponding to The growth of compaction contributes to the
the straight line separating the two states from soils increase of speeds, especially when the moisture
for then knowing a phase of progressive increase content comes close to the Proctor optimum. Let
with the growth of this ratio. us note that curves corresponding to 60 blows
Similarity of previous curves illustrates the present more important speed values compared to
existence of a similar behavior of the collapsible other energies of compaction, especially with the
soils with respect to the penetration and that a increase in the moisture contents. This proves a
limit characteristic value separates the collapsible good state of compactness due to the humidifica-
soils of the non collapsible soils. tion and the rearrangement of the grains; it is the
case of non collapsible soils.
6.3 Ultrasonic tests
6.3.2 Prediction of collapse by ultrasonic test
6.3.1 Influence of the moisture content Figures 15 and 16 concretize a vital relationship
and the energy of compaction between ultrasonic speed and potential collapse; the
The results of the ultrasonic tests show that ultra- decrease of one is synchronized with the increase of
sonic speed varies according to variation of the other. In Figure 15, curves have the same shape.
energy of compaction and/or moisture content They pass by three phases, in the beginning paral-
the Figures 13 and 14. For the same value of the lel straight lines representing an important fall of
energy of compaction, whatever the soil, the ultra- the CP with very close speed values. Then, two suc-
sonic speed is increasing with the growth of the cessive slopes of the curves are noted; in the first,
moisture content. a reduction of the CP corresponds to an increase

535

ISGSR2013.indb 535 10/18/2013 9:45:16 AM

 Figure 15. Variation of collapse potential with ultra- Figure 17. Variation of limit penetration with ultra-
sonic speed for w% (Soil 1). sonic speed.

In the second case the susceptibility of collapse

depends on the water content and the state on com-
pactness on the soil. This procedure can be applied
to the restructured or intact soil, at the laboratory
and even on site. Considering its advantages, the
results of the ultrasonic sounding can be general-
ized with various types of collapsible soils such as
the loesses and other unsaturated soils.

6.3.3 Relation limit penetration ultrasonic speed

Results of the compression tests using the oedom-
eter show that the energy of compaction which
corresponds to 60 blows makes the soils non col-
Figure 16. Variation of collapse potential with ultra-
sonic speed for E (Soil 1). lapsible, whatever the percentage of the fine parti-
cles and the water content. This deduction agrees
with the representation of the limit penetrations
speeds, in the second, the stabilization of collapse against the ultrasonic speeds (Fig. 17), where it
is explained by great values speeds and very close is noted that the reduction of limit penetrations is
collapse potentials. increasing with the increase of ultrasonic speeds.
The curves of collapse potential according to For energies of compaction varying between 10
speed show that the compaction and the water and 40 blows, the soils remain likely to collapse
content take part with the reduction of collapse and the lines have almost the same slope and the
and the increase of ultrasonic speeds. Figure 16 same tendency equations. A remarkable slope of
shows that the energy of compaction contrib- these slopes is visualized by applying energy of
utes more effectively than the water content to compaction equal to 60 blows; which explains the
the reduction of collapse. For a higher energy of existence of a similar behavior specific to the col-
compaction, making the non collapsible soils, it is lapsible soils and which differed from the behavior
noted a low variation in the state of compactness of the non collapsible soils.
soil, this for various proportions of water content,
while speeds variation is more important. 7 CONCLUSIONS
From these observations, values of ultrasonic
speeds are compared against various water content The principal conclusions which one can draw
and energy of compaction values. Since the ques- from this study can be summarized as follows:
tioned soils have the possibility of collapsing when
1. The suggested experimental program allows
they are in a loose state; one propose a prediction
to characterize in a satisfactorily manner the
method of collapsible soils based on ultrasonic
behavior of the collapsible soils.
tests (non destroyed), fast and easy to realize.
2. Collapsible soils can be reconstructed in the
Values of ultrasonic speed are limited as follows:
laboratory by mixing with various propor-
• V ≤ 400 m/s → Collapse appears; tions kaolin, such fine particles with sand, led
• 400 m/s < V < 1000 m/s → Collapse can occur. to water contents lower than the optimum of
• V >1000 m/s → Risk of Collapse is not likely to Proctor and than compacted to moderate ener-
occur. gies of compaction.

536

ISGSR2013.indb 536 10/18/2013 9:45:18 AM

 3. The results obtained clearly show the influence Booth A.R. 1975. The Factors Influencing Collapse Set-
of certain parameters such as; kaolin content, tlement in Compacted Soils. In: Proceedings of the 6th
water content and energy of compaction on the Regional Conference for Africa on Soil Mechanics and
collapse potential, the limit penetration and the Foundation Engineering. Dublin, South Africa, 57–63.
Booth A.R. 1977. Collapse Settlement in Compacted
ultrasonic speed. Soils. CSIR Research Report 321 NITRR, Pretoria,
4. The collapse potential can be excessive if the South Africa, Bulletin 13.
initial water content is low. For water content CTC Biskra 2002. Intervention du directeur du CTC
lower than the optimum of Proctor, there is an de Biskra. 2eme Colloque Maghrébin de Génie Civil,
energy of compaction beyond which collapse Biskra, Algérie. 10,11 Decembre.
does not occur. Cui Y.J., Delage P., Alzoghbi P. 1998. Microstructure and
5. The use of the cone penetrometer as an identi- Collapse Behaviour of Loess from Northern France.
fication mean of the collapsible soils is possible. In: Proceeding of Unsat 98, Beijing, Vol. 1 31–36.
This makes it possible to follow the evolution Cui Y.J., Delage P., Schlosser F., Wonarowcz M. 1999.
Etude du comportement volumique d’un lœss du nord
of collapse and to propose a limit penetration, de la France. Xlléme congrès Européen de Mécanique des
separating the collapsible soils from the noncol- sols et de Géotechnique Amsterdam, Vol. 1 p. 337–342.
lapsible soils. Cui Y.J., Magnan J.P. 2000. Affaissement locaux dus à
6. A new experimental approach of prediction l’infiltration d’eau en géomécanique environnemen-
of collapsible soils based on ultrasonic tests, tale. Chapitre n°6 «Risques naturels et patrimoine».
easy and fast, is proposed. The results obtained Ed. Hermes 139–164.
depend on grains size distribution, state of Dudley J.H. 1970. Review of Collapsing Soils [J]. Journal
compactness of the soil and water content. of Soil Mechanics and Foundation Div ASCE 96(3):
Ultrasonic speeds are limited as follows: 925–947.
Feda J. 1994. Mechanisms of collapse of soil Structure.
– V ≤ 400 m/s, then collapse appears; In Derbyshire (ed) Genisis and properties of collaps-
– V >1000 m/s, then the risk of collapse is ible soils. NATO Series C. Mathematical and Physical
isolated. Sciences, Vol. 468 pp. 149–172. Dordrecht: Kluwer.
– Between these two limits collapse can occur, Fredlund D.G., Rahardjo H. 1993. Soil mechanics for
it depends then on the water content and the unsaturated soils. New York, John Wiley and Sons,
state of compactness. Inc.
7. The ultrasonic test can be carried out in the lab- Ganeshan V. 1982. Strength and Collapse Characteristics
oratory or in situ, on intact or altered samples of Compacted Residual Soils. Thesis (M.E), Asian
of an unspecified form. Institute of Technology Bangkok Thailand.
Holtz W.G., Hilf J.W. 1961. Settlement of Soil Founda-
8. The ultrasonic speed of metastable soils tion due to Saturation In: Proceeding 5th International
gives an idea of the state of compactness; it Conference on Soil Mechanics and Foundation Engi-
is inverse proportion with the potential of neering, Vol. 3 673–679.
depression. Jennings J.E., Knight K. 1975. The Additional Settle-
ment of Foundation due to Collapse of Sandy Soils on
Wetting. In: Proceeding. 4th International Conference on
REFERENCES Soil Mechanics and Foundation Engineering, 316–319.
Lawton E.C., Fragaszi R.J., James H. 1989. Collapse of
Abbeche K., Mokrani L., Boumekik A. 2005. Compacted Clayey Sand. Journal of Geotech. Eng
Contribution à l’identification des sols effondrables. ASCE, Vol. 155(9):1252–1267.
Revue Française de Géotechnique 110:85–90. LNHC 2000. Projet Extension des Gazoducs à Hassi
Abbeche K., Hammoud F., Ayadat T. 2007. Influence of Messaoud. Rapport de reconnaissance de sol.
Relative Density and Clay Fraction on Soils Collapse. Laboratoire National de l’Habitat et de la Construction
Experimental Unsaturated Soil Mechanics. Springer de Batna, Algérie.
Proceedings in Physics, 112:3–9. Markin B.P. 1969. Discussion on standard criteria of Sag
Abelev M.Y. 1988. Loess and its Engineering Problems in in Loess soils. Soil Mechanics and Foundations Engi-
the USSR, Proc. of the Int Conf. Engineering Problems neering, no. 2, p. 137.
of Regional Soils. Beijing, China. Morgenstern N., De Matos M. 1975. Stability of slopes
Ayadat T., Ouali S. 1999. Identification des sols affaissa- in residual soils. Proc. 5th. Pan American Conf. on
bles basée sur les limites d’Atterberg. Revue française soil Mech. and Found Eng. Buenos Aires Argentina.
de géotechnique. 3, p. 367–383.
Ayadat T., Bellili F. 1995. Sols susceptibles d’affaissement: Qian H.J., Lin Z.G. 1988. Loess and its Engineering
Identification mécanique et traitement. Revue Algérie Problems in China. In: Proceeding of the International
Equipment, no. 20, p. 18–23. Conference Engineering Problems of Regional Soils.
Ayadat T., Belouahri B., Ait Ammar R. 1998. La migra- Beijing, China, 136–153.
tion des particules fines comme approche d’explication Ting W.A. 1979. Consolidation of a partilly saturated
du mécanisme de l’effondrement des sols. Revue residual soil. Proc. 6th Asian Reg. Conf. on Soil Mech.
Française de Géotechnique 83:1–9. and Found. Eng., Singapore, vol. 1, p. 95–98.

537

ISGSR2013.indb 537 10/18/2013 9:45:20 AM

This page intentionally left blank
 Geotechnical Safety and Risk IV – Zhang et al. (eds)
© 2014 Taylor & Francis Group, London, ISBN 978-1-138-00163-3

Effect of riverbed scouring on bridge piles during earthquake

Z.H. Khan, M.R.I. Khan, M.F. Raiyan & K.M. Amanat

Bangladesh University of Engineering and Technology, Dhaka, Bangladesh

ABSTRACT: This study represents an approach for investigating the effect of riverbed scouring
on piles of a reinforced concrete bridge in the event of an earthquake. A finite element model of box
girder bridge with 9 spans was developed for analysis of this phenomenon. Loading was generated
due to earthquake in along the traffic and perpendicular to the traffic direction. Scouring depth of
the pier and rived bed scour percentage with main river width are taken as parameter of the study.
Flexural stress in pile increases by 137% and axial stress was found increasing by 11% about original
no scoured condition for earthquake along the traffic direction while the flexural stress increases by
87% and axial load increases by 5% about original no scoured condition for earthquake in parallel
direction of traffic.

1 GENERAL INTRODUCTION on the structure. This study tries to focus on this

phenomenon exclusively.
Numerous current studies have indicated that only
static analysis is not enough to provide required
structural safety of any structure. So structural 2 OBJECTIVES AND SCOPES
response for dynamic disturbance must be evalu-
ated with static disturbance of the structure. In The major objective of this study and analysis is
safety assessment, the effect of bridge pier scour- to enhance our knowledge about pier scouring and
ing on pier design parameters such as moment relevant changes in the design parameters in the
and axial load is observed for the event of earth- event of an earthquake.
quake. Excessive scouring leads to rapid increase The salient features of this study are briefly
in pier moment reducing the factor of safety of mentioned below.
the structure. A screening and evaluation pro-
gram in the United States reports that, of a total I. To model a reinforced concrete cantilever
of almost 500,000 bridges, there are more than box girder bridge with nine continuous spans
26,000 scour-critical and more than 26,000 scour- using beam element.
susceptible bridges (Pagàn-Ortiz 2003). Again in II. To apply earthquake in perpendicular direc-
recent years, in Bangladesh, Meghna Bridge situ- tion to the traffic and in the parallel direction
ated on the most important route of the coun- to the traffic of the bridge.
try (National Highway N1), was detected to be III. Perform modal analysis and single point
severely scoured threatening to cripple the eco- response spectrum analysis with 5% damping
nomic trade route of the country with the outside ratio.
world. Therefore it is necessary to assess the effect IV. To observe the changes in the axial load and
of scouring in the overall integrity of the bridge moment in piles by varying the scouring depth
structure. Generally in most of the bridge design and number of piers scoured.
it is assumed that the riverbed level will remain
within the level of pile-cap. For this support con-
dition, relevant formulas in AASHTO bridge 3 THE MEGHNA BRIDGE
design were used and the member size and shape
of the structure was calculated with some factor Bangladesh is a riverine country and differ-
of safety. But with large amount of scouring in ent parts of this country are divided from each
around any pier increases its unsupported length other by a number of small and large rivers.
and also changes the design parameters. With these Meghna is one of the large rivers, which has
changes in moment or axial load if earthquake is divided the southeast part from the other parts
induced it might have some negative consequences of the country, especially from the central part.

539

ISGSR2013.indb 539 10/18/2013 9:45:20 AM

 Meghna Bridge was constructed between 1986– an updated assessment of the magnitude of a 1 in
1991 (Figs. 1 and 2). It is a prestressed con- 200 year seismic event. The area in which Meghna
crete cantilever box section bridge with hinges Bridge is situated is classified as zone 2 in the
at the middle of every span. The overall length BNBC zoning map. Before the BNBC revision the
of the bridge is 930 m. The Bridge is situated corresponding ground acceleration of zone 2 was
on the 25th Km of Dhaka-Chittagong highway 0.05g and after the revision the value was updated
N1 (Latitude 23°36'20.98"N and Longitude to 0.15g. This represents a threefold increase in
90°36'50.32"E). The Bridge not only connects seismic ground accelerations on the structure com-
the capital city Dhaka with port city Chittagong pared with the values adopted in the original design.
but also acts as a gateway of other countries of The amount of confining reinforcements currently
South-Asia through Myanmar. provided in the bridge piers also do not comply
with the requirements of the AASHTO standard
3.1 Present scenario of Meghna Bridge specifications 16ED. Details in this regard can be
found at Final design Report on Rehabilitation of
Construction of Meghna bridge was completed
Existing Meghna Bridge, Roads and Highways
in 1991 to contemporary Bangladesh seismic
Department (RHD), 2011.
standards. From the relevant documents of the
Toll collection and Operation and Maintenance
bridge, it was found that, for design, seismic hori-
Operator currently responsible for the regular
zontal loading was taken as a static load equiva-
maintenance of the bridge.
lent to 5% of the vertical dead loads. The seismic
They have reported that due to the last three
zoning map in the Bangladesh National Building
consecutive floods riverbed erosion has taken
Code (BNBC) was revised in 1994, which reflects
place.
Underwater survey conducted in May 2010
revealed severe scouring of bridge foundation,
especially in pier locations P5 to P10 of Meghna
bridge with more than 15 m of scouring at P8.
Scouring is directly related to the shear stress or
velocity. As per the past report the velocity for 1
in 50 and 1 in 100 year flood event are 1.31 and
1.36 m/s at the vicinity of the Meghna bridge.
According to FAP 9B, design roughness and the gra-
dient between Bhairab Bazar and Meghna bridge
location was determined and these were Manning’s
coefficient Km = 40 and slope S = 2.5 × 10−5. Using
these values, for maximum water level of 22.5 m at
the Meghna bridge location of 5.20 m PWD, the
velocity was calculated to be 1.58 m/sec.
Water level variation of Meghna Bridge along
the year is shown in Figure 3.
However detailed bathymetric surveys show that
near the Meghna bridge water depth is over 35 m,
which results in a flow velocity of 2.15 m/sec. So
this increased velocity is one of the reasons behind
Figure 1. The Meghna Bridge. this pile scouring.

Figure 2. Bed profile of Meghna Bridge (all dimensions are in mm).

540

ISGSR2013.indb 540 10/18/2013 9:45:20 AM

 developed. Meghna Bridge is 930 m long with a
width of 9.2 m. It is a prestressed concrete box
girder bridge with total thirteen spans. The first
two spans of the bridge are 48 m long while the
next nine are 86.840 m long. The rest two spans
are 24 m long. Typical deck slab of the bridge
is hinged on both sides. The deck slab portion
of the Meghna bridge is curved at the bottom,
with its dimension thickest near the pier and
thinnest at the hinge side. For simplicity a model
was developed with 9 equal spans each having a
Figure 3. Maximum water level at Ferry Ghat Station length of 87 m. Pier height and pile length were
of the Meghna Bridge. also assumed to be of uniform dimension. Pier
height of 27 m and pile Length (L) of 48 m were
chosen for the developed model. To incorporate
the curved architectural feature of the deck slab
4 METHODOLOGY
portion of the bridge, the deck slab was divided
into 10 equal segments on both sides of the each
Finite element analysis and engineering simulation
pier. A typical cross section of the deck slab near
software ANSYS 11.0 was used for the complete
the pier location is shown in Figure 4. For crea-
analysis in this study. A model of nine spans of
tion of the model, beam type element was used
box girder cantilever reinforced concrete bridge,
for deck slab portion and pier. Shell type element
with simply support condition at both ends was
was used for pile cap and pipe type element was
developed using this software. Earthquake load
used for pile creation.
was applied in perpendicular direction to the traf-
The effect of earthquake on the bridge structure
fic and in the parallel direction to the traffic of the
was assessed for the dead load of the structure
bridge by means of single point response spectrum
only. The necessary material property used in the
analysis with 5% damping ratio. As a prerequisite to
model is given in Table 1.
the single point response spectrum analysis, modal
analysis of the structure was also performed. With
these loading conditions, scouring depth of the
pile along with number of scoured pier was varied
to observe the change in pile design forces.

4.1 Assumptions
The investigation was carried out on the basis of
some assumptions to avoid complications. These
assumptions are as follows:
I. Materials involved in the study were elastic.
II. Only lateral force that was encountered by the
structure is the Earthquake.
III. In analysis for earthquake only dead load of
the structure was considered. No live load was
taken into consideration.
IV. Linearly increasing soil subgrade modulus
was considered.
Figure 4. A typical cross section of deck slab (all
V. Soil scouring profile around the bridge piers
dimensions are in mm).
was assumed to be parabolic.
VI. All the calculations were made on the basis of
pure loading. No scaling of load was done for
single point response spectrum analysis. Table 1. Material property.

Modulus of Soil subgrade

5 MODEL DEVELOPMENTS Poison’s elasticity, E Density modulus, k
ratio (N/mm3) (Ton/mm) (N/mm3)
In this study, a balanced cantilever bridge with 0.2 20000 2.4 × 10−9 0.02
similar attributes like Meghna Bridge was

541

ISGSR2013.indb 541 10/18/2013 9:45:22 AM

 5.1 Boundary condition
Like the support condition in the Meghna Bridge,
the abutments in the both ends of the developed
model were simply supported. The piles were fixed
supported but fixing length of the pile had to be
identified. In Figure 5 it can be seen that both free
and fixed headed pile are subjected to a lateral
load Hu, in these circumstances, piles act as a sim-
ple cantilever. Here e is the height from the ground
surface to the point of application of the load and
zf is the depth to virtual fixity. Together (e + zf)
is called equivalent height. With the increase in
riverbed scouring, the value of e will increase and
so will the unsupported length. To calculate the
value of zf two stiffness factors will be required,
depending on whether the soil modulus is constant
or increases linearly. For constant soil modulus,
stiffness factor R is used and for linearly increasing
soil modulus, stiffness factor T is used.

EI
Stiffness factor, R = 4 (1)
kB

In stiffness factor R, k is the soil modulus.

EI
Stiffness factor, T = 5 (2)
nh

In stiffness factor T, nh is the coefficient of

horizontal subgrade reaction.
ACI recommends that for long pile the value of zf
can be taken either 1.4R or 1.8T. But for granular
soils and normally consolidated clays, generally, the
best practice is to assume linearly increasing soil
modulus. That is why to calculate the value of zf
1.8T has been adopted. For the developed model,
the coefficient of the horizontal subgrade reaction Figure 6. Fully developed model with proper boundary
was assumed to be 0.01078 N/mm3. condition in no scouring condition.
Table 2 shows the relation between pile type and
soil modulus. The formula that was used as per
ACI to calculate zf is for long elastic piles. So it is
Table 2. Relation between soil modulus and pile type.

Soil modulus

Pile type Linearly increasing Constant

Rigid (free head) L ≤ 2T L ≤ 2R

Elastic (free head) L ≥ 4T L ≥ 3.5R

important to make sure that the pile in the model

was a long and elastic one. From the developed
model it was observed that L ≥ 4T.
Figure 5. Piles under horizontal loads are considered as The vertical and lateral Degrees of Freedoms at
simple cantilevers. both ends of the balanced cantilever portion were

542

ISGSR2013.indb 542 10/18/2013 9:45:23 AM

 is for earthquake in parallel direction of traffic and
Figure 9 is for earthquake in perpendicular direc-
tion of traffic. The Black, Golden and Blue graphs
are for 10%, 40% and 60% riverbed scouring
respectively. Peak moment in the pile is achieved
at different scouring depths in case of earthquake
in parallel direction of traffic. With no scouring
condition, the moment is around 1500 kN-m and
maximum value of moment is around 2800 kN-m
at 12 m scoured depth for 60% river bed scouring.
The percentage increment is around 87%. At 18 m
of scouring depth, the value of moment becomes
almost same regardless the percentage of riverbed
Figure 7. Segments of bridge deck slab (isometric scoured. For earthquake in perpendicular direction
view). of traffic, moment increases with the increment
of scouring depth. In this condition, maximum
moment for 10% riverbed scouring is achieved at
Table 3. Parametric change employed in the study. 12 m scouring depth and for 40% of the riverbed
scouring, the maximum moment is achieved at 15 m
Parametric of scoured depth. Among the moments for earth-
Parameter name change Consequence quake in perpendicular direction of traffic, maxi-
mum value is found to be at 18 m of scoured depth
Percentage 10% One pier
of riverbed scoured
scoured (as a 40% Three piers
percentage scoured
of river width) 60% Five piers
scoured
For each percentage of scoured river width
Scoured depth 0 m, 3 m, 6 m, Change in the
9 m, 12 m, fixing length
15 m and 18 m of pile

coupled with the adjacent deck slabs. In Figure 7,

the small red portion indicates the coupled DOF’s.
Coupling of DOF’s in vertical and lateral direc-
tion means any displacements in those directions,
between the adjacent deck slabs, occur together Figure 8. Moment variation for earthquake in parallel
but the displacement is free in longitudinal direc- direction of traffic.
tion of the bridge.

6 STUDY PARAMETERS

For this particular study and model, some param-

eters were chosen to evaluate their impact on
the structure on an occasion of earthquake. The
parameters that were studied summarized in
Table 3.

7 STUDY OBSERVATION

7.1 Moment variations in corner pile

First the flexural variation in the corner pile of the Figure 9. Moment variation for earthquake in
maximum scoured pier will be discussed. Figure 8 perpendicular direction of traffic.

543

ISGSR2013.indb 543 10/18/2013 9:45:24 AM

 Figure 10. Axial load variation for earthquake in Figure 11. Axial load variation for earthquake in
parallel direction of traffic. perpendicular direction of traffic.

for 60% of the riverbed scouring and the value is scouring and the change in moment and axial load
almost 1850 kN-m. With no scouring condition in the piles. As the riverbed scouring progresses,
the moment is around 820 kN-m and the percent- the fixing length of the pile also moves down-
age increment of moment is 137%. Although the ward making the bridge more susceptible to
percentage increment of moment is more in earth- lateral loading. For that reason, in the event of an
quake in perpendicular direction of traffic but earthquake with scoured riverbed condition, the
earthquake in parallel direction of traffic is more moment in the pile increases and the axial load
vulnerable because of the higher moment value in decreases.
the latter case.

8 CONCLUSIONS
7.2 Axial load variation in corner pile
Figures 10 and 11 shows axial load variation, in From above results and discussions it can be con-
corner pile of maximum scoured pier, for earthquake cluded that bridge pier scouring incorporated with
in parallel direction of traffic and for earthquake earthquake will have adverse effect in the overall
in perpendicular direction of traffic respectively. integrity and health of the structure. So proper
It has been found that for 10% and 40% of river- protective measures should be taken for the soils
bed scouring, axial load in pile decreases for earth- around the bridge piers to prevent scouring.
quake both in perpendicular and parallel direction
of traffic. For 60% of riverbed scouring, axial load
in pile increases by a small margin. The increment REFERENCES
percentage is about 5% for earthquake in parallel
Pagán-Ortiz, J.E. 2003. Scour program update. Proc.,
direction traffic and 11% for earthquake in perpen- Midwestern Hydraulic Engineering Conf., U.S.
dicular direction of traffic about the axial load in Department of Transportation, Federal Highway
original no scouring condition. Administration, Washington, D.C.
From the above numerical results, it can be Tomlinson, M.J. 1994. Pile Design and Construction
inferred that there is a relation between the riverbed Practice. 4th Edition, E & FN Spon.

544

ISGSR2013.indb 544 10/18/2013 9:45:26 AM

 Geotechnical Safety and Risk IV – Zhang et al. (eds)
© 2014 Taylor & Francis Group, London, ISBN 978-1-138-00163-3

Risk minimization by an adapted dewatering scheme at the construction

of the new ship lock in Minden

H. Montenegro, B. Odenwald & R. Kauther

BAW German Federal Waterways Engineering and Research Institute, Karlsruhe, Germany

ABSTRACT: The excavation design for the new ship lock in Minden resulted challenging due to the
immediate vicinity of the 14 m deep excavation pit to existing structures and surface water bodies. An
alternative to a fully braced variant consisted of anchored support walls and slopes to the opposite side.
The hydrogeological setting consists of fill materials and Quaternary sands above bed-rock. Borehole pulse
and pumping tests revealed preferential flow along a system of fissures in the claystone in which river stage
fluctuations can propagate very fast. The excavation design in a setting with a fair hydraulic connection
to the river demanded an effective groundwater containment system and a robust dewatering. In order
to minimize the level of risk an active-passive dewatering scheme was developed. The excavation design
proved to be robust and yielded significant cost savings compared to a fully braced support walls variant.

1 INTRODUCTION which was constructed in 1914. In the course of

the adaptation of the Weser Valley to larger ves-
The Midland Canal traverses the Weser Val- sels, the lock is being replaced by a new structure.
ley near Minden in Westphalia, Germany. At Benchmark data for the new ship lock are the
the crossing point, the water table of the canal length of 139 m, the width of 12.5 m and the 4 m
lies about 13 m above the mean water level of access way depth. Figure 1 shows the setting of
the Weser. The rising and the lowering for ship- the existing lock, the excavation site for the new
ping is currently being managed by a ship lock, ship lock as well as the adjacent water surfaces of
the Weser and the upper pool which connects to
the Midland Canal.
The new excavation site is no longer situated on
“a wet meadow”, which was the case during the
construction of the lock, but is rather in the imme-
diate vicinity of existing facilities. Therefore it had
to be ensured that the existing lock and the ongo-
ing navigation would not be affected or impaired
due to the adjoining excavation. This required pro-
found examinations of soil-structure-interactions
as well as proof of serviceability of the existing
lock during any excavation stage. As the pit cuts
into the surface water of the upper pool and the
Weser (13 m–18 m above the pit base), any leak in
the canal lining of the upper pool as well as poten-
tial leakages from ship impact at the pit walls need
to be considered explicitly.

2 GROUND CHARACTERIZATION

The ground in the planned excavation site can

be divided into three strata. Directly under the
ground surface there are up to 12 m natural fill
material. Those underlie alluvial silt and loam as
Figure 1. Aerial view of the existing lock in Minden well as sands and gravels of a predominantly high
and the construction site for the new lock. density of up to 3.5 m layer thickness. Below these

545

ISGSR2013.indb 545 10/18/2013 9:45:27 AM

 Quaternary layers lies tertiary claystone in a thick- 4 EXCAVATION PIT CONCEPT
ness beyond the target depth of the ground explo-
ration. The claystone belongs to the group of weak With the objective to control extensive excavation-
rocks and has a distinctive orthogonal system of induced deformations and to avoid any damage
discontinuities. This consists of steep joints with in the adjacent structures the pit design evolved
an inclination between 45° and 90° and a low dip- around braced or anchored excavation support
ping bedding. The strength parameters for the walls. First calculations revealed however limited
compact rock and the rock and interfaces were anchor forces in the claystone. An alternative to a
determined on the basis of in-situ and laboratory fully braced variant consisted of anchored support
shearing tests (uniaxial compressive strength on walls and slopes to the opposite side. This alter-
average 7.7 MN/m2 in vertical and 9.2 MN/m2 in native demanded a robust control of groundwater
horizontal direction). inflow by an effective dewatering.
On basis of preliminary groundwater flow cal-
culations an excavation concept was developed.
3 HYDROGEOLOGICAL CONDITIONS Rigid excavation support walls were established
on the western side as a bored pile wall anchored
In the hydrogeological analysis of deep excavations back fourfold. At the opposite side the pit was
in close proximity to rivers the appropriate estima- sloped with an inclination of 70° in the claystone.
tion of effective conductivity and hydraulic connec- The Quaternary soil layers above were sloped at an
tivity to adjacent surface water bodies is essential. angle of 1:1.5 to 1:2 as shown in Figure 2.
The analysis of pore pressure measurements in the
claystone revealed variations in response to the
filling and emptying of the existing lock, even 5 GROUNDWATER CONTAINMENT
at a considerable distance (see Fig. 2). The rapid SYSTEM
(in relation to the conductivity of the claystone)
changes of effective stress due to the filling and Inflows from adjacent surface waters through the
emptying of the chamber induce an immediate rise permeable Quaternary sands was limited by the
in pore pressure beneath the lock which can effec- construction of a impervious containment cutting
tively propagate in space to the installed pressure
transducers. These observations suggested fissured
flow in the claystone.
Hydraulic properties of the claystone had to be
determined at the appropriate scale. Borehole pulse
tests and pumping tests revealed preferential flow
along a system of fissures. The effective hydraulic
conductivity of the claystone was characterized with
correspondingly large values in the range of 10−5 m/s.
Responsible for such unanticipated high values is
the spatially varying fissure interconnectivity.
The hydraulic conductivity of the quaternary
sediments was determined on the basis of labora-
tory tests in the expected range of 10−3 m/s. These
results led to the conclusion that a fair hydraulic
connection to the Weser was present, not only
in the permeable Quaternary sediments but also
through the fissured claystone.

Figure 2. Cross section of the excavation pit for the new Figure 3. Arrangement of the containment system
ship lock in Minden. around the excavation pit.

546

ISGSR2013.indb 546 10/18/2013 9:45:28 AM

 into the claystone surrounding the excavation pit.
Components of the containment system are the
existing lock (based in the claystone), the sheet-pile
walls of the cofferdams for the prospective mole
and the upper and lower protective structures and
a surrounding sheet-pile wall.
The map in Figure 3 shows the arrangement of
the components of the containment system, the
existing ship lock, the bored pile wall on the west-
ern side, the slopes on the eastern side and the pro-
tective cofferdam structures in the ponds.
The fissure system of the claystone in which
river stage fluctuations can propagate very fast
demanded required an effective and robust
groundwater lowering system. Different dewater-
ing approaches were studied to reliably control
groundwater heads and reduce actions from pore
pressure on the pile wall and seepage forces on the
slopes. Figure 5. Schematic illustration of the dewatering sys-
tem during passive operation.

6 DEWATERING SCHEME

Due to the limited anchor forces of 1000 kN accept- through the bored pile wall. Figures 7 and 8 show
able groundwater heads (according to the static schematically the active and passive dewatering
loading design of the bored pile wall) had to be approach respectively.
met. There are essentially two approaches to con- The passive operation has a considerable advan-
trol actions from groundwater on excavation sup- tage in the case of failure of the pump system
port walls. On the one hand, “actively” by means (e.g. by lightning strike etc.) as no inadmissible
of groundwater lowering through wells outside the head differences can build up behind the pile wall.
excavation pit, and on the other hand, “passively” In the worst case the pit is inundated in contrast
discharging the groundwater into the excavation pit to the active operation mode where a pumping
taking advantage of the groundwater head differ- breakdown can result in a rapid rise of groundwa-
ence outside and inside the pit. The groundwater ter heads and loadings on the excavation support
discharge and so the lowering can be considerably walls.
enhanced if groundwater inflow is short-circuited Further, a passive dewatering provides opera-
tional benefits since no redundant pump system
and permanent monitoring is needed as is the
case during active groundwater lowering to mini-
mize the risk of inadmissible static heads/loading
behind the support wall due to unanticipated fail-
ure of the pumps.
On the other hand a passive dewatering scheme
right from the start of the excavation on was not
appropriate. This would require groundwater
discharge through the bored pile wall at several
excavation levels. Therefore an active dewatering
based on vertical extraction wells (see Fig. 6) was
established during the excavation. After reach-
ing the final excavation level, the dewatering was
shifted from active to passive operation. To short-
circuit groundwater through the pile wall inclined
boreholes were drilled from the excavation base
to the vertical extraction wells and supplied with
a casing. Taking advantage of the head difference
outside and inside the pit, groundwater flows is
Figure 4. Schematic illustration of the dewatering sys- captured outside the pit by the vertical wells, dis-
tem during active operation at the final excavation level. charged through the wall in the obliquely-installed

547

ISGSR2013.indb 547 10/18/2013 9:45:29 AM

 Figure 6. Installation of the vertical extraction wells for
the groundwater lowering during active dewatering.

casings and ultimately conducted to the drainage

system at the base of the excavation. Both dewa-
tering operational modes were examined based on
spatial groundwater flow calculations. Figure 7. 2D-large-scale model (green) and the
3D-model (red). Lines in the area of the excavation pit:
calculation sections.
7 GROUNDWATER FLOW
CALCULATIONS

The necessary control of extensive deformations

in the existing structures demanded for reliable
groundwater flow and pore water pressure calcula-
tions, particularly behind the pile walls and in the
slopes. For this purpose, groundwater flow models
at different scales and spatial resolutions were built.
At first, the effects of the containment system on
the regional groundwater flow were calculated in
a 2D-model (large-scale model, see Fig. 7). Then Figure 8. Cross section of the 3D-excavation pit model.
Ochre: Quaternary sediments, blue: claystone, violet:
the above outlined excavation pit and a series of
groundwater containment system.
extraction wells required for the active dewatering
was modeled in a 3D-model (excavation pit model
see Fig. 8). Finally, detailed calculations along rel-
evant cross sections in 3D-models of high spatial
resolution were performed to design the dewa-
tering system as well as to calculate the actions
from pore water pressure and seepage forces on
the pile wall and the slopes. Within this modelling
approach, the detailed model represents a cut-
out from the excavation pit model and the latter,
in turn, corresponds to a snippet from the large-
scale model. Such a method has the advantage that Figure 9. 3D-detailed model for the dewatering design.
the initially unknown boundary conditions can be Blue: groundwater surface during active dewatering.
adopted from the respectively larger scaled model
(Montenegro & Kauther 2010).
Though the topography was captured in detail in 8 DEWATERING SYSTEM DESIGN
the excavation pit model (see Fig. 8), the spatial reso-
lution and so the accuracy of the results were not The dewatering system during both active and pas-
satisfactory as the radial components of groundwa- sive operation was designed according to the static
ter flow in the vicinity of the extraction wells cannot loading in terms of heads/pore pressure distribu-
be adequately represented (Montenegro & Odenwald tions on the pile wall. Calculations considered a
2009). These limitations demanded for the detailed 100-year flood by applying the corresponding river
models of higher spatial resolution (see Fig. 9). stage as boundary condition. Further damage of

548

ISGSR2013.indb 548 10/18/2013 9:45:31 AM

 the protective structure in the upper pool by ship slip surfaces were evaluated. The groundwater
collision was taken into account by specifying leaks heads halfway between two wells are design rel-
along the water-side sheet pile wall in the excava- evant because the potentials are higher there as
tion pit model (see Fig. 8). compared to the position of the wells where the
Based on the detailed models (see Figs. 9 and 10) largest draw down occurs. Finally failure of single
seepage forces behind the bored pile wall and in the pumps was considered which led to an extraction
slopes of the eastern side were calculated at relevant well distance of 8 meters.
cross sections. For different excavation stages pore The arrangement of the extraction wells and the
water pressure distribution at the corresponding withdrawal rate [m3/h] at the beginning of active
operation (well drawdown in all extraction wells at
NN+27 m) is shown in Figure 11. The irregular dis-
tribution of withdrawal rates indicates higher inflows
in the northern area which is reasonable due to the
shorter distance to the river. The extraction wells
behind the short eastern bored pile wall show higher
withdrawal rates as well which can be ascribed to the
comparably larger catchment area of those 8 wells
compared to the 27 wells on the western side.
Due to the fundamental importance of the
Figure 10. Groundwater flow at a well during active groundwater lowering for the static loading of the
operation (groundwater surface: blue, groundwa- anchored bored pile walls a redundant pumping
ter head distribution: yellow) from detailed model system was mounted consisting of two independ-
3D-computations. ent pumps in each extraction well.

Figure 11. Initial withdrawal rates [m3/h] at the start of Figure 12. Well drawdown [NN+m] at the start of the
active dewatering. Well drawdown at NN+27 m. passive operation.

549

ISGSR2013.indb 549 10/18/2013 9:45:33 AM

 control extensive excavation-induced deformations
and to avoid any damage in the adjacent structures
the pit design evolved around fully braced exca-
vation support walls. An alternative consisted of
anchored support walls and slopes to the opposite
side.
The particular hydrogeological setting consists
of fill materials and Quaternary sands above bed-
rock. Pore pressure measurements and hydraulic
tests in the claystone revealed that river stage fluc-
tuations can propagate very fast and may increase
groundwater heads even at large distances due to
flow along a system of fissures. The pit design in
a setting of permeable sands above fissured clay-
Figure 13. Aerial view of the pit for the new ship in
stone, both in fair hydraulic connection to the river
Minden at the final excavation stage. demanded an effective groundwater containment
system and a robust dewatering design.
An extensive groundwater flow analysis at vary-
Further a groundwater monitoring system ing scales was carried out. Groundwater head
was installed in every well and at other relevant distributions for relevant design situations were
stations to supervise the groundwater lowering. predicted based on a 3D model of the entire excava-
Measurement were instantly available via a server tion pit. Based on detail models at a higher spatial
application. For selected stations warning and alert resolution the arrangement of vertical extrac-
groundwater heads were established and automatic tion wells was optimized considering admissible
alarms via E-mail and SMS message were set up. groundwater heads behind the bored pile wall.
After 6 months the final excavation level was In order to minimize the level of risk during
reached and dewatering mode was shifted to pas- excavation an active-passive dewatering scheme
sive operation establishing so a much lower risk was developed. The excavation design proved to
level. Observed water levels in the wells after the be robust and yielded significant cost savings com-
transition to passive operation are depicted in pared to a fully braced support walls variant.
Figure 12. The drawdown in the wells reflects the
level in the drainage system at the base of the pit
maintaining so the heads behind the wall signifi- REFERENCES
cantly below the established limit values for that
Montenegro, H. & Odenwald, B. 2009. 3D groundwater
particular excavation stage. Only near the southern flow modelling at the excavation pit for a ship lock
gate increased drawdown values appear as in that in Minden, Germany. In 14th Danube-European
sector no relief casings were installed due to lack- Conference on Geotechnical Engineering DECGE 2010,
ing static necessity. A view of the excavation at the Bratislava, 2–4 June 2010.
final stage just after changing to passive mode is Montenegro, H. & Odenwald, B. 2009. Analysis of spa-
shown in Figure 13. tial groundwater flow for the design of the excava-
tion pit for a ship lock in Minden Germany. In 2nd
International FEFLOW User Conference, Potsdam,
14–18 September 2009.
9 SUMMARY
Montenegro, H. & Kauther, R. 2010. A multi-scale
approach for the consideration of spatial groundwater
The design of the 14 m deep excavation pit for flow in the stability analysis of a large excavation pit.
the new ship lock in Minden resulted demanding In 7th European Conference on Numerical Methods in
due to the immediate vicinity of the pit to existing Geotechnical Engineering NUMGE 2010, Trondheim,
structures and to surface water bodies. In order to 2–4 June 2010.

550

ISGSR2013.indb 550 10/18/2013 9:45:35 AM

 Geotechnical Safety and Risk IV – Zhang et al. (eds)
© 2014 Taylor & Francis Group, London, ISBN 978-1-138-00163-3

Geotechnical instrumentation monitoring for the construction

of the West Kowloon Terminus of the Express Rail Link

A.K.O. So, P.W.L. Ko & V.K.W. Man

MTR Corporation Limited, Hong Kong, China

ABSTRACT: In Hong Kong, it is a statutory requirement to provide a monitoring plan where the con-
struction works may affect any nearby building, structure, land, street or services. The traditional use of
the triple-A response values has many drawbacks when the projects are large and complicated. A modified
instrumentation monitoring plan is therefore proposed. Instruments are grouped to serve two purposes.
The first grouping is for the protection of the external sensitive receivers. They are categorized into types
for different response actions as different sensitive receivers can tolerate different extent of settlement
and angular distortion. The second grouping is for the performance review of the excavation and lateral
support systems. Another set of response values are determined from the predicted movements of the sys-
tems at various critical stages. Failure mechanisms are identified from the geological models and PLAXIS
analysis. Mitigation and contingency measures are considered in advance.

1 INTRODUCTION to be agreed with the relevant government

department or utility company and to be based
In Hong Kong, it is a statutory requirement that on the amount of movement that services could
a monitoring plan is to be provided where the tolerate (BD 2004). However, the response levels
construction works may affect any nearby build- imposed by the relevant government department
ing, structure, land, street or services (BD 2004). or utility company are generally conservative and
Monitoring check points for the ground settle- engineers are obliged to follow, aiming for easy
ment, pavement settlement, utility settlement, approval by the BD/GEO. In February 2012, BD
building and structure settlement and tilting, and provided details of the monitoring requirements
geotechnical instruments such as various types in APP-18 and APP-137 in the Practice Notes
of piezometers, extensometers and inclinometers for Authorized Persons (AP), Registered Struc-
are installed. These check points and instruments tural Engineers (RSE) and Registered Geotech-
are monitored against three triggering or triple-A nical Engineers (RGE) or the PNAP. The AAA
(AAA) levels for response actions. They are to be levels are defined as “Alert, Alarm and Action”.
approved by the Buildings Department (BD) and Reference example of contingency measures for
Geotechnical Engineering Office (GEO) based piling works for each level to be carried out by
on the distance of the sensitive receivers from the RSE and Registered Specialist Contractor is
the construction works. In the case of the first given (PNAP APP-18). Ground settlement limits
A level is reached, the Contractor shall inform resulting from the piling and similar operations
the Engineer within 24 hours of exceedance of are recommended (Appendix B in PNAP APP-
the first A level and shall increase the monitor- 137). Reference AAA values for the ground set-
ing frequency and/or monitoring check points tlement markers, services settlement markers and
and instruments. When the second A level is building titling markers are given based on serv-
reached, the Contractor shall review the construc- iceability requirements. As different structures
tion method in order to mitigate the detrimental will have different tolerance in accommodating
effects arising from the ground settlements and/ the movements of their foundations, acceptance
or movements. If the third A level is reached, the of estimated ground settlements should be con-
construction works have to cease and a thorough sidered on a case-by-case basis with respect to
investigation report with an action plan shall be their integrity, stability and functionality of the
submitted to the Engineer and BD/GEO before supported structure.
resumption of the works. The traditional use of AAA values and the
Before February 2012, BD only required the recent monitoring requirements given in the PNAP
third A level for a particular type of service APP-18 and APP-137 are generally acceptable for

551

ISGSR2013.indb 551 10/18/2013 9:45:36 AM

 small projects. However, they have many draw- Hong Kong and the Futian Station of Mainland
backs if the projects are large and complicated, China. The Hong Kong section of the link com-
like the West Kowloon Terminus (WKT) of the prises an underground terminus (the WKT),
Express Rail Link constructed by the MTR Cor- an approach tunnel and about 26 km of tunnel.
poration Limited, which are to be discussed as The WKT site is located on a reclaimed land sur-
follows. rounded by the Austin Station on the east, Victoria
Harbour on the south, Kowloon Station and the
above-station development on the west and Jordan
2 THE WEST KOWLOON TERMINUS Road on the north.
The site has a complex history of reclamation
As shown in Figure 1, the Express Rail Link pro- and use. Figure 2 shows the geological stratigraphy
vides an efficient means of mass transport between of the site. The ground is generally flat with an exist-
ing ground level at about +4 mPD to +5.5 mPD and
a groundwater level varying at about +0.8 mPD to
+1.8 mPD. However, the groundwater level close
to the Harbour is influenced by tidal fluctuations.
The geological stratigraphy comprises a reclama-
tion fill overlying the marine deposits, alluvium
and residual soils (completely and highly decom-
posed granite).
Figure 3 shows the cross section and schematic
construction of the WKT station structure which
is about 550 m long, 250 m wide and 30 m deep.
A partial top down construction at the central
and full top down at the two ends of the station
is adopted. Bulk excavation takes place with a
cut slope supporting the diaphragm wall with a
Figure 1. Geographical location of WKT. slope of 1 in 2. Once the final excavation level is

Figure 2. Geological stratigraphy of the site (N-S direction).

Figure 3. Cross section across the station (W-E direction).

552

ISGSR2013.indb 552 10/18/2013 9:45:36 AM

 reached, construction of the central portion of instrumentation plan is explained and illustrated
the WKT structure takes place by the bottom-up with practical examples.
method. Progressive top-down construction of the
remaining WKT structure then takes place with
the floor slabs providing lateral support to the dia-
4 THE PROPOSED INSTRUMENTATION
phragm wall.
MONITORING PLAN

4.1 Basic concept

3 PROBLEMS PROBABLY ENCOUNTERED
IN TRADITIONAL USE OF AAA VALUES MTR is very proactive in instrumentation moni-
toring. According to the standard Material and
As the construction of the WKT station structure Workmanship Specification, the AAA values are
was commenced in March 2011, the traditional defined as “Alert, Action and Alarm”. The Con-
use of AAA values would be followed because tractor is required to submit a Generalized Plan of
detailed monitoring requirements in the PNAP Action Plan within 45 days of the date for com-
APP-18 and APP-137 were not available at that mencement of the Works. This Plan is a plan of
time. Because of the size and complication of the action determined in advance by the Contractor
project, over 1,000 monitoring check points and and approved by the Engineer for general response
instruments have to be installed. This becomes a actions to be taken by the Contractor in the case
very demanding daily task for the project team to of response levels being reached for each instru-
interpret the monitoring data, check the behav- ment type and certain groups of instruments and
iour of the Excavation and Lateral Support (ELS) key locations. It shall include emergency response
systems and validate the design assumptions and actions (i.e. Contingency Plan) to be taken by the
design parameters. Due to the variance in the Contractor in case of Alarm level is being reached.
ground geology, the great depth and extent of the A Detailed Plan of Action based on the Generalized
excavation, the nature and interfaces of the con- Plan of Action may be requested by MTR at Alert
struction works, large settlements of ground and level before reaching Action level status. However,
movements of ELS systems are anticipated. This these requirements may still be impractical for a
may imply that an extensive amount of investiga- large and complicated but fast track project. Thus,
tion reports are to be submitted to BD/GEO with MTR has streamlined the procedures so that some
many suspensions of works and reviews of ELS of the information as required by the Generalized
performance. Furthermore, the AAA values for the Plan of Action can be submitted through design
ground settlements are set traditionally as 15 mm, submissions and method statements when the con-
20 mm and 25 mm respectively before the revision struction activities become apparent. Furthermore,
of the PNAP APP-18 and APP-137 in February MTR has carried out more critical design checking
2012. These AAA values do not bear any direct in advance which is to be explained as follows.
relationship with the predicted settlements and Prior to the implementation of the modified
are quite misleading to use as the control levels. instrumentation monitoring plan, the Engineer has
Likewise, the estimated ELS system movements carried out detailed assessment of the construction
and the associated ground settlements in the ELS effect to the external sensitive receivers such as the
design are normally governed by the strength and existing buildings, structures, facilities, pavements,
serviceability requirements. They may not bear ground and utilities based on the method of con-
any direct relationship with the AAA values tra- struction, ground geology and geotechnical design.
ditional adopted for the ELS system movements The movements of the external sensitive receivers
and ground settlements. Despite many check and ELS systems are predicted. In different to the
points and instruments could have far exceeded traditional instrumentation monitoring plan, AAA
the traditional third A level, the works are repeat- levels are set based on the predicted values for the
edly permitted to continue construction after per- response actions. The instruments are grouped to
formance reviews. As a result of this large number serve two main purposes. One group is for the pro-
of exceedance of AAA levels, the alertness of the tection of the external sensitive receivers and the
project team to watch over the monitoring read- other group is for the performance review of the
ings may be reduced. ELS systems (could be extended to pile driving and
In order to overcome these drawbacks, a similar operations). Prescribed mitigation measures
modified instrumentation monitoring plan is are provided if the tolerable limits of the ground set-
introduced. This modified plan supplements the tlements and ELS system movements are exceeded.
site impact assessment report and is accepted This modified instrumentation monitoring plan is
by the BD/GEO and the concerned utility com- introduced and supplements the site impact assess-
panies in advance. In this paper, the modified ment report for the approval by the BD/GEO.

553

ISGSR2013.indb 553 10/18/2013 9:45:39 AM

 Table 1. Categories of sensitive receivers and response actions.

Sensitive receivers Response actions for exceedance of AAA

Categories Examples Settlement Angular distortion

Category A (important Buildings/structures Investigation reports and Investigation reports and

structures) on piles and mitigation measures if mitigation measures
the seawall required if required
Category B (semi- Gas main, fresh and No investigation reports or Investigation reports and
flexible utilities) salt water main mitigation measures except mitigation measures
Storm water main at junctions to buildings if required
and sewer
Category C (flexible Cables such as No investigation reports or No investigation or mitigation
utilities and electricity, telephone mitigation measures except measures unless the change
pavements) and cable TV at junctions to buildings is abrupt
Carriageway and Investigation report and Investigation report and
footpath resurfacing if the change resurfacing if the
is abrupt change is abrupt

4.2 For the protection of the external sensitive and contingency measures are provided in advance
receivers by the Contractor and approved by the Engineer
and BD/GEO. During excavation, the design
For the external sensitive receivers, settlements
assumptions and design parameters are validated.
and angular distortions are the basic criteria
Monitoring instruments along a design section are
for the AAA values, which are the same as the
grouped for easy interpretations. Should the moni-
traditional AAA system and that recommended
tored ELS system movements and ground settle-
in the PNAP APP-137 (February 2012 revision).
ments deviate from the predicted trends, mitigation
However, different types of sensitive receivers can
measures are implemented. Should the monitored
tolerate different extent of settlement and angular
ELS system movements and ground settlements
distortion, and require different response actions.
reach the tolerable limits, contingency measures are
They can be categorized as in Table 1.
implemented and the Work is ceased immediately.
More thorough investigation report with further
4.3 For the performance review action plans are submitted to BD/GEO for approval
of the ELS systems before resumption of the Works.
In most situations, the estimated ground settle-
ment caused by the movement of the ELS system
5 MANAGEMENT OF THE
may not bear any direct relationship with the tra-
INSTRUMENTATION MONITORING
ditional AAA values for the ground settlements.
If, for example, the adjacent ground is bare or the
5.1 Management team
utilities are supported, the tolerable limit of ground
settlement can be larger than the traditional third In order to safeguard the unawareness of probable
A level, and the excavation can still be carried out catastrophic failure, it is the contract requirement
in a safe manner. Despite this, investigation reports to establish a joint monitoring team. The team
with mitigation measures are still required to be comprises the instrumentation team, geotechni-
submitted to BD/GEO and a new third A level has cal engineers, and the designer if necessary, of the
to be approved before resumption of the Works. Contractor, and the geotechnical engineers, con-
In the modified instrumentation monitoring struction engineers and detailed design consultant
plan, the concept of minimum expected value, from the Client, and an Independent Monitoring
most probable value, and maximum expected Consultant (IMC).
value in the MTR Practice Note PIM/PN/11-22
is extended such that a new set of AAA values for
5.2 Management process
design review are set from the predicted movements
at several critical stages. This is similar to the tol- During construction, the Contractor would
erable ground settlement limits given in the PNAP measure the instrument readings and upload
APP-137 (February 2012 revision). All probable them to a unified web-based instrumentation
failure mechanisms are identified, and mitigation database daily at a specified format agreed with

554

ISGSR2013.indb 554 10/18/2013 9:45:39 AM

 the IMC. The IMC would carry out a guarantee 6.1 Monitoring for the protection of the external
check to ensure all the uploaded monitored data sensitive receivers
are correct. Should there be any exceedance of
The works along a carriageway are chosen for
the AAA value, a message would be sent to the
illustration. These involved the installation of dia-
responsible persons. The geotechnical engineers
phragm walls, temporary cutting of a JGC stabi-
of the Contractor, the geotechnical engineers of
lized slope, shoring and deep excavation for the
Client and the IMC would meet on Monday and
station structure and the installation of socket
Wednesday to have a quick review of the moni-
H-piles, sheet piles, shoring and deep excavation for
tored data. The joint monitoring team would
the taxi lay-by. External sensitive receivers affected
meet on Friday to have a more thorough review
by the construction were given in Table 2.
of the monitored data and the weekly geotechni-
For the instruments monitoring the Austin Sta-
cal instrumentation monitoring report prepared
tion, they were interpreted in traditional manner.
by the Contractor. The geotechnical engineer of
However, for the instruments monitoring the fresh
the Contractor would report the main areas of
water main, they were grouped in string to deter-
concern of the report. Mitigation actions would
mine the angular distortion. As shown in Figure 4,
be confirmed to implement if the measured data
despite the settlement at a utility marker increased
exceeded the Action level is reached. Works would
with time and reached 36 mm, the angular distor-
be suspended temporary if the measured data
tion at that point was 1:4712 only which is still
exceeded the Alarm level. The RGE and BD/GEO
within the tolerable limit. Thus, submission of the
would be notified immediately. Remedial measures
investigation report to BD and implementation of
would be proposed by the Contractor, reviewed by
mitigation measure are not necessary. As a result,
the Client, RGE and BD/GEO and implemented
there was no interruption to the works and the BD
before resumption of the Works.
was notified by RGE instantly through email and
then by RGE’s T5 weekly and monthly reports. For
5.3 Management reports the instruments monitoring the carriageways and
pedestrian footpaths, their surfaces were main-
All monitoring results would be uploaded daily by
tained regularly to avoid any abrupt change in lev-
the Contractor to the unified web-based instru-
els, and the measured readings could therefore be
mentation database which is assessable by the joint
reset and maintained within the tolerable limit all
monitoring team and designated persons. Detailed
the time. Likewise, submission of the investigation
investigation reports would be submitted by the
report to BD is not required. As a result, there was
Contractor for any AAA exceedances. A weekly
no interruption to the works and the BD was noti-
geotechnical instrumentation monitoring report
fied by RGE instantly through email and then by
as described in the previous section would be sub-
RGE’s T5 weekly and monthly reports.
mitted by the Contractor to the Client for review.
A weekly summary report would be prepared by
the Technical Competent Person (TCP) grade T5 Table 2. Sensitive receivers at the selected carriageway.
in the RGE Stream (RGE’s T5) to the top man-
agement of the Client, the key TCPs and BD. Category Sensitive receivers
A monthly report would be prepared by the RGE’s
T5 to the RGE, RSE and BD/GEO as traditionally A Austin station and MTR tunnel
required. B 200 mm diameter fresh water main
C Road D1 and D1A carriageway and a
pedestrian pavement
6 ILLUSTRATIVE EXAMPLES

In the WKT site, over 1,000 instruments, includ-

ing ground settlement markers, utility settlement
markers, standpipes, piezometers, inclinometers in
diaphragm walls, slopes and Jet Grout Columns
(JGC), were installed to monitor the construction
effects to the adjacent sensitive receivers. They
were named and grouped according to their loca-
tions for easy review and reporting. Furthermore,
an Automatic Deformation Monitoring System
(Tse & Luk 2011) was also installed to monitor the
MTR tunnel which is not to be discussed in this Figure 4. String of instruments to determine the angu-
Paper. lar distortion of the utility.

555

ISGSR2013.indb 555 10/18/2013 9:45:39 AM

 Figure 5. Group of instruments to study the performance of the ELS system.

6.2 Monitoring for the performance review

of the ELS system
The bulk excavation for the construction of the
station structure is chosen for illustration. As
shown in Figure 5, instruments at each design sec-
tion were grouped to study the performance of the
ELS system which involved the interactions of the
diaphragm wall, JGC and slope. The movements
of the diaphragm wall, JGC, slope and the settle-
ments of surrounding grounds, roads, footways,
buildings, utility services and other features that
would occur at each and every critical stage of
works were estimated for monitoring. Figure 6. Plotting of deflection against excavation at an
inclinometer.
Inclinometers in diaphragm wall at each design
sections are chosen as bench marks for close
monitoring and weekly reporting. In principle, Investigation and design review were carried out.
the Alarm level at a location is taken as the maxi- Mitigation measures including lowering the water
mum movement or settlement predicted in the ELS table at the back of the wall were implemented
design and associated settlement analysis. AAA and excavation in front of the east slope could be
values are set as 50%, 80% and 100% of the predic- resumed within a short period. The maximum wall
tion corresponding to the excavation level with a movement was found reducing from the original of
factor of safety not less than 1.2. For traditional 80.7 mm at excavation depth of −17 mPD to later
projects, they are normally set as discrete values 71.3 mm at excavation depth of −23 mPD, which is
corresponding to the final excavation depth and below the Action level corresponding to this exca-
profile for simplicity. vation depth.
As shown in Figure 6, because of the scale and
complexity in this project, they were set as discrete
values corresponding to 4 critical excavation depths 7 CONCLUSIONS
and profiles; i.e. −9 mPD, −17 mPD, −23 mPD
and −32.6 mPD. For intermediate stages, they A modified instrumentation monitoring plan was
should be the interpolated values represented by proposed following the concept of the General-
the yellow, green and red curves respectively. The ized Plan of Action and Detailed Plan of Action
measured deformed profile of the diaphragm wall in the MTR Standard Material and Workman-
is also examined. If the measured profile deviates ship Specification. Instruments are grouped to
very much from the predicted profile, the geological serve two purposes. The first grouping is for the
design parameters and model have to be reviewed. protection of the external sensitive receivers. They
For example, the trend of the maximum wall move- are categorized into types for different response
ment at an inclinometer was found increasing rap- actions as different sensitive receivers can toler-
idly to 80.7 mm and approaching the Alarm level ate different extent of settlement and angular
(i.e. 100% of the prediction) for −17 mPD. Excava- distortion. The second grouping is for the per-
tion in front of the east slope stopped immediately. formance review of the ELS systems. A new set

556

ISGSR2013.indb 556 10/18/2013 9:45:40 AM

 of AAA levels are determined from the predicted BD, 2012b. PNAP APP-137-Ground-borne vibrations and
movements of the ELS systems at various critical ground settlements arising from pile driving and simi-
stages. Probable failure mechanisms are identified lar operations. Buildings Department, Hong Kong
from the geological models and PLAXIS analysis Government.
MTR, 2009. Section 23-geotechnical instrumentation.
such that mitigation and contingency measures Materials and Workmanship Specification for Civil
can be considered in advance. The daily task for Engineering Works, Volume 3, MTR Corporation
the project team to interpret the monitoring data, Limited.
check the behaviour of the ELS systems and vali- MTR, 2010. PIMS/PN/11-22/A1-Reporting of exceed-
date the design assumptions and design param- ances in instrumentation monitoring readings to senior
eters is alleviated. The amount of investigation management. Project Integrated Management System
reports to be submitted to the BD/GEO with many PIM Practice Note, MTR Corporation Limited.
suspensions of works and performance reviews are Tse, C. & Luk, J. 2011. Design and implementation of
reduced. Large deviation from the prediction can automatic deformation monitoring system for the
construction of railway tunnel: a case study in West
be rectified at early stage with lesser efforts and lit- Island Line. In Proceeding of Joint International Sym-
tle interruption to the work progress. posium on Deformation Monitoring, Session 3H-02.

REFERENCES

BD, 2004. Code of practice for foundations. Buildings

Department, Hong Kong Government.
BD, 2012a. PNAP APP-18-Requirements for an excava-
tion and lateral support plan building (administration)
regulation 8(1)(bc). Buildings Dep., Hong Kong
Government.

557

ISGSR2013.indb 557 10/18/2013 9:45:41 AM

This page intentionally left blank
 Geotechnical Safety and Risk IV – Zhang et al. (eds)
© 2014 Taylor & Francis Group, London, ISBN 978-1-138-00163-3

The impact of geometry bedding toward slope stability in coal mining

Supandi
Mining Engineering Department, STTNAS University, Yogyakarta, Indonesia

H. Hidayat
Geotechnical Department, PT Borneo Indobara, South Kalimantan, Indonesia

ABSTRACT: Bedding geometry is important in supporting the slope stability, particularly on low-wall
part of coal mining. Elements that are part of geometry are bedding slope, slope height and bedding ratio.
The key to this analysis is inserting geology conditions mainly bedding geometry. Bedding ratio value is
calculated based on modeling geology in which more detailed modeling geology factor can be identified
better analysis in slope stability. Slope stability analysis is assembled based on correlation between dip
of bedding, bedding height and factor of safety. The result from this study is increasing bedding ratio
value will be followed by increasing the factor of safety and bedding height. Changes in dip of bedding
slope will impact on the slope stability and respond changes on factor of safety and bedding height. The
final of the study is providing geo-mechanic parameters purposed for mine planning design based on site
characteristic geology.

1 INTRODUCTION Through the structure of the sediment, it appears

the term bedding. Bedding is similar to a layer of
Slope stability analysis is done routinely in rock, in which two contact areas in the two lay-
mining, where it has some excavation and dump- ers confine bedding. Both of contact layers could
ing activities. Analysis needs to be done by an be a thin layer that has important role in the slope
appropriate method based on characteristics and stability. Generally, the contact material is at plastic
the existence of geology conditions to facilitate material and it has a bad geo-mechanic properties.
geometry of the slope on which the design can be While the material forming the layers, the potential
optimum. In general the purposed of slope stabili- experience became thinning and thickening, thus
ties analysis are: leading to the form of bedding ratio. Each material
has its optimum bedding height.
1. Determining slope stability;
Through lithology factor, weak zone or carbona-
2. Defining the landslide mechanism;
ceous zone, slope geometry and optimum bedding
3. Evaluating the support system and stabilization;
height information, drive in doing this analysis so
4. Optimizing the existing resources with a safety.
that all components are correlated into a single
Geotechnical investigation must be done to sup- unit in the slope stability. With a unity slope stabil-
port the parameter to improve quality analysis. ity analysis, the value of slope stability in specified
One of the components that should be investigated conditions is important in delivering this study.
is the geological condition in which the lithology
and structural geology factor became an impor-
tant component in the slope stability analysis. 2 BATULAKI BACK LIMIT
The limit equilibrium concept is to compare EQUILIBRIUM ANALYSIS
between the driving force and resisting force.
Driving force is controlled by slope geometry, in Analysis was performed on the mining concession
which material load, and resisting force is control- of PT Borneo Indobara—(Golden Energy Mine)
led by rock mechanics properties. Refers to this con- located in the Barito basin formation Warukin
dition, it is important to determine the mechanical in southern Borneo. In general, mudstone, clay
properties of the constituent stratigraphic along stone, sandstone and coal materials dominate the
the slope. The performance of Geological struc- material with hardness <1 MPa. Groundwater lev-
ture and sedimentary structures is also important els tend to be high and almost up to the surface
to know in order to determine the orientation from (∼3 m below surface). In the open mine mainly
discontinue and daylight structure. on coal mining, there are two terms high-wall and

559

ISGSR2013.indb 559 10/18/2013 9:45:41 AM

 low-wall (Fig. 1). High-wall slope is perpendicular geo-mechanic test material was resulted 18 kN/m3
to the layers and low-wall is in line to rock lay- unit weight, cohesion of 87 KPa and friction angle
ers (bedding). Mining activities conducted at PT of 22 degrees. Due to the fact that the failure
Borneo Indobara with open pit mine system by occurred at the carbonaceous zone but because of
mining along the bedding dip. This study focuses the limitation of thickness it is hard to do testing
on the slope stability analysis at the low-wall part so this value is not included in previous analyzes.
of open pit coal mine. With the landslide, it is carried the bulk sampling
Slope failure is occurred at the low-wall part at to the carbonaceous material and doing the geo-
Batulaki pit in early May 2012. The failure occurs mechanic test. The test results on this material
throughout the 200 m long, 40 m high and 5 m and the back analysis results obtained carbona-
thick of mudstone with around 80 Kton mate- ceous geo-mechanic properties are Cohesion = 0,
rial slides from low-wall part (Fig. 2). No injury, Phi = 15o and unit weight = 13 kN/m3.
fatality and damage caused by it. Contact areas In addition to the material properties factor
between carbonaceous zones about 20 cm of thick- used for subsequent analysis performed also an
ness with a 20 degrees dip of bedding control this analysis of the slope geometry. From the results of
failure. Another thing that encourages to the fail- field measurement data showed that the landslide
ure is the depletion downward where the obtained occurred at the height of 40 m and no instability
value bedding ratio <1. occurs at a height below 40 m. Under these condi-
To determine the slope stability at other loca- tions, the correlation between the bedding geom-
tions, then carried back analysis based on these etry to slope stability is important. With reference
incidents. From the available data, there are several to the conditions, in order to redesign and to
factors appear, that are dip of bedding, bedding review the slope stability at the site and surround,
height, bedding ratio and carbonaceous. According the analysis and modeling is applied to improve
to those factors, back analysis is done so it can be slope stability analysis on other location.
used as improvement in the future analysis and
design. Back analysis is done by involving all com-
ponents above with slip surface in the contact area 3 SLOPE STABILITY ANALYSIS
of the two materials. Back analysis conducted using
the Slides 6.0 with limit equilibrium-based, it can be Slope stability analysis is done based on charac-
calculated the actual value of the material’s mechan- teristic geo-mechanics properties based on labora-
ical properties. Based on the acquired properties, tory test and back analysis results. Samples taken
by drilling method and take bulk samples after
the incident so that the geo-mechanic parameters
are sufficient in this analysis. In concern, geologi-
cal conditions in this analysis are dip of bedding,
bedding height and bedding ratio. It is driven by
the slip surface pattern using full specified on
the carbonaceous zone, which assumed as a field
delimiter in the bedding ratio calculation. Based on
the explanation above, there are three main com-
ponents in this analysis: dip of bedding, bedding
height and bedding ratio.
Figure 1. Overview definition low-wall & high-wall. Analysis is done based on existing geological
condition, which relatively in common has similar
dip of bedding. However, in some places obtained
higher dip of bedding so the analysis conducted with
another dip of bedding value. Modeling carried out
on any dip of bedding so each of that can be obtained
at the optimum value of bedding ratio. Bedding ratio
used in the analysis was 1.5, 1 and 0.5.
The first modeling is done to make any correla-
tion to bedding ratio toward factor of safety and
bedding height. Any changes would have impact
to other condition, so that it needs to do modeling
in order to facilitate differences deviate from exist-
ing geological condition. Other modeling applied
to discover optimum geometry where the optimum
Figure 2. Failure analysis back area for this study. geometry is calculated at factor of safety 1.2 so

560

ISGSR2013.indb 560 10/18/2013 9:45:41 AM

 that all the information is in the model reflects the
optimum factor of safety. Modeling applied to the
optimum value to discover the correlation of dip of
bedding, bedding ratio and bedding height. When
creating the designs with existing field conditions,
it can refer to both of this model so it can be used
as guidance in the slope design.
Analysis is done is to discover the optimum
height for each slope value which in any bed-
ding slope was analyzed for each ratio. In one
bedding ratio is analyzed 3 times by changing
the bedding height. The bedding height used for Figure 5. Correlation slope height vs FoS for 35° of dip.
this study in the analysis for all ratios are every
20 m, 35 m and 50 m. From the differences bed-
ding height will get different factor of safety so it Table 1. Summary optimum bedding height and bedding
can be predictable the optimum bedding height for ratio.
another condition. This analysis was repeated for Bedding Bedding Bedding
each distinct bedding ratio value using the same ratio 0,5 ratio 1,0 ratio 1,5
analysis method to produce charts as Figures 3–5.
From the images above obtained information Height Dips Height Dips Height Dips
that the larger bedding ratio the greater factor of (m) (degree) (m) (degree) (m) (degree)
safety and the smaller bedding ratio followed by
the decline of the factor of safety. One bedding 14 20 19 20 24 20
ratio value can be obtained the correlation toward 10 25 16 25 19 25
factor of safety and bedding height where the 8 30 13 30 17 30
higher bedding height obtained the down of fac-
tor of safety. Based on the analysis contained in
Figures 3–5, it shows the optimum bedding height
for any bedding ratio. If using the stability opti-
mum values at 1.2, it can be obtained at any bed-
ding height for each bedding ratio in Table 1.

Figure 6. Correlation dip of bedding vs. optimum

bedding height for factor of safety 1.2.

From the above table, it can be used as a basis

model that correlates between bedding ratio and
Figure 3. Correlation slope height vs FoS for 20° of dip. the bedding height on each slope. On this mode-
ling chosen at factor of safety optimum so that the
values are not optimum excluded from modeling.
The final result is obtained by using factor of safety
value of 1.2 so that all information contained in
the model shows the value of factor of safety 1.2.
From the data in Table 1 can be made the model
shown in Figure 6.
Figure 6 shows the bedding ratio value on the x
and the bedding height on the y. The lines shown
in the model show the slope bedding value at the
optimum conditions (factor safety 1.2). Based on
the picture, it is concluded that the increasing bed-
Figure 4. Correlation slope height vs FoS for 25° of dip. ding ratio responded by the greater bedding height.

561

ISGSR2013.indb 561 10/18/2013 9:45:42 AM

 The greater dip of bedding driven decreasing bed- Modeling simply in predicting the slope stability
ding height and the greater dip of bedding push value based on the geology characteristics. From
down the factor of safety. the geology data obtained from the field or from
the geological model can be put in the model in
order to obtain the optimum geometry to give
4 APPLICATION MODEL
optimum bedding height guidance for each layers
or bedding. Model is effective in efforts to translate
Sediment-layered material made failure patterns
technical language into geo-mechanic common
that exist at the low-wall tend to be controlled by the
language which easy to understand.
contact zone. Contact layer usually as a formed of
carbonaceous so that when the stratigraphic profile
encounter this layer, then must be careful in determin-
ing the slip surface. The carbonaceous field stands as 5 CONCLUSION
the interface between layers so that the overall stabil-
ity low-wall built by individual bedding. Research conducted on the sedimentary basin with
On the fact, many lithology of rock layers and consist same rock formations that have relatively
carbonaceous zones where there is thickening the same geo-mechanic and geology conditions.
below, thinning above and parallel layers. All gave In this study tended to emphasize the relation-
impact on slope stability and should be consid- ship between the dip of bedding, bedding height
ered in slope analysis. When doing slope stability and bedding ratio. The three factors have greatly
analysis, first thing to do is making the geologic effect to the slope stability. Each material in spe-
cross section based on geologic model. From the cific slope has optimum bedding height which to
cross section will appear stratigraphy and slope be considered in the mining design.
geometry. After getting the stratigraphic profile, Analysis was conducted to determine the corre-
the main thing to do is identifying the contact lation between the factor of safety value towards
zone or carbonaceous. In this case both materials the bedding height and dip of bedding at a certain
referred to a weak zone. bedding ratio. From the study, it is concluded that
Weak zones identified then proceed to calculate the increase of bedding ratio followed by a declin-
the thickness of the layer where the weak zone is ing the factor of safety value, while the higher of
underneath. Measurement is done perpendicular to bedding height reducing the factor of safety value.
the layer or bedding and performed at top and bot- With those influence, it is necessary to find the
tom of the layer/bedding. The upper part measured optimum conditions.
on crest and the lowest point measured at the loca- The optimum conditions obtained from the
tion to be mined. The lowest point to be mined has modeling based on the analysis outcomes. The
many of factors that should be done based on the first model is performed to find the correlation
mining plan. From the thickness calculation of the between bedding ratio value toward dip of bedding
top and the bottom will obtain the ratio bedding slope and optimum bedding height. Other model
value. Bedding ratio 1 is the layer thickness should is finding the optimum geometry using a factor of
be equal between the top and the bottom. Bedding safety 1.2 decreasing from the previous model that
ratio <1 indicates the thickness of the bottom is is doing the correlation between dip of bedding
smaller than the top while the ratio >1 indicates the toward bedding ratio and bedding height.
thickness in the bottom is greater than the top. From the existing modeling assist a lot in iden-
Based on the explanation above, it should be tifying mining risk and can be used as guidance
considered in found bedding ratio is <1. It is based in constructing mining design. The field data or
on rock mechanic concept that pressure will get geology model data obtained can be put in the
greater at the toe to sustain more loads. Bedding model so that it can quickly gather the geotechni-
ratio <1 had a greater risk because it receives a cal risk.
greater pressure for a particular area.
By knowing bedding ratio <1, it is known that
there is a risk to the slope stability so that at the REFERENCES
analysis should pay attention to this condition.
Slip surface field followed the weak zones so that Eberhardt E. 2003. Rock slope stability analysis-utilization
the analysis must apply the fully specified in order of advanced numerical techniques. Geological Engi-
neering, UBC, Vancouver.
to force slip surface in this zone. Each layer in a Hoek, E. & Bray, J.W. 1981. Rock slope engineering,
particular slope at a certain point has optimum third edition. Institution of Mining and Metallurgy,
bedding height, so that applying the analysis and Vancouver.
modeling can help providing guidance in geotech- Hoek, E., Carranza-Torres, C. & Corkum, B. 2002.
nical risk or provide mine design parameters. Hoek-Brown failure criterion. Vancouver.

562

ISGSR2013.indb 562 10/18/2013 9:45:44 AM

 Geotechnical Safety and Risk IV – Zhang et al. (eds)
© 2014 Taylor & Francis Group, London, ISBN 978-1-138-00163-3

Risk identification and mitigation for construction of a subway

transfer station in Beijing

H. Wang & Z. Wu
Ranken Railway Construction Group Co., Ltd., Chengdu, China

D. Wang
Tianjin Institute of Urban Construction, Tianjin Key Laboratory of Soft Soil and Engineering Environment,
Tianjin, China

ABSTRACT: More risks associated with subway transfer station construction than typical
excavations due to larger depth and more interactions. Besides elaborate design and careful arrangements
for constructions, monitoring is a critical process to control, predict and mitigate risks emerging possibly
during construction, as well provides updated information to revise the design in a timely manner. This
paper introduces a famous subway construction project-Jiaomen West Station, a transfer station for the
existing subway Line 4 (M4) and the proposed new subway Line 10 (M10) in Beijing. Based on the project
information, this paper identifies the risks associated to the project, introduces the methods to reduce
and mitigate the risks; monitoring data are presented and discussed in association with the mitigation
methods. Information presented in this paper and lessons learned from this case study are useful to the
similar projects.

1 INTRODUCTION to immediate component (ground loss) and the

consolidation component. In order to reduce the
Since the excavation depth and the intensive inter- traffic congestion in metropolitan areas, China is
actions with existing buildings, tunnels and other experiencing substantial development of subways
human-built environment, subway transfer sta- or so-called Mass Transit Railways (MTR). Due to
tion construction has more risks associated than the complex of the project and the numbers of the
interim station and the tunnel construction with undergoing projects, it is not surprised to see many
Tunnel Boring Machine (TBM). Ground move- accidents reported. Deng et al. (2010) conducted a
ment control is the principle factor to be considered research to statistically analyze the numbers, fre-
in design and construction phases for reliable pre- quencies and patters of the subway construction
diction of ground settlement; other issues, which accidents based on a total of 126 subway construc-
are more or less related to the ground movement, tion accidents which occurred in China from 1999
such as the deflection of diaphragm walls, the con- to 2008. Their statistical analysis indicates that
sequent risk of the collapse of excavation as well the numbers of the accidents in China increased
as the damage to the existing structures are also 5 times in 2008 from 1999, and the occurrence rate
important to the project. in January and July is the highest among the twelve
Numerous literatures studied the risks associ- months. Analysis by Li et al. (2006) indicated that
ated with tunneling and subway station excavations. the frequency of factors causing risk accidents in
Sejonaha et al. (2009) categorized four types of China is: 16.4% (complex or unrevealing geological
failure, namely, (1) cave-in collapse; (2) significant and geotechnical conditions), 29.1% (poor design),
exceeding of expected deformation of the tunnel 36.4% (casual construction), 1.8% (poor monitor-
tube; (3) exceeding of acceptable progress subsid- ing and supervising), 6.4% (poor management)
ence trough and (4) disturbance of water regime of and 10% (other reasons).
in the surroundings. Gatti & Cassani (2007) stud- Based on a real and complex case study, this
ied the measures to control the ground loss in EPB paper indentifies the potential risks associated with
(Earth Pressure Balance) TBM tunneling excava- a famous subway construction project-Jiaomen
tions. Hulme et al. (1990) reported and analyzed West Station, which is a transfer station proposed
the ground settlement caused by construction of for the existing Beijing Subway Line 4 (M4) and the
Singapore MTR, and classified the settlement new Subway Line 10 (M10). This paper introduces

563

ISGSR2013.indb 563 10/18/2013 9:45:44 AM

 the corresponding risk mitigation measures proposed bottom slab of Jiaomen West Station for
adopted in the project, such as dewatering, sta- M10 is 27 m bgs, and a twin tunnels with a dimen-
bilization, pre-support measures and excavation sion of 10.25 m (width) by 9.15 m (height) will be
sequences as well as the construction monitoring excavated below the station for M4. The length of
plans. Finally, monitoring data are presented and the covered undercutting is 35.1 m. Three transfer
discussed to validate such measures. Information passages with different lengths were also to be exca-
presented in this paper with respect to this case vated using covered undercutting technique.
study is useful to the similar projects. The geological conditions of the site were shown
in Figure 2 and summarized in Table 1 with more
details. According to the geological investigation
2 RISKS IN CONSTRUCTION A SUBWAY report of the project, the ground surface elevation
TRANSFER STATION IN BEIJING is about 40.0 m, stable groundwater level varies
greatly for each layer. Layers and are aquifers,
As shown in Figures 1 and 2, the subway transfer layers 1 and 1 are aquitards. Fortunately, the
station, Jiaomen West, is located in the intersection groundwater in layers and is phreatic water.
of Majia Pu West Road and Jiahe Road, Beijing. It
was proposed to build for Beijing Subway Line 4 2.2 Risk identification
(M4), which is extended from north to south, and
Beijing Subway Line 10 (M10), which is extended Based on the project documents, the potentials of
from east to west. The main structure for M4 is the risks for this project were studied. The follow-
located below the intersection of the surface roads, ing factors are identified as the sources to poten-
and the new proposed covered undercutting for tially cause problems during construction:
M10 needs to be constructed below the station of 1. Geologic conditions
M4. High-rising residential and commercial build- The excavation depth requires the design per-
ings are congested around the corners of the trans- formed correctly based on the reliable geologic
fer station. Moreover, a lot of utility pipelines were data, the construction methods and risk mitiga-
crossed over the excavation pit. tion measures evaluated accordingly. Moreover,
the layers and consist of highly perme-
able cobbles with a very highly risk of water
2.1 Project description
leaking during excavation and uplifting the
Station halls were designed at the two end sides of bottom ground. Such leaking/uplifting may raise
the stations both for M4 and M10, while the middle additional construction difficulties, or cause the
parts were proposed utilizing covered undercutting. stability problems of the bracing systems and
Jiaomen West station for M4 was already built using the bottom ground of the excavation pit.
a large span tri-arch cross section with a width of 2. Engineering work
22.1 m, the depth is about 8 m (to the top) and 18 m The structural design is not risky as it is very
(to the bottom) below the ground surface (bgs). The mature already, the risks may exist on excava-

Figure 1. Jiaomen West Station for Beijing Subway M4 and M10.

564

ISGSR2013.indb 564 10/18/2013 9:45:44 AM

 Figure 2. Cross-section of M4 and longitudinal section of M10.

Table 1. The geological conditions of Jiaomen West 4. Construction management and supervision
transfer station. Poor construction management may destroy a
good project. Not only it is likely to reduce the
Layer Descriptions project profits but also may cause project acci-
Backfill, clay mixed silt, bricks, cobbles, loose dents or failure. Good supervision, however,
2
and inhomogeneous may prevent this happening.
Silty clay, yellow-brown, soft to stiff, wet, 5. Construction monitoring
2
highly compressible Underground work has high potential to cause
Cobble and gravels mixed with fine sands, settlement of ground, displacement of vertical
medium dense, wet, size 2–5 cm, max. size strut, deflection of bracing system, damage to
10 cm, the utility pipes, etc. Rigorously monitoring is
1 Clayey silt and silty clay, yellow, medium stiff, very helpful to indentify and control such risk,
medium dense, medium compressible finally may make them avoidable. However, poor
2 Fine to medium sands, yellow, dense, wet, low monitoring would not achieve the expected role.
compressible
Cobbles mixed with fine and medium sands,
mixed color, dense, wet, low compressible.
Size 2–6 cm, max. size 10 cm, 3 MITIGATION MEASURES:
1 Clayey silt and silty clay, yellow, medium stiff, CONSTRUCTION AND MONITORING
wet, low to medium compressible
PLANS
2 Fine to medium sands, yellow, dense,
saturated, low compressible
Cobbles mixed with fine and medium sands, 3.1 Construction plans
mixed color, dense, wet to saturated, low Bench-cutting method was proposed to undercut the
compressible, size 4–8 cm, max. size 10 cm soils below the bottom slab of Jiaomen West Station
for M4. The maximum progress length for each
tion bracing system design, strut selection and excavation cycle was limited to 4 m. In order to sta-
design as well as dewatering, particularly, the bilize and facilitate the soil excavation, the site was
work mentioned above are more associated with dewatered with stages described below; a shaft in east
geological and geotechnical conditions as well side of the station was built to provide construction
as construction workmanship. space and mobilize personnel and equipment to the
3. Construction experience and workmanship designed ground level; surroundings soils were then
These involve in every process of the project stabilized by grouting cement-silicate mortar into the
construction. Particularly, the construction ground along the long PVC sleeve valve pipes and
workmanship for the engineering work men- pre-support systems, utilizing short perforated pipes
tioned in item 2 is more important, even critical, to form umbrella vault, were installed. The details of
to the entire project. the construction plan are described below:

565

ISGSR2013.indb 565 10/18/2013 9:45:46 AM

 1. Dewatering steel trusses; steel trusses were locked to the
Three dewatering areas were scheduled: locking anchor pipes which were inserted and
1) Dewatering for shaft and covered undercut- grouted to the ground; 3) cut the cement-silica-
ting construction below the station for M4; soil material in Area III, install the preliminary
2) Dewatering for excavation of east hall of the lining and steel trusses; 4) apply and repeat the
station, and 3) dewatering for excavation of west above construction sequence for areas IV, V and
hall of the station. The depth of the wells uses VI respectively; 5) apply and repeat the above
36 m bgs evenly with a spacing of 6 m along the procedures for the left tunnel.
circumferences of the excavations. The length of the bench used in above proce-
2. Shaft construction dure is about to 5–10 m so that the progress for
A shaft was proposed to provide the con- each construction cycle is limited to 4 m. The
struction space and mobilize personnel and excavation for the right and left tunnels shall be
equipment to the designed ground level for apart at least 10 m along the longitudinal direc-
undercutting. The shaft is 26.5 m deep and is in tion to avoid interactions during excavation.
a rectangular shape with a dimension of 25.9 m
As shown in Figure 5, umbrella vault system was
by 9.6 m. The bracing system for the shaft con-
installed to provide pre-support for bench-cutting
sists of bored piles, horizontal struts and a cap-
each area of the cross section. Grouting materials
ping beam which was designed in conjunction
utilized 1:1 cement-silica mortar and the mortar
with those designed for excavation of the east
were pressured into the soils via the perforated pipes
and west halls of the station.
with a diameter of 32 mm. Such pipes were installed
3. Stabilizing the surrounding soils
with different upward angles (45°and 15°) due to
Prior to undercutting the soils below the sta-
the limitations of the bottom slab of the station for
tion for M4, grouting techniques are proposed
M4 and the short PVC sleeve valve pipes. The spac-
to stabilize the soils surrounding the tunnels of
ing for small perforated pipes utilizes @ = 1000 mm
M10. The area of the stabilization is shown in
in longitudinal direction while @ = 400 mm along
Figure 3. PVC sleeve valve pipes with a diameter
the circumference of the undercutting.
of φ = 56 mm and varying lengths from 10 to 13 m
are used to pressure cement-silica mortar materi-
als into ground. As shown in Figure 5, the spacing 3.2 Monitoring plans
of such long PVC pipes utilized 1.0 m by 1.0 m
and applied perpendicularly to the cross-section Monitoring the construction is important as it can
of the undercutting. The designed 3-day uncon- predict potential risks such as excessive settlement,
fined compressive strength of cement-silica-soil horizontal deflections and collapse, etc, as well
material is no less than 1.0 MPa with a perme- provide the data in a timely manner to modify the
ability coefficient less than 0.1 m/day.
4. Bench-cutting sequence and pre-support system
After the soils surrounding the tunnels for
M10 are stabilized, excavation proceeds with
bench-cutting method. As shown in Figure 4,
the sequence of bench-cutting is as follows:
1) Proceed with the right tunnel, install umbrella
vault system using grouting technique with small
perforated pipes to pre-support excavations,
cut the improved cement-silica-soil material in
area I; 2) cut the cement-silica-soil material in
area II, and install the preliminary lining and
Figure 4. Bench-cutting sequence and umbrella vault
installed with excavation progressing.

Figure 3. Stabilizing the soils surrounding the twin tun- Figure 5. Umbrella vault system and soil stabilization
nels for M10. (longitudinal section view).

566

ISGSR2013.indb 566 10/18/2013 9:45:47 AM

 Table 2. Monitoring items and the control criteria.

No. Classification Monitoring targets Monitoring item Accuracy Control criterion

1 Environment Buildings around the station Settlement of buildings 1.0 mm 20 mm

2 around the Utility pipes, including φ 1000 Settlement of the 1.0 mm Water and gas pipes:
undercutting and φ 1800 storm sewer pipelines 10 mm; storm and
pipes, φ 1000 sewage pipes, sewage pipes:
φ 400 gas pipes and φ 400 20 mm; rate:
water supply pipes, etc 2 mm/d
3 Roads and ground surface Settlement of ground 1.0 mm 30 mm; rate 2 mm/d
surface
4 Structures of existing Vertical and horizontal 0.3 mm 10 mm (absolute);
station for M4 displacements 5 mm (differential)
in structures
5 Vertical and horizontal 0.3 mm 10 mm (absolute);
displacements in 5 mm (differential)
track-bed structures
6 Deformation of 0.1 mm 3 mm
deformation joints
in structures
7 Distance of rail tracks 1.0 mm +4 mm, −2 mm
Geometry of rail tracks 1.0 mm 4 mm
8 Geometry of seamless 0.3 mm 2 mm
rail track
9 Connecting passage Settlement of interface 1.0 mm 3 mm
and gateways for M4 structures
10 Settlement of gateway 1.0 mm 5 mm
structures
11 Bracing Excavation pits Deflection of bracing 1.0 mm 30 mm
system for main structures pile heads
12 and auxiliary Deflection of bracing 1.0 mm 30 mm
structures piles
13 Forces in struts 1.0% F⋅S –

design. Table 2 summarizes the items to monitor Table 3. Criteria for alarm warning issuance.
and the control criterion utilized for monitoring in
construction. Due to the length limitation of the Alarm
paper, the plan of the monitoring points is shown warning level Criterion for alarm warning issuance
in Figure 1 partially. In the plan, DB means the Yellow Both absolute value and the rate of the
points monitoring the settlement of ground, JCJ monitored data exceeding 70% of
means the points for monitoring the building set- the control criterion, or one of them
tlement and GXC means the points for monitoring exceeding 85% of the control criterion.
the settlement of the utility pipes. Orange Both absolute value and the rate of the
Project management classified the monitor- monitored data exceeding 85% of
ing data to three levels to issue alarm warning, the control criterion, or one of them
1) yellow; 2) orange; and 3) red. The classification exceeding 100% of the control criterion.
is based on the absolute value and the rate of the Red Both absolute value and the rate of the
monitored data. Table 3 summarizes the criterion monitored data exceeding 100% of
the control criterion, or the rate of the
of the three levels for warning issuance. monitored data varies dramatically.

4 MONITORING DATA EXCEEDING

CRITERION
alarm warnings at different levels were issued in
The monitoring work was performed in accordance these monitoring points and construction meas-
with the frequency requirement associated with ures were taken accordingly, and the trend of the
the construction progress. Table 4 summarizes the deformation thus suspended.
monitoring data that exceeds the control criterion When excavating for shaft and covered under-
or alarm warning levels. During the constructions, cutting for the twin tunnels, ground surface shows

567

ISGSR2013.indb 567 10/18/2013 9:45:49 AM

 Table 4. Monitoring items for main structures of the station exceeding the control criterion.

Location description Accumulative Monitoring

No. Item description (monitoring point #) deformation (mm) conclusions

1 Settlement of ground Above the undercutting −24.6 Exceeding the alarm

surface (DB-05-03) warning level
2 Deformation of track Right alignment stationing −15.2 Exceeding the control
bed structures (K1+536.2) criterion
3 Distance of tracks Right alignment stationing −3.0 Exceeding the control
(K1+563.0) criterion
4 Settlement of ground Northwest transfer passage −27.5 Exceeding the alarm
surface (DB-44-01) warning level
5 Settlement of utility Northeast transfer passage −16.8 Exceeding the control
pipelines (GXC-37-02) criterion

subsidence accordingly, the maximum settlement construction arrangement are important to a safe
occurred right behind the cutting face at point and successful project; 2) monitoring is a critical
DB-05-3, which is above the right twin tunnel. process to construction; 3) covered undercutting
Monitored data also shows the subsidence of has more risk potential that may lead to more
ground around the east and west exaction pit. ground subsidence than open excavations; and
However, the rate of the settlement around excava- 4) with respect to covered undercutting, construc-
tion is small and the maximum settlement occurred tion quality and overburden depth may play a role
at DB-13-03 (east part of the excavation pit of the for ground subsidence. Lessons learned from this
east hall), but not exceeding the control criterion. project are: 1) every underground project is unique,
Analysis indicated that the backfill behind the even in the same project, covered undercutting for
bracing system is not compacted well and is more transfer passage and twin tunnels may lead to dif-
influenced by the excavation, the consolidation ferent settlement; 2) attention should be paid to
of the backfill also contributed to the settlement more details no matter how small the project is.
observed. As for the movement of track and track Actually, auxiliary parts of the project may deserve
bed structures, it was caused by the subsidence of more attention though people usually do to big
the soils surrounding the twin tunnels. Such move- and main part.
ment is probably inevitable as long as the ground
loss occurred, and probably, the control criterion REFERENCES
for such movement will govern the stabilization
design for the covered undercutting. Deng, X.P., Li, Q.M. & Zhou, Z.P. 2010. Statistical analy-
One of the interests is the monitoring data for sis of the accidents of subway construction in China.
transfer passage (which was not shown in Fig. 1 Chinese J. Statistics and Decision 9: 87–89.
due to length limitation of this paper) exceeds the Gatti, M.C. & Cassani, G. 2007. Ground loss control in
criterion. The dimension of the cross-section, the EPB TBM tunnel excavations. In J. Bartak, I. Hrdina,
treatment and excavation methods for the transfer G. Romancov & J. Zlaml (eds), Underground Space-
passage are similar with those applied in twin tun- the 4th Dimension of Metropolis; Proceedings of the
nels, except the overburden depth is almost half of World Tunnel Congress 2007 and 33rd ITA/AITES
Annual General Assembly, Prague, May 2007. London:
the covered undercutting for the twin tunnel. Thus, Taylor and Francis Group.
the reasonable explanations could be 1) shallower Huang, M.T. 2011. Dewatering practice for a deep sub-
overburden depth lead to more obvious ground way station excavation. Chinese J. Modern Urban
subsidence observations and 2) poorer construc- Transit 3: 58–61.
tion quality lead to more subsidence observed. Hulme, T.W., Shirlaw, J.N. & Hwang, R.N. 1990.
Settlement during the underground construction of
the Singapore MTR. In Proc. 10th South East Asian
5 CONCLUSIONS Geotechnical Conference, Taipei, 1990.
Li, T., Xu, L.P., Chen, H. & Gu, G.R. 2006. Quantitative
Analysis of risks of foundation excavations. Chinese J.
This paper identifies the risks associated with a Geotech. Eng. 11(28): 1917–1920.
subway transfer station construction in Beijing. Sejonaha, J., Jaruskova, D., Spackova, O. & Novotna, E.
Mitigation plans are introduced and the monitor- 2009. Risk quantifications for tunnel excavation
ing data are presented. Conclusions are made from process. World Academy of Sciences, Engineering, and
the project that 1) elaborate design and careful Technology 58: 393–401.

568

ISGSR2013.indb 568 10/18/2013 9:45:49 AM

 Geotechnical Safety and Risk IV – Zhang et al. (eds)
© 2014 Taylor & Francis Group, London, ISBN 978-1-138-00163-3

Ground anchor considerations for crane footings adjacent

deep rock excavations

N.R. Wightman
Sinclair Knight Merz (HK) Limited
Formerly SMEC Asia Ltd.

A.D. Mackay
Nishimatsu Construction Company Limited
Formerly SMEC Asia Ltd.

ABSTRACT: To increase power output from the Magat Pumped Storage Power Station, Luzon Island in
the Philipines, additional power generating equipment will be installed immediately adjacent to the existing
pump storage system. A concept design for the open cut excavation needed to install the equipment involved
rock excavation to an average of 60 m below the existing ground level with a 22 m span. To allow equip-
ment placement and future maintenance access into the excavation a 325 tonne load capacity gantry crane
spanning the excavation was proposed. The crane will run on rails placed onto rock with sufficient bearing
capacity to support the dead and live loads exerted during operation. Based on a site inspection, literature
search and preliminary ground investigation the rock mass bearing capacity will depend upon the presence
and characteristics of the rock joints. These included potentially persistent joints, continuing the length
of the excavation, with calcite coating and an adverse orientation with respect to the excavation, trending
parallel to the sides of the excavation and dipping about 60 degrees into the excavation. The calcite poten-
tially reduces the joint surface shear strength to a friction angle of about 20 degrees. To increase the shear
resistance it is intended to install six high strength proprietary rock anchors (each with 16 strand multi-core
tensioned cables and capable of imposing loads over 2,200 kN) perpendicular to the joint orientation. This
paper summarizes the anticipated ground conditions, the potential bearing capacity of the ground support-
ing the crane and the ground anchor installation proposed to improve the rock mass strength.

1 INTRODUCTION stability, supporting the gantry crane footings, will

be improved by the installation of six proprietary
The Magat Pumped Storage Power Station com- high capacity multi-strand cables beneath the rail
prises an existing dam, spillway and penstocks footings. The installations will increase the fric-
constructed in 1976. Due to the increased power tional resistance against sliding along the joints.
demand in the area additional power generation is
now required which will be provided by additional
power generating equipment to be accommodated 2 POWER STATION SETTING
in an open cut rock excavation. The current pro-
posal is to install the power house equipment into 2.1 Location
the excavation using a 325 tonne gantry crane span- The existing power station is located at Magat,
ning the excavation. In order to reduce the additional north Luzon in the Philippines. Refer to Figures 1
volume of reinforced concrete required to support and 2 for the general location.
the gantry crane and keep space for power station It is intended that the additional power generation
operation, it is proposed to place the crane rail plant will be a stand-alone structure located
footings onto competent rock away from the exca- in close proximity to the existing powerhouse
vation perimeter. The regional geology comprises accommodated in an open cut excavation. Refer to
volcanic rock, with pyroclastic breccia and lapilli Figure 3 for the Dam location.
tuff inclusions and a fault zone running through
the proposed excavation location. The joints antici-
2.2 Geology
pated during the excavation are adversely orientated
with respect to the excavation gradient and align- The geology at Magat comprises volcanic rock from
ment. As part of the concept design the rock mass the Lower Miocene to Upper Miocene geological

569

ISGSR2013.indb 569 10/18/2013 9:45:49 AM

 Figure 1. Magat, Philipines (Google Earth, 2011).

Figure 4. General view of the dam looking north west.

Figure 2. Magat, Luzon Island (Google Earth, 2011).

Figure 5. Regional active fault locations (Mines and

Geoscience Bureau, Philippines, 2010).

The structural geology includes faults and

shear zones. Major structural features which have
influenced the surrounding area include the active
Major Philippines Fault, located about 75 km to
the west of the site, trending North-North-East
(NNE) to South-South-West (SSW), and a large
active transverse fault, trending West-North-West
(WNW) to East-North-East (ENE), located 50 km
to the south of the site.
Figure 3. Aerial view of the existing power station and
proposed site location. 2.3 Engineering requirements
A pre-requisite for the construction of the
periods (Geological map, Mines and Geoscience additional power generation plant was that no dis-
Bureau, Philippines, 2010). The volcanic rock ruption to the on-going power generation opera-
comprises ‘Agglomerate’. This mainly includes tions would be allowed. Of the available options
“pyroclastic breccia”, which is very coarse grained the cable anchor installation was preferred to sta-
and coloured dark grey to greenish grey with clast bilise the ground beneath the crane footing. This
sizes up to 150 mm diameter in a welded matrix, provided a reduction in the quantity of concrete
and “Agglomeratic Sandstone”, which is a biotitic required for the gantry crane footing and, as a
crystalline lapilli ash tuff, comprising welded result, a reduction in the excavation quantity,
volcanic and igneous rock fragments from silt which was the main anticipated cause of disruption
0.06 mm up to about 4 mm. to the power generation operation. The open cut

570

ISGSR2013.indb 570 10/18/2013 9:45:49 AM

 Figure 6. General layout plan of the planned extension
(SMEC, 2011).

Figure 7. General layout section of the planned

extension showing crane location (SMEC, 2011).

excavation was therefore positioned with sufficient Figure 8. Areas of major jointing in the vicinity of the
site.
space to allow the installation of the cable anchors.
Figures 6 and 7 present the proposed open cut
excavation in relation to the existing dam.

2.4 Ground conditions

The rock condition in the vicinity of the gantry
crane footings was determined by site inspection
and a preliminary ground investigation.
These revealed the presence of inter-bedded
pyroclastic breccia and lapilli tuff rock with
discontinuities comprised bedding and closely
spaced joints. The average major joint set
orientations were revealed to be 32°, N166° and 56°, Plate 1. Calcite veining in fresh volcanic rock.
N214°. These approximately corresponded with the
two main faults identified west of the site, trend-
ing NNE to SSW, and south of the site, trending
WNW to ENE, respectively. The joints also exhib-
ited slickensiding and calcite deposits along the
surfaces. Major zones of jointing, identified during
the site inspection are presented in Figure 8.
The rock mass was moderately strong to strong
(British Standard, BS5930:1999 and A2:2010,
2010) and, based on laboratory testing, had
unconfined compressive strength values ranging
from 13.4 to 92.0 MPa, averaging 40.4 MPa
(moderately strong). Frictional resistance along
rock joints with rock surface contact, with a rough
undulating profile, typically range from 42° to 45°
(Hencher et al., 2011 & Richards & Cowland,
1982). However, due to the presence of calcite
veining and slickensiding, “polishing” along the Plate 2. Calcite veining at outcrop.

571

ISGSR2013.indb 571 10/18/2013 9:45:52 AM

 joints (refer to Plates 1 and 2) the estimated joint
frictional resistance was reduced to 22° (Lembo
Fazio et al., 1990). As a result of this reduced
frictional resistance robust stabilization measures
were needed to improve the shear strength resist-
ance along the joint surfaces.

3 CABLE SUPPORT TO CRANE

FOOTINGS

3.1 Position of crane footings

The proposed crane footings are to be positioned
either side of the open cut excavation, see Figure 9.
To support the loading exerted from the crane the
footings will be keyed into the rock with additional
support provided by the installation of ground
anchors immediately beneath the footings. The
ground anchor heads will be incorporated into the
reinforced concrete footings to ensure a robust tie
back of the concrete block into the rock mass.

3.2 Anchor support to the crane footings

Figure 10. Permanent anchors example by Dywidag
The aim of the rock anchor installation will be to (2012).
increase the load across the joint surface thereby
increasing the shear strength. To ensure the load
capacity and long-term rock anchor strength is
achieved, proprietary ground anchors with a 16 strand
multi-core cable anchoring system, were proposed in
the concept design. Based on the data available 6 No.
anchor cables, each loaded to approximately 1,500 kN
and increased to 2,165 kN to account for the factor
of safety (BS8081, 1989), will be required. Based on
the manufacturer’s proprietary literature each anchor
would need to have a diameter of 115 mm and a
200 mm formed hole to allow anchor installation.
Each anchor had an estimated total length of 30 m,
a minimum anchor length of 10 m (allowing for full
rock to grout contact) and minimum sheathed length
of 7 m (See Figure 10).

Figure 11. Grouting of permanent anchors (Dywidag,

2012).

Following rock anchor installation post grout

tensioning devices, jacking off and bearing onto
heavily reinforced sections of the crane footings,
are anticipated. Figure 10 shows a typical multi-
strand cable system (Dywidag, 2012), with cor-
rosion protection by epoxy coating for long-term
Figure 9. Location of crane footings and grouted anchor installation. Figure 11 (Dywidag, 2012)
multi-strand cables. shows a schematic of the anticipated two phase

572

ISGSR2013.indb 572 10/18/2013 9:45:55 AM

 grout injection to ensure the rock to anchor bond those solely of the authors and not of any other
is complete following installation of the anchor. party. The authors would like to thank Ronald
The crane will be positioned on solid ground at Chan for his assistance in the presentation.
the end of the crane rail alignment, away from the
excavation when not in use. This avoids continual
loading of the ground anchors when the crane not REFERENCES
in operation.
British Standards (BS) 5930:1999 + A2:2010 (2010). Code
of Practice for Site Investigations. British Standards
4 CONCLUSIONS Institution (BSI). London. 147p.
BS8081:1989 (1989). Code of practice for ground
The use of high strength anchors provided a solu- anchorages. British Standards Institution (BSI).
tion to the potential foundation instability along London. 180p.
adversely orientated joints beneath the gantry Dywidag Systems International (2012). Dywidag Strand
crane footings at the Magat power station site. The Anchors Systems brochure.
solution reduced the excavation required to form Google Earth (2011). Searchable map/satellite view of
the Republic of the Philippines.
the concrete footings thereby reducing disruption Hencher, S.R., Lee, S.G., Carter, T.S. & Richards, L.R.
to on-going power generating operation. Based on (2011). Sheeting Joints: Characterisation, Shear
the available data, six high capacity proprietary Strength and Engineering. Rock. Mech. Eng., 44:1–22,
multi-core cable anchor, with approximately 16 no Springer.
strands in each anchor for each footing will be Lembo Fazio, A., Ribucchi, R. & Tommani, P. (1990).
required. A maximum load capacity of 2,165 kN is Sheared bedding joints in rock. engineering: Two case
required for each multi-strand cable anchor. This histories in Italy. Rock Joints, Barton & Stephansson
provids sufficient resisting force to increase the (eds) A.A. Balkem, Rotterdam, pp. 83–90.
shear strength across the rock joints. Mines & Geoscience Bureau, (2010). Geology of the
Philippines (Second Edition), Mines and Geoscience
Bureau, Department of Environment and Natural
ACKNOWLEDGEMENTS Resources, Quezon City, Pilippines, 532p.
Richards, L.R. & Cowland, J.W. (1982). The effects of
surface roughness on field shear strength of sheeting
The authors wish to thank the SMEC design team, joints in Hong Kong Granite. The Engineer, Vol. 10,
particularly Mr. Andreas Neumaier for feedback No. 10, pp. 39–43.
on the submission and the project sponsor SN SMEC (2011). Feasibility Design Report, Magat Pumped
Aboitiz. The opinions expressed in this paper are Storage Project, July 2011.

573

ISGSR2013.indb 573 10/18/2013 9:46:00 AM

This page intentionally left blank
 Geotechnical Safety and Risk IV – Zhang et al. (eds)
© 2014 Taylor & Francis Group, London, ISBN 978-1-138-00163-3

Upper bound finite element method for seismic stability limit analysis
of rockfill dams

X.G. Yang & S.C. Chi

School of Hydraulic Engineering, Dalian University of Technology, Dalian, China

ABSTRACT: In this study, an upper bound limit analysis Finite Element Method (FEM) is developed
to study the seismic stability of rockfill dams. Considering the large value of the internal friction angel
and the non-linear shear strength parameters of rockfill materials, a static form, which is corresponding
dual Second-Order Cone Programming (SOCP) problem of the upper bound limit analysis, is formulated
with constraints based on the yield criterion, flow rule, boundary conditions and the energy-work balance
equation. The upper bound solution of critical seismic coefficient is then obtained by a state-of-the-art
interior-point algorithm iteratively. Compared with the limit equilibrium method, the proposed method is
applied to the seismic stability problem of a typical rockfill dam. The results demonstrate the correctness
and practical value of the proposed method. From the displacement field obtained by solving the SOCP,
we can predict the failure surface of the dam slope easily.

1 INTRODUCTION Sloan & Kleeman 1995) have made significant

progress in developing the Finite Element (FE) limit
Numbers of rockfill dams are built in the southwest analysis for stability and bearing capacity problems.
area of China, which is also an earthquake-prone In order to solve the mathematical programming
zone. Since Wenchuan earthquake, the safety easily and conveniently, a Linear Programming (LP)
assessment of rockfill dams in earthquakes has problem of lower or upper bound limit analysis is
attracted growing concern and wide attention of formulated based on the linearization of the yield
geotechnical engineering. function in their works. However, the error of cal-
Landslide is a common failure mode of rockfill culation caused by the linearization increases with
dams in earthquakes. For this reason, the analysis the value of the internal friction angle ϕ . Therefore,
of seismic stability of dam slopes is a very impor- the LP mathematical model of the limit analysis is
tant issue. Many researchers are attempting to suitable for the state of ϕ = 0 or smaller ϕ . For the
elaborate and develop new calculation methods to simplicity, the linear approximation of the failure
analyze the seismic stability problems. However, the envelope has been widely accepted for over half a
conventional pseudo-static approach is still widely century in which the shear strength of a material
used in engineering design (Loukidis et al. 2003). is described by two parameters, namely the cohe-
In this study, the seismic stability of the dam slope sion and friction angle. It has been argued that due
is expressed in terms of a single parameter, the to the nonlinearity, these two parameters are not
critical (or yield) seismic coefficient, kc . The criti- independent of stress level and consequently not
cal seismic coefficient, kc , is the ratio of the seismic constant (Li & Cheng 2012).
acceleration, ac , yielding a factor of safety equal to Unfortunately, a substantial number of experi-
unity, to the acceleration of gravity g (= 9.81 m/s 2). ments (Indraratna et al. 1993, Yamaguchi 2008)
For the rockfill dam, ac is also the maximum anti- have clearly shown that the rockfill materials proc-
seismic capability. ess two main properties: the large value of the
The pseudo-static approach is traditionally internal friction angle ; and the nonlinear nature
implemented in limit equilibrium methods to assess of the shear strength parameters, which are com-
the seismic stability of the slope. However, nei- monly expressed in terms of stresses. Consequently,
ther static nor kinematic admissibility is necessar- the FE limit analysis of the stability problems of
ily satisfied in limit equilibrium. Limit analysis, as rockfill dam slopes can’t be formulated as an LP
an alternative approach, has been a powerful tool problem.
that provides rigorous lower and upper bounds During the last decade there has been consid-
to the exact collapse load in recent years. Sloan erable progress in the application of nonlinear
and his cooperators (Sloan 1988, Sloan 1989, programming, which allows the yield function to

575

ISGSR2013.indb 575 10/18/2013 9:46:00 AM

 be treated in its native form. Recent research by e.g. the seismic loads in this paper; and constant
Makrodimopoulos & Martin (2006, 2007a, b) has loads q0, e.g. the gravity and other loads in this
concentrated on FE limit analysis using Second- paper. Thus, an upper bound on kc of the dam
Order Cone Programming (SOCP) to solve stabil- slope can be calculated by solving the following
ity or bearing capacity problems with the native optimisation problem:
form of the yield function. Based on the work by
Makrodimopoulos & Martin (2007a), this paper min Dp ( ) − q T0 u
develops an upper bound FEM to analyse the seis-
s.t. ε( ) ∈ E in
i V
mic stability of rockfill dams. With the native form
of the yield function, the error caused by the large u = u 0 on S
value of the internal friction angle can be reduced. q1T u = 1 (6)
Moreover, a static form, which is corresponding
dual SOCP problem of the upper bound limit anal- The six-node triangle finite element with straight
ysis, can be formulated. This means we can get the sides that is used in the upper bound analysis is
stress field of the dam in a weak sense, which can shown in Figure 1.
be used to determine the nonlinear shear strength Because the nodal displacement field can be
parameters of rockfill materials iteratively. expressed as a quadratic form within a triangular
element, any strain component may vary linearly.
2 UPPER BOUND THEOREM Moreover, if the element sides are straight, the
strain tensor at any point within the triangle can
Consider a rigid, perfectly plastic construction be expressed as a combination of those at three
V with boundary S; according to the upper bound vertices, i.e.,
theorem, when the structure collapses, there exists 3 3
a kinematically admissible displacement field such
that energy dissipation is no more than the work
( ) ∑ ( ) , ( x)) , ∑ Li ( ) = 1 (7)
i =1 i =1
of external force:
where the coefficients Li Ai /( A1 A2 + A3 ) are
∫V σ : ε( u )dV ≤ q
T
u (1) area coordinates. Because any strain tensor can
be expressed by ε i at the vertices, if the flow rule
is enforced at the three vertices of the element, it
where q are equivalent nodal loads; u should satisfy
holds at all points within the elements. Thus, the
the boundary conditions of S, u u 0 ; σ should sat-
isfy yield function, ∀σ ∈ F = { ≤ }. Because
flow rule constraint only needs to be enforced
at a finite number for it to hold throughout the
displacement discontinuities between elements
structure.
do not exist, the power to be dissipated by plastic
As mentioned by Makrodimopoulos & Martin
deformation is only permitted to occur within each
(2007a), the main restriction of the proposed
triangular element. The energy-work balance con-
method is that the yield function must be expressed
dition in Equation 1 can be written as
as a conic quadratic form. The Mohr-Coulomb
criterion in plane strain conditions is a typical
Dp ( ) q T u a d ε( ) E (2) example of a yield restriction with a conic quad-
ratic form. It can be written in the following form:
where
s red + aσ m − k ≤ 0 (8)
Dp ( ) ∫V d p ( ) dV (3)

p∑σ : ε
d p ( ) sup (4)
σ ∈F

and the set of plastically admissible strains is

E { d +∞ } (5)

The plastically admissible strains are those that

satisfy the associated flow rule. We divide equiva- Figure 1. Six-node linear strain element for upper
lent nodal loads into two parts: collapse loads q1, bound analysis.

576

ISGSR2013.indb 576 10/18/2013 9:46:01 AM

 where NP
min ∑ ciT zi − q0T u
s red = [ ]T (9) i =1
s.t. zi ∈ Ki (i = 1, …, NP)
A i z i − Bi u = 0 (i = 1, …, NP)
1 D
σm ∑ σ ii , sij = σ ij
D i =1
σ mδ ij (10) q1T u = 1 (14)

where
where D is the dimension of the tensors and δ is
Kronecker’s δ . For plane strain conditions of the ci ℜ3 c iT = ⎡⎣ ki 0 0 ⎤⎦
Mohr-Coulomb criterion, D = 2, a = sin ϕ , and
k c cos ϕ , where c is the cohesion and ϕ is the z i ∈ℜ3 z iT = ⎡⎣ λi ( e ired )T ⎤⎦
ℜ3×3 A i = diag [ai 1 1]
internal friction angle.
Ai
T
ℜ3× NZ Bi = ⎡⎣ B m,i
Bi ∈ℜ B d,i ⎤⎦ (15)
3 FEM FORMULATION FOR UPPER
BOUND LIMIT ANALYSIS NZ is the total number degrees of freedom and
z i ∈ K i is the second-order cone constraint. Sets K i
Considering a plane strain structure divided into are quadratic cones of the following form:
NE finite elements, the optimisation problem
Equation 6 can be transformed as follows: K { ℜd : ≤ x1, x1 ≥ } (16)
2 :d
NE
min ∑ ∫A i
λdAi − q T0 u Equation 14 is a standard SOCP problem, which
can be transformed into the corresponding dual
i =1

s.t. λi ≥ e ired (i = 1, …, NE) problem as follows:

ai λi = B m
m,i u (i = 1, , NE) max kc
e ired B d,i u (i = 1, …, NE) s.t. ( ym,i , sired ) ∈ K i (i = 1, …, NP)
q1T u =1 (11) ym,i + ai m ,i = ki* (i = 1, …, NP)
NP NP
where ∑ G m,
m σ m ,,ii ∑ Gd,i,i sired β q1 = q 0 (17)
i =1 i =1
θi = B ,i u (12)
where
T
⎡2exx 2exy ⎤⎦ = B dd,i u (13) G m, B Tm,i
⎣
G d, B dT,i
The matrices B m,i and B d,i can easily be obtained
from the typical strain-displacement relations. In ki* ki / ηi
geotechnical engineering, the displacement bound- ηi = Aee,,i / 3 (18)
ary is typically that u = 0 on S. For conciseness, we
assume that the above-mentioned conditions have It has been shown that the dual formulation
already been satisfied in Equation 11 and in what Equation 17 can be considered as a static form
follows. of upper bound analysis by Makrodimopoulos &
As mentioned above, the flow rule only needs to Martin (2007a). The main advantages of solving
be enforced at the three vertices of each six-node the dual problem of the upper bound limit analysis
triangular element to ensure that it holds through- are not only the dual form is robust and stability
out the element. For this reason, we take the ver- but also the stress distribution of structure can be
tices of each triangle as flow rule points. Thus, for obtained, which is used to describe the nonlinear
an NE-element structure, the total number of the failure criterion of rockfill materials.
flow rule points must be NP, where NP = 3NE. At present, large-scale SOCP problems can be
Now, Equation 11 can be formulated as follows, solved effectively using the interior-point method.
and more details can be found in the work by Based on this method, various state-of-the-art
Makrodimopoulos & Martin (2007a): algorithms are developed by researchers in

577

ISGSR2013.indb 577 10/18/2013 9:46:06 AM

 mathematical programming. In this paper, we use 5 NUMERICAL EXAMPLES
the algorithm software SDPT3 (Tütüncü et al.
2003) to solve SOCP problems. The LP/SOCP/SDP Consider the seismic stability problem of a
algorithm SDPT3 employs an infeasible primal- completed rockfill dam in an earthquake with dam
dual predictor-corrector path-following method. It height H = 100 m, gradient n = 1:1.4, and width of
has proven to be robust and efficient because vari- the dam crest Dc = 8 m. It is suggested by Speci-
ous techniques to improve efficiency and stabil- fications for seismic design of hydraulic structures
ity are incorporated. Moreover, the basic code of (DL5073-2000) in China that the seismic stabil-
SDPT3 is written in MATLAB (MATLAB 2011), ity of rockfill dams can be analysed by a pseudo-
so it is extremely simple and convenient to use by static approach. The concept of the pseudo-static
YALMIP (YALMIP 2012), which is a modelling approach relies on the representation of the
language implemented as a toolbox for MATLAB earthquake-induced loading by statically applied
for advanced modelling and solution of convex inertial forces. Meanwhile, to assess the distribu-
and nonconvex optimisation problems. tion of the response acceleration along the height
of dams, the pseudo-static loads are calculated in
terms of the dynamic distribution coefficient α i ,
4 UPPER BOUND ANALYSIS BASED ON
which is suggested by the specifications mentioned
THE NONLINEAR SHEAR STRENGTH
above. As shown in Figure 2, when the seismic
design intensity is of degrees VII, VIII, and IX, the
Considered cohesionless soils, rockfill materials
maximum dynamic distribution coefficient α m on
naturally evidence a nonlinearity of failure crite-
the top of the dams is 3.0, 2.5, and 2.0, respectively.
rion with c = 0. As clearly shown by triaxial tests
Thus, the horizontal seismic loads acting on each
(Indraratna et al. 1993; Yamaguchi 2008), rockfill
element can be calculated as
materials are susceptible to particle breakage under
high confining pressure, which leads to the redistri-
bution of intergranular stress and a decrease in the Q khξα iW (20)
internal friction angle. Consequently, for rockfill
materials, the Mohr-Coulomb envelopes are typi- where ξ is the seismic reduction coefficient
cally curved, and their variation can be expressed (generally ξ = 0.25), W is the weight of each ele-
in the form described by Duncan et al. (1978) ment, and kh is the horizontal seismic coeffi-
cient, which is the ratio of the horizontal seismic
ϕ ϕ 0 − Δϕ l (σ 3 /Pa ) (19) acceleration ah to the acceleration of gravity g
(= 9.81 m/s 2 ): kh ah /g.
g The vertical seismic loads
where ϕ 0 is the friction angle at the unit atmos- V can be determined as 2/3 of horizontal seismic
pheric pressure Pa of confining pressure σ 3 , and loads approximately:
Δϕ is the reduction in ϕ for a 10-fold increase
in σ 3 . We apply iterative approach to upper bound
analysis in order to consider the nonlinearity of the V 2 /33Q (21)
failure criterion. Combined with Section 3, the cal-
culating procedures for upper bound limit analysis In this paper, we assume the design intensity
using nonlinear failure criterion are as follows: is of degree IX. The critical seismic coefficient
kc (k
kc kh in this paper) of the dam slope can be
Step 1: Set the initial value of ϕ ini for each
obtained by maximising Equation 17 for the upper
element.
bound analysis.
Step 2: Assemble the constraint data in
Equation 17 and formulate the data as a dual
SOCP problem.
Step 3: Solve the problem using the algorithm
SDPT3 and obtain the stress field of the dam.
Step 4: According to the stress field, deter-
mine ϕ sec by Equation 19 and compare it to
ϕ ini. If | | ≤ ζ for 99.9% of the total ele-
ment numbers, then go to next step; otherwise,
set ϕ ini ϕ sec and go to Step 2, where ζ is the
tolerance.
Step 5: Obtain the exact upper bounds and out-
put the results.
Generally, converged rigorous upper bounds
can be obtained after several iterations. Figure 2. Dynamic distribution coefficient α i .

578

ISGSR2013.indb 578 10/18/2013 9:46:13 AM

 Three projects is considered in this paper:
Project 1, consider the horizontal seismic loads
only; Project 2, consider the horizontal and vertical
seismic loads and assume the direction of the verti-
cal seismic loads is upward; Project 3, consider the
horizontal and vertical seismic loads and assume
the direction of the vertical seismic loads is down-
ward. In these three projects, we assume the direc-
tion of the horizontal seismic loads is downstream. Figure 4. Failure surface and displacement field.
The nonlinear shear strength parameters can be set
as ϕ 0 = 52.3
and Δϕ = 11.0
(Chen & Chen 2007).
The mesh of the dam is shown in Figure 3, and the proposed method. From the displacement field, the
results are presented in Table 1. failure surface is predicted easily. The comparison
From results, we see that the solutions of kc of failure surface obtained from Bishop’s simpli-
obtained by upper bound FEM is close to those fied method and the proposed method (fine meth)
obtained by Bishop’s simplified method (Bishop is shown in Figure 4. It is noted from Figure 4 that
1955). The average relative difference of kc obtained the failure surfaces predicted by the two mentioned
by the writers is 3.8% and 2.8% larger using coarse methods are very close.
and fine mesh respectively than Bishop’s method. It
is noted that we can still get much more rigours upper
bound solution by formulating the correspond- 6 CONCLUSIONS
ing dual SOCP problem of the upper bound limit
analysis, though rockfill materials process larger In this paper, an upper bound limit analysis FEM
internal friction angel. Because of the high-order is developed to study the seismic stability of rock-
element type, the solutions are not very sensitive to fill dams. In order to reduce the errors caused by
mesh density. By determining the internal friction the linearization of the yield function, the upper
angle iteratively, the nonlinear shear strength prop- bound limit analysis is based on the native form
erties of the rockfill materials are considered in the of the Mohr-Coulomb criterion and formulated
as a second-order cone programming program. It
is noted that though the value of internal friction
angle of the rockfill materials is very large, we still
get rigorous upper bound solutions. Moreover, the
nonlinear shear strength properties of the rockfill
materials are also considered by using an itera-
tive approach for the friction angle. The influence
of mesh density on solutions is also studied. The
results show that the solutions are not very sensi-
tive to mesh density. This means the solutions can
still be accuracy even though the finite element
mesh is coarse. Comparisons of Bishop’s simpli-
fied method illustrate the correctness and the prac-
tical value of the proposed method.

ACKNOWLEDGMENT

The authors gratefully acknowledge the finan-

Figure 3. Mesh of the rockfill dam. cial support provided by the National Nature
Science Foundation of China through Grants
Nos. 50979014 and 51179024.
Table 1. Results of critical seismic efficient kc.

kc
REFERENCES
Project 1 Project 2 Project 3
Bishop, A.W. 1955. The use of the slip circle in the
Coarse mesh 1.139 0.948 1.369 stability analysis of slopes. Géotechnique 5(1): 7–17.
Fine mesh 1.128 0.945 1.349 Chen, L. & Chen, Z. 2007. Effect of nonlinear strength
Bishop’s method 1.09 0.92 1.32 of rockfill on slope stability of high earth-rock dam.
Rock and Soil Mechanics 28(9): 1807–1810.

579

ISGSR2013.indb 579 10/18/2013 9:46:20 AM

 Duncan, J.M., Byrne, P.M. & Wong, K.S. 1978. Strength MATLAB 2011. The Mathworks, Natic, MA, USA.
stress-strain and bulk modulus parameters for finite ele- Standard of Ministry of Electric. Power, PRC. 2001.
ment analysis of stress and movements in soil masses, Specifications for seismic design of hydraulic structures
Berkeley University of California, Berkeley. (DL5073-2000). Beijing: China Electric Power Press.
Indraratna, B., Wijewardena, L.S.S. & Sloan, S.W. 1988. Lower bound limit analysis using
Balasubramaniam, A.S. 1993. Large-scale triaxial finite elements and linear programming. International
testing of greywacke rockfill. Géotechnique 43(1): Journal for Numerical and Analytical Methods in
37–51. Geomechanics 12(1): 61–77.
Li, D. & Cheng, Y. 2012. Lower bound limit analysis Sloan, S.W. 1989. Upper bound limit analysis using
using nonlinear failure criteria. Procedia Earth and finite elements and linear programming. International
Planetary Science 5: 170–174. Journal for Numerical and Analytical Methods in
Loukidis, D., Bandini, P. & Salgado, R. 2003. Stability Geomechanics 13(3): 263–282.
of seismically loaded slopes using limit analysis. Sloan, S.W. & Kleeman, P.W. 1995. Upper bound limit
Géotechnique 53(5): 463–479. analysis using discontinuous velocity fields. Computer
Makrodimopoulos, A. & Martin, C.M. 2006. Lower Methods in Applied Mechanics and Engineering 127:
bound limit analysis of cohesive-frictional materials 293–314.
using second-order cone programming. International Tütüncü, R.H., Toh, K.C. & Todd, M.J. 2003. Solving
Journal for Numerical Methods in Engineering 66(4): semidefinite-quadratic-linear programs using SDPT3.
604–634. Mathematical Programming 95(2): 189–217.
Makrodimopoulos, A. & Martin, C.M. 2007a. Upper Yamaguchi, Y. 2008. Strength evaluation of rockfill
bound limit analysis using simplex strain elements materials considering confining pressure dependency.
and second-order cone programming. International Journal of Japan Society of Dam Engineers 8(3):
Journal for Numerical and Analytical Methods in 166–181.
Geomechanics 31(6): 835–865.
Makrodimopoulos, A. & Martin, C.M. 2007b. Upper
bound limit analysis using discontinuous quadratic
displacement fields. Communications in Numerical
Methods in Engineering 24(11): 911–927.

580

ISGSR2013.indb 580 10/18/2013 9:46:22 AM

 Editors
Zhang
Wang
Wang
Li
Geotechnical Safety and Risk IV contains the contributions presented at the
4th International Symposium on Geotechnical Safety and Risk (4th ISGSR,
Hong Kong, 4-6 December 2013), which was organised under the auspices of
Geotechnical Safety and Risk IV
the Geotechnical Safety Network (GEOSNet), TC304 on Engineering Practice
of Risk Assessment and Management and TC205 on Safety and Serviceability

Geotechnical Safety and Risk IV

in Geotechnical Design of the International Society of Soil Mechanics and
Geotechnical Engineering (ISSMGE). Geotechnical Safety and Risk IV
covers six themes:
(1) Geotechnical uncertainty and variability
(2) Geohazards such as landslides, earthquakes and climate changes
(3) Reliability and risk analysis
(4) Reliability-based design and limit-state design in geotechnical
engineering
(5) Risk assessment and management in geotechnical engineering and
infrastructural projects, and
(6) Practical applications.

Geotechnical Safety and Risk IV is an excellent source of reference on the

latest developments and practices of geotechnical risk and safety, and will
prove to be invaluable for civil engineers (particularly geotechnical engineers),
geologists, and natural hazards reduction agencies.

ISGSR IV is a continuation of a series of symposiums and workshops

on geotechnical risk and reliability, starting with LSD2000 (Melbourne,
Australia), IWS2002 (Tokyo and Kamakura, Japan), LSD2003 (Cambridge,
USA), Georisk2004 (Bangalore, India), Taipei2006 (Taipei, Taiwan), 1st
ISGSR (Shanghai, China, 2007), 2nd ISGSR (Gifu, Japan, 2009) and 3rd
ISGSR (Munich, Germany, 2011).

Editors: L.M. Zhang, Y. Wang, G. Wang and D.Q. Li

an informa business