Soil variability and its consequences

in geotechnical engineering

Von der Fakultät für Bau– und Umweltingenieurwissenschaften

der Universität Stuttgart
zur Erlangung der Würde eines Doktors der Ingenieurwissenschaften (Dr.-Ing.)
genehmigte Abhandlung,

vorgelegt von

D IPL .-I NG . D IPL .-I NG . M AXIMILIAN H UBER

aus Leonding (Österreich)

Hauptberichter: Prof. Dr.-Ing. Pieter A. Vermeer

Mitberichter: Prof. Dr.techn. Dr.-Ing. Andras Bardossy
Prof. Dr. Michael A. Hicks

Tag der mündlichen Prüfung: 19. Juni 2013

Institut für Geotechnik der Universität Stuttgart

2013
[email protected]
Mitteilung 69
des Instituts für Geotechnik
Universität Stuttgart, Germany, 2013

Editor:
Univ.-Prof. Dr.-Ing. habil. Christian Moormann

c
Maximilian Huber ([email protected])
Institut für Geotechnik
Universität Stuttgart
Pfaffenwaldring 35
70569 Stuttgart

All rights reserved. No part of this publication may be reproduced, stored in a retrieval
system, or transmitted, in any form or by any means, electronic, mechanical,
photocopying, recording, scanning or otherwise, without the permission in writing of
the author.

Keywords: soil variability, spatial variability, correlation length, Random Finite

Element Method, uncertainty quantification.

Printed by e.kurz + co, druck und medientechnik GmbH, Stuttgart, Germany, 2013

ISBN 978-3-921837-69-6
(D93 - Dissertation, Universität Stuttgart)
[email protected]
 Preface of the editor
With issue no. 69 of the proceedings the Institute of Geotechnical Engineering at the
University of Stuttgart (IGS) the dissertation study of Dr.-Ing. Maximilian Huber is pub-
lished. Dr. Huber’s thesis is related to soil and rock heterogeneity, to the mathematical
approaches for describing soil variability and to its consequences in geotechnical engi-
neering.
Soils and rocks as natural geological materials exhibit spatial variability of material
properties. The spatial variability, frequently referred to as heterogeneity, is anisotropic,
often depth-dependent and occurs at multiple scales: at the very small scale, also called
grain scale, as seen in the arrangement of solid particles of sand or in the arrangement of
pellets in clay, at the decimeter to meter scale, as observed in specimen testes in labora-
tory or in CPT/SPT-soundings in soil layers; and at larger scales too, like the geotechni-
cal scale relevant for geotechnical structures like foundation, excavations or tunnels, or
at the geological scale of several hundreds of meters, which is affected by the layering
of soils of different types. Spatial variability of soils and rocks influences the material
behaviour in mechanical and hydraulic sense, the flow of groundwater water and the
performance of geotechnical structures. Traditional deterministic analyses based on sin-
gle "representative" soil property values lead to partial or global factors of safety, which
provide no information regarding probability of failure and which does not consider the
uncertainty that arises through having incomplete information about the subsoil con-
ditions. In contrast probabilistic approaches allow defining soil properties in statistical
terms and simulating geotechnical performance probabilistically for example on terms
of reliability. The reliability, defined as the complement of the failure probability, is a
rational measure of safety.
The scientific work of Dr. Huber focuses on the evaluation of the effects of soil vari-
ability by using advanced mathematical models within the framework of probabilistic
methods. The evaluation of stochastic soil properties including spatially correlated soil
properties is investigated on a wide theoretical basis. These results are used in case
studies on the evaluation of the spatial correlation of different measurement data sets.
Considering these outcomes finally the effects of soil variability are evaluated in typi-
cal problems in tunnelling and foundation engineering demonstrating the effect of soil
heterogeneity and spatial variability at different scales.
The thesis of Dr. Huber shows that probabilistic analyses are powerful and versatile
tools for investigating the influence of uncertainties on a given geotechnical problem, but
indicates the meaning of the given input data and limitations of probabilistic analyses
also.

[email protected]
 Safety, reliability and risk are key issues in situations with continuously increasing
complexity especially in geotechnical engineering. In this regard the doctoral thesis of
Dr. Huber proves as a valuable contribution.

Christian Moormann
Stuttgart, July 2013

[email protected]
 Preface of the supervisor of the PhD thesis
As yet probabilistic analyses are not common in geotechnical engineering and will prob-
ably never be introduced for regular structures and foundations, but it is expected that it
will become more and more used for major engineering projects in difficult ground. As
usual in most branches of engineering, loads on structures are stochastic, but the natu-
ral variability of soil properties exceeds by far the variability of man-made engineering
materials. As a consequence, a special need of probabilistic design exist in geotechni-
cal engineering. For this reason, I have encouraged Maximilian Huber to do a doctoral
study on the application of probabilistic methods in geotechnics, leaving the choice of a
more precise topic entirely to him.
The first idea he came up with was to perform borehole jacking tests in a particular
layer of mudstone to measure its stiffness and to assess the corresponding correlation
length. I was very happy to learn that my colleague Prof. Andras Bardossy was will-
ing to support these field measurements even financially. These experiments had to be
done in the Fasanenhof tunnel, being at that time under construction at Stuttgart by the
companies Weiss & Freitag and Max BÃűgl. I am not only indebted to the support from
these companies, but also for the support of the owner of the tunnel, i.e. the Stuttgarter
Strassenbahnen AG.
The candidate has chosen to focus this study on soil variability, being usually taken
into account by distinguishing between different soil layers. In probabilistic studies soil
properties within a layer are based on a probabilistic density function, as defined by
a mean value, a standard deviation and possibly skewness. In such an approach the
spatial variability within a soil layer may be disregarded on assuming homogeneity. On
approaching reality more closely, spatial variability within soil layers may be modelled
in combination with a correlation length, i.e. "a scale of fluctuation" for the soil property
considered. I consider such developments as fascinating and would like to know to
which extent such approaches are already applicable in engineering.
This dissertation study convinced me of the matureness of probabilistic design in
geotechnical engineering. On the other, it made me realise that variability within soil
layers is still a topic of research, as the assessment of correlation lengths is not straight
forward. The assessment of this length is probably the most important scientific achieve-
ment of this dissertation study on soil variability. Many of the case studies may also
serve as a manual for the application of probabilistic design in advanced geotechnical
engineering. As a consequence, his research has already attracted good attention in the
international research community.

[email protected]
 I got to know Maximilian Huber not only as a young researcher and teacher before
my retirement from Stuttgart University, but also later in Delft when he worked with his
external advisor Prof. M. Hicks. These visits to Delft also provided the opportunity of
additional social contacts, which I enjoyed very much. So I had the pleasure of learning
to know him not only as a talented researcher with creative ideas, but also as gifted
musician with a wide field of interests.

Pieter A. Vermeer
Nederhorst den Berg, Netherlands, July 2013

[email protected]
 Acknowledgments
The research presented in this thesis is the result of the work carried out during the
years 2007 - 2013 at the Institute of Geotechnical Engineering of Stuttgart University. Be-
hind this work there are important contributions and the significant support of a certain
number of persons, whom I would like to thank sincerely.
I wish to express first of all my deep gratitude to Prof. Pieter Vermeer for his extensive,
invaluable and steadfast guidance, for his support and patience, sharing his knowledge
and experience, throughout my PhD studies.
Likewise, I thank Prof. Andras Bardossy and Prof. Michael Hicks for their expert
advise and feedback as my co-supervisor throughout the time of my studies.
Throughout the years, I benefited from the cooperative and friendly support, advise
and guidance of many others. With this regard, I would like to express my gratitude
to Patrick Arnold, Dr. Robby Caspeele, Ed Calle, Anneke Hommels, Dr. Jon Nuttall,
Prof. Christian Moormann, Dr. Massimiliano Moscatelli, Dr. Thomas Most, Dr. Dirk
Proske and to longtime colleagues at the University of Stuttgart. Amongst them are Prof.
Thomas Benz, Dr. Lars Beuth, Fursan Hamad, Dr. Issam Jassim, Dr. Annette Lächler,
Dr. Axel Möllmann, Dr. Marcus Schneider, Peter Ströhle, Prof. Bernhard Westrich and
Bernd Zweschper.
Of those mentioned here and many others whom I met during the exciting years of
my PhD studies I maintain invaluable good memories. I am very grateful to my parents,
to my sister and to friends, who gave me indispensable support through their company
and advise.

Maximilian Huber
Weimar, July 2013

[email protected]
[email protected]
 Contents

Summary xv

Zusammenfassung xix

1 Introduction 1
1.1 Background and rationale . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Research aim . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.3 Thesis scope . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3

2 Characterizing soil variability at different scales 5

2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.2 Probabilisitic site characterization . . . . . . . . . . . . . . . . . . . . . . . . 5
2.3 Sources of soil variability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.4 Describing spatial variability . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.4.1 Scales of variability in soil science . . . . . . . . . . . . . . . . . . . 10
2.4.2 Mathematical description of spatial correlation . . . . . . . . . . . . 11
2.5 Quantifying spatial dependence . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.5.1 Method of Moments . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.5.1.1 Variogram approach . . . . . . . . . . . . . . . . . . . . . . 14
2.5.1.2 Autocorrelation function approach . . . . . . . . . . . . . 15
2.5.1.3 Local average theory . . . . . . . . . . . . . . . . . . . . . 15
2.5.2 Maximum likelihood method . . . . . . . . . . . . . . . . . . . . . . 17
2.5.2.1 Consideration of a trend within the ML approach . . . . . 18
2.5.2.2 Residual maximum likelihood approach . . . . . . . . . . 18
2.5.3 Additional approaches . . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.5.4 Sampling strategies . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
2.5.5 Estimation of the different scales of spatial variability . . . . . . . . 21
2.6 Variogram calculation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.6.1 Theoretical variogram models . . . . . . . . . . . . . . . . . . . . . 24
2.6.2 Estimation of the theoretical variogram function . . . . . . . . . . . 25
2.6.3 Alternative approaches . . . . . . . . . . . . . . . . . . . . . . . . . 27
2.6.4 Model selection using the A KAIKE Information criterium . . . . . . 27
2.7 On the correlation length . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
2.7.1 Uncertainty of the correlation length . . . . . . . . . . . . . . . . . . 29
2.7.2 Anisotropy with respect ot the correlation length . . . . . . . . . . 30
[email protected]
2.8 Synopsis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31

i
Contents

3 Case studies on the evaluation of spatial variability 33

3.1 Case study A – Random function . . . . . . . . . . . . . . . . . . . . . . . . 33
3.1.1 Generation of the random process . . . . . . . . . . . . . . . . . . . 34
3.1.2 Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
3.1.3 Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
3.2 Case study B – CPT data evaluation . . . . . . . . . . . . . . . . . . . . . . 40
3.2.1 Site description & measurement technique . . . . . . . . . . . . . . 40
3.2.2 Analysis of the data . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
3.2.3 Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
3.3 Case study C – Fasanenhof-tunnel . . . . . . . . . . . . . . . . . . . . . . . 51
3.3.1 Site description & measurement technique . . . . . . . . . . . . . . 51
3.3.2 Analysis of the data . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
3.3.3 Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
3.4 Case study D – Sheikh Zayed road in Dubai . . . . . . . . . . . . . . . . . . 57
3.4.1 Site description . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
3.4.2 Analysis of the data . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
3.4.3 Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
3.5 Literature database on the spatial correlation of soil properties . . . . . . . 61
3.6 Evaluation of the vertical correlation length using CPT databases of dif-
ferent soil types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
3.7 Combining expert knowledge and measurements . . . . . . . . . . . . . . 65
3.8 Summary and consequences of the case studies . . . . . . . . . . . . . . . . 66

4 Safety and uncertainties 69

4.1 Safety in engineering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
4.2 Basics of uncertainty quantification . . . . . . . . . . . . . . . . . . . . . . . 77
4.2.1 Definition of the limit state function . . . . . . . . . . . . . . . . . . 78
4.2.2 Random number generation . . . . . . . . . . . . . . . . . . . . . . 79
4.2.3 Generation of correlated random numbers . . . . . . . . . . . . . . 79
4.2.4 Generation of random fields . . . . . . . . . . . . . . . . . . . . . . . 80
4.3 Computation of failure probabilities . . . . . . . . . . . . . . . . . . . . . . 81
4.3.1 First Order Reliability Method . . . . . . . . . . . . . . . . . . . . . 83
4.3.2 Monte Carlo simulation . . . . . . . . . . . . . . . . . . . . . . . . . 84
4.3.3 Stochastic Finite Element Method . . . . . . . . . . . . . . . . . . . 86
4.3.4 Response surface methods . . . . . . . . . . . . . . . . . . . . . . . . 91
4.4 Sensitivity analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
4.4.1 Local sensitivity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
4.4.2 Global sensitivity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
4.5 Synopsis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94

5 Selected case studies in uncertainty quantification using random properties 95

5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
5.2 Tunnel lining . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
5.2.1 Mechanical description of the problem . . . . . . . . . . . . . . . . 96
[email protected]
5.2.2 Stochastic variables and parametric studies . . . . . . . . . . . . . . 98

ii
 Contents

5.2.3 Sensitivity analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99

5.2.4 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
5.3 Bearing capacity of vertically loaded strip footings . . . . . . . . . . . . . . 102
5.3.1 Mechanical description of the system . . . . . . . . . . . . . . . . . 102
5.3.2 Stochastic variables and parametric study . . . . . . . . . . . . . . . 104
5.3.3 Sensitivity analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
5.3.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
5.4 Tunnel face stability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
5.4.1 Mechanical description of the problem . . . . . . . . . . . . . . . . 110
5.4.2 Stochastic variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
5.4.3 Parametric studies . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
5.4.4 Sensitivity analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
5.4.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
5.5 Synopsis of the case studies . . . . . . . . . . . . . . . . . . . . . . . . . . . 118

6 Selected case studies in uncertainty quantification including spatial variability119

6.1 Estimation of tunnelling induced settlements . . . . . . . . . . . . . . . . . 119
6.1.1 Mechanical description of the problem . . . . . . . . . . . . . . . . 120
6.1.2 Stochastic variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
6.1.3 Parametric study on tunnelling settlement in a single-layered subsoil122
6.1.4 Parametric study on tunnelling settlement in a two-layered soil . . 125
6.1.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128
6.2 Slope stability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
6.2.1 Mechanical description of the problem . . . . . . . . . . . . . . . . 130
6.2.2 Parametric studies of single-layered soil slopes . . . . . . . . . . . . 131
6.2.3 Parametric studies of two-layered soil slopes . . . . . . . . . . . . . 138
6.2.4 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141
6.3 Risk based characterisation of an urban site . . . . . . . . . . . . . . . . . . 144
6.3.1 Site description . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144
6.3.2 Calculation of fragility curves . . . . . . . . . . . . . . . . . . . . . . 146
6.3.3 Evaluation of the ultimate loads according to the state-of-the-art . 149
6.3.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150
6.4 Synopsis of the case studies . . . . . . . . . . . . . . . . . . . . . . . . . . . 154

7 Summary and conclusions 155

7.1 Potential and limitations of probabilistic methods . . . . . . . . . . . . . . 155
7.2 How well do we need to know soil variability? . . . . . . . . . . . . . . . . 156
7.3 What is the best way to evaluate the effects of soil variability? . . . . . . . 157
7.4 Recommendations for research . . . . . . . . . . . . . . . . . . . . . . . . . 158

Bibliography 160

A Database on stochastic soil properties 191

A.1 Database on the variability of rock . . . . . . . . . . . . . . . . . . . . . . . 191
[email protected]
A.2 Database on the variability of frictional soil . . . . . . . . . . . . . . . . . . 193

iii
Contents

A.3 Database on the variability of cohesive soil . . . . . . . . . . . . . . . . . . 197

B Measurement data of CPT - databases 207

C Basic definitions & statistical background 217

C.1 Moments of a distribution function . . . . . . . . . . . . . . . . . . . . . . . 217
C.2 Distribution functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 218
C.2.1 Normal distribution function . . . . . . . . . . . . . . . . . . . . . . 218
C.2.2 Lognormal distribution function . . . . . . . . . . . . . . . . . . . . 219
C.2.3 Multivariate normal distribution . . . . . . . . . . . . . . . . . . . . 220
C.2.4 Estimation methods for probability distribution functions . . . . . 221

D Polynomial Chaos Expansion 223

D.1 Hermite polynomials . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 223
D.2 Computation of the expectation of products of bivariate Hermite polyno-
mials . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 223
D.3 Two variables PCEs used for different values of the PCE order . . . . . . . 225

E Sequential simulation algorithms 227

E.1 Sequential Gaussian Simulation algorithm . . . . . . . . . . . . . . . . . . . 228
E.2 Kriging algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 229
E.3 Sequential Indicator Simulation algorithm . . . . . . . . . . . . . . . . . . . 230
E.4 Comparison of SGSIM and SISIM algorithms . . . . . . . . . . . . . . . . . 231

F Failure criteria in constitutive modelling of soil 237

G Application of non-intrusive SFEM to a reliability problems 239

H Pluri-Gaussian Simulation Method 245

H.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 245
H.2 Pluri-Gaussian simulation approach . . . . . . . . . . . . . . . . . . . . . . 246
H.2.1 Geological constraints . . . . . . . . . . . . . . . . . . . . . . . . . . 246
H.2.2 Pluri-Gaussian simulations . . . . . . . . . . . . . . . . . . . . . . . 255

[email protected]

iv
 List of Figures

1.1 Data on the frequency of natural and industrial disasters against num-
ber of lost humans lives. In colour shading data from Dutch Govern-
ment Group Risk Criteria [395] for floods, data on other natural disaster
as reported by the N ATURAL D ISASTERS D ATABASE NATURAL DISASTERS
DATABASE and Christian & Baecher [23]. . . . . . . . . . . . . . . . . . . . . 2
1.2 Average annual damages ($US billion) caused by reported natural disas-
ters 1990 - 2011 by the I NTERNATIONAL D ISASTERS D ATABASE [378]. . . . 2

2.1 Mapping of the same area of Canada in 1958 (left) and in 1923 (right) using
the same data [22]. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.2 Uncertainties in Reliability Based Design modified from Baecher & Chris-
tian [23], Honjo [166], Phoon[286] and Phoon & Kulhawy [282, 283]. . . . . 9
2.3 Illustration of the multi-scale nature of soil after Borja [48], Chen et al.
[73], Christakos [82] and Wackernagel [401]. . . . . . . . . . . . . . . . . . . 10
2.4 Nested variogram from Jaksa [184]. . . . . . . . . . . . . . . . . . . . . . . . 22
2.5 Comparison for computing a variogram from regular sampling on a tran-
sect:
(a) with a complete set of data, indicated with • and
(b) with missing values indicated by ◦ from [403]. . . . . . . . . . . . . . . 23
2.6 The geometry for discretizing the lag into bins by distance and direction
in two dimensions from [403]. . . . . . . . . . . . . . . . . . . . . . . . . . . 24
2.7 Theoretical variogram functions (a) and comparison of semivariance func-
tion and autocorrelation function (b). . . . . . . . . . . . . . . . . . . . . . . 25
2.8 Dependency of the correlation length on the sample size, taken from De-
Groot & Baecher [90]. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29

3.1
Lognormal distribution with a mean value of Ȳ = 1 and different coeffi-
cients of variation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
3.2 Quantile-quantile transformation of normally N (µ = 0, σ = 1) into log-
normally distributed data log-N (µ = 1.5, σ = 10). . . . . . . . . . . . . . . 35
3.3 Evaluated correlation lengths using MoMvar , MoMLA and ML approach
(a,c) and resulting nugget/sill ratios (b,d) using MoMvar and ML approach
for different error levels ε = 0.01 (a,b) and ε = 0.0001 (c,d). . . . . . . . . . 36
3.4 Case study A: Analysis of the indicator correlation lengths for the different
quantiles of the CDF (log-N (µ = 1.5, σ = 10)) and a measurement noise
of N (µ = 0, σ = 0.01). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
3.5 Case study A: distributions of the relative errors εtarget of the MoMvar ,
[email protected]
MoMLA and ML- approaches for COV = 1 % (a) and COV = 100 % (b). . . 39

v
List of Figures

3.6 Case study B: plan view of the 138 CPT measurement locations. . . . . . . 40
3.7 Case study B: Average, minimum and maximum measurement values of
the CPT data. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
3.8 Steps of the stochastic site characterisation and involved methods. . . . . . 42
3.9 Histogram and fitted normal distribution function of the detrended cone
resistances in sand of CPT 4 N (µ = 0.10 kN/m2 |σ = 2.51 kN/m) with a
skewness γ1 = 0.40( kN/m2 )3 and an excess kurtosis γ2,excess = −0.51( kN/m2 )4 . 44
3.10 (a) Cone resistance qc (z) measured with depth z and soil layer
identification of CPT 4 data using RI and Bartlett statistics and
(b) semivariogram on the raw and detrended data of the silty clay layer
at 10 m depth. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
3.11 Case study B: probability density function of the MoMvar , MoMLA and ML
approach and the combined probability density functions using Bayesian
Model Averaging of the silty clay layer. . . . . . . . . . . . . . . . . . . . . 48
3.12 Casestudy B: Analysis of the correlation lengths for indicators of the CDF
using MoMvar , MoMLA and ML with fitted lognormal distribution func-
tions for silty clay layer. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
3.13 Plan view and cross section of the experiments in the Fasanenhof tunnel-
ing project. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
3.14 Experimental equipment of the borehole jacking probe according to the
DIN 4094-5 [104]. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
3.15 Loading, unloading and reloading cycles performed in the boreholejack-
ing tests according to DIN 4094 − 5 [104]. . . . . . . . . . . . . . . . . . . . 53
3.16 Measurement results within the Fasanenhoftunnel. . . . . . . . . . . . . . 53
3.17 Casestudy C: Analysis of the indicator correlation lengths using MoMvar ,
MoMLA and ML of the EU,3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
3.18 Case study C: combination of the results using BMA. . . . . . . . . . . . . 55
3.19 Case study D: Plan view of the four sites at the Sheikh Zayed road in Dubai. 57
3.20 (a) Observations of the uniaxial compressive strength,
(b) histogram and fitted lognormal probability distribution function of
the UCS at the Sheikh Zayed road in Dubai. . . . . . . . . . . . . . . . 58
3.21 Case study D: Combination of the MoMvar , the MoMLA and the ML ap-
proach using BMA. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
3.22 Case study D: Analysis of the indicator correlation lengths using using the
[email protected]
MoMvar , the MoMLA and the ML approach. . . . . . . . . . . . . . . . . . . 59

vi
 List of Figures

3.23 Soil classification into 9 soil behaviour types according to Robertson [308]:
zone 1: sensitive fine grained
zone 2: organic soil-peat
zone 3: clay-silty clay
zone 4: clayey silt - silty clay
zone 5: silty sand - sandy silt
zone 6: clean sand to silty sand
zone 7: gravelly sand to sand
zone 8: very stiff sand to clayey sand and
zone 9: very stiff to fine grained soil. . . . . . . . . . . . . . . . . . . . . . . 63
3.24 Combination of measurement data of silty clay, results of the literature
database and results of the CPT databases using Bayesian Model Averaging. 65

4.1 Calibration of the partial safety factors according to E UROCODE [71]. . . . 72

4.2 Safety concepts in the context of structural engineering from Proske [295]. 74
4.3 Chart showing average annual risks posed by a variety of traditional civil
facilities and other large structures or projects proposed by Baecher &
Christian [23]. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
4.4 Lifetime probabilities of stabilitiy failures and comparative human risks
from Meyerhof [251]. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
4.5 Steps in Uncertainty Quantification from Sudret [364]. . . . . . . . . . . . . 78
4.6 Random field discretisation after Li & Kiureghian [219]. . . . . . . . . . . . 81
4.7 Non-deterministic approaches for uncertainty quantification modified from
[57, 176, 191, 262]. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
4.8 First order reliability method in Honjo [168] and Sudret [364]. . . . . . . . 83

5.1System of tunnel lining from Erdmann [120]. . . . . . . . . . . . . . . . . . 97

5.2Influence of the variability in parametric studies 1 and 2 on the tunnel lining. 99
5.3Accuracy estimation of the PCE in for the global sensitivity analysis. . . . 100
5.4Results of the local (a) and global (b) sensitivity analysis in parametric
study 2 of a tunnel lining in a soil with COVϕ’ = 20% and a COVc’ = 10%. . 100
5.5 Failure mechanism of the classic bearing capacity theory from Prandtl [293].103
5.6 (a) Variability of the theoretical results of the width stability number Nb
from
Perau [277] and
(b) geometry of the rigid strip footing. . . . . . . . . . . . . . . . . . . . . . 103
5.7 Different failure criteria of Mohr-Coulomb, Matsuoka-Nakai and Lade-
Duncan. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
5.8 Stability numbers Nb (a) and Nc (b) for different constitutive failure criteria. 105
5.9 Fragility curves for different constitutive failure criteria and different COVs
of the cohesion c′ and of the friction angel ϕ’. . . . . . . . . . . . . . . . . . 106
5.10 Influence of the COV of the cohesion c′ , of the friction angel ϕ’ and of the
footing pressure on the probability of the failure by using the MC-criterion
[email protected]
in parametric study 2. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107

vii
List of Figures

5.11 Local sensitivities of the cohesion c′ and the friction angle ϕ′ with respect
to the deterministic footing pressure in parametric study 1. . . . . . . . . . 108
5.12 Local (a) and global (b) sensitivity factors of the full probabilistic analysis
in parametric study 2 for COVϕ′ = 20%, COVc′ = 40% and COVq = 20%. . 108
5.13 Geometry of the tunnel. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
5.14 Flow area at collapse (a) and typical pressure displacement curve (b). . . . 111
5.15 Variability of the stability numbers Nc (a) and Nγ (b) for different consti-
tutive failure criteria (Mohr-Coulomb, Matsuoka-Nakai, Lade-Duncan) in
comparison to experimental results from literature [196]. . . . . . . . . . . 112
5.16 Influence of the COV of the cohesion c′ and of the friction angle ϕ′ on face
pressure versus probability of failure using the Mohr-Coulomb criterion. . 114
5.17 Limit state surfaces for different levels of the face pressure qt in (a) and
for probability of failure versus correlation coefficient between cohesion
c′ and friction angle ϕ′ in (b) for the MC- criterion. . . . . . . . . . . . . . . 115
5.18 (a) Variation of the COVs for the cohesion c′ , the friction angle ϕ′ and
the tunnel face pressure qt together with the corresponding probabili-
ties
of failure pf .
(b) Variation of the coefficients of variation for the cohesion c′ and the
friction angle ϕ′ for a given coefficient of variation COVqt = 10%
the tunnel face pressure qt , together with the corresponding probabil-
ities
of failure pf . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
5.19 Reliability indices βi and local sensitivity αi (a) as well as the global sen-
sitivity values δiP C values (b) for parametric study 2 on the tunnel face
stability using different constitutive failure criteria MC, MN, LADE. . . . 116

6.1
(a) Evaluation scheme of the probability of damage pdamage due to
differential settlements,
(b) cumulative distribution function and
(c) histogram with fitted probability density function of the normally
distributed limit state function g, based on an underlying random field
(B/D = 2, H/D = 1, µ = 60 MN/m2 , COV = 50 %, θh = θv = 2D). . . . 123
6.2 (a) Influence of the width B of the building,
(b) influence of the coefficient of variation,
(c) correlation lengths and
(d) influence of the anisotropy of the correlation structure on the surface
settlements due to tunnel excavation. . . . . . . . . . . . . . . . . . . . . 124
6.3 Effects of the ratio of anisotropy θh /θv on the maximum pdamage . . . . . . . 125
6.4 FEM mesh and location of the boreholes (B1,B2 B3). . . . . . . . . . . . . . 126
6.5 Results of parametric study on tunnelling in a layered soil. . . . . . . . . . 127
6.6 (a) Strength reduction factor versus maximum settlement of a 2D
homogeneous slope (c′ = 50 kN/m2 , ϕ′ = 0) and
(b) FEM mesh of the investigated slopes with 8-noded,
[email protected]
quadrilateral elements. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129

viii
 List of Figures

6.7 Effect of varying COVc on the reliability index β and on the probability
of failure pf using lognormally distributed random variables for µc = 50
kN/m2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132
6.8 (a) Effects of varying COVc on the probability of failure using different
random field generators with a Θ = θ/H = 2 and random variables
for µc = 50 kN/m2 ;
(b) effects of varying Θ = θ/H on the probability of failure for different
random field generators in comparison to random variables with
a COVc’ = 20%. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
6.9 Influence of the spherical, exponential and Gaussian correlation function
on the probability of failure of a slope. . . . . . . . . . . . . . . . . . . . . . 134
6.10 Different measures for the accuracy of the PCE to the system response of a
2D slope stability analysis (a,b) and (c) global sensitivity measures of the
input variables. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136
6.11 Different failure modes derived in Hicks & Spencer [158]. . . . . . . . . . . 137
6.12 Geometry of the layered slope including one borehole using random vari-
ables (a) and random fields (b). . . . . . . . . . . . . . . . . . . . . . . . . . 138
6.13 C ASE STUDY I: Reliability evaluation of a two layered slope stability prob-
lem (a) and the corresponding global sensitivity measures (b). . . . . . . . 140
6.14 C ASE STUDY II: Normalized reliability index of different RFEM approaches
for a two layered slope with respect to a varying Θboundary = θboundary /H. . 141
6.15 C ASE STUDY III: Different determination coefficients for the PCE accuracy. 142
6.16 C ASE STUDY III: Global sensitivity measures δiP C for the isotropic correla-
tion length of the upper layer (1) and the lower layer (2) as well as for the
layer boundary (3) for the PCE expansion order M = 5. . . . . . . . . . . . 142
6.17 Principles of the Pluri-Gaussian Simulation approach, [14]:
Two uncorrelated Gaussian random fields Y1 (a) and Y2 (b) with different
anisotropies are combined via the lithotype rule (c); the red square high-
lights the way the lithotype rule is used to construct the Pluri-Gaussian
random field simulation (d). . . . . . . . . . . . . . . . . . . . . . . . . . . . 147
6.18 Part of the realisation of the Pluri-Gaussian random field (200/200/75m)
including all soil types. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147
6.19 2D FEM model used in this case study. . . . . . . . . . . . . . . . . . . . . . 148
6.20 Sample fragility curve for ultimate differential settlements
αultimate = 1/1, 000 and the fitted lognormal cumulative distribution func-
tion. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150
6.21 Map of allowable loads pSLS on a footing with αulitmate = 1/300 (a) and
αulitmate = 1/500 (b). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
6.22 Map of allowable loads pSLS on a footing with αulitmate = 1/600 (a) and
αulitmate = 1/1, 000 (b). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152

A.1 Statistics of the available properties in the literature database. . . . . . . . 203

A.2 Frequency of the correlation functions in the database for frictonal and
[email protected]
cohesive soils. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 203

ix
List of Figures

A.3 Histogram of vertical correlation lengths in cohesive soils for all scales (a),
for small(b), medium (c) and large scales (d). . . . . . . . . . . . . . . . . . 204
A.4 Histogram of horizontal correlation lengths in cohesive soils for all scales
(a), medium (b) and large scales (c). . . . . . . . . . . . . . . . . . . . . . . 205
A.5 Histogram of vertical correlation lengths in frictional soils for all scales
(a), medium (b) and large scales (c). . . . . . . . . . . . . . . . . . . . . . . 206
A.6 Histogram of horizontal correlation lengths in frictional soils for all scales
(a), medium (b) and large scales (c). . . . . . . . . . . . . . . . . . . . . . . 206

B.1 Plan view and typical CPT profile of NGES site A LAMEDA. . . . . . . . . . 208
B.2 Plan view and typical CPT profile of NGES site E VANSVILLE A REA. . . . . 209
B.3 Plan view and typical CPT profile of NGES site L ANCESTER. . . . . . . . . 210
B.4 Plan view and typical CPT profile of NGES site S AN B ERNADINO C OUNTY. 211
B.5 Plan view and typical CPT profile of NGES site S AN L UIS O BISPO C OUNTY.212
B.6 Plan view and typical CPT profile of NGES site S ANTA C LARA C OUNTY. . 213
B.7 Plan view and typical CPT profile of NGES site S OLANO C OUNTY. . . . . 214
B.8 Plan view and typical CPT profile of PEER site A NSSALL. . . . . . . . . . . 215
B.9 Plan view and typical CPT profile of PEER site B ERKELEY. . . . . . . . . . 216

E.1 Random field realisations of the SGSIM (a) and the SISIM (b) algorithm
with an isotropic, spatial correlation (θver = θhor ). . . . . . . . . . . . . . . . 232
E.2 Cumulative probability distribution (a) and variograms (b) of the SGSIM
and SISIM random field in figure E.1. . . . . . . . . . . . . . . . . . . . . . 233
E.3 Indicator correlation lengths for each threshold of the SGSIM and SISIM
algorithm. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 233
E.4 Main characteristics of various random field simulation methods from
Chiles & Delfiner [79]. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 234
E.5 Main characteristics of various random field simulation methods from
Chiles & Delfiner [79]. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 235

F.1 The SMP concept of Matsuoka & Nakai.

(a) Three mobilized planes where the maximum shear stress to normal
stress ratio is reached for the respective principal stresses.
(b) SMP in principal stress space from Benz [38]. . . . . . . . . . . . . . . . 238

G.1 Input variables and system response of an PCE with order M = 5 and a
face pressure of qt = 40 kN/m2 . . . . . . . . . . . . . . . . . . . . . . . . . . 240
G.2 Comparison of non-intrusive SFEM and FORM in terms of the probability
of failure for a COVϕ′ = 5 % and COVϕ′ = 10 % . . . . . . . . . . . . . . . . 241
G.3 Acccuracy of the PCE fitting for a face pressure qt = 40 kN/m2 . . . . . . . . 241
G.4 Collocation points in Gaussian space and physical space for several PCE
expansion orders M. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 242
G.5 Statistical moments of the approximated system response for a face pres-
sure qt = 40 kN/m2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 243
[email protected]
G.6 Probability of failure as a function of the PCE expansion order. . . . . . . . 243

x
 List of Figures

G.7 Sobol indices for several PCE expansion orders for the face pressure qt =
40 kN/m 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 244

H.1 (a) Histogram of a standard normal distribution (i.e. N (µ = 0, σ = 1))

showing the two thresholds -0.6 and 0.5,
(b) simulated greytone image with N (µ = 0, σ = 1) and
(c) same image after being truncated at the thresholds -0.6 an 0.5.
Values below -0.6 have been shaded dark grey, those between -0.6 and 0.5
are coloured light grey while values above 0.5 are shown in white, [14]. . . 246
H.2 (a) Simplified geological sketch map of the study area from Raspa et al.[301].
Legend:
(1) upper Pleistocene - Holocene alluvial deposits,
(2) middle-upper Pleistocene volcanic bedrock,
(3) Plio-Pleistocene sedimentary bedrock,
(4) boreholes with geotechnical samples,
(5) boreholes with samples endowed with the full set of
geotechnical information and
(6) track of the geological cross section in figure H.3 (a).

(b) conceptual geological model of the Tevere valley by [238]. . . . . . . . 248

H.3 (a) Geological cross section of the recent alluvial deposits filling the
Tevere valley from [301].
(b) 3D plot of the selected boreholes with location of the geotechnical
samples (black points) for location of the boreholes in figure
H.2 from [301]. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 249
H.4 Plan view investigated area and area of the case study (400 m x 600 m)
in the city of Rome with locations of the boreholes and soil investigations
from [238]. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 250
H.5 Variogram map from Marconi [238]. . . . . . . . . . . . . . . . . . . . . . . 251
H.6 Horizontal indicator variograms of the three different soil types from [238]. 252
H.7 Vertical indicator variograms of the three different soil types from [238]. . 253
H.8 Vertical proportion curve of an area 500 m x 500 m (a) and map of the
vertical proportion curves of the soil types 1, 2 and 3, [238]. . . . . . . . . . 254
H.9 Used lithotypes for the Pluri-Gaussian Simulation [238]. . . . . . . . . . . 255
H.10 Plan view of the Pluri-Gaussian mesh from Marconi [238]. . . . . . . . . . 256
H.11 Part of the realisation of the pluri-Gaussian random field for soil type 1
(a), soil type 2 (b). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 257
H.12 Part of the realisation of the pluriGaussian random field including soil
type 3 (a), soil type 4 (b), antropogenetic top soil layer (c). . . . . . . . . . . 258

[email protected]

xi
[email protected]
 List of Tables

2.1 Approximate guidelines for coefficients of variation of some design soil

parameters taken from [183, 282, 283]. . . . . . . . . . . . . . . . . . . . . . 8
2.2 Different scales of spatial variability of soil properties of sedimentary de-
posits by Koltermann & Gorelick [204]. . . . . . . . . . . . . . . . . . . . . 12
2.3 Types of sampling strategies defined by two sources of randomness ac-
cording to Brus & de Gruijter [56]. . . . . . . . . . . . . . . . . . . . . . . . 20
2.4 Relationship between the scale of fluctuation δ and the correlation dis-
tance θ for various autocorrelation functions from Vanmarcke [384]. . . . . 28
2.5 Typical ratios of the horizontal θhor and vertical correlation lengths θver
collected from literature [27, 79, 96, 97, 190, 204, 401, 408]. . . . . . . . . . . 30

3.1 Lognormally distributed results of the experiments in the Fasanenhof tun-

nel. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
3.2 Case study C: Comparison of the evaluated correlation lengths θ. . . . . . 54
3.3 Correlation lengths using the J ACKKNIFE approach for the MoMvar ap-
proach in comparison to the mean value and the standard error of the ML
approach. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
3.4 Properties of the lognormally distributed horizontal and vertical correla-
tion lengths of CPT data for frictional and cohesive soils. . . . . . . . . . . 62
3.5 Summary of the different CPT databases [50, 129, 275, 380]. . . . . . . . . . 64
3.6 Mean value and COV of the vertical correlation lengths of the different
CPT databases and the BMA combination results. . . . . . . . . . . . . . . 64

4.1 Factors of global safety in engineering according to Visodic [397]. . . . . . 70

4.2 Factors of global safety in geotechnical engineering after Terzaghi & Peck
[373]. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
4.3 Target reliabilities according to E UROCODE [71]. . . . . . . . . . . . . . . . 75
4.4 Reliability based design: different levels of accuracy from Honjo et al. [168]. 77

5.1 Properties of parametric studies on the tunnel lining. . . . . . . . . . . . . 98

5.2 Stochastic variables of the silty sand in two parametric studies. . . . . . . 105
5.3 Properties of parametric study 1 on the tunnel heading using the Mohr-
Coulomb failure criterion. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
5.4 Properties of parametric study 2 on the tunnel heading using the Mohr-
Coulomb failure criterion. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115

6.1 Material properties for the parametric study on tunnelling settlement in

[email protected]
an single-layered subsoil. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121

xiii
List of Tables

6.2 Material properties for the parametric study on tunnelling settlement in a

layered soil. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
6.3 Lognormal distributed input parameters of the parametric studies. . . . . 131
6.4 Lognormal distributed input parameters of the parametric study. . . . . . 135
6.5 Lognormal distributed input parameters of the C ASE STUDIES I, II and III. 139
6.6 Mean values and standard deviations of the lognormally distributed mea-
surement and literature data. . . . . . . . . . . . . . . . . . . . . . . . . . . 145
6.7 Number of measurements inside each soil type. . . . . . . . . . . . . . . . 146

A.1 Database on the variability of rock . . . . . . . . . . . . . . . . . . . . . . . 192

A.2 Database on the variability of frictional soils. . . . . . . . . . . . . . . . . . 194
A.3 Database on the variability of cohesive soils. . . . . . . . . . . . . . . . . . 198

C.1 Summary of the normal distribution . . . . . . . . . . . . . . . . . . . . . . 219

C.2 Summary of the lognormal distribution . . . . . . . . . . . . . . . . . . . . 220
C.3 Summary of the multivariate normal distribution . . . . . . . . . . . . . . 220

H.1 Input parameters of the soil model. . . . . . . . . . . . . . . . . . . . . . . . 259

[email protected]

xiv
 Summary
The core competences of civil engineers are designing, building and maintaining struc-
tures and buildings to enable life and business for society. This includes the prevention
against natural hazards such as climatological, hydrological, meteorological and geo-
physical disasters. Such complex hazards asks for sophisticated techniques to ensure
appropriate safety standards for society. Too low safety standards can result in many
casualties and much economic damage, whereas too high standards results in overly ex-
pansive systems. Therefore, these phenomena ask for sophisticated methods to consider
their impacts on structures. Especially geotechnical engineers are asked for integrated
concepts to design structures withstanding the above mentioned hazards.
This research is focusing on the evaluation of the effects of soil variability within the
framework of probabilistic methods. The evaluation of stochastic soil properties includ-
ing spatially correlated soil properties is deeply investigate on a theoretical basis. These
results are used in case studies on the evaluation of the spatial correlation of different
measurement data sets. These outcomes are also used in different case studies, which
are focusing on the effects of soil variability for typical geotechnical problems.

This thesis is organized thematically in the following parts:

Chapter 2: The basics of probabilistic site characterization are described within this
chapter including a comparison to the state-of-the-art. This involves the description of
the basics of statistics and geostatistics in order to describe the variability and the spatial
correlation of soil properties. Besides this, the main sources of error in geotechnical
engineering are summarized, which influence probabilistic site characterization. Also an
overview of sampling schemes is provided. Three main methods for the mathematical
characterization of spatial variability are derived. These methods are used to analyse
and quantify spatial variability of soil properties.

Chapter 3 The possibilities and limitations of the three different approaches on the
evaluation of spatial variability can be deduced from the results of four different case
studies. At first, an analytically defined random process is used to show the capabilities
of the three different methods. On basis of this, three case studies are presented on the
evaluation of spatial variability of equally spaced and irregular spaced data sets. These
datasets are used to explain the concepts of probabilistic site characterization, which
includes the identification of trends and layers as well as a novel scheme for the combi-
nation of different models for spatial variability by means of statistical approaches.
The results of these case studies are compared to a large literature study on the spa-
[email protected]
tial correlation of soil properties, which is enriched by the results of an extensive study

xv
Summary

on cone penetration tests in different soil types. Finally, these different sources of in-
formation on the correlation length are combined via the Bayesian Model Averaging
methodology.

Chapter 4: This chapter focuses on the basics of safety and reliability in engineering. It
provides a description of common approaches for dealing with uncertainties and safety
in geotechnical engineering. Global and partial safety factors as well as the basics of
uncertainty quantification and reliability based design are described and compared with
a special focus on the basics of the generation of random numbers and random fields as
well as on the computation of failure probabilities.
Moreover, an introduction into local and global sensitivity analyses is given.

Chapter 5: Typical geotechnical design problems are investigated within the frame-
work of uncertainty quantification.
At first, an analytical limit state equation (LSE) is used to investigate the effects of soil
variability on the design of a tunnel lining.
In another case study the bearing capacity of a strip footing is investigated by means
of probabilistic methods. Within this, semi-analytical LSEs are derived from 2D FEM
simulations, which considers the Mohr-Coulomb, Matsuoka-Nakai, Lade-Duncan fail-
ure criteria to model the soil behaviour. These LSEs are used to investigate the effects
of the soil variability, of the geometry of the problem and of the load uncertainty. Sen-
sitivity analyses evaluate the contribution of the random properties to the probability of
failure.
The tunnel face stability problem is investigated in a similar way. The semi-analytical
LSEs are derived from parametric 3D FEM simulations using the Mohr-Coulomb, Mat-
suoka-Nakai, Lade-Duncan failure criteria. The effects of soil variability, geometry and
construction processes on the probability of failure are quantified by means of uncer-
tainty quantification and sensitivity analyses.

Chapter 6: The effects of spatial soil variability are evaluated in typical problems in
tunnelling and foundation engineering by using the framework of uncertainty quantifi-
cation.
In the first case study, the effects of spatially correlated soil properties on surface set-
tlements, which are induced by the construction of a tunnel, are investigated. Starting
from traditional approaches of a single-layered subsoil, a novel concept for considering
the spatial variability at multiple scales is presented.
The effects of soil variability at different scales are also investigated within slope sta-
bility problems. These studies also include sensitivity analyses to investigate the contri-
bution of spatial variability to the probability of failure.
Finally, the effects of large scale spatial variability is focused in the case study on the
risk-based characterisation of an urban site. The macro-scale variability of the subsoil is
simulated via the Pluri-Gaussian simulation approach. This approach captures the un-
[email protected]
certainty of the boundaries of the spatially distributed soil types, incorporating expert

xvi
 judgements, soil investigations and stochastic properties of the soil types. These simula-
tion results analysed using fragility curves. Herein, fragility curves suggest admissible
footing pressure due to differential settlements. These results are used for the generation
of risk maps of the investigated urban site.

Chapter 7: This chapter comprises a summary of the objectives of this thesis and a
summary of the conclusions. Further research topics are announced in the outlook.

[email protected]
[email protected]
 Zusammenfassung
Die Kernkompetenzen des Bauingenieurwesens sind das Planen, Erstellen und Erhal-
ten von Bauwerken, um die Existenz und Wirtschaftsleben einer Gesellschaft zu er-
möglichen. Dies umfasst auch Maßnahmen gegen Naturkatastrophen wie z.B. Klima-
wandel, Hochwasser oder Massenbewegungen. Diese komplexen Gefahren erfordern
hochentwickelte Technologien, um ein angemessenes Level an Sicherheit für die Gesell-
schaft zu sichern. Zu niedrige Sicherheiten können zu großen Verlusten und hohen
ökonomischen Schäden führen, wohingegen zu hohe Sicherheitsanforderungen in einem
teuren und einer nicht wirtschaftlichen Bemessung von Bauwerken enden. Daraus kann
abge-leitet werden, dass hochentwickelte Methoden notwendig sind, um die Einwirkun-
gen auf Bauwerke zu simulieren. Speziell in der Geotechnik ist es in Folge der hohen
Variabilität des Untergrundes erforderlich, integrierte Konzepte für die Bemessung von
Bauwerken gegen die zuvor genannten Einwirkungen anzuwenden.
Im Rahmen dieser Dissertation wird die mathematische Beschreibung von Boden-
variabilität und die Bestimmung der Konsequenzen dieser Variabilität in verschiede-
nen geotechnischen Fragestellungen untersucht. Hierbei werden probabilistische Meth-
oden angewendet, um die Folgen von unsicheren und räumlich streuenden Bodenken-
ngrößen zu bestimmen. Neben der Zusammenstellung der theoretischen Grundlagen
der Quantifizierung von räumlichen Varaibliät wird die Anwendung exemplarische in
mehreren Fallbeispielen aufgezeigt. Diese Resultate werden in weiterführenden Unter-
suchungen verwendet, in welchen die Folgen von Bodenvariabilität in geotechnischen
Problembestellungen beispielhaft analysiert werden.

Diese Dissertation ist thematisch in die folgenden Teile unterteilt:

Kapitel 2: In diesem Kapitel werden die Grundlagen für eine probabilistische Charak-
terisierung des Untergrundes zusammengestellt. Dies umfaßt die theoretischen Grund-
lagen der Statistik und Geostatistik, welche für die Beschreibung von Variabilität und
räumlicher Korrelation von Bodeneigenschaften erforderlich sind. Drei gängige Meth-
oden für die Analyse von räumlicher Variabilität in der Geotechnik werden hergeleitet
und einander gegenübergestellt. Diese ist notwendig, um die Ergebnisse der Analyse
von Messdaten in Hinblick auf die räumliche Variabilität zu verstehen und zu inter-
pretieren.

Kapitel 3: In vier verschiedenen Fallstudien wird die Bestimmung der räumlichen

Variabilität von Bodeneigenschaften erläutert.
In der ersten Fallstudie werden drei verschiedene Analysemethoden von räumlicher
Variabilität an einem analytische definiertem Zufallsprozess angewandt, um so plakativ
[email protected]
die Vor- und Nachteile der verschiedenen Methoden aufzuzeigen.

xix
Zusammenfassung

Auf Basis dieser Ergebnisse wird in der zweiten Fallstudie eine Datenbank von CPT
Feldversuchen analysiert, welche im Rahmen einer Großbaustelle durchgeführt wur-
den. Exemplarisch wird hier das Erstellen eines stochastischen Modells des Baugrundes
gezeigt. Hierbei wird mittels empirischer und statistischer Methoden die Variabilität des
Untergrundes auf verschiedenen Skalen charakterisiert. Exemplarisch wird das Kom-
binieren verschiedener Analysemodelle von räumlicher Variabilität aufgezeigt. Dies
Herangehensweise beruht auf dem Satz von Bayes.
In einer weiteren Fallstudie wird die räumliche Variabilität von Steifigkeiten in einer
homogenen Bodenschicht untersucht. Anhand dieser Analyse wird eine Methodik zur
Bestimmung der Unsicherheit von Modellen der räumlichen Variabilität abgeleitet.
In der vierten Fallstudie wird die räumliche Variabilität einer sehr großen Anzahl von
Festigkeitsmessungen analysiert, welche auf verschiedenen Baustellen durchgeführt wur-
den. Die Ergebnisse werden mit den Daten einer umfangreichen Literaturstudie ver-
glichen und verifiziert.
Abschließend wird eine Methodik für die Kombination von gemessener räumlicher
Varaibität mit einer Expertenmeinung vorgestellte. Das Bayes’sche Prinzip wird ver-
wendet, um verschiedene Vorinformationen, Expertenwissen, Literaturergebnisse und
Messdaten miteinander zu kombinieren. Dieses Vorgehen wird exemplarisch an einem
Datensatz aufgezeigt.

Kapitel 4: In diesem Kapitel sind die Grundlagen für Sicherheit und Zuverlässigkeit
zusammengestellt. Ausgehend vom Stand der Technik, bei dem globale und partielle
Sicherheiten verwendet werden, werden die Grundlagen der Zuverlässigkeitsanalyse
von Bauwerken erklärt. Hierbei wird besonders auf die in der Geotechnik gängigen
Näherungsverfahren zur Ermittelung der Versagenswahrscheinlichkeit eingegangen, wo-
bei auch das Generieren von Zufallszahlen und -feldern beschrieben wird.
Die mathematischen Grundlagen für die Konzept zur lokalen und globalen Sensitiv-
itätsanalyse werden am Ende dieses Kapitels erklärt.

Kapitel 5: In diesem Kapitel wird das Konzept der Quantifizierung von Unsicherheit
anhand von verschiedenen Fallstudien in der Geotechnik erläutert. Hierbei wird die
Bodenvariabilität durch Zufallszahlen simuliert.
In der ersten Fallstudie im Tunnelbau werden die Folgen von streuenden Bodenken-
ngrößen in der Bemessung einer Tunnelschale untersucht. Hierfür wird eine analytische
Versagenszustandsgleichung für die Bestimmung der Änderung der Wahrscheinlichkeit
eines Versagens in Folge von streuenden Bodenkenngrößen untersucht. Lokale bzw.
globale Sensitivitäten quantifizieren den Einfluss der streuenden Eigenschaften mit Hilfe
statistischer Methoden.
Die Folgen von Bodenvariabilität auf die Tragfähigkeit von vertikal belasteten Streifen-
fundaments mittels probabilistischen Methoden untersucht. Darüber hinaus werden
neben der Variabilität der Festigkeitseigenschaften des Untergrundes auch die Geome-
trie und die Belastung des Streifenfundamentes analysiert. Dies wird mit verschiedenen
[email protected]
Versagenszustandsgleichungen beschrieben, welche aus 2D FEM Untersuchungen mit

xx
 verschiedenen stofflichen Versagenskriterien (Mohr-Coulomb, Matsuoka-Nakai, Lade-
Duncan) abgeleitet werden. Für die Analyse dieser Fragestellung werden Fragilitäts-
graphen verwendet. Fragilitätsgraphen beschreiben in diesem Zusammenhang zwis-
chen aufgebrachter Last und Versagenswahrscheinlichkeit.
In einer weiteren Fallstudie wird Stabilität der Ortsbrust im Tunnelbau untersucht.
Das Versagen der Ortsbrust wird mit verschiedenen stofflichen Versagenskriterien (Mohr-
Coulomb, Matsuoka-Nakai, Lade-Duncan) in Kombination mit Methoden der Zuverläs-
sigkeitsanalyse untersucht, wobei in diesem Zusammenhang neben der Untergrund-
varibilität auch die Geometrie und der Herstellungsvorgang mittels Zufallsvariablen
berücksichtigt werden. Der Vergleich der verschiedenen Ergebnisse wird mit einem Ver-
gleich der Sensitivitäten der verwendeten Zufallsvariablen abgerundet.

Kapitel 6: In diesem Kapitel werden die Auswirkungen von räumlicher Variabilität in

Fallbeispielen aus dem Tunnelbau und Grundbau dargestellt.
Der Auswirkungen von räumlich streuenden Bodenkenngrößen wird in einer Fall-
studie zur Setzungen infolge Tunnelbau untersucht. Ausgehend von der konventionellen
Herangehensweise von homogenen Zufallsfeldern wird eine Erweiterung für die Berück-
sichtigung von räumlich korrelierten Bodeneigenschaften auf mehreren Skalen präsen-
tiert und die Auswirkungen exemplarisch aufgezeigt.
Auch in der Fallstudie zur Böschungsstabilität werden die Auswirkungen von räum-
licher Variabilität auf mehreren Skalen untersucht. Im Rahmen dieser Untersuchungen
wird ein Schema zur Bestimmung der Sensitivität der Eingangsparameter für räumliche
Variabilität auf verschiedenen Skalen aufgezeigt.
In einer letzten Fallstudie werden die Folgen von makro-skaliger Bodenvariabilität
untersucht. Hierfür werden die Ergebnisse einer gestatistischen Simulation verwen-
det, welche Expertenmeinung, Bodenaufschlüsse und stochastische Bodeneigenschaften
verschiedener Bodentypen berücksichtigt. Diese Ergebnisse werdenfür die Erstellung
von Risikokarten verwendet. Die erstellten Risikokarten geben eine zulässige Last an,
welche für ein definiertes Gebäude und eine zulässige Verdrehung möglich ist.
Dieses Kapitel schließt mit einer Zusammenfassung der Erkenntnisse, welche aus den
präsentierten Fallstudien abgeleitet werden.

Kapitel 7: Im abschließenden Kaptiel werden die Themen dieser Dissertation zusam-

mengefasst und die Schlußfolgerung werden präsentiert.

[email protected]
[email protected]
 Chapter 1

Introduction

1.1 Background and rationale

The core competence of civil engineers are designing, building and maintaining struc-
tures and buildings to enable life and business for society. This includes the prevention
against natural hazards. Natural hazards can be subdivided into climatological, hydro-
logical, meteorological and geophysical disasters. As defined in the I NTERNATIONAL
D ISASTERS D ATABASE, climatological disasters are (bush, forest, scrub and grassland)
fires, meteorological disasters are local storms, extra-tropical and tropical cyclones. Hy-
drological disasters are general floods, flash floods, mudslides, storm surges and coastal
floods. The complexity of natural hazards becomes more evident in figure 1.1. Herein,
the annual frequency of floods, mass movements, seismic activities, storms, vulcanos,
and wildfires is plotted with respect to the lost lives. These results from the I NTERNA -
TIONAL D ISASTERS D ATABASE [378] show good agreement with data from Christian &
Baecher [23] on nuclear power plants, dam failures, explosions air crashes and man-
caused fatalities. The Dutch Government Group Risk Criteria [395] indicate that these
natural hazards are not acceptable risks and have to be mitigated with advanced con-
cepts. This asks for sophisticated techniques to ensure appropriate safety standards for
society.
Too low safety standards can result in many casualties and much economic damage,
whereas too high standards results in overly expansive systems. Therefore, it is impor-
tant to evaluate the safety of structures with appropriate approaches, which allow to
consider variability and uncertainty in a proper way. These complex phenomena cause
tremendous economical damage as shown in figure 1.2. One can deduce that sophisti-
cated methods to consider their impacts on structures are urgently needed. Especially
geotechnical engineers are asked for integrated concepts to design structures withstand-
ing the above mentioned hazards.
In international conferences and workshops the state-of-the-art safety concepts are
continuously improved. The global and partial safety factor concepts are mainly driven
by experience and now enriched and extended by the results of probabilistic analy-
ses. Theses new developments help to contribute to more economic and safe design
[email protected]
approaches.

1
Chapter 1 Introduction

annual frequency
1
10
floods
0
mass movements
10 seismic activities
exp storms
l osi
-1 on vulcano
10
wildfire
10
-2 acceptable risk
reduction desired risk
-3
not acceptable risk
10

da
m
f a il
-4
10 ai

ure
nuc
lea
rc
-5 rp
ra
10 ow
s
he
er
st
pl
a
ot
-6
10
al
nt
s

-7
10 1 2 3 4 5 6
10 10 10 10 10 10
lives lost

Figure 1.1: Data on the frequency of natural and industrial disasters against number of
lost humans lives. In colour shading data from Dutch Government Group
Risk Criteria [395] for floods, data on other natural disaster as reported by the
N ATURAL D ISASTERS D ATABASE NATURAL DISASTERS DATABASE and Chris-
tian & Baecher [23].

1.0
Average annual estimated damages

geophysical
Proportion of average annual
damages per disaster group
($US billion) per disaster group

hydrological
40 meteorological 0.8
climatological
30 0.6

20 0.4

10 0.2

0 0.0
A s

A s
a

a
ia

ia
a

a
a

a
pe

pe
ic

ic
si

si
an

an
ic

ic
fr

fr
ro

ro
er

er
ce

ce
A

A
Eu

Eu
m

m
O

O
A

Figure 1.2: Average annual damages ($US billion) caused by reported natural disasters
1990 - 2011 by the I NTERNATIONAL D ISASTERS D ATABASE [378].
[email protected]

2
 1.2 Research aim

1.2 Research aim

The research is focusing on the evaluation of the effects of soil variability within the
framework of probabilistic methods. To achieve this aim, several topics are studied to
quantify the effects of spatially correlated soil properties. Topics studied include evalua-
tion of spatially correlated properties, geostatistical simulation methods, and probabilis-
tic methods within geotechnical and civil engineering. In addition to this, case studies
are elaborated to demonstrate the application of probabilistic methods in tunnelling and
foundation engineering. These established calculation procedures for uncertainty quan-
tification are seen as guidelines for reliability based design in geotechnics.

1.3 Thesis scope

In addition to this introduction, the thesis is arranged in six chapters as indicated below:
At first, an introduction into probabilistic site characterization is given in chapter
Chapter 2. This includes the description of the mathematical framework to quantify
spatial variability at different scales.
Subsequently, this framework is applied in four different case studies within Chapter
3, which shall help the reader to understand and apply these concepts of spatial variabil-
ity in geotechnical engineering. The results of these case studies compared to the results
of an extensive literature review of contributions on spatial variability of soil properties.
On top of this, these findings are compared with the results of a CPT database analy-
sis. The evaluation of spatial variability is improved by merging different sources of
knowledge and information together with measurement data.
Chapter 4 focuses on the basics of safety and reliability and provided a description of
common approaches for dealing with uncertainties and safety in geotechnical engineer-
ing. Global and partial safety factors as well as the basics of uncertainty quantification
and reliability based design are described and compared to each other. This includes
the basics of the generation of random numbers and random fields as well as the com-
putation of failure probabilities. This chapter also provides an introduction to local and
global sensitivity analysis of systems.
In Chapter 5, typical geotechnical case studies dealing with tunnelling and footing
analyses are investigated within the framework of uncertainty quantification. Herein,
soil variability is represented by random variables. Starting from semi-analytical limit
state equations (LSE) for the design of tunnel-linings, the ultimate limit state of a footing
is investigated in 2D. This investigation encloses the effects of soil variability by means of
fragility curves and probabilistic methods. Moreover, the effects of the Mohr-Coulomb,
Masuoka-Nakai and Lade-Duncan failure criteria are investigated by means of proba-
bilistic methods and compared to each other. In addition to this, the ultimate limit state
of the tunnel face is investigated by using three different constitutive failure criteria. The
results of these 3D FEM studies are used for the formulation of the LSE, which is used
for the uncertainty quantification of the tunnel face stability.
Chapter 6 presents three different case studies in tunnelling and foundation engineer-
[email protected]
ing focusing on the effects of spatial soil variability. The effects of spatial variability on

3
Chapter 1 Introduction

the estimation of tunnelling induced settlements are investigated for a single- and two
layered subsoil using probabilistic methods. In addition to this, the effects of multi-scale
soil variability are also investigated for soil slopes. The quantification of these effects
by means of global sensitivity measures offers insight into the contribution of uncertain
properties of single- and two-layered soil slopes. In the case study on the risk-based site
characterisation the effects of macro-scale soil variability are investigated by means of
the Pluri-Gaussian simulation method, expert judgement and the framework of uncer-
tainty quantification.
Chapter 7 is a summary of the most relevant findings of this research, drawing con-
clusions and including recommendations for further research.

[email protected]

4
 Chapter 2
Characterizing soil variability at different scales

2.1 Introduction
Interpretation of site exploration data is illustrated in figure 2.1 [22]. The presented geo-
logical maps were drawn 30 years apart and are quite different from each other, although
the sample data are the same. According to Baecher [22], the theory of geology was re-
evaluated in the profession and this led to a different reinterpretation of the data.
One can conclude from [22] that for site characterization only parts of exploration can
be statistically modelled. Statistical analysis of data can at most indicate how to logically
modify what was thought before to what should be thought after.
For this reason, it is useful to use different statistical approaches and schemes to de-
scribe various phenomena in detail. Within this chapter, the basics of probabilistic site
characterization are described, which involves a summary of uncertainties and main
sources of error in geotechnical engineering. The focus of the author is to enlighten
the mathematical framework to describe spatial variability at different scales. Herein, a
basic description of the uncertainty of the spatial correlation is offered together with a
summary of the anisotropy ratios of the spatial correlation.

2.2 Probabilisitic site characterization

According to Baecher & Christian [23], site characterization can be defined as a set of
activities (e.g. processes), which will lead to information about site geology. In this way
one can get estimates of parameters about site geology and finally one can get estimates
of parameters to be used in modelling engineering processes. This has been previously

Figure 2.1: Mapping of the same area of Canada in 1958 (left) and in 1923 (right) using
[email protected]
the same data [22].

5
Chapter 2 Characterizing soil variability at different scales

thought of as an entirely intuitive process based on engineering judgement and without

any analytical or mathematical consideration, but meanwhile approaches in site charac-
terization have been developed in the oil, gas and mining industries. In these industries
the importance of accurate estimates of oil reservoirs or mineral deposits are of greater
economic consideration than, e.g., finding weak soil layers beneath a dam.
Baecher & Christian [23] developed a site characterization program including three
stages: reconnaissance, preliminarily investigation and detailed investigation. In the re-
connaissance, which is also called desk study, the engineer tries to find information on
site geology from different sources. During the preliminarily investigation, the quantita-
tive estimates of properties are used to verify the assumptions from the first stage. In
[23], the authors stress that not too many tests are to be performed. After this, a detailed
investigation is carried out to verify and refine the model of the site geology. One can eas-
ily conclude that for this reason a broad test program is required for accurate estimation
of geometry and material properties.
For the description of the minimum requirements, for the extent and content of ground
investigation, design analysis and site supervision, and the risks to property and life,
three different geotechnical categories are described in the EN 1997-1 [119]:

• Small and relative simple structures, for which basic stability and performance
requirements can be fulfilled by experience and qualitative ground investigation
and for which risks are negligible. For this category, it is assumed that ground
conditions are known by experience.

• Conventional structures and foundations that can be designed by routine geotech-

nical procedures. For this category, quantitative site investigation is required and
should normally include quantitative analyses; it is to be verified that the design
requirements are satisfied.

• This category refers to major civil engineering projects in difficult ground where
both large sampling and large modelling is justified.

Within the concept of probabilistic site characterization the knowledge of standards

like EN 1997-1 [119] is extended by using statistical and probabilistic methods to com-
bine expert knowledge with the uncertainties of geology and the soil properties recom-
mended [23, 286]. After setting up a conceptual model of the geology, one has to refine
this model using a strategic programme of works to progressively reduce uncertainty
and to provide the information necessary to assess the risk of the site as proposed by
Chaplow [72].
According to Jaksa et al. [187], various factors influence the effectiveness of a site in-
vestigation program. Within the geological and geotechnical characterisation of a site
one can use different techniques to describe soil layering and soil variability. The geo-
logical and geotechnical characteristics of a site (number of layers, stratigraphy of the layers
and variability of the geotechnical properties) and the expected response to external loads
(structure and foundation type) can also be used to judge the effectiveness of additional
[email protected]
site investigations. A consequence of a first probabilistic characterization can also be

6
 2.3 Sources of soil variability

the definition of uncertain regions of a site and to optimize any additional characteriza-
tion efforts with the objective of reducing the amount of uncertainty. Journel and Alabert
[189] report that the amount of material that is sampled in site characterization boreholes
is typically only 10−6 to 10−9 of the total site volume, which creates a need for statistical
techniques for characterizing uncertainty.
Several authors [320, 337, 341] used methods of artificial intelligence, e.g. artificial
neural networks or support vector machines to characterize soil properties within a sta-
tistically homogeneous soil layer. Also geostatistical interpolation simulation techniques
can be used to evaluate the uncertainty of a conceptual geological-geotechnical model
[79, 250].

2.3 Sources of soil variability

Soils are geological materials formed by weathering, erosion and sedimentation pro-
cesses and, save for residual soils, transported by physical means to their present lo-
cations [23]. They have been subjected to various stresses, pore fluids, and physical
and chemical changes. Thus, it is hardly surprising that the physical properties of soils
vary from place to place within resulting deposits. The scatter observed in soil data
comes both from this spatial variability and from errors in testing as shown in table
2.1. Different coefficients of variation of soil parameters are summarized in this table to
provide an overview of the variability of design soil parameters presented in literature
[183, 282, 283].
Uncertainty in geotechnical engineering can be categorized into aleatoric, epistemic
and decision model uncertainty [23, 286, 382]. Aleatoric uncertainty consists of physical
uncertainty. Physical uncertainty is also known as inherent uncertainty and intrinsic un-
certainty and is a natural randomness of a quantity such as the variability in the soil
strength from point to point within a soil volume. Such physical uncertainty or natural
variability is a type of uncertainty, which cannot be reduced on increasing site investiga-
tion according to guidelines presented in [106]. Epistemic uncertainty consists of model
uncertainty and measurement uncertainty, which can be related to incomplete knowl-
edge. This implies that epistemic uncertainty can be reduced by more data as a type of
uncertainty associated with limited, insufficient or imprecise knowledge, as described in
[23, 106]; model uncertainty involves imperfections and idealizations made in applied
engineering models. Epistemic uncertainty also includes the uncertainty of the chosen
distribution as well as the parameter uncertainty, as enlightened in [382], to represent a
phenomenon as well as the choice of statistical distribution. In addition to this, Honjo
[166] adds the decision model uncertainty. This type of uncertainty includes objectives,
values and time preferences, which are related within a project management process
within the design of structures. Other types of uncertainties exist such as workmanship,
human errors and gross errors, but they are seldomly taken into account, as described in
[23, 106, 286].
The different kinds of uncertainty and errors in geotechnical engineering are summa-
rized in figure 2.2. Epistemic and aleatoric uncertainty arise, while setting up a concep-
[email protected]
tual model for the real ground. The probabilistic characterization of the subsoil model

7
Chapter 2 Characterizing soil variability at different scales

Table 2.1: Approximate guidelines for coefficients of variation of some design soil pa-
rameters taken from [183, 282, 283].

design test b soil type point spatial avg. correlation

property a COV [%] COV [%] c equation e
su (UC) direct (lab) clay 20 − 55 10 − 40 -
su (UU) direct (lab) clay 10 − 35 7 − 25 -
su (CIUC) direct (lab) clay 20 − 45 10 − 30 -
su (field) VST clay 15 − 50 15 − 50 14
su (UU) qT clay 30 − 40 d 30 − 35 d 18
su (CIUC) qT clay 35 − 50 d 35 − 40 d 18
su (UU) N clay 40 − 60 40 − 55 23
su e KD clay 30 − 55 30 − 55 29
su (field) PI clay 30 − 55 d - 32
ϕ direct (lab) clay, sand 7 − 20 6 − 20 -
ϕ (TC) qT sand 10 − 15 d 10 d 38
ϕ (CV) PI clay 15 − 20 d 15 − 20 d 43
K0 direct (SBPMT) clay 20 − 45 15 − 45 -
K0 direct (SBPMT) sand 25 − 55 20 − 55 -
K0 KD clay 35 − 50 d 35 − 50 d 49
K0 N clay 40 − 75 d - 54
EPMT direct (PMT) sand 20 − 70 15 − 70 -
ED direct (DMT) sand 15 − 70 10 − 70 -
EPMT N clay 85 − 95 85 − 95 61
ED N silt 40 − 60 35 − 55 64
a
su = undrained shear strength
UU = unconsolidated-undrained triaxial compression test
UC = unconfined compression test
CIUC = consolidated isotropic undrained triaxial compression test
su (field) = corrected su from vane shear test
ϕ = effective stress friction angle
TC = triaxial compression
ϕ (CV) = constant volume ϕ at critical state
K0 = in-situ horizontal stress coefficient
EPMT = pressure-meter modulus
ED = dilatometer modulus
b
VST = vane shear test
qT = corrected cone tip resistance
N = blow counts in the standard penetration test
KD = dilatometer horizontal stress index
PI = plasticity index
c
averaging over 5 m
[email protected]
d
COV is a function of the mean; refer to COV equations in [283] for details
e
equation numbering in Phoon & Kulhawy [283]

8
 2.3 Sources of soil variability

shall take this into account; the transformation error is introduced through the design
processes [166]. Uncertainties of the load, of the (mechanical) model and of the decision
model are finally ’hidden’ in the design result, as described in detail in Honjo & Kuroda
[166]. As stated in the introduction, the focus of this chapter is in the description of
the uncertainty estimation, which is introduced through the setting up of the conceptual
geological ground model.

After doing site investigations, the engineer’s aim is to estimate the values at unsam-
pled locations. One would intuitively choose a procedure like interpolation in order to
take the sampled properties in the neighbourhood of an unsampled location. A vast va-
riety of methods for interpolation can be found in different textbooks, which are very
present in fields like hydrology, statistics or geostatistics. These interpolation methods
are based on the concept of spatial correlation. In contrast to the interpolation approach,
there are the geostatistical simulation approaches, which are described in chapter 4.

load modelling
uncertainty uncertainty

real model design use design

ground ground ground result

knowledge uncertainty natural variability transformation decision model

epistemic uncertainty aleatoric uncertainty error uncertainty

model temporal objectives

parameters spatial
values

time preferences
SPT values SPT values j’-values

z z z
m±s m±s

Figure 2.2: Uncertainties in Reliability Based Design modified from Baecher & Christian
[23], Honjo [166], Phoon[286] and Phoon & Kulhawy [282, 283].

[email protected]

9
Chapter 2 Characterizing soil variability at different scales

2.4 Describing spatial variability

2.4.1 Scales of variability in soil science
Many scientists have investigated the spatial variability of soil properties in different
fields ranging from hydrology, soil sciences, reservoir engineering up to geotechnical
engineering. Some of these findings have been collected and assembled in a database, as
described in detail in section 3.5.
Among these authors, Koltermann & Gorelick [204] point out that the spatial vari-
ability of soil, and especially in subsurface flow, has to be treated at different scales
(table 2.2). Depending on the scale of the study, several direct and/or indirect field
investigation methods are thus applied in order to define the main characteristics of the
variability of the site. Considering subsurface flow, Koltermann & Gorelick [204] and
Matti [246] write that this includes geophysics, bore hole surveys, hydro-chemical anal-
ysis, hydraulic well and infiltrations tests, displacement measure, geotechnical labora-
tory testing. These methods allow a qualitative as well as quantitative description of the
heterogeneity in the field of subsurface flow and reservoir engineering. Some may de-
lineate large scale features such as permeable channels, whereas others may detect finer
scale transitions. According to the author’s knowledge, a comprehensive summary of
similar techniques is not available in geotechnical engineering .
It can be clearly seen that there are different scales of variability, ranging from the
micro level at the grain size scale to the geological scale of several tens and hundreds
of meters as shown in figure 2.3. In this context, heterogeneity can be defined as the
opposite of homogeneity and is further used as a synonym of spatial variability at large
scales variability.

sa
failure compacted
surface P grain
zone

sf

dilative
“homogeneous” shear zone void
soil band

(m) 10,000 1,000 100 10 1 0.1 0.01 0.001

geological geotechnical specimen phenomeno- grain

scale scale scale logical scale scale

Figure 2.3: Illustration of the multi-scale nature of soil after Borja [48], Chen et al. [73],
[email protected]
Christakos [82] and Wackernagel [401].

10
 2.4 Describing spatial variability

The geotechnical level is between the specimen scale and the geological scale; there-
fore, it is important to keep in mind that there is not a single spatial scale, but multiple
spatial scales contributing to soil variability. Of course, this plays a role in the evalua-
tion of spatial variability of soil properties as well in the evaluation of the effects of soil
variability.

2.4.2 Mathematical description of spatial correlation

As pointed out above, soils are geological materials formed by weathering processes
and, save for residual soils, transported by physical means to their present locations [23].
They have been subjected to various stresses, pore fluids, and physical and chemical
changes. Thus, it is hardly surprising that the physical properties of soils vary from
place to place within resulting deposits. Scatter observed in soil data comes both from
this spatial variability and from errors in testing.
This can be mathematically described in a smooth way by using random variables
and random functions. In comparison to this, a deterministic variable can just model
one outcome; this outcome is either known or unknown leaving no flexibility for un-
certainty [304]. Conversely, a random variable is an independent variable that can take a
series of possible outcomes, each with a certain probability or frequency of occurrence. A
random variable is typically denoted with the capital letter Z and its possible outcomes
are denoted with the corresponding small case letter zi , i = 1, ..., n. Most applications of
geostatistics involve mapping, which is the joint consideration of variables at several lo-
cations in space and/or time. For this reason, random functions can be used to describe
the joint spatial distribution of variables. A random function Z(X) is a set of dependent
random variables, each marked with a coordinate vector x. The variable X = (x, y, z)
can involve space coordinates, but also both space and time as e.g. atmospheric pres-
sure, in which case X = (x, y, z, t), which is rather unusual in geotechnical engineering
in comparison to other sciences like earth sciences or meteorology [79, 82].
Using well known means of univariate statistics, one can describe theses measure-
ments using a mean value µ and the standard deviation σ, a coefficient of variation
COV = σ/µ and a probability distribution function, as described in Appendix C. Uni-
variate statistics is not able to describe the spatial structure of the data.
As stated by several authors [31, 79, 143, 144, 189], the simplest way of describing spa-
tial variability is to choose the multi-Gaussian way. Within the multi-Gaussian approach,
a random process or random field can be described by a mean value µ, a standard devi-
ation σ and a covariance function C. For n pairs of a random variable Z of two different
locations Xi and Xj , the covariance C(Xi , Xj ) of the random function Z(Xi ) and Z(Xj )
is given by equation 2.1. Herein, E denotes the expectation, as described in detail by
equation C.1 in the Appendix C.

C(Xi , Xj ) = E [ ( Z(Xi ) − E(Z(Xi ) ) ( Z(Xj ) − E(Z(Xj )) ) ] (2.1)

It has to be pointed out that the means of univariate statistics are not influenced by the
covariance function. The probability density function can follow e.g. a normal distri-
[email protected]
bution or a lognormal distribution whether or not the investigated data have a spatial

11
12
Scale name: Basin Depositional envi- Channels Stratigraphical Flow regime fea- Pores
ronments features tures
Approximate 3 km– 100 km 80 m–3 km 5 m–80 m 0.1 m–5 m 2 mm–0.1 m < 2 mm
length scale

Geologic features Basin geometry, Multiple facies, Channel geome- Abundance of Primary sedimen- Grain size, shape,
strata geometries, facies relations, try, bedding type sedimentary tary structures: sorting, packing, ori-

[email protected]
structural fea- morphologic and extent, lithol- structures, strat- ripples, cross- entation, composi-
tures, lithofacies, features ogy, fossil content ifcation type, bedding, parting tion, cements, inter-
regional facies upward fning/ or lineation, lamina- stitial clays
trends coarsening tion, soft sediment
deformation

Heterogeneity Faults (sealing) Fractures (open Frequency of Bed boundaries, Uneven diage- Provenance, dia-
by Koltermann & Gorelick [204].

affected by folding, external or tight), intra- shale beds, minor channels, netic processes, genesis, sediment
controls (tec- basinal controls sand and shale bars, dunes sediment trans- transport mecha-
tonic, sea level, (on fuid dynamics body geometries, port mechanisms, nisms
climatic history), and depositional sediment load bioturbation
thickness trends, mechanism) composition
unconformities

Observations/ Maps, seismic Maps, cross- Outcrop, cross- Outcrop, lithologic Core plug, hand Thin section, hand
measurement profles, cross- sections, litho- well tomography, and geophysical sample, outcrop lens, individual
Chapter 2 Characterizing soil variability at different scales

techniques sections logic and geo- lithologic and logs clast, aggregate
physical logs, geophysical logs analysis
seismic profles

Support volume of Shallow crustal Regional (long Local (short term Near-well Core plug anal- Several pores
hydraulic measure- properties term pumping or pumping or tracer (non-pumping ysis (permeame- (mini-permeameter)
ments tracer tests) tests) tests-height of ter)
screened interval)
Table 2.2: Different scales of spatial variability of soil properties of sedimentary deposits
 2.5 Quantifying spatial dependence

correlation. In the case of a multivariate distribution, all random variables are linked
through a covariance matrix, as defined in equation C.1 in appendix C.
The basic assumptions for the description of spatial variability are as follows:

• S TATIONARITY is defined as when the mean value µ and the standard deviation σ
are constant over the whole domain. Moreover, the covariance C(τ ) is only depen-
dent on the separation τ and not on the absolute position [403].

• H OMOGENEITY is defined as stationarity of the variance statistically spoken [284],

which is closely linked to the definition of the stationarity.

• E RGODICITY is also closely related to stationarity. A random process is said to be

ergodic, when the moments of the single observable realization in space approach
those of the ensemble as the regional bounds expand towards infinity. According
to Webster & Oliver [403], it is of mainly theoretical interest rather than of practical
value because the regions studied in geotechnical engineering are finite.

2.5 Quantifying spatial dependence

Different researchers have focused on the description and estimation of spatial variabil-
ity. Probably the most well known among them are Chiles & Delfiner [79], Journel &
Huijbregts [190] and Deutsch & Journel [97] from the field of geostatistics; in the field
of geotechnical engineering Vanmarcke [385], Fenton & Griffiths [127] and Phoon [286]
have published important contributions. The effects of spatially variable soil properties
are still topics of ongoing research in soil sciences [403] or hydrology [82] amongst other
disciplines.
Generally speaking, there are two broad approaches in estimating the spatial correla-
tion: the method of moments (MoM) and the maximum likelihood (ML) approaches.

2.5.1 Method of Moments

According to Phoon [286] and Baecher & Christian [23], the most common method of
estimating spatial variability is the Method of Moments (MoM). Herein, the statistical
moments of the observations (e.g. sample means, variances and covariances) are used
as estimators of the corresponding moments of the population being sampled.
The Method of Moments is a non-parametric approach, which means that no assump-
tions are needed about the mathematical shape of the autocovariance function; there is
only a need to assume that the second moments exists. The moment estimator is con-
sistent and asymptotically unbiased. Therefore, it is a desirable method as stated in
[23, 27, 79, 182].
The Method of Moments can be subdivided into the variogram technique (MoMvar ),
[email protected]
the autocorrelation function approach and the local average (MoMLA ) approach.

13
Chapter 2 Characterizing soil variability at different scales

2.5.1.1 Variogram approach

The semivariance γ b(τ ) of a random function Z(X) can be computed using equation 2.2,
which is also called the empirical variogram or semivariogram (figure 2.7). The lag vec-
tor τ is generally a vector describing the mutual distance between the points. The lag
vector τ becomes a scalar τ = |τ | in case of an isotropic variogram. The isotropic vari-
ogram describes the spatial correlations as being the same in all directions, for which γ
b
can be computed only at integral multiples of the sampling interval.
m(τ )
1  2 1 X
b(τ ) ≡ E (Z(X) − Z(X + τ )) ≈
γ (Z(Xi ) − Z(Xi + τ ))2 (2.2)
2 2 m(τ ) i=1

The variogram does not require the knowledge of the mean of the random function Z(X)
because the squared difference in equation 2.2 eliminates the mean value. Moreover,
small variations are filtered out [79].
It has to be pointed out that the variogram approach describes the spatial dependence
as an integral of the whole distribution of parameter values. The spatial correlation of
the extreme values of the random function Z(X) cannot be investigated separately. In
contrast to this is the indicator approach. In equation 2.3 a threshold cutk is used for
truncating the random function Z(X), which is investigated by the variogram approach.
The truncation of the random function introduces difficulties in the estimation of the
spatial correlation because the truncated random function suffers from a sometimes very
skewed distribution. As shown in the following case studies in chapter 3, this causes
additional difficulties in the estimation of the spatial correlation.
Choosing different percentile values (appendix C) of the cumulative distribution func-
tion of Z(X), one can analyse the different spatial correlation of extreme values. In fact, it
has long been recognized [29, 189] that different percentile values for example extremes
can have a different spatial dependence structure from the central values. Indicator vari-
ograms can be used to express the difference in dependence as a function of the observed
values. This requires the specification of the random function Z(X) and threshold cutk
to create the indicator transform I(X; cutk ) as defined in equation 2.3. Different authors
[29, 176, 189] noted that with the help of indicator variograms in many cases the depen-
dence between variables departs considerably from the variogram approach.
(
1 if Z(X) ≤ cutk
I(Z(X; cutk ) = (2.3)
0 otherwise

The expected value of the indicator random function I(X; cutk ) identifies the cumulative
probability, i.e. the proportion of the property no greater than cutk [29]:

E ( I( X; cutk ) ) = 1 × Prob { Z(X ) ≤ cutk } + 0 × Prob { Z(X) > cutk }

= Prob { Z(X) ≤ cutk } = F (cutk ) (2.4)

Similarly, the indicator cross-covariances for two thresholds cutk and cut′k and for a sep-
[email protected]
aration distance τ identifies the bivariate (two-point) cumulative distribution function

14
 2.5 Quantifying spatial dependence

F (·) [144].

E ( I( X; cutk ); I( X + τ ; cut′k ) ) ≡ Prob { Z(X) ≤ cutk , Z(X + τ ) ≤ cut′k }

≡ KI (τ ; cutk , cut′k ) (2.5)

The indicator cross-covariance describes the covariance between cutk and cut′k and is
written as:

CI (τ ; cutk , cut′k ) ≡ E ( [I(X; cutk ) − F (cutk )] [I(X; cut′k ) − F (cut′k )] )

≡ KI (τ ; cutk , cut′k ) − F (cutk ) F (cut′k ) (2.6)

For cutk = cut′k the indicator cross-covariance becomes the indicator covariance, which
can be interpreted as a two point connectivity function [144].

CI (τ ; cutk ) ≡ Ki (τ ; cutk ) − F 2 (cutk ) (2.7)

It has to be stressed that connectivity is here used in a probabilistic sense; a high value of
the indicator covariance indicates that there is a high probability that any two locations
separated by the vector τ be jointly below the threshold cutk , but this is not enough to en-
sure the existences of a continuous path between two points. The indicator covariances
are symmetric with respect to the median threshold as derived in [144].
Illustrative results of an indicator analysis of an analytically generated random func-
tion can be found in chapter 3.1. The indicator varigoram analysis of measurement data
and the involved difficulties are also shown in the case studies in chapter 3.

2.5.1.2 Autocorrelation function approach

The autocorrelation function approach is coming originally from mathematicians and
is also applied by engineers to describe spatial variations in soil properties [23]. The
formula to compute the (empirical) autocorrelation function ρb(τ ) at lag τ is shown in
equation 2.8.
E (Z(X), Z(Xi + τ ))
ρb(τ ) ≡ (2.8)
var (Z (X))
Herein, τ stands for the lag separating the pairs of the random function Z(X). Linking
the semivariogram approach shown in figure 2.7 with the autocorrelation function one
has to follow equation 2.9.

b(τ ) = var (Z (X)) (1 − ρb (τ ))

γ (2.9)

2.5.1.3 Local average theory

Central to the development of robust random field models is the concept of the "local
average" of a random field. It is seldom useful or necessary to describe in detail the local
"point-to-point" variation occuring on a micro-scale in time or space [385]. Even if such
[email protected]
information were desired, it may be impossible to obtain. Vanmarcke [385] states that

15
Chapter 2 Characterizing soil variability at different scales

Heisenberg’s principle of uncertainty asserts that true patterns of point–to–point varia-

tion cannot be known: there is a basic trade-off between the accuracy of a measurement
and the time or distance interval within which the measurement is made. Strain-gauges,
stress cells, heat sensors or anemometers (owing to size, inertia, etc.) all measure some
kind of local average over space and time. Moreover, through information processing,
so called raw data are often transformed into average or aggregate quantities (e.g. one-
minute averages, daily or annual totals).
The spatial correlation can be derived from the variance function Γ, which adequately
explains the effects of spatial averaging as pointed out by Vanmarcke [385] as well as
Wickremesinghe & Campanella [404] among others. Within this thesis this local average
approach will be abbreviated with MoMLA .
In general, measurement data are skewed and follow a non-symmetric probability
density function. This implies difficult estimation of the variance function. Therefore,
these data are transformed to a normal distribution N (µ = 0, σ = 1) by using a quantile-
quantile transformation [79] as shown in figure 3.2.
These data are first considered in pairs (n = 2); a moving average series for the data is
obtained, where the length of averaging will be equal to the spacing of the data points
Z2 (X). The standard deviation σ2 of this series is also calculated, which will be lower
than the standard deviation of the original data set σ due to the cancelling out of fluc-
tuation due to spatial averaging. The above procedure is then extended to n = 3, and
the corresponding standard deviation σ3 is calculated with the spacing Z3 being equal
to twice the spacing of the original data points. This procedure is continued until n
approaches the total number of data N . The effect of spatial averaging will be more sig-
nificant with increasing n which σ b1 > σb2 > σb3 > σb4 > . . . σn as described by [385, 404].
For each n the variance function can be calculated from:

bn2
σ
Γ2 (Zn ) = (2.10)
b2
σ
Herein, σbn2 is the variance of the derived moving average series of degree n, and σ b2 is
the variance of the original data. If the spacing of the data is d, Zn in equation 2.10 will
be equal to (n − 1)d. The variance function Γ2 (Zn ) in equation 2.10 can be determined
for different sizes of the averaging window, which is used for the caluculation of σ bn .
Wickremesinghe & Campanella [404] derive the scale of fluctuation for large values of Z
and very large values of n.
θ = max( Γ2 (Zn ) Z ) (2.11)
Wickremesinghe & Campanella [404] recommend to pick the value of Γ2 (Z) from the
curve of Γ2 (Zn ) Z vs. n at a reasonably high value of Z, where there is a distinct change
in the curve. MoMLA offers nearly the same results like other methods of the state-of-
the-art for equally spaced and normally distributed data. It can be deduced from the
case studies B, C and D in chapter 3 that the MoMLA approach does not offer reliable
results in the case of non equally spaced measurement data. This can be related to the
different number of measurement data, which have been used to calculate the variance
of different distance classes. This results in a not clearly detectable maximum of equation
[email protected]
2.11.

16
 2.5 Quantifying spatial dependence

2.5.2 Maximum likelihood method

The Maximum Likelihood (ML) method of estimating the the unknown parameters Θ b is
a parametric method assuming that the distribution of the data is known. ML takes the
value of Θ as an estimate of the unknown parameters Θb that provides the greatest prob-
ability of having measurements Z , as calculated from the joint probability distribution
of the observations conditioned on Θ.
The possible outcomes z(X) of the random function Z(X) with mean value Z̄ and
covariance matrix CZZ are assumed to be described by a n-dimensional multivariate
normal distribution in equation 2.12.
 
1 1
fz (z) = p exp − (z − Z̄)T C−1
ZZ (z − Z̄) (2.12)
n
(2π) |CZZ | 2

The covariance matrix CZZ contains the values of the auto-covariance function C(Zi , Zj )
of each possible pair of measurements. Selecting the unknown parameters in a vector
Θ = [Z̄, σr , θh , θv ]T the log-likelihood for Θ is given in equation 2.13.

n 1 1
L(Θ|z) = − ln(2π) − ln |CZZ | − (z − Z̄)T C−1
ZZ (z − Z̄) (2.13)
2 2 2

By maximizing the likelihood, the optimal parameter set Θ can be obtained by stan-
dard optimization strategies, for example the simplex method. The advantage of the
simplex algorithm is that the results are independent of the initial parameters, hence
only depending on data.
De Groot & Baecher [90] state that the maximum likelihood estimators Θ b are asymp-
totically jointly normally distributed:

b ∼ N (Θ, B−1 )
Θ (2.14)

where the information matrix B can be obtained as

B = diag(BZ̄ , BΘ ) (2.15)

where Θ = [σr , θh , θv ]T is a vector containing the parameters of the theoretical variogram

model. The entries of the information matrix are given in [90] as follows:

BZ̄ = 1T C−1 1
ZZ 
1 −1 ∂CZZ −1 ∂CZZ
(2.16)
BΘij = tr CZZ CZZ
2 ∂Θi ∂Θj

where 1 is a unit vector of length n. Using the information matrix B, the accuracy of the
[email protected]
obtained parameters is estimated.

17
Chapter 2 Characterizing soil variability at different scales

2.5.2.1 Consideration of a trend within the ML approach

Assuming a linear regression model to remove a certain trend from the data, the random
function z(x) is represented as

z(x) = pT (x) β + ε(x) (2.17)

where p(x) is the basis vector of the regression model and β contains the regression
coefficients. If the correlation structure of the measurements is known in advance, the
regression coefficients could be estimated.

β̂ = (PT R−1
ZZ P)
−1
PT R−1
ZZ z (2.18)

where P is the so-called level matrix in the regression model containing the basis vector
terms for the measurement positions, and RZZ is the correlation matrix, which is related
to the covariance matrix by the residual variance CZZ = σr2 RZZ .
The correlation matrix, which can be directly calculated from the correlation lengths
and the measurement positions, is generally not known in advance. Thus, the regression
can be done only by assuming an initial guess and updating the required parameters
iteratively by using either the moment estimator or the maximum likelihood formulation
from equation (2.13). Another possibility is to incorporate the regression coefficients
directly in the maximum likelihood approach as proposed in [90].
n 1 1
L(Θ|z) = − ln(2π) − ln |CZZ | − (z − P β)T C−1
ZZ (z − P β). (2.19)
2 2 2
The accuracy of the estimated parameters can be estimated again using the inverse of the
information matrix B where only a slight modification of equation (2.16) is necessary:

B = diag(Bβ , BΘ )
Bβ = PT C−1 P
ZZ  (2.20)
1 −1 ∂CZZ −1 ∂CZZ
BΘij = tr CZZ CZZ
2 ∂Θi ∂Θj

2.5.2.2 Residual maximum likelihood approach

It has been established [100, 199, 200, 239] that the simultaneous estimation of drift and
covariance parameters produces biased estimates of the covariance. The Residual Max-
imum Likelihood avoids this problem by using special linear combinations of the data
(called generalized increments) instead of the original observations. These generalized
increments filter the drift and only the covariance parameters are estimated. The gener-
alized increments can be represented as

z̄ ≡ Λ z (2.21)

where the matrix Λ is constructed from the projection matrix P̂

[email protected] P̂ ≡ I − P (P PT )−1 PT (2.22)

18
 2.5 Quantifying spatial dependence

dropping out p rows, because among the generalized increments

z̄ = P̂ z (2.23)

p increments are linearly dependent on the others [197]. The matrix P̂ has the property
that:
P̂ P = 0 (2.24)
then
P̄ z̄ = P̂ P β + P̂e = P̂ ε (2.25)
and the drift is filtered out, whatever the coefficients β are. Here, the increments z are
assumed to be normally distributed N (µ = 0, σ = 1) and the covariance parameters are
estimated by the minimization of the negative log-likelihood function similar to equation
2.19. Herein, z is substituted by z̄, CZZ by Λ CZZ ΛT and P β is dropped [270].
m m m
σ 2 , θ|z̄) =
L(b ln(2 π) + − ln(m) +
2 2 2
1 1

+ ln(Λ CZZ ΛT ) + ln z̄T (Λ CZZ ΛT )−1 z̄ (2.26)
2 2
where m = n − p. The same method of minimization can be used to obtain the Residual
Maximum Likelihood estimates of Θ.
The accuracy of the estimated parameters can be estimated by using equation 2.27 as
shown in [403].
1
−1
σ= zT Λ CZZ ΛT z (2.27)
n−p

2.5.3 Additional approaches

Hybrid approaches: The Akaike Information Criterion, which is described in detail in
section 2.6.4, was merged with the ML-approach by Honjo [165]. Due to efficiency, this
was extended with a Bayesian approach to the Extended Bayesian Method.
Bayesian inference has not been widely used in geotechnical or geostatistical appli-
cations [42, 110, 198, 400] for the evaluation of the correlation length. It is less well
developed than the MoM or ML approach.

Transition probability & Markov chain: Especially in capturing the uncertainty in

subsurface flow simulation as well in soil pattern simulation, the concept of Markov
chains is used to describe heterogeneity as pointed out in Elfeki [115] and Carle & Fogg
[69]. A Markov chain is a probabilistic model that exhibits a special type of dependence
[79, 115]. In formulae, let Z0 , Z1 , . . . , Zm be a sequence of random variables taking values
in the state space S1 , S2 , . . . , Zn . The sequence is a Markov chain model, if
[email protected]
Prob(Zi = Sk |Zi−1 = Sl |Zi−2 = Sn |Zi−3 = Sr , . . . Z0 = Sp ) ≡ plk (2.28)

19
Chapter 2 Characterizing soil variability at different scales

where the symbol ”|” is the symbol for conditional probability. In one dimensional prob-
lems a Markov chain is described by a single transition probability matrix. Transition
probabilities correspond to relative frequencies of transitions from a certain state to an-
other state. Theses transition probabilities can be arranged in a square matrix form.
 
p11 p12 . . . . . . p1n
 . . 
 p11 . . . . . . . . .. 
p=  ... . . . p . . . ... 
 (2.29)
 lk 
.
pn1 . . . . . . . . pnn
where plk denotes the probability of transition from state Sl to state Sk ; n is the number
of states in the system. Thus the probability of a transition from S1 to S1 , S2 , . . . Sn is
given by p1l , l = 1, 2, . . . n in the first row and so on. The matrix p has to fulfil specific
properties. Its elements are non-negative, plk ≥ 0 and the elements of each row sum up
to one. For a more detailed explanation and extension of this approach (e.g. to more
dimensions, to multiple steps or to conditioning to measurements) and application of
this concept the reader is referred to Elfeki [115] and Caers [64].

Copula approach: Geostatistical literature [4, 29, 31, 151] offers via the copulas statis-
tics another means to describe spatial dependence. Copula describe the dependence
structure between random variables in a more general but also complex way than the
variogram. Therefore, copulas are useful tools to describe the spatial dependence. A de-
tailed explanation of copulas and their relationship to indicator variograms is presented
by Bardossy [29].

2.5.4 Sampling strategies

It is stated by various authors like Webster & Oliver [403] or Brus & de Gruijter [56]
amongst others [90, 142] that the classical statistical theory of independent, identically
distributed samples is not applicable in geosciences because in this case data are spatially
correlated. For this purpose, new methods have been developed taking into account two
sources of randomness: the location of the samples and the measurement error as shown
in table 2.3.
Brus & de Gruijter [56] define the fully deterministic, design based and model based
as well as fully random strategies via the fixed or random locations as well as values.

Table 2.3: Types of sampling strategies defined by two sources of randomness according
to Brus & de Gruijter [56].

VALUES AT GIVEN LOCATIONS

fixed random
SAMPLE LOCATIONS fixed fully deterministic strategies model base stategies
random
[email protected] design based strategies fully random strategies

20
 2.5 Quantifying spatial dependence

In the design based approach, stochasticity is introduced at the stage of sampling [56];
the sample locations are selected by predefined procedures such as pure random sam-
pling. Stratified sampling, clustered sampling or nested sampling offer a smooth way of
describing a heterogeneous population [21, 23, 56, 84]. Basically spoken, homogeneous
groups are set up inside the heterogeneous population, in which properties can be linked
to the variance of the population. This approach is widely used in geostatistics and will
be described in detail later.
In the model-based approach the soil forming process, which has led to the field of val-
ues of a property in the study area, is modelled as a stochastic process. The difference
of these two approaches is how the sample data are weighted. In the design-based ap-
proach the weights are derived from the sampling design; in the model-based approach
the weights are derived from the chosen model and the actual configuration of the sam-
ple locations. According to [56], the model based approach is used in the field of geo-
environmental engineering in the context of contaminated soils.
Chiles & Delfiner [79] state that the choice of the sampling concept depends on the
objective of the sampling (e.g. exploration or placement of a new well in petroleum
exploration). Of course, one has also to take practical constraints such as accessibility
and costs into account. They emphasize that the randomness as introduced by sampling
on a regular grid is smaller compared to stratified random sampling and pure random
sampling.
In geotechnical literature, Baecher [21] provides an overview of search theory and its
implications in probabilistic site description. Later, Tang [370] and Halim & Tang [150]
worked on geometric models of anomalies using the Bayesian principle. Herein, single
stage search and grid search concepts are used in a Bayesian context leading finally to a
sequential search approach.
Recently, Schweckendiek & Calle [331] used the Bayesian approach in a risk based con-
cept. Li et al. [220] proposed an alternative approach using copula statistics in this con-
text. In the field of geography [146] an alternative approach based on the fuzzy method
is presented. Sampling schemes, which are optimized by adaptive sampling [237] pro-
cedures or simulated annealing [359] are presented in the soil sciences as promising new
developments.

2.5.5 Estimation of the different scales of spatial variability

As pointed out in chapter 2.4.1, spatial variability of soil properties occurs at different
scales depending on the type of problem. Among others [61, 82, 125, 184, 401], Vanmar-
cke [384] recognizes the multiple scales of soil variability. He suggests that geotechnical
properties may exhibit two or more superimposed scales of fluctuation, depending on
the modelling scale. In the one dimensional case, Vanmarcke [384] suggests to combine
different correlation functions by simply summing them up. Jacksa [184] uses this con-
cept in an illustrative case study to combine two scales, as shown in figure 2.4.
In geostatistics [79, 190, 401], a nested models refers to a model of spatial variation,
which may be described by the simple addition of spatial variograms with different
[email protected]
parameters and possibly by different forms. A nested variogram γ(τ ) is set up by adding

21
Chapter 2 Characterizing soil variability at different scales

semivarience function g(t)

g1 + g2
300
g2
200
g1

100

0
0.20.5 3 6 25
mutual distance t in meter

Figure 2.4: Nested variogram from Jaksa [184].

n elementary variograms (figure 2.4). This approach is not very common in geotechnical
literature [184, 384].
Xn
γ(τ ) = γi (τ ) (2.30)
i=1

An alternative way is offered by the so called fractal approach. Herein, fractals are used
to model the self-similarity of natural phenomena. This offers the opportunity to de-
scribe the phenomena at different scales and fractals can portray an exact self-similarity.
Fenton [126] suggested that such fractal or long memory behaviour is likely to be present
due to the large scale mixing processes (e.g. erosion, transportations, deposition and
weathering) that are involved in the formation of soils.
According to Webster & Oliver [403], Bellehumeur & Legendre [33], Cheng [74] and
Burrough [59–61], the fractal dimension of transects and surfaces can be related to the
variogram. One big advantage of this approach is the modelling of irregular spaced
data, but a vast amount of data are needed for this method.
In environmental and agricultural sciences, the nested sampling approach is used to
characterize multi-scale phenomena. Nested sampling refers to a form of multi-stage
sampling, because the higher stage units are ”nested” within the lower stage units [281,
403].
Apart from these nested structures and the nested sampling approach, various authors
offer hybrid approaches combining the different approaches mentioned above or adding
new promising schemes like adaptive schemes [237], sequential schemes [187], Bayesian
approaches [150] or copula statistics [220].

2.6 Variogram calculation

The simplest case for calculating the variogram is an equally spaced dataset as shown in
[email protected]
figure 2.5 (a). First, the squared differences between neighbouring pairs of values z1 and

22
 2.6 Variogram calculation

z2 , z2 and z3 are determined for each position and averaged.

If there are missing values at some locations, then there will be fewer neighbouring
pairs as indicated in figure 2.5 (b).
If data are irregularly scattered then the average semivariance of a particular lag can
be derived only by grouping the individual lag distances between pairs of points, as
depicted in figure 2.6 and as shown in equation 2.31. In typical geostatistical literature
[30, 97], the classical equation of determination of the semi-variogram is

1 X
γ̂(τ ) = [Z(Xi ) − Z(Xj )]2 (2.31)
2N (τ )
(ij)∈R(τ )

where
R(τ ) = {τ − w/2 ≤ ds (ui , uj ) ≤ τ + w/2} (2.32)

dS (Xi , Xj ) is the spatial distance between the spatial point sets Xi , Xj , N (τ ) is the num-
ber of pairs in R(τ ) and w is the width of the spatial distance class as shown in figure
2.6. The bigger the w becomes, the smoother is the semi-variogram because the ε filters
out the very high and low values.
The variograms of second-order stationary processes reach upper bounds, at which
they remain constant after their initial increases as shown in figure 2.7. A variogram
may reach its sill at a finite lag distance, in which case it has a range, also known as the
correlation length; since this is the range at which the autocorrelation becomes 0 (figure
2.7). This separation marks the limit of spatial dependence. Places further apart than
this are spatially independent. For practical purposes their effective ranges are usually
taken as the lag distances at which they reach 95% of their sills [79, 97, 190, 401].
Some semi-variograms may approach their sills asymptotically, and so they have no
strict ranges. This can indicate a trend in the data. In some instances the variogram
decreases from its maximum to a local minimum and then increases again, figure 2.7.
This maximum is equivalent to a minimum in the covariance function, which appears
as a hole. This form arises from fairly regular repetition in the process. A variogram

(a) (b)
lag 1 lag 1

lag 2 lag 2

lag 3 lag 3

Figure 2.5: Comparison for computing a variogram from regular sampling on a transect:
(a) with a complete set of data, indicated with • and
[email protected]
(b) with missing values indicated by ◦ from [403].

23
Chapter 2 Characterizing soil variability at different scales

Figure 2.6: The geometry for discretizing the lag into bins by distance and direction in
two dimensions from [403].

that continues to fluctuate with a wave-like form with increasing lag distance signifies
greater regularity.
If there is only a neglectable or just a little range in the semivariogram, this is called
a nugget effect. This discontinuity at the origin of the semivariogram is used to charac-
terize the residual influence of all variabilities, which have a range much smaller than
the available distances of observation [190]. The nugget effect is equivalent to the well
known phenomenon of white noise in physics.
A semivariogram is said to display a hole effect when its growth is not monotonic
and shows bumps, which reflects the tendency for high values to be systematically sur-
rounded by low values and vice versa [79]. Journel & Huijbregts [190] attribute the hole
effect to different reasons. They recommend in case of periodic sampling distances that
the hole effect may result in a refined investigation scheme.

2.6.1 Theoretical variogram models

It is necessary to know the variogram γ(τ ) at any value of τ , if one wants to use the
variogram in terms of geostatistical simulation or interpolation. For this reason it is
necessary to fulfil the continuity, the differentiability condition and the conditional pos-
itiveness as shown in detail in Chiles & Delfiner [79]. Amongst others, Gascuel-Odoux
& Boivin [134] specify several sources of error: firstly only one realization is generally
available in nature and it is considered as representative; also errors in the experimental
variogram due to sampling and measurement must be considered; secondly, errors may
result from the choice of the model and estimation of the theoretical variogram [134].
The semivariogram has to be approximated to be able to simplify further work like
e.g. performing a stochastic simulation or interpolation between measurements. For
this reason, the behaviour of the variogram model has to be defined at the origin and
over the entire range. Different theoretical variogram models can be classified, according
[email protected]
to [190], into models with a sill (bilinear model equation 2.33), spherical variogram (linear

24
 2.6 Variogram calculation

(a) g(t) sill g(t) (b)

r(t)
exponential variogram g(t)

Gaussian variogram
spherical variogram
nugget
bilinear variogram r(t)
effect
range lag t
lag t

Figure 2.7: Theoretical variogram functions (a) and comparison of semivariance function
and autocorrelation function (b).

behaviour at the origin, equation 2.34), exponential variogram (linear behaviour at the
origin, 2.35), Gaussian variogram (parabolic behaviour at the origin, equation 2.36), models
without a sill (power functions, fractal model, logarithmic variogram). Other models
like the cubic model, generalized Cauchy models, K-Bessel model, power-law model,
pentaspherical model, Matern model or logarithmic model can be found in standard
geostatistical textbooks [79, 190, 401, 403].
(
τ /a for τ ≤ a,
γ(τ ) = (2.33)
a for τ > a

(

τ 3
c{ 23τb − 1
} for τ ≤ b ,
γ(τ ) = 2 b
(2.34)
c for τ > b
 τ
γ(τ ) = 1 − exp − (2.35)
c
 
τ2
γ(τ ) = 1 − exp − 2 (2.36)
d

2.6.2 Estimation of the theoretical variogram function

To compare different theoretical variogram models, it is necessary to fit theoretical vari-
ogram models in a standard way and automatically to the sample variogram in order to
avoid judgement errors. In the course of a detailed geostatistical analysis, an automatic
fit rarely provides definitive results [79]. Chiles & Delfiner [79] as well as Deutsch &
Journel [97] point out that this can be only the first step of a manual fit.
Generally, we look for a variogram γ(τ ; b) where b represents a vector of the parame-
ters of the variogram (e.g. range, sill, nugget effect,...) of n available pairs of data. This
[email protected]
vector b can be evaluated by minimizing the following equations:

25
Chapter 2 Characterizing soil variability at different scales

• O RDINARY LEAST SQUARES:

X
n
Q(b) = γ (τj ) − γ(τj ; b)]2
[b (2.37)
j=1

• G ENERALIZED LEAST SQUARES :

By minimizing equation 2.38, one can take into account the different correlations
between the different values of the sample variogram.

γ (τj ) − γ(τj ; b)]T V−1 [b

Q(b) = [b γ (τj ) − γ(τj ; b)] (2.38)

Herein, γ
b(τj ) is the vector of the empirical variogram, V is the variance-covariance
matrix of γ
b(τ ). The calculation of the variance covariance matrix is rather compli-
cated as highlighted by Ortiz & Deutsch [268].

V(τ ) = E [ Z(Xi ) − Z(Xi + τ ) ]2 · [ Z(xj ) − Z(Xj + τ ) ]2 − [ 2 γ
b(τ ) ]2 (2.39)

• W EIGHTED LEAST SQUARES :

A compromise between efficiency and simplicity is the weighted least squares,
namely the minimization of Q(b):

X
n
Q(b) = γ (τ ) − γ(τj ; b))]2
wj2 [b (2.40)
j=1

Herein, w is the weight, which can be the reciprocal of the number of pairs at each
lag, as proposed by Matheron [240] or also the variance at each point [79, 268]. The
variance-covariance matrix in equation 2.39 can be used to calculate the variance
of the variogram. The expression 2.41 tells us that the uncertainty in the variogram
at a distance τ is the average covariance between the pairs of the pairs used to
calculate the variogram for that particular lag assuming a multivariate Gaussian
distribution of the variables.
n(τ ) n(τ )
1 XX
w= 2
= n(τ )/ Vij (τ ) (2.41)
σ2γ(τ ) i=1 j=1

There are different approaches for choosing the weights w. Cressie [87] shows
that for equally spaced Gaussian variables, the variance of the estimates can be
approximated by equation 2.42:

N (τ )
w≈ (2.42)
b(τ ) ]2
2[γ

where γ̂ 2 (τ ) is the value of the theoretical variogram and N (τ ) is the number of

pairs at a mutual distance of τ . It is argued by several authors [79, 87, 272] that this
[email protected]
estimation is too crude to construct confidence intervals.

26
 2.6 Variogram calculation

McBratney & Webster [247] redefined this further:

b3 (τ )/m(τ ) γ
w≈γ b(τ )2 (2.43)
where γb(τ ) is the observed value of the semivariance at τ . This is usually desirable
for kriging , though it might be less desirable if the aim is to estimate the spatial
scale of variation. The process of fitting must iterate even where all the parameters
are linear because the weights in the two schemes depend on the values expected
from the model.

2.6.3 Alternative approaches

The theoretical variogram can also be fitted to the experimental variogram by using dif-
ferent techniques. This can be done by maximizing the likelihood function of the residual
between the theoretical variogram function and the semivariogram values, as explained
in Chiles & Delfiner [79] in detail. Another possibility is offered by Ecker & Gelfand
[110], offering a Bayesian approach to conduct the fitting of the theoretical variogram.
Using the ML approach, the theoretical variogram function is fitted to the available data
as pointed out above.

2.6.4 Model selection using the A KAIKE Information criterium

The selection of the most appropriate model is done via the A KAIKE I NFORMATION C RI -
TERITUM (AIC) [6], which is defined for a finite sample set n:

AIC = 2 k − 2 ln(L) (2.44)

where k is the number of parameters in the statistical model, and L is the maximum
value of the likelihood function for the estimated model. The first term is a measure
of the quality of fit of a model and the second is a penalty factor for the introduction
of additional parameters into the model. AIC is a measure of the loss of information
incurred by fitting an incorrect model to the data. Therefore, given a set of different
models for the data, the preferred model is the one with the minimum AIC value. Hence
AIC not only rewards goodness of fit, but also includes a penalty that is an increasing
function of the number of estimated parameters. This penalty discourages overfitting.
The preferred model is the one with the lowest AIC value. Assuming that the model
errors are normally and independently distributed, the AIC can be rewritten for a fitting
by least squares. Herein, the residual sum of the squares (RSS) are defined:
AIC = 2 k − n [ ln( 2π RSS/n ) + 1 ] (2.45)
Xn
RSS = εbi (2.46)
i=1
One can clearly see by looking at equations 2.45 and 2.44 that for a big sample size n the
AIC is independent of n.
Alternatively, different information criteria amongst others like BIC or KIT can be used
for small sample sizes or other boundaries, which is described in detail in literature
[email protected]
[7, 52, 403].

27
Chapter 2 Characterizing soil variability at different scales

Table 2.4: Relationship between the scale of fluctuation δ and the correlation distance θ
for various autocorrelation functions from Vanmarcke [384].

Autocorrelation function δ θ
δ
(
1 − |τ | /a for |τ | < a
ρ(τ ) = bilinear model a 1
0 for |τ | ≥ a
ρ(τ ) = e −|τ |/b
Markov model 2b 1/2
2 √ √
ρ(τ ) = e−(|τ |/c) Gaussian model πc 1/ π

2.7 On the correlation length

Within the description of spatial variability one has to distinguish between the correla-
tion length and its related model parameters. The definition of the correlation length
was introduced by Vanmarcke [385], which is refereed to by other authors [27, 68, 127,
156, 184, 288, 408]. They often call it scale of fluctuation.
Vanmarcke [385] defines the correlation length in equation 2.47. The correlation length
θ is the distance within which points are significantly correlated (i.e. by more than about
10 %), as described by Fenton [127]. Conversely, two points separated by a distance more
than θ will be largely uncorrelated.
Z ∞ Z ∞
θ= ρ(τ ) dτ = 2 ρ(τ ) dτ (2.47)
−∞ 0

The correlation length is defined without the factor of 2 shown to the right-hand side
of equation 2.47 especially in the geostatistical literature e.g. Journel & Huijbregts [190].
Equation 2.47 implies that θ has to be finite; otherwise, alternative concepts like fractal
processes have to be used [127]. Another consequence for the application in engineering
sciences as well as in earth sciences is that the correlation function is only meaningful
for strictly non-negative correlation functions.
The correlation length can also be defined in terms of the variance function in the local
averaging context as a limit [127, 385], whereas the correlation length θ is assumed to be
finite.
θ = lim T γ(T ) (2.48)
T →∞

DeGroot & Baecher [90] state that the Method-of-Moments (MoM) is unbiased in the
case of infinite samples. Otherwise, the correlation length is dependent on the sample
size. It can be clearly seen in figure 2.8 that for a finite record length L of the data, which
are sampled with a mutual separation distance τ , the estimation of the correlation length
is strongly biased. These findings coincide with the suggestions of Journel & Huijbregts
[190] or Chiles & Delfiner [79]. They recommend to sample with a distance between the
[email protected]
measurement points that is at least smaller than 1/5 to 1/4 of the correlation length.

28
 2.7 On the correlation length

1.0

autocovariance function E[C(t)]/s²]

0.8

Normalized expected 0.6

0.4

0.2
L/t →∞
0.0 20
10
5
-0.2 3
1

-0.4
0 1 2 3 4
Normalized lag distance t/t0

Figure 2.8: Dependency of the correlation length on the sample size, taken from DeGroot
& Baecher [90].

2.7.1 Uncertainty of the correlation length

In [79, 403], the uncertainty of the correlation length is connected with the application
e.g. spatial interpolation or (geostatistical) simulation. Within the concept of interpo-
lation, cross-validation is used to investigate the influence of the correlation length, as
described in detail in Webster & Oliver [403].
As mentioned above, different sources like measurement and modelling errors (statis-
tical model, nested structures,... etc.) cause an uncertainty of the evaluated correlation
length. Focusing on the Method of Moments, the main source of uncertainty is the defi-
nition of the distance and direction classes (figure 2.6). In the case of too wide classes, the
resulting variogram will be smoothed too much, whereas in the other case the variogram
will be too noisy. As pointed out in section 2.6.2, several authors express the uncertainty
of the correlation length via the experimental variogram γ b(τ ) to calculate the variance
and the correlations of the values of the experimental variogram γ b(τ ). Herein, they use
an approximation of the variance-covariance matrix of the experimental variogram by
assuming a normal distribution for the variance-covariance matrix, which is difficult to
verify in any application. The variances of γ b(τ ) can be used as weights for fitting the
theoretical variogram to the values of the experimental variogram. This offers also a link
to the uncertainty of the correlation length.
[email protected]
In the Maximum Likelihood approach, the assumption of normal distributed variables

29
Chapter 2 Characterizing soil variability at different scales

Table 2.5: Typical ratios of the horizontal θhor and vertical correlation lengths θver col-
lected from literature [27, 79, 96, 97, 190, 204, 401, 408].

geological anisotropy ratio qhor/qver

processes 10:1 100:1 1000:1

point bars
fluvial
eolian
estuariane
deepwater
deltaic
carbonates platform

is governing the whole evaluation of the correlation length. The estimation of the uncer-
tainty of the evaluated spatial correlation is a by-product of the ML procedure as shown
in equation 2.16 and for data with a linear trend equation 2.19. Via these equations it is
possible to estimate the error of the estimated variogram model-parameters and conse-
quently also of the correlation length.

2.7.2 Anisotropy with respect ot the correlation length

Following the introduction to geostatistics presented in Chiles & Delfiner [79], Deutsch
& Journel [97] and Wackernagel [401], one will encounter the anisotropy in the spatial
correlation of measurement data. Anisotropy, being defined as the ratio of horizontal θhor
and vertical correlation length θver , can be classified in zonal and horizontal-to-vertical
anisotropy.
Zonal anisotropy is related to stratification. This implies that the sill value of the hori-
zontal and vertical variogram are different, which can attributed to different sample dis-
tributions. Deutsch [96] points out the importance of the conceptual geological model to
describe this in detail (e.g. figure 2.7).
Horizontal-to-vertical anisotropy can be related to the geological processes which formed
the investigated statistically homogeneous layer.
Different anisotropy ratios from literature [27, 79, 96, 97, 190, 204, 401, 408] are summa-
rized in table 2.5. Most of these sources are from geostatistics as well as from petroleum
engineering. Therefore, the names of the different categories in table 2.5 come from en-
gineering geology and petroleum engineering. It can be deduced from table 2.5 that
the bigger the geological process is the bigger will be the anisotropy ratio θhor /θver . The
anisotropy ratios range from from 10 : 1 up to 1, 000 : 1. Therefore, it can be concluded
that it is of major importance to set up a conceptual geological model before working on
[email protected]
the spatial variability.

30
 2.8 Synopsis

2.8 Synopsis
Within this chapter different approaches to quantify spatial variability are presented,
which are used to quantify the different scales of soil variability. The literature review
on this showed that the most used approaches are the variogram approach, the local
average approach and the Maximum Likelihood approach. Therefore, the author is con-
centrating on these methods to evaluate purely the spatial variability of soil properties
without considering a trend.

[email protected]

31
[email protected]
 Chapter 3
Case studies on the evaluation of spatial
variability
This chapter provides a comparison of the Method-of-Moments (MoM) and the Maxi-
mum Likelihood (ML) approaches by applying them to four different case studies. The
Method-of-Moments methods include the variogram (MoMvar ) in chapter 2.5.1.1 and the
local average approach (MoMLA ) in chapter 2.5.1.3. These approaches are used to anal-
yse data fulfilling the basic assumptions of stationarity, homogeneity and ergodicity. Via
this, the possibilities and limitations of these approaches are discussed while applying
them to analytically generated random sequence as well as to equally and non equally
spaced data.
Moreover, the indicator approach is used to analyse the correlation structures of mea-
surement data to compare theses results to the indicator correlation lengths of an analyt-
ically defined random process.
The results of these case studies are compared to the literature database, which is cov-
ering the author’s knowledge of publications on the spatial variability of soil properties.
On top of this, the results of a study on the spatial correlation of different soil types
is presented, which is based on the analysis of CPT databases. Finally, these different
sources of information on spatial variability are merged via the Bayesian Model Averag-
ing approach.

3.1 Case study A – Random function

The aim of this case study is the investigation of the possibilities and limitations of
MoMvar , MoMLA and ML through the analysis of a random process with a predefined
correlation length. All three approaches offer the same correlation length for a symmet-
rical probability density function (pdf) of the random process. It is shown that the more
skewed the lognormal pdf becomes, the more differ the correlation lengths evaluated by
the MoMvar , MoMLA and ML approach. In addition to this, different levels of random
noise are added to the random process. This adds additional difficulty to the calculation
of the correlation length using MoMvar , MoMLA and ML, which offers additional insights
into the strengths of each approach.
Besides this, the indicator approach is used to investigate the spatial correlation of the
quantiles of the lognormal random process with a low coefficient of variation, which
implies a nearly symmetric probability density function. This investigation of the indi-
[email protected]
cator correlation lengths of very low and high values allows a more detailed picture of

33
Chapter 3 Case studies on the evaluation of spatial variability

analytically defined, multivariate-Gaussian random processes, which can be extended

to random fields.

3.1.1 Generation of the random process

The simplest model to describe spatially correlated numbers is the multivariate Gaussian
distribution. A multivariate Gaussian distribution can be described only by its mean
value and its covariance function. If the data are one dimensional, then it is called a
random process and otherwise random field. Due to the simple definition of the spatial
covariance by only one correlation function, the entropy or spatial disorder of the ran-
dom process or random field is assumed maximal [144]. As described by Bucher [57], the
joint probability density function of a multidimensional Gaussian distribution function
is defined in equation 2.12.
Different sources of error discussed in section 2.3 are considered by adding a normal
distributed measurement noise ε. Theses sources of error are used to compare the ro-
bustness of MoMvar , MoMLA and ML approaches.
The covariance matrix CZZ is symmetric and non-negative definite (e.g. it does not
have any negative eigenvalues). Therefore, the covariance matrix can be factored by a
Cholesky decomposition [57].
CZZ = LLT (3.1)
In equation (3.1), L is a non-singular lower triangular matrix. The Cholesky factor L
can be utilized for a representation of the random function Z in terms of zero-mean
uncorrelated random variables Y by applying a linear transformation in equation (3.2).
Z = L−1 Y + Ȳ (3.2)
In this case study, artificially generated data are investigated, which are represented by
a random function. This artificially generated data represent measurement data. This
random function with a mean value of Ȳ = 1, a standard deviation of σ̂ = 1, a spatial
correlation of θ = 10 m, a length of 300 m and a spacing of the data every 1 m is trans-
formed by a q-q transformation into a lognormal distribution, as shown in figure 3.2.
This lognormal distribution has a mean value of X̄ = 1.5 m and has different values of
the coefficient of variation COV = σ/µ = 1 − 100%. It can be clearly seen in figure 3.1
that the higher COV the more skewed and asymmetric is the probability density func-
tion. In order to show the possibilities of the different approaches to quantify spatial
variability, different intensities of noise are added to the lognormal distributed random
sequence. The noise ε is following a normally distributed standard normal distribution
N (µ = 0, σ = 0.01 − 1.0).

Ẑ = F F −1 (X) + ε (3.3)

3.1.2 Analysis
The MoMvar and the MoMLA approach as well as the ML approach are used to anal-
[email protected]
yse a set of 300 independent, identically distributed (iid) random functions Ẑ. MoMvar

34
 3.1 Case study A – Random function

2.0
probability density

COV = 1 %
COV = 50 %
function

1.5
COV = 100 %

1.0

0.5

0.0
0 1.0 2.0 3.0 4.0
generated data in [m]

Figure 3.1: Lognormal distribution with a mean value of Ȳ = 1 and different coefficients
of variation.

standard normal distribution lognormal distribution

1.0
Cumulative probability

Cumulative probability

0.6 0.6

0.3 0.3

0.0 0.0
-2 -1 0 1 2 1 2 3 4 5
Data Data

Figure 3.2: Quantile-quantile transformation of normally N (µ = 0, σ = 1) into lognor-

mally distributed data log-N (µ = 1.5, σ = 10).

[email protected]

35
Chapter 3 Case studies on the evaluation of spatial variability

(a) (b)

nugget effect / sill in %

30 ε = 0.0001
|1-q/qtarget| in %

2
1.5
15 1
0.5
0 0 1 2
0 0 1 2
10 10 10 COV in % 10 10 10 COV in %
0 3 30 g1 0 3 30 g1

ML MoMLA MoMvar
(c) (d)
nugget effect / sill in %
40 ε = 0.01
|1-q/qtarget| in %

4
20
2

0 0 0
1 2
10 10 10 COV in % 10
0
10
1
10
2
COV in %
0 3 30 g1 0 3 30 g1

Figure 3.3: Evaluated correlation lengths using MoMvar , MoMLA and ML approach (a,c)
and resulting nugget/sill ratios (b,d) using MoMvar and ML approach for dif-
ferent error levels ε = 0.01 (a,b) and ε = 0.0001 (c,d).

and MoMLA have been implemented into a M ATLAB program to analyse the data in an
efficient way for all case studies in this thesis. The program of Pardo-Iguzquiza [270]
has been modified and applied for the ML approach. The ML approach was chosen for
calculating the correlation length of data without a trend in all following case studies;
otherwise, the Residual Maximum Likelihood method (chapter 2.5.2.2) would be more
difficult to compare because the trend of data is eliminated also. The bilinear, spheri-
cal, exponential and Gaussian variogram functions have been fitted to the experimental
variogram values using the weighted least squares method using the weight defined by
McBratney & Webster [247]. The same theoretical variograms have been used within the
ML approach to evaluate the range of noisy random sequences. After this, the Akaike
Information Criterion approach has been applied to select the best fitting correlation
length of the variogram and the ML approach.
The evaluated correlation lengths of the iid random functions ẑ are fitted to a lognor-
mal distribution function using the maximum likelihood method. In figure 3.3 one can
see the influence of the skewness of the distribution as well as consequences of the noise.
[email protected]
The deviation |1 − θ/θtarget | of the evaluated correlation length from the target correla-

36
 3.1 Case study A – Random function

MoMVar MoMLA ML
100 %

cumulative correlation function %

percentile values of the
80 %

60 %

40 %

20 %

0 % 20 % 40 % 60 % 80 % 100 %

scaled indicator correlation length qind / qtarget

Figure 3.4: Case study A: Analysis of the indicator correlation lengths for the different
quantiles of the CDF (log-N (µ = 1.5, σ = 10)) and a measurement noise of
N (µ = 0, σ = 0.01).

tion length θtarget = 10 m is shown as a function of the skewness and asymmetry of the
b
underlying distribution of Z(X). Moreover, the ratio of the nugget effect and the sill is
shown as a function of the COV in figure 3.3.
The bigger the skewness of the underlying distribution and the bigger the measure-
ment noises, the worse is the evaluated spatial correlation. It was also observed by Kerry
& Oliver [193, 194] that the variogram approach becomes unreliable when the data are
strongly asymmetric or skewed as well as in presence of outliers or extreme values. Sim-
ilar to the findings of Webster & Oliver [403], it can be seen in figure 3.3 that as COV and
the skewness increase the nugget and the sill variances also increase. It is also shown
that the ratios of the nugget to sill increase as the skewness increases even though this
is not considered in the generation of the random process. Webster & Oliver [403] also
observed this and point out that the ratio of nugget to sill is a combination of the degree
of asymmetry and the spatial distribution of data points of the tail of the distribution.
By comparing the three approaches in figure 3.3, MoMvar and MoMLA show more sta-
ble results in comparison to the ML approach, which can be traced back to the basic
assumptions of normally distributed data. All three approaches become more unreliable
the more noisy the data are.
The indicator approach is employed to evaluate the correlation lengths of 300 iid ran-
dom processes ẑ. The different percentiles of the cumulative distribution function of
[email protected]
ˆ are used as thresholds cutk , as described in section 2.5.1.1. The random process
Z(X)

37
Chapter 3 Case studies on the evaluation of spatial variability

Z(X) is truncated by the thresholds cutk . MoMvar , MoMLA and ML approaches are used
to calculate the indicator correlation lengths θind . One would expect that theses indicator
correlation lengths would be the same, but this is not the case as shown in figure 3.4.
This becomes clear, if one recalls the generation of the random process: only one correla-
tion length is used to define the spatial correlation. The indicator correlation lengths of
the different thresholds are symmetric towards the median value, as reported in [143].
Moreover, the extreme values have significantly lower correlation lengths than the me-
dian value, as shown in in figure 3.4.

3.1.3 Remarks
The non-parametric MoMvar technique has proven its strengths in comparison to the
MoMLA and ML approach. It can be seen in figure 3.3 that the variogram technique
is more robust than the MoMLA and ML approaches in the presence of a high noise
as well as when the underlying distribution is highly skewed and asymmetric, which
coincides with the findings in [193, 194]. The MoMLA and ML approaches use the q-q
transformation to convert the underlying distribution into standard normal distributed
variables N (µ = 0, σ = 1). This transformation causes the inaccuracies shown in figure
3.3, especially in the presence of an asymmetric distribution.
It has to be pointed out that the variogram methodology is working in the best way
for normally distributed values N (µ = 0, σ = 1). If the underlying distribution becomes
skewed the outcome is less reliable, which is also true for the MoMLA and the ML ap-
proach. For this reason, robust techniques are offered in a geostatistical framework as
described in detail amongst other by Chiles & Delfiner [79], Bardossy & Kundzewicz
[30] and Marchant & Lark [236]. Following publications summarized in [403], differ-
ent estimators like the Cressie & Hawkins estimator, median variogram estimators or
Genton’s estimator can overcome this problem in a complicated way. The form of these
estimators does not allow explicit computing of their correlation structure as stated by
Genton [137].
There are differences in the evaluated correlation length using the non-parametric
MoMvar and MoMLA approaches as well as the parametric ML approach. Moreover, one
has to keep in mind that there has been performed a q-q transformation for the MoMLA
and the ML. The difference for the ML approach can be related to the smoothing due to
the definition of lag classes in the variogram approach. In this context Webster & Oliver
[403] stress the benefit of the ML approach not to smooth the spatial structure because
there is no ad hoc definition of lag classes; the model parameters are calculated directly
from the variance-covariance matrix of the full data. The ML and MoMvar approach as-
sume that the data follow a multivariate Gaussian distribution, which is a simplification
of the data and very difficult to verify in practice [403].
The relative error εtarget is shown for the MoMvar , MoMLA and ML approaches for a
lognormally distributed random function with a COV= 1 % in figure 3.5 (a) and COV=
100 % in figure 3.5 (b). A measurement noise of N (µ = 0, σ = 0.01) was added to show
the influence of the asymmetry of the underlying lognormal distribution function on the
[email protected]
evaluated correlation length. The shifted, lognormal distribution functions show in the

38
 3.1 Case study A – Random function

case of a nearly symmetric distribution with COV = 1 % comparable results. The MoMvar
and ML approaches offer more or less the same result for the correlation length. The
mean value of the distribution function is below εtarget < 10 %; the results of the MoMLA
approach offer a slightly higher relative error εtarget . In the case of a COV= 100 % of the
underlying distribution, the robustness of the MoMvar approach is strengthened again
in comparison to the ML and MoMLA methods, which are very sensitive to asymmetric
and skewed distributions.
For the correlation lengths in general, it can be concluded for that one has to pay atten-
tion to the univariate distribution of the data: the higher the skewness and asymmetry,
the more attention has to be paid in evaluating the correlation length. One has to pay
more attention when analysing skewed data because then the results of the three meth-
ods scatter significantly.

(a) MoMvar mε = 8.6 % MoMvar mε = 3.1 % (b)

target target
MoMLA approach mε = 17.9 % MoMLA approach mε = 17.6 %
target target
ML approach mε = 1.8 % ML approach mε = -34.7 %
target target

0.14
probability density funciton
probability density funciton

0.1
0.12
0.08
0.1
0.06 0.08
0.06
0.04
0.04
0.02
0.02
0.00 0.00
-50 % 0 % 50 %100 % -50 % 0 % 50 %100 %
εtarget = (θ - θtarget )/θtarget εtarget = (θ - θtarget )/θtarget

Figure 3.5: Case study A: distributions of the relative errors εtarget of the MoMvar , MoMLA
[email protected]
and ML- approaches for COV = 1 % (a) and COV = 100 % (b).

39
Chapter 3 Case studies on the evaluation of spatial variability

0 37.6 150 225 300

meters

Figure 3.6: Case study B: plan view of the 138 CPT measurement locations.

3.2 Case study B – CPT data evaluation

This case study deals with a stochastic site description, involving the identification of
soil layering and the detection of trends in the measurement database, in which the engi-
neering judgement is enriched by using mathematical and statistical methods. The var-
iogram approach (MoMvar ), the local average approach (MoMLA ) and Maximum Likeli-
hood (ML) approach are combined to analyse a big CPT dataset with the focus on the
vertical spatial variability. The evaluation of the vertical spatial variability within each
layer is conducted within the developed scheme of probabilistic site description.
The results of the MoMvar , the MoMLA and ML approach are combined on the basis of
B AYES principle, which allows the statistical combination of different models of spatial
variability.

3.2.1 Site description & measurement technique

An extensive field investigation has been carried out for a big industrial construction
project in South America. Within an extensive exploration campaign, 67 standard pene-
tration tests and 138 cone penetration tests with measurement of pore pressure according
to the DIN 4094 − 1 [103] have been carried out [139], as shown in figure 3.6. For this
case study 138 closely spaced CPT measurements have been selected for evaluation.
Under a crust with a thickness of a few decimetres up to one meter a very soft soil
layer exists down to a depth of about 2.2 to 5.9 m, generally in depth of 3.6 to 4.7 m
under ground level. The soft layer consists mainly of silty clay. The crust at the ground-
level is more solid due to roots, consolidation and other influences. Beneath the first
layer of 1.8 m, there is layer of clayey sand with a mean value of 2.4 m. At the base of
this first sand layer a second clay layer is found down to a depth of about 12 to 15 m
[email protected]
below ground-level. The layer is largely similar to the clay layer above. Up to the total

40
 3.2 Case study B – CPT data evaluation

tip resistance in MN/m²

0 10 20 30 40 50
0
lower and upper bound
5 of the measurements
average tip resistance
depth in m 10

15

20

25

30

Figure 3.7: Case study B: Average, minimum and maximum measurement values of the
CPT data.

depth of the field investigations sand can be found again with different properties: fine
to coarse sand, sometimes even sand with a silty part. The density has been detected
over a range from loose to dense. The groundwater level has been measured at a depth
of 0.6 to 1.4 m below ground level. Due to the high precipitation, the groundwater level
is expected to be at the surface of the area.
The summary of the measurement results is shown in figure 3.7. The upper and lower
bounds of the measurement data are shown together with the mean value of the CPT
measurements.

3.2.2 Analysis of the data

By looking at the upper and lower bounds of the measurements in figure 3.7, one can
easily see the need for a concept in the probabilistic site characterization as described in
chapter 2.
The basic steps of the probabilistic site characterization are summarized in figure 3.8.
The first step is the engineering judgement on soil layering and setting up of preliminary
boundaries. After this, the measurement data have to fulfil the homogeneity and station-
arity criteria. Therefore, the measurement data have to undergo statistical tests. In the
presence of a significant trend of the data, one has to detrend the measurements inside
each layer. Now each layer is analysed by the MoMvar , MoMLA and ML approaches to
evaluate the correlation lengths.
Of course, one has to check now the sensitivity of the correlation length inside each
layer to the small changes of the layer boundaries. This is important because the subdi-
vision of a soil profile into layers and the detrending inside each layer has a significant
[email protected]
impact on the correlation length.

41
Chapter 3 Case studies on the evaluation of spatial variability

engineering judgement on soil layering

setting up layer boundaries

for each layer
detrending measurements Mann-Kendall test
testing homogeneity & stationarity Bartlett test, RI test
evaluating spatial variability MoMvar, MoMLA & ML approaches

Figure 3.8: Steps of the stochastic site characterisation and involved methods.

Using this scheme, one can separate different scales of spatial variability (figure 2.3):
the large geologically based spatial variability are separated from the meso-scale phe-
nomena, which can be investigated without injuring the basic assumptions of the theory
pointed out in section 2.4.2. Via this separation of the different scales of spatial variabil-
ity, it is possible to reduce the noise significantly, which was simulated in section 3.1 via
a normally distributed random noise ε because the homogeneity and stationarity criteria
are fulfilled.
By looking at the measurement data CPT(z) in figure 3.7, one can clearly indicate a
trend of the measurements with depth, which can be described by equation 3.4 where
m(z) is a deterministic function giving the mean measurement value at a depth z below
the surface level; and ε(z) are the random residuals.

CPT(z) = m(z) + ε(z) (3.4)

To apply the theory to evaluate the spatial variability, one has to check the basic assump-
tions of this theory in order to use the MoMvar MoMLA and ML approaches. If a trend
is not removed from the measurement data, the evaluation of the correlation lengths is
more difficult or even impossible as shown in figure 3.10 (b). Therefore, a selection of
techniques to identify and test measurement data on stationarity and homogeneity are
presented in this section.
S TATIONARITY can be defined by a constant mean value and variance within the test
data. Various authors have proposed techniques for detecting the stationarity of the
data. Only some are enlightened in this section. Amongst others, Jacksa [184] as well as
Bennett [36] propose different methods to detect a trend.

• V ISUAL INSPECTION: Mere inspection of the raw data is often sufficient to detect
non-stationarity. Visual inspection, however, is not sufficient to detect the form of
the trend, nor most cases of non-stationary variance.

• H ISTOGRAM PLOTS: A simple, though crude technique, is to split the random field
into a number of subsections and to plot each of their histograms. Comparison of
these histograms enables shifts in the means and variances to be detected.

• I NSPECTION OF THE EMPIRICAL VARIOGRAM: Box and Jenkins [51] as well as Ben-
[email protected]
nett [36] investigated this approach amongst others. It can be deduced from look-

42
 3.2 Case study B – CPT data evaluation

ing at equation (2.2) that non-stationary data will impose also a trend in the em-
pirical variogram. The range cannot be determined. This feature can be used to
indicate a trend of measurement data as shown in figure 3.10.
• S IGNIFICANCE TESTS ON TRENDS: Another smooth way to detect stationarity of
measurements is offered by the non-parametric method of Mann-Kendall’s τ [234].
Alternatively, Sachs [316] offers a different test for testing the significance of trends,
namely the Cox-Stuart test or Neumann test. The Mann-Kendall test is based on
the statistics S. Each pair of observed values yi , yj (i > j) is inspected to find out
whether yi > yj or yi < yj . Let the number of the former type of pairs be P , and the
number of the latter type of pairs be M . Then S is defined as
S =P −M (3.5)
For n > 10, the sampling distribution of S is as follows. Z follows the standard
normal distribution, where

(S − 1)σs if S > 0

Z= 0 if S = 0 (3.6)


(S + 1)σs if S < 0
r
n(n − 1)(2n + 5)
σs =
18
The null hypothesis that there is no trend is rejected if the computed Z value is
greater than Zα/2 in absolute value. Herein, the significance α = 1% is chosen
according to [287].
After identifying the trend via the described approaches, the least squares method [34]
is used to fit a linear trend to the data as recommended by [23, 286]. Alternatively, one
could also use different approaches based on the Bayesian principle or on a Maximum
Likelihood method [23, 286].
After removing the trend of the measurement data by fitting a linear function by least
squares, Maximum-Likelihood or Bayesian approach to the measurement data, one has
to check the data on homogeneity. This can be done be engineering judgement; but also
statistical can support this judgement on a mathematical basis.
Homogeneity can be defined as stationarity of the variance as presented in [284, 316].
The intra-class correlation coefficient and the Bartlett statistics are used herein for check-
ing the homogeneity of the measurements.

Intra-class correlation coefficient: The intra-class correlation coefficient RI is reported

as a useful statistical method for detecting layer boundaries using CPT soundings, Wick-
remesinghe [404]. The RI profile is generated by moving two continuous windows con-
taining m data points each over a measurement profile and computing the following
index at the centre of the double window.
 
1
[email protected] RI = 1/ 1 + (3.7)
(m − 1)/m + (µ̂1 − µ̂2 )2 /2/(ŝ21 + ŝ21 )

43
Chapter 3 Case studies on the evaluation of spatial variability

probability density function

0.20

0.15 histogram
fitted pdf
0.10

0.05

0.00
-10 -5 0 5 10 15
residual cone resistance in MN/m²

Figure 3.9: Histogram and fitted normal distribution function of the detrended cone re-
sistances in sand of CPT 4 N (µ = 0.10 kN/m2 |σ = 2.51 kN/m) with a skew-
ness γ1 = 0.40( kN/m2 )3 and an excess kurtosis γ2,excess = −0.51( kN/m2 )4 .

Basically, ergodicity of the mean value and the variance within the moving window
is assumed. This is only valid for symmetric distributions according to various au-
thors [154, 417]. The critical RI value RIcrit is estimated according to Hegazy et al.
[154]. The boundaries are identified quantitatively at locations where RI exceeds the
empirical relationship of the mean µRI and standard deviation σRI of the RI profile:
RIcrit = µRI + 1.65σRI ; it is recommended to check the computed results visually and to
judge the evaluated soil layer. The critical RI according to Hegazy et al. [154] is slightly
higher than the recommendation of Zhang & Tumay [417] RIcrit = 0.7, which is also an
empirical rule. Others [154, 417] also point out that the choice of RIcrit does not seem to
depend on the underlying correlation structures of the profile, which is also discussed
by Phoon et al. [284].

Bartlett statistics: This classical test is used to test the equality of multiple sample vari-
ances for independent data sets. This has not to be taken into account in this case study
because a normal distribution function can be fitted to the residuals of the CPT measure-
ments as depicted in figure 3.9.
For the case of two sample variances, s21 and s22 , the Bartlett test statistic reduces to:

2.30295(m − 1)
Bstat = [2 log s2 − (log s21 + log s22 )] (3.8)
1 + 2/(1(m − 1))

where m is the number of data points used to evaluate s21 or s22 . The total variance s22 is
defined as:
s2 + s22
s= 1
2
While using the Bartlett statistics, one has to keep in mind that this procedure is very
sensitive to non-normally distributed and skewed variables. According to Sachs [316],
[email protected]
he implies that in the case of a small deviation from the symmetric normal distribution,

44
 3.2 Case study B – CPT data evaluation

the procedure will not offer reliable results. Especially in the presence of a skewness
γ1 6= 0( kN/m2 )3 and an excess kurtosis γ2,excess 6= 0( kN/m2 )4 , which can be observed
very often in the case of measurement data.
A continuous Bartlett statistic profile can be easily generated by moving a sampling
window over the simulated soil profile. Campanella et al. [68] as well as Wickremesinghe
[405] recommend a window width of approximately the scale of fluctuation in the layer.
This is also pointed out by Phoon et al. [284]. This implies an iterative approach. The
sampling window is divided into two equal segments and the sample variance s21 and s22
is calculated from data points lying within each segment. The Bartlett statistic basically
indicates the difference between the sample variances in these two adjacent segments.
As shown in equation 3.8, the Bartlett statistic is zero if s21 and s22 are equal. Phoon et al.
[284] offer a critical value Bcrit under the framework of the MODIFIED B ARTLETT STATIS -
TICS taking into account the spatial correlation using an exponential model. Herein,
I1 = n/k ranges between 5 and 50 and I2 = m/k where k is the number of points in one
scale of fluctuation; n is the total number of points in the entire soil record and m is the
number of points in one sampling window.

Bstat,crit = (0.23k + 0.71) log(I1 ) + 0.91k + 0.23 (3.9)

This critical value Bcrit is calculated for every layer to take the different correlation
lengths into account. The lowest critical value Bcrit is used for the whole CPT profile.
In figure 3.10 the detrended measurement data are used to evaluate the RI as well as
the Bartlett statistics. The width of the sampling window is chosen as big as the correla-
tion length. The critical values RIcrit and Bcrit indicate the boundary between both soil
layers.

Application of the Intra-class correlation coefficient and the Bartlett statistics: RI

and Bstat offer help in the soil layer identification using statistical methods, but these
approaches can just support the engineering judgement, due to the above mentioned
assumptions and simplifications as stated by Phoon et al. [284]. The above mentioned
methods are well suited to normally distributed data. Therefore, it is has to be stressed
that these methods support the engineer in detecting different soil layers from CPT mea-
surements. This statement becomes more clear when looking at figure 3.10 (a). By look-
ing at the geological profile, which is derived by an engineer [139], one can clearly see
the correlation between the soil layering and the cone resistance. The changes in the
cone resistance offer a reasonable basis for the soil layering. These soil layer bound-
aries are detected by the B ARTLETT STATISTICS, whereas the RI-profile does not show all
boundaries in a reliable way. Therefore, it can be deduced that the B ARTLETT STATISTICS
offers a more detailed insight and is more suitable to detect layer boundaries by means
of statistics. The RI concept shows poor results in this study and is not suitable for a
clear detection soil layers from CPT measurement data.

Data processing: The CPT measurement data are detrended by the a least square fitting
[email protected]
of a linear function within one layer.

45
Chapter 3 Case studies on the evaluation of spatial variability

(a) CPT data geological Bartlett statistics Bstat Intra-class correlation

tip resistance Profile coefficient RI
0 10 20 150 100 50 0.6 0.8
1,35
silty clay

5 clayey sand 5

10 silty clay 10
Depth

15 15
silty sand
20 20

silty clay
25 25
silty sand
30 30

35 Bstat, crit 35 RIcrit

raw cone resistance measurement data
detrended data

(b)
g(t)
raw data
20
10
detrended data

0
0 2 4 6 8
lag distance t

Figure 3.10: (a) Cone resistance qc (z) measured with depth z and soil layer
identification of CPT 4 data using RI and Bartlett statistics and
(b) semivariogram on the raw and detrended data of the silty clay layer
at 10 m depth.

[email protected]

46
 3.2 Case study B – CPT data evaluation

After identifying the homogeneous section of the CPTs and the detrending by least
square fitting of a linear function to these homogeneous sections, the MoMvar , MoMLAS
and ML approaches are used to evaluate the spatial variability of the site for each layer.
Different theoretical variogram functions (namely spherical, exponential bilinear and
Gaussian functions) are fitted to the experimental variogram values using the weighted
least squares and Akaike Information criterion to identify the best fitted theoretical var-
iogram function. Also within the ML approach the Akaike Information criterion was
used to identify the best suitable theoretical variogram model.
The analysis of the correlation lengths and of the indicator correlation lengths is car-
ried out for the different soil types. For this reason, the charts of Robertson [308] were
used to evaluate the the soil types from the normalized cone resistance and the friction
ratio of a CPT test as shown in figure 3.23. For theses soil layers the the cumulative
distribution function was used for the evaluation of the different thresholds cutk,i . This
has been used for each of the 138 CPT measurements to evaluate the indicator correla-
tion length of each threshold cutk,i . It is found that these indicator correlation lengths
follow a lognormal distribution function. The mean values θind of each threshold cutk,i
are shown in figure 3.12, which is scaled by the correlation length evaluated by the vari-
ogram approach θ.

3.2.3 Remarks
Combination of different models: One can clearly see in figure 3.11 that the MoMvar ,
the MoMLAS and the ML approaches offer comparable results. This can be deduced from
the nearly symmetric distribution of the residual values shown in figure 3.9, because
there is only a very small skewness of the measurement data. These three probabil-
ity density functions of the MoM and ML approaches are equally probable and can be
merged by using the so called B AYESIAN MODEL AVERAGING (BMA) scheme [162, 300],
which is based on the Bayes’ theorem.
The Bayes’ theorem updates a subjective, prior probability distribution f (θ) with a
likelihood function L(θ|z1 , z2 , ..., zn ), which is the conditional probability function of z1 , z2 , ..., zn
e.g. measurement values.
f (θ|z1 , z2 , ..., zn ) ∝ f (θ) · L(θ|z1 , z2 , ..., zn ) (3.10)
The resulting posterior pdf f (θ|z1 , z2 , ..., zn ) of the variable of interest θ is conditioned on
the prior probability f (θ) and on the Likelihood function L(θ|z1 , z2 , ..., zn ), [70, 136, 382].
If θ is the quantity of interest, then its posterior distribution given data D is
K
X
Prob(θ|D) = Prob(θ|Mk , D) · Prob(Mk |D) (3.11)
k=1

This is an average of the posterior distributions under each of the models considered,
weighted by their posterior model probability. In equation 3.11 M1 , M2 , ..., Mk are the
models considered. The posterior probability for the model Mk is given by
Prob(D|Mk ) · Prob(Mk )
[email protected](Mk |D) = PK (3.12)
l=1 Prob(D|Ml ) Prob(Ml )

47
Chapter 3 Case studies on the evaluation of spatial variability

probability density function

3.5
3
2.5 MoMvar (m = 0.88 , COV = 121 % )
2 MoMLA (m = 0.62 , COV = 204 % )
ML (m = 0.75 , COV = 89 % )
1.5
BMA (m = 0.32 , COV = 58 %)
1
0.5
0
0 0.5 1 1.5 2 2.5 3
vertical correlation length q in meter

Figure 3.11: Case study B: probability density function of the MoMvar , MoMLA and ML
approach and the combined probability density functions using Bayesian
Model Averaging of the silty clay layer.

where Z
Prob(D|Ml ) = Prob(D|θk , Mk ) Prob(θk |Mk ) dθk (3.13)

is the integrated likelihood of model Mk , θk is the vector of parameters of model Mk ,

Prob(θk |Mk ) is the prior density of θk under model Mk , Prob(D|θk , Mk ) is the likelihood
and Prob(Mk ) is the prior probability that Mk is the true model under the assumption
that one of the models considered is true. All probabilities are implicitly conditional on
the set of all models being considered.
Hoeting et al. [162] report that BMA presents several difficulties: the specification
of the prior distribution over competing models has to be carried out with attention.
Another fundamental task is to choose the models over which the averaging can be
performed; one of the most challenging tasks is the evaluation of the integral in equation
3.13. Within [70, 162, 300, 382], it is recommended to use Markov Chain Monte Carlo
methods to evaluate this integral for the very general combination of different functions.
In the simple case of the combination of two (log)normal distribution functions equation
3.13 can be solved by numerical integration [382].
The combination of the three distribution functions using BMA is shown in figure
3.11. The resulting distribution function has a lower variance as well as a lower mean
value of the correlation length. Via BMA all information from the three different models
(MoMvar , MoMLA and ML) have been incorporated and this higher level of information
is leading to a significant reduction of the coefficient of variation of the combined pdf.
This pdf of the correlation lengths is an extension of the state of the art. It is stated in
standard geostatistical textbooks [79, 401] that a correlation length is needed for interpo-
lation or simulation of spatial variability. The correlation length is assumed as one single
value without uncertainty. This case study shows that the analysed correlation lengths
of CPT measurement data follow lognormal distributions. These results extend the state
[email protected]
of the art presented in literature [23, 79, 184, 286]. Herein, one single correlation length is

48
 3.2 Case study B – CPT data evaluation

cumulative distribution function

80 %
MoMvar
percentile of the MoMLA
60 % ML

40 %

20 %

0.4 0.6 0.8 1 1.2

normalised correlation length qind / q

Figure 3.12: Casestudy B: Analysis of the correlation lengths for indicators of the CDF
using MoMvar , MoMLA and ML with fitted lognormal distribution functions
for silty clay layer.

used to interpolate or simulate spatial variable data under the assumption of one single,
known correlation length. On basis of the presented results, the author recommends to
take the most probable value e.g. the mean value from the distribution in figure 3.11 for
this issues. Moreover, the effects of a distributed correlation lengths is investigated in
chapter 6.2.

Indicator correlation lengths analysis: The results of the indicator correlation length
analyses are shown in figure 3.12. Around the median value the indicator correlation
lengths are slightly higher in comparison to the other thresholds, but have nearly the
same indicator correlation lengths for all quantiles of the CDF. Comparing these re-
sults to figures 3.4 and 3.12, one can conclude that the correlation structure of these
CPT measurements cannot be fully characterized by the mean value, standard devia-
tion and covariance function; it can be concluded that this measured structure of spa-
tial variability cannot be described by only one correlation function as it is done in the
well known multi-Gaussian model ,[143]. This was also found out by other researchers
[143, 144, 189]. The extreme low and high values have indicator correlation lengths. The
results of the MoMvar , MoMLA and ML approaches show a comparable behaviour: the
indicator correlation lengths are nearly the same for all thresholds. The scattering of the
indicator correlation lengths can be related to the indicator approach: The symmetrically
distributed data shown in figure 3.9 are transformed through the indicator approach in
equation 2.3 into skewed and asymmetric datasets. As shown before, this asymmetric
distribution causes less reliable results for the correlation length analysis. The MoMvar ,
MoMLA and ML methods are very sensitive to deviations from the normal distribution
(e.g. skewness γ1 > 1, excess kurtosis γ2,excess 6= 0). Finally, it has to be emphasised that
[email protected]
the indicator correlation length analyses clearly show a non-multi-Gaussian behaviour,

49
Chapter 3 Case studies on the evaluation of spatial variability

which cannot be fully captured by standard geostatistical simulation approaches (e.g.

LU-decomposition or Sequential Gaussian Simulation algorithm) as used section 3.1.
According to the author’s knowledge, this has not been observed by others in geotech-
nical engineering, but mentioned in geostatistics [79, 144].

Figure 3.13: Plan view and cross section of the experiments in the Fasanenhof tunneling
project.

[email protected]

50
 3.3 Case study C – Fasanenhof-tunnel

3.3 Case study C – Fasanenhof-tunnel

This case study on the evaluation of spatial variability describes the evaluation of the
spatial variability of equally spaced measured data within a homogeneous soil layer.
The variogram approach (MoMvar ), the local average approach (MoMLA ) and the Max-
imum Likelihood (ML) approach are used for this. The evaluation of the uncertainty
of the correlation length from a relatively small number of measurement data is shown
focusing on the uncertainty of the spatial correlation.
In addition ot this, the indicator correlation lengths are investigated using the MoMvar ,
MoMLA and ML approaches.

3.3.1 Site description & measurement technique

During the tunnelling construction process of the Fasanenhoftunnel in Stuttgart (Ger-
many) 45 horizontal core borings have been carried out as shown in figure 3.13. These
horizontal core-borings were grouped within a geologically homogeneous layer of mud-
stone with a separation distance of 2.5 m. The elevation of the boreholes was varied
according to the gradient of the tunnel. Therefore, the first and the last borehole have a
difference in the elevation of approximately 2.5 m. 45 borehole jacking tests have been
carried out at an approximate depth of 1.35 m as shown in figure 3.13.
Figure 3.14 shows the equipment of the borehole jacking test according to the German
standard DIN 4094 − 5 [104]. Two half-shells are pressed diametrically against the walls
of each borehole. Three different loading cycles are executed. The pressure is raised up to
three different levels of 1 MN/m2 , 2 MN/m2 and 3 MN/m2 . During this loading process
the deformation of the pressure plates is measured as illustrated by the schematic curve
in figure 3.15. The stiffnesses E1 , E2 and E3 for loading and reloading are evaluated
using equation 3.14.
Herein, d is the initial borehole diameter, p the pressure of the plates into the soil and
f (ν) is a constant, which depends on the Poisson’s ratio (e.g. f (ν = 0.30) = 0.972).

∆p
E = f (ν) d (3.14)
∆d

3.3.2 Analysis of the data

Correlation length analyses: The measurement data have been analysed in terms of
homogeneity and stationarity as described in chapter 3.2. Fulfilling these basic assump-
tions, the univariate statistics of the measurement data have been analysed. The results
are summarized in table 3.1. A χ2 –goodness-of-fit-test showed that the results of the
loading as well as the reloading experiments can be described by a lognormal distribu-
tion as shown by Huber et al. [172]. The mean values of the initial loading stiffness EL,i
are lower than the mean value of the reloading stiffness ER,i , whereas the COV is just
[email protected]
slightly higher in the case of reloading.

51
Chapter 3 Case studies on the evaluation of spatial variability

pressure
data acquisition
guide rod control
hydraulic
pump
pressure hose

sediment catcher A-A

110 mm 10°

A A borehole jacking probe

Figure 3.14: Experimental equipment of the borehole jacking probe according to the DIN
4094-5 [104].

The MoMvar , MoMLA and ML approach have been used to analyse the spatial variabil-
ity of the measurements. The Akaike-Information-Criterion has been used to identify the
best–fit in the variogram and ML approaches. The results are summarized in table 3.2.
The different approaches offer different results: the correlation lengths for loading and
reloading are different for the loading and reloading data within the variogram method,
which is not present in the other approaches; for the initial loading measurements, the
MoMvar and the MoMLA approach have more or less the same results, whereas the ML
method has longer correlation lengths.
The ML approach offers in contrast to MoMvar and MoMLA methods a possibility to
make a statement on the accuracy of the estimated correlation lengths as described in
equation 2.16. To overcome this drawback of the MoMvar and MoMLA methods, the non-
parametric J ACKKNIFE method is introduced to evaluate the reliability of the correlation
length. The basic idea of the J ACKKNIFE approach lies in systematically recomputing the
statistic estimate, leaving out one or more observations at a time from the sample set as

Table 3.1: Lognormally distributed results of the experiments in the Fasanenhof tunnel.

mean value COV = σ/µ

EL,1 loading 124 MN/m2 54 %
ER,1 reloading 660 MN/m2 64 %
EL,2 loading 229 MN/m2 53 %
ER,2 reloading 432 MN/m2 57 %
EL,3 loading 229 MN/m2 59 %
ER,3 reloading
[email protected] 397 MN/m2 61 %

52
 3.3 Case study C – Fasanenhof-tunnel

Figure 3.15: Loading, unloading and reloading cycles performed in the boreholejacking
tests according to DIN 4094 − 5 [104].

EL,1
6
EL,1 , EL,2 , EL,3

EL,2
in MN/m²

EL,3
4

2
0 20 40 60 80 100
location in the tunnel in m

ER,1
EU,1 , EU,2 , EU,3

8
in MN/m²

ER,2
ER,3
6

4
0 20 40 60 80 100
location in the tunnel in m

Figure 3.16: Measurement results within the Fasanenhoftunnel.

[email protected]

53
Chapter 3 Case studies on the evaluation of spatial variability

cumulative distribution function

100 %
MoMvar
80 % MoMLA
percentile of the

ML
60 %

40 %

20 %

0%
0.85 0.90 0.95 1.0 1.05
normalised correlation length qind / q

Figure 3.17: Casestudy C: Analysis of the indicator correlation lengths using MoMvar ,
MoMLA and ML of the EU,3 .

Table 3.2: Case study C: Comparison of the evaluated correlation lengths θ.

Method of Moments Maximum likelihood approach

MoMvar MoMLA ML
selected model θ [m] θ [m] selected model θ [m]
EL,1 exponential 4.59 4.72 exponential 44.98
EU,1 Gaussian 19.66 6.10 Gaussian 46.82
EL,2 exponential 5.96 4.67 exponential 42.44
EU,2 exponential 12.49 5.37 Gaussian 31.27
EL,3 Gaussian 8.67 4.62 Gaussian 45.21
EU,3 exponential 18.86 6.31 Gaussian 44.13

elaborated by Efron [111] and Journel [188].

An estimate for the bias and an estimate for the variance of the statistic can be cal-
culated from these 300 new sets of replicates of the statistic. The benefit of this non-
parametric approach is obvious: one can evaluate the consequences of the outliers in a
non-parametric way; moreover, it is also possible to estimate the variability of the esti-
mated correlation length using the MoMvar and MoMLA approaches. Within the J ACK -
KNIFE approach the selection of the theoretical variogram is carried out using the Akaike
Information criterion. The results are shown in table 3.3. Comparing the mean values in
table 3.3 with the correlation lengths in table 3.2, one encounters big differences, which
can be explained by the influence of the extreme values. This influence is be detected by
[email protected]
the J ACKKNIFE approach in the results in table 3.3.

54
 3.3 Case study C – Fasanenhof-tunnel

probability density function

0.04 MoMvar (μ = 22.59 m , COV = 73 %)
ML (μ = 44.98 m , COV = 59 %)
0.03
BMA (μ = 26.15 m , COV = 49 %)
0.02

0.01

0.00
0 20 40 60 80 100 120 140 160
correlation length q in meter

Figure 3.18: Case study C: combination of the results using BMA.

Indicator correlation length analyses: The J ACKKNIFE method is used in a similar way
for the analysis of the indicator correlation lengths. The thresholds are the percentile
values of the measurement data. MoMvar , MoMLA and ML methods offer comparable
results in figure 3.17. The indicator correlation lengths are nearly the same for all thresh-
olds; comparing theses results with the ones of case study A, one can clearly see that
the extreme high and low values show indicator correlation lengths that are more or less
the same as for the median value, as shown for EU,3 in figure 3.17. Although there is
a scattering, the distribution of the indicator correlation length at different thresholds
does not show the same distribution as in the analytical case shown in figure 3.4. This
can be attributed to a non Multi-Gaussianity of the measurement data, which cannot be
described by knowing solely the mean valued, the standard deviation and the spatial
correlation of measurements.

Table 3.3: Correlation lengths using the J ACKKNIFE approach for the MoMvar approach
in comparison to the mean value and the standard error of the ML approach.

MoMvar ML BMA
µ ± σ [MN/m2 ] µ ± σ [MN/m ] µ ± σ [MN/m2 ]
2

EL.1 22.59 ± 16.60 44.98 ± 26.33 26.15 ± 12.81

EU,1 25.30 ± 22.30 46.82 ± 27.42 28.28 ± 12.16
EL,2 9.31 ± 23.24 42.44 ± 25.91 22.76 ± 12.44
EU,2 29.27 ± 16.71 31.27 ± 16.71 24.83 ± 9.35
EL,3 11.31 ± 22.38 45.21 ± 27.69 23.98 ± 12.86
[email protected]
EU,3 37.29 ± 11.58 44.13 ± 27.96 34.73 ± 9.50

55
Chapter 3 Case studies on the evaluation of spatial variability

3.3.3 Remarks
A main concern in this measurement setting has been the different sources of error,
which are summarized in chapter 2.3. To keep the measurement error as small and
constant as possible, the same people carried out all the tests with the same equipment.
The knowledge error and description error were kept minimal by interviewing different
experts to judge the geological and geotechnical circumstances in order to identify the
homogeneous layer as shown in figure 3.13. After this, the spatial variability inside a
homogeneous layer was assumed to be the main source of uncertainty. But one has to be
aware that the mutual distance between the samples itself is also a source of error.
In tables 3.2 and 3.3 the results of the correlation length are presented; the results for
each loading and reloading cycle show an almost similar spatial correlation within the
MoMvar , MoMLA and ML approaches. The MoMvar , MoMLA and ML approaches show
different results. The differences between the three approaches can be related to the dif-
ferent evaluation schemes of the spatial correlation. By looking at the results, one can
conclude that the results of the MoMLA approach are not reliable. If one compares the
results of the MoMLA approach to the sampling scheme of the test results every 2 m, the
results of the MoMLA approach are not reliable because there are just 2 measurements
within the correlation length. Therefore, MoMLA is not considered for J ACKKNIFE ap-
proach.
The J ACKKNIFE approach is employed to evaluate the uncertainty of the spatial cor-
relation, which is new to the reader’s knowledge. Moreover, this offers the possibility
to combine via the Bayesian Model Averaging Scheme the results of the MoMvar and
the ML approach. Different assumptions and different approaches are merged in the
Bayesian Model Averaging scheme, which enables a more precise characterization of the
spatial correlation length with a lower COV = 49% in figure 3.18.
When looking at the results of the indicator correlation length analysis, one can clearly
see that the normalised correlation lengths is nearly the same for all investigated thresh-
olds. The normalised correlation lengths differ significantly from the analytically de-
fined case in case study A (section 3.1) From these new findings, one can conclude that
the simple methods for the simulation of spatial variability (by using a mean value, a
standard deviation and only one single covariance function) do not fully acknowledge
the measured spatial correlation structure.

[email protected]

56
 3.4 Case study D – Sheikh Zayed road in Dubai

3.4 Case study D – Sheikh Zayed road in Dubai

This case study deals with vertically, non equally spaced soil properties following a
skewed lognormal distribution. The resulting problems and challenges are presented
while using the variogram approach (MoMvar ), the local average approach (MoMLA ) and
the ML approaches.

3.4.1 Site description

Within this case study measurements from Wolff [406] are used to identify the spatial
properties of the uniaxial compressive strength (UCS) in limestone. Lab tests have been
carried out on 198 samples from different depths in 31 boreholes at four different con-
struction sites at the Sheikh Zayed road in Dubai. To make a clear statement on the
sources of error for these measurement data is not easy, as data have been collected from
different sites and the UCS tests have been conducted by different persons using differ-
ent measurement equipment.
The plan view of the four sites is given in figure 3.19. A χ2 goodness-of-fit-test showed
that the measured UCS in figure 3.20 (a) may be assumed to have a lognormal distribu-
tion with a mean value µ = 2.49 MN/m2 and a coefficient of variation of COV = 44 % as
shown in figure 3.20 (b).

3.4.2 Analysis of the data

The measurement data have been tested for their homogeneity and stationarity. After
this, the MoMvar , the MoMLA and the ML approaches have been used to explore the
spatial variability. As shown in figure 3.19, there are just four sites for the investigation
of the vertical correlation length. The horizontal correlation length is not investigated,
because one would need more data to get reliable results as pointed out in section 3.8.
The uncertainty of the vertical correlation length is investigated by using MoMvar ,
MoMLA and ML approaches ntogether with the J ACKKNIFE approach. The Akaike-

D
B
C

[email protected]
Figure 3.19: Case study D: Plan view of the four sites at the Sheikh Zayed road in Dubai.

57
Chapter 3 Case studies on the evaluation of spatial variability

(a) UCS in MN²/m (b)

0 2 4 6 8
5 histogram

probability density
0.30 fitted pdf
10
depth in m

15
0.20
20
25
0.10
30 measurements
35 mean value
trend 0.00
0 2 4 6 8
UCS in MN/m²

Figure 3.20: (a) Observations of the uniaxial compressive strength,

(b) histogram and fitted lognormal probability distribution function of
the UCS at the Sheikh Zayed road in Dubai.

Information-Criterion is employed to select the best fitting theoretical variogram model

within these approaches. The results of these investigations is shown in figure 3.21.
The data have been also analysed in terms of their indicator correlation lengths for
different thresholds cutk as shown in figure 3.22.

3.4.3 Remarks
The data, which have been evaluated in this section are non equally spaced, as shown in
figure 3.20. Therefore, the evaluation of the variogram has to follow equation 2.31 and
figure 2.6.
The strength of the MoMvar approach can be clearly seen in figure 3.21. It has by far
the lowest variations of the resulting correlation length, which is not the case for the
ML and MoMLA approaches. This can related to the subdivision of the non-equidistant
measurements into distance classes. If there are outliers in these distance classes, they
can influence the correlation length significantly. The problem of outliers is also cap-
tured by the J ACKKNIFE approach as a side effect; via this algorithm the influence of
these outliers on the correlation length is evaluated through the variance of the correla-
tion length as shown in figure 3.21. This challenge of non-equidistant data is not really
present for the ML approach because all data and their mutual correlation are captured
trough the correlation matrix. The ML approach is also just slightly influenced by the
q-q transformation.
In comparison to the variogram approach, the ML and MoMLA approaches show a
high COV of the correlation length. This can be related to the q-q transformation and
to the relative high skewness of the distribution, which is influencing the reliability of
these methods, as shown in case study A.
When looking at the results of the indicator correlation length analysis in figure 3.22,
[email protected]
one has to keep in mind the high skewness and asymmetry introduced by the indica-

58
 3.4 Case study D – Sheikh Zayed road in Dubai

probability density function

5 MoMvar (μ = 1.97 m , COV = 5 %)

4 MoMLA (μ = 1.17 m , COV = 23 %)
ML (μ = 1.67 m , COV = 24 %)
3
BMA (μ = 1.90 m , COV = 3.6 %)
2
1
0
0 0.5 1 1.5 2 2.5 3
correlation length q in meter

Figure 3.21: Case study D: Combination of the MoMvar , the MoMLA and the ML approach
using BMA.
cumulative distribution function

80 %
MoMvar
percentile of the

60 % MoMLA
ML
40 %

20 %

0%
0.95 1.0 1.05 1.10
normalised correlation length qind / q

Figure 3.22: Case study D: Analysis of the indicator correlation lengths using using the
MoMvar , the MoMLA and the ML approach.

[email protected]

59
Chapter 3 Case studies on the evaluation of spatial variability

tor approach. As in case studies B and C, the results are different from the findings of
the generated random process following a multi-variate Gaussian distribution in case
study A. The lower extreme values show a significantly higher correlation length than
the lower extreme values and the median value. Similar findings for permabaility mea-
surement data are presented in literature [144]. In this context, the indicator correlation
lengths are more or less the same for all thresholds [144].

[email protected]

60
 3.5 Literature database on the spatial correlation of soil properties

3.5 Literature database on the spatial correlation of soil

properties
Geostatistical textbooks [79, 97, 190, 401] always refer to expert knowledge, experience
or expert intuition in the interpretation of measurement results. No doubt, which is
even necessary in the presence of an automated fitting procedure as pointed out in sec-
tion 2.6.2. But it is very difficult to develop this experience. For this reason, the author
studied a vast number of publications (127 journal papers, technical reports, PhD theses
and conference proceedings) to gain more knowledge in the spatial variability of soil
properties. Afterwards the author set up three databases describing the spatial variabil-
ity of rock as well as frictional and cohesive soils. The database in Appendix A has the
following categories: property, soil type, vertical θver and horizontal correlation length
θhor and applied theoretical correlation function.
The main test types can be grouped into three classes; most of the investigations on
spatial variability are base on CPT measurements; another source of information are
observations of permeability among other measured properties. The uncertainties of
different sources e.g. sampling, measurement or statistical models are not investigated
within this database.
In Appendix A the different techniques can be seen in detail, which have been used
for the investigation of rock, frictional soils and cohesive soils. Additional investigations
have been carried out using the databases of frictional and cohesive soil as well as rock
in order to describe the horizontal correlation length θhor and the vertical correlation
lengths θver by means of probability density functions. For this reason the entries in
the database have been grouped into cohesive soils and frictional soils. For cohesive
soils 99 entries for θver and 64 entries for θhor and for frictional soils 63 for the θver and
52 entries for θhor have been found. Several authors offered only a lower and upper
bound instead of one single value for θhor and/or θver . To perform a proper analysis of
the collected data, these bounds have been used to generate variables, which follow a
uniformly distributed random variable between the lower and upper bound. One could
argue that the uniform distribution is quite a simple and conservative way to describe
the probability distribution of the correlation length. This is true, but in the absence of
more detailed information, it is possible to use it. The histograms of the θver and θhor for
frictional and cohesive soils are shown in figures A.3, A.4, A.5 and A.6.
Figures A.3 and A.4 show the empirical histogram of the correlation length for co-
hesive soils. In figures A.5 and A.6 the empirical histogram the correlation length for
frictional soils are depicted. In all four figures it can be seen clearly that the correlation
length can be described at different scales with different mean values and correspond-
ing standard deviations of fitted lognormal distribution functions. Table 3.4 summarize
the findings of the literature database. It offers mean values and standard deviations for
vertical as well as horizontal correlation length. It was found that frictional soils have
two different scales of correlation in the horizontal and vertical directions. Something
similar was found for cohesive soils. These findings can be explained by the different
scales of soil variability as mentioned in section 2.4.1.
[email protected]
The literature database on the spatial correlation of soil properties is offering an insight

61
Chapter 3 Case studies on the evaluation of spatial variability

Table 3.4: Properties of the lognormally distributed horizontal and vertical correlation
lengths of CPT data for frictional and cohesive soils.

frictional soils cohesive soils

scale 0 ∼ 10 m 10 ∼ 30 m 0 ∼ 15 15 ∼ 40 40 ∼ 60
vertical 0.18 ± 0.78 m 30 m 0.29 ± 1.14 m 23.30 ± 1.6 m 50 m
scale 0 ∼ 50 m 50 ∼ 100 m 0 ∼ 100 m 100 ∼ 500 m
horizontal 2.35 ± 0.43 m 90 m 2.27 ± 1.72 m 25.54 ± 0.30 m

into correlation lengths of different soil. The presented database is used to derived upper
and lower limits of vertical and horizontal correlation lengths of different soil types. This
can be used to set up efficient sampling schemes because the spatial correlation length
is essential to investigate the micro-, meso- or macro scale soil variability. Moreover, the
presented database links soil types and spatial variability presented in literature.

3.6 Evaluation of the vertical correlation length using CPT

databases of different soil types
One can clearly deduce from the literature database in Appendix A that there is need for
more experimental investigations on the correlation lengths in different soils. But most
experiments for the detection of the correlation length are cost intensive. Therefore, not
many case studies are carried out. One way out of this is offered by the cone penetration
test (CPT). This relatively cheap and simple technique allows one to measure nearly
continuously soil properties, which can be evaluated for their correlation length; a big
number of publications on this has been collected in the database in appendix A .
A database of the N ATIONAL G EOTECHNICAL E XPERIMENTATION S ITES (NGES) [50,
129, 380] and of the PACIFIC E ARTHQUAKE E NGINEERING R ESEARCH C ENTER (PEER)
[275] offer a large number of vertical CPT measurements of different sites together with
soil profiles. The key-points of the different CPT sites such as number of CPTs, minimum
and maximum depth and sampling rate are summarized in table 3.5. In Appendix B a
plan view of the CPTs together with a typical CPT profile is added.
Just like the CPT measurements presented in case study 3.2, all 671 CPT measure-
ments are analysed by the stochastic site characterization scheme described in case study
3.2. The subdivision of each CPT profile into homogeneous soil layers combines expert
judgement and statistical techniques; the vertical correlation length was investigated in-
side each layer.
After this, the charts of Robertson [308] in figure 3.23 are used to classify each layer into
one of the 9 proposed soil types. Herein, qt is the cone resistance, σv0 is the initial, total
vertical stress, σv0
′
is the initial, effective vertical stress, u0 the pore water pressure and uz
the measured pore water pressure and fs the sleeve friction. These quantities are used
to calculate the normalized cone resistance Qt , the friction ratio Fs and the normalized
[email protected]
pore water pressure Bq .

62
 3.6 Evaluation of the vertical correlation length using CPT databases of different soil types

1000 1000
7 8 sv0
qt
7
no u
rm 9
all
100 100 6
6
yc
on

Qt Qt 5
sol
ida

5 4
dte

10 4 10 3

1 3
1
2
1 1
0.1 1 10 -0.4 0.0 0.4 0.8 1.2
Fr [%] Bq
q t -σ v0 u z -u 0 fs
Qt = Bq = Ft = × 100%
σ'v0 q t -σ v0 q t -σ v0

Figure 3.23: Soil classification into 9 soil behaviour types according to Robertson [308]:
zone 1: sensitive fine grained
zone 2: organic soil-peat
zone 3: clay-silty clay
zone 4: clayey silt - silty clay
zone 5: silty sand - sandy silt
zone 6: clean sand to silty sand
zone 7: gravelly sand to sand
zone 8: very stiff sand to clayey sand and
zone 9: very stiff to fine grained soil.

The vertical correlation lengths for the soil types are statistically analysed and the
mean values and coefficients of variation of the fitted lognormal distribution functions
are summarized in table 3.6. This table also contains the results of the BMA scheme,
which were used to combine the results of the MoMvar , the MoMLA and the ML approach.
The above cited databases do not cover the investigated CPT measurements of the soil
type 7 "gravelly sand to sand", soil type 8 "very stiff sands to clayey sand" and soil type
9 "very stiff, fine grained sand".
One can clearly see that the difference between "sensitive fine grained soils" and "clean
sands to silty sands" is not that big. For the MoMvar , the MoMLA and the ML approach,
the mean values of the vertical correlation length ranges between θvert = 0.4 m and
[email protected]
θvert = 0.8 m. Also the COV values do not show a big scatter. Maybe this can be linked

63
Chapter 3 Case studies on the evaluation of spatial variability

Table 3.5: Summary of the different CPT databases [50, 129, 275, 380].

site number of CPTs depth sampling rate

min max
m m #/m
NGES Alameda 195 9.65 50.75 0.025
NGES Evansville 40 12.00 33.25 0.025
NGES Lancaster 41 8.35 23.25 0.025
NGES San Bernardino County 88 10.00 27.00 0.050
NGES San Luis Obispo 37 13.50 30.00 0.050
NGES Santa Clara County 163 3.15 38.50 0.050
NGES Solano County 12 13.60 25.00 0.050
PEER Anssall 8 10.60 30.83 0.050
PEER Berkeley 87 2.40 10.20 0.050

Table 3.6: Mean value and COV of the vertical correlation lengths of the different CPT
databases and the BMA combination results.

soil types MoMLA MoMvar ML BMA

µ COV µ COV µ COV µ COV
in m in % in m in % in m in % in m in %
sensitive, fine grained 0.78 170 0.47 160 0.54 101 0.53 64
organic soil to peat 0.53 145 0.46 120 0.51 79 0.49 53
clay to silty clay 0.43 121 0.45 127 0.48 81 0.46 52
clayey silt to silty clay 0.53 126 0.50 135 0.56 80 0.54 53
silty sand to sandy silt 0.75 201 0.51 140 0.57 79 0.56 57
clean sand to silty sand 0.45 116 0.48 100 0.29 108 0.47 54

to the genesis of the tested soils because all of the tested soils can belong to sediment
soils. It is very difficult to compare the presented results to other works because no
comparable study is available to the author’s knowledge.
The presented results are an extension of the literature database in section 3.5 using
measurement data. On basis of CPT measurements, the vertical correlation lengths of
various soil types are evaluated. These results offer a detailed insight into the spatial
correlation of CPT measurements, which have been performed in 6 different soil types.
These results are contributing to a general description of stochastic soil properties of 6
different soil types.

[email protected]

64
 3.7 Combining expert knowledge and measurements

literature database (m = 1.20 m , COV = 66 %)

probability density function

CPT database (m = 0.53 m , COV = 64 %)

2.5
MoMvar (m = 0.88 m , COV = 120 %)
2.0 MoMLA (m = 0.62 m , COV = 204 %)
ML (m = 0.75 m , COV = 89 %)
1.5
BMA (m = 0.40 m , COV = 34 %)
1.0

0.5

0.0
0 0.5 1 1.5 2
vertical correlation length q in meter

Figure 3.24: Combination of measurement data of silty clay, results of the literature
database and results of the CPT databases using Bayesian Model Averaging.

3.7 Combining expert knowledge and measurements

Before making use of data collected at the site [369, 382], the engineer can express his in-
formation from a literature study or similar measurement or even his expert judgement
concerning the set of uncertain parameters.
This combination of prior (subjective) information and "objective" data can be carried
out via the B AYESIAN M ODEL AVERAGING (BMA) scheme, which is already explained
in case study B in section 3.2.3. The BMA scheme enables the combination of different
models as well as different sources of information in a relativly simple manner, namely
the B AYES theorem. Via this approach it is possible to derive a conditional probability
distribution, which considers different sources of information.
This scheme is been applied to the results of the CPT data of a silty clay from C ASE
STUDY B. As shown in figure 3.11, the the MoMvar , the MoMLA and the ML approaches
can be combined sequentially via the BMA scheme. This is proceeded by adding ad-
ditional information to the results: the results from the literature database and CPT
database offer additional expert knowledge of comparable soil types. By looking at fig-
ure 3.24, one can see that the combined model has a lower mean value as well as a lower
COV. The results of the literature database have a relatively flat and wide pdf, which
does not have a big influence on the combined model. Those pdfs have a high influence
on the combined model, which have a low COV. This can be explained by the availability
of information: the more information is available, the more precise is the description of
the probability density function.

[email protected]

65
Chapter 3 Case studies on the evaluation of spatial variability

3.8 Summary and consequences of the case studies

The application of different methods for evaluating the correlation length to an analytical
random process shows the strength and weaknesses of Mthods-of-Moment and ML ap-
proaches, which are even more highlighted by the evaluation of equally and non-equally
spaced measurement data.

Case study A: The simplest boundary conditions are used in CASE STUDY A; artificially
generated, homogeneous, stationary data with a known correlation length θtarget and
an added error are analysed by the MoMvar , the MoMLA and the Maximum Likelihood
approaches. The data are quasi continuously defined in order to exclude a sampling
induced errors. It is shown that all three approaches offer nearly the same results under
ideal conditions of a symmetric probability density function.
In addition to this, the indicator approach is employed to analyse the spatial corre-
lation of extreme low and high values. The indicator correlation length of the extreme
low and high values of the random process show a significantly lower correlation than
the median values. Via this, the indicator correlation length of a theoretical multivariate-
Gaussian random process is analysed and shown, which help to understand the results
of the case studies in the latter.

Case study B: A scheme for stochastic site characterization is used to describe the soil
tested by 138 vertical CPT measurements in CASE STUDY B. This procedure combines
engineering judgement and these statistical techniques to identify soil layering. More-
over, the spatial variability of each soil layer is identified and statistically described by a
lognormal distribution function. The presented scheme enables the engineer to separate
different scales of variability, e.g. the geological macro-scale and and smaller geotechni-
cal meso-scale inside each layer. By separating the macro- and the meso-scale, the evalu-
ation of the correlation lengths becomes more easy because measurement data from the
observed layers do contain less erroneous parts of e.g. large scale fluctuations.
The results of the the MoMvar , the MoMLA and the ML approaches offer independent
interpretation of the measurement data. Due to the big number of CPT measurements
it is possible to identify a lognormally distributed vertical correlation length, which is a
novel insight into this context.
The results of the MoMvar , the MoMLA and the ML approaches are combined using
the B AYESIAN M ODEL AVERAGING scheme, which combines the information of differ-
ent models using the B AYES theorem. This procedure results in a probability density
function of the correlation length, which contains different models and assumptions in
one results. The resulting probability density function of the correlation length is a new
insight into spatial variability.
The indicator correlation length analyses clearly show that the measurement data have
a particular correlation of the extreme low and high values. This implies that these spa-
tial correlations cannot be modelled by conventional simulation approaches using one
[email protected]
single correlation function.

66
 3.8 Summary and consequences of the case studies

Case study C: Equi-distant in-situ measurement of stiffnesses are analysed within this
case study. In order to analyse the uncertainty of the correlation length, the well known
the MoMvar and the MoMLA approach approach is extended by using the J ACKKNIFE
procedure. Via this novel approach, a comparison with the uncertainty measures of the
ML approach is possible.
Furthermore, the results of the the MoMvar , the MoMLA and the ML approaches can be
combined with the introduced Bayesian Model Averaging scheme.
In addition to this, the indicator correlation length analyses indicate a spatial structure
of the measurement data, which cannot be fully described by standard approaches using
one single correlation length used in case study A.

Case study D: In case study D the non-equidistant data are investigated. These mea-
surement results of the soil strength follow a skewed, lognormal distribution. Apart
from the challenges of analysing these data, the B AYESIAN M ODEL AVERAGING ap-
proach offers an elegant tool to combine the results of the the MoMvar , the MoMLA and
the ML methods.
The results of the presented case studies show that the presented methodologies can
evaluate the correlation length of equally and irregularly spaced measurement data. The
results in the case studies show different ratios of the nugget effects to sill values. This
implies that different scales of variability are involved. But one has to keep in mind that
the the MoMvar , the MoMLA and the ML approaches can only identify the spatial corre-
lation of one scale. Therefore, the sampling concept of the measurement data is essential
and has a major impact on the results. One has to know in advance the scale in order to
measure it correctly. But there are no satisfying solutions offered in literature, as pointed
out in section 2.5.4. The big drawback of the presented design-based sampling concepts
of geostatistics is the high sampling effort, which implies high costs. Model based design
approaches like nested-sampling design [403] or adaptive sampling approaches offer a
promising alternative; but this is suffering from the assumption of normally distributed
variables.
Due to the multi-scale nature of spatial soil variability no general answer for the min-
imum sampling effort to describe spatial variability can be given. Barnes [32] states that
the upper bound of required sampling for spatially dependent samples has to be lower
than for identically, independently distributed samples. Different authors focused on
this but only some of them like Journel & Huijbregts [190] or Lark [214] offer rules of
thumb for a one dimensional spatial correlation. Here, the number of pairs should be at
least 30 to 50 pairs. In the field of ecology and environmental engineering, Marchant &
Lark [236] recommend to have at least 100 to 150 samples for a two dimensional spatial
analysis, whereas Webster & Oliver [402] recommend at least 200 samples. Other publi-
cations [90, 197–200, 270, 271] refer to a minimum sampling size of 100 − 150 samples in
a two dimensional analysis of the spatial correlation using the ML approach. DeGroot &
Baecher [90] provide in figure 2.8 a relationship between the correlation length and the
sample size, which also gives comparable suggestions for a minimum required sample
size.
[email protected]
If the number of pairs is lower, the uncertainty of the correlation length becomes very

67
Chapter 3 Case studies on the evaluation of spatial variability

dominant and complicated for the ML approach as well for the presented J ACKKNIFE
method.

Literature database and CPT databases: At the end of an evaluation of measurement

data, one has to compare the resulting correlation lengths with knowledge, which re-
flects experience, the state-of-the-art or the state-of-science. This can be done by using
the presented databases of literature, which presents data on spatial variability of differ-
ent soil types in literature.
In addition to this, presented results of the CPT databases offer a more specific insight
to this problems. The vertical correlation lengths of eight different soil types ranging
from clay to sand are presented by indicating mean value and standard deviation of the
lognormal distribution.
At the end of this chapter, the B AYESIAN M ODEL AVERAGING procedure is used to
fuse different model results. In a similar way, these results can be updated with addi-
tional knowledge coming form expert judgement, literature knowledge or results from
comparable tests in similar soils. Via this, the probabilistic site characterisation can be
enriched by expert knowledge or other sources of information within a mathematically
defined framework.

[email protected]

68
 Chapter 4
Safety and uncertainties
Within this chapter the basics of safety and reliability are highlighted. It provides a
description of common approaches for dealing with uncertainties and safety in geotech-
nical engineering. Global and partial safety factors, as well as the basics of uncertainty
quantification and reliability based design, are described and compared with a special
focus on the basics of the generation of random numbers and random fields, as well as
on the computation of failure probabilities. Moreover, an introduction to the local and
global sensitivity analysis of systems is presented.

4.1 Safety in engineering

The goal of safety is the preservation of existence of an individual or a community. Since
many general requirements of humans have found their way into laws, the preservation
of psychological and physiological functioning of humans can be found in many consti-
tutions and in the UN Charter of Human Rights as stated in Proske [295]. Although the
term safety can be found in many laws, this does not necessarily mean that the content
of the term is clearly defined. Many people have a different understanding of the term.
Some common descriptions are presented in [163, 261, 295]:

• Safety is a state in which no disturbance of the mind exists, based on the assump-
tion that no disasters or accidents are impending.

• Safety is a state without threat.

• Safety is a feeling based on the experience that one is not exposed to certain hazards
or dangers.

• Safety is the certainty of individuals or communities that preventive actions will

function reliably.

Safety requirements and safety concepts have a long history in some technical fields
[295]. Nearly 4,000 years ago, this can be seen in the code of H AMMURABI, in which
strong penalties in the case of collapsing structures were fixed. Probably the first ap-
plication of a global safety factor in structural engineering dates back to P HILO from
Byzantium [338] in 300 B.C., who introduced the global safety factor η in terms of:

[email protected] resistance
η= (4.1)
load

69
Chapter 4 Safety and uncertainties

Table 4.1: Factors of global safety in engineering according to Visodic [397].

Factor of safety knowledge of load knowledge of material knowledge of environment

1.2 - 1.5 excellent excellent controlled
1.5 - 2.0 good good constant
2.0 - 2.5 good good normal
2.5 - 3.0 average average normal
3.0 - 4.0 average average normal
3.0 - 4.0 low low unknown

Table 4.2: Factors of global safety in geotechnical engineering after Terzaghi & Peck [373].

item factor of safety

earthworks 1.3 – 1.5
earth retaining structures 1.5 – 2.0
excavations, offhore foundations 2.0 – 3.0
foundations on land
uplift, heave 1.5 – 2.0
piping 2.0 – 3.0
load tests 1.5 – 2.0
dynamic formulas 3.0

Only in the last few centuries have the application of global safety factors become widespread.
Over time many different values were developed for different materials [295]. In most
cases, the values dropped significantly during the last century. Proske et al. [297] report
that in 1880 in brick masonry, a factor of η = 10 was required, whereas 10 years later, a
factor between η = 7 − 8 was required. In the 20th century, the values have changed from
a factor of η = 5, then to η = 4, and now for the recalculation of historical structures a
factor of η = 3 is chosen [297].
Especially with the development of new materials, an increase in concern over the
safe application of these materials has arisen. At the beginning of the 20th century the
development of safety factors for different materials led to initial efforts in developing
material-independent factors, such as those shown in table 4.1. Visodic [397] shows
that the knowledge of load, material and environment have an influence on the global
safety factor depending on the state of knowledge: if there is more knowledge on the
load, material and environment, the factor of global safety can be reduced to a certain
limit [297]. A similar tendency can also be observed for other materials like steel. In
geotechnical engineering, Meyerhof [251] reports that the factor of global safety for the
stability of a retaining wall remained the same, since being introduced in the 18th century
by Belidor and Coulomb. The global factors of safety for different geotechnical problems
are summarized in table 4.2. These values by Terzaghi & Peck [373] do not take into
[email protected]
account the variability of the soil properties or additional knowledge on the soil.

70
 4.1 Safety in engineering

More advanced changes might be considered to meet the demanding requirements of

economic and safe structures in order to include the different levels of knowledge. Such
developments include special safety factors for the different column heads in table 4.1,
for example a safety factor for load and a safety factor for material. Indeed, this is the
basic idea of the partial safety factor concept. A more precise consideration of the different
uncertainties in loads and resistances yields a more homogeneous level of safety of a
structure [297]. The proof of safety is carried out by the simple comparison of the load
event Ed with the resistance of the structure Rd .

E d ≤ Rd (4.2)

Subsequently, the load event can be built from several single elements, such as the char-
acteristic dead load Gk,j connected with a special safety factor γG,i only for dead load,
and the characteristic life load Qk,j connected with a special safety factor γQ,i and ψ0,i for
the combination of different life loads:
X X
Ed = γG,i Gk,i + γQ,i Qk,i γQ,i ψ0,i Qk,i ≤ Rd (4.3)

According to Proske [297], although the partial safety factor concept was first introduced
in structural engineering after World War II, it took quite some time to become practi-
cally applicable. In geotechnical engineering Talyor [372] was one of the first amongst
others to introduce separate factors of safety on the components of shear strength in
slope stability estimation. It is reported in [251] that Brinch-Hansen generalized this
approach and initiated partial safety factors in geotechnical engineering. In civil engi-
neering, these partial factors were chosen to give about the same design estimate as con-
ventional total factors of safety as stated by Schuppener [325, 330]. These factors have
been refined subsequently by semi-probabilistic methods on the basis of the variability
of the loads, soil strength parameters and other design data in practice [251]. Therefore,
the development of partial safety factors is strongly connected to the development of the
probabilistic safety concept in structural engineering [297].
The calibration of partial safety factors is described in E UROCODE [71] as shown in
figure 4.1. The deterministic as well as the probabilistic approach can be used for cal-
ibration of the partial safety factors. The deterministic approach includes historical as
well as empirical methods in civil engineering, which have shown their strengths over
several years or even decades [297], for the calibration of partial safety factors. The
probabilistic approach can use reliability methods and fully probabilistic methods (e.g.
Monte Carlo approach) for the calibration of the partial safety factors. The probabilistic
methods offer the basis for the calibration of the partial safety factors, which are used
in the semi-probabilistic approach. In the semi-probabilistic approach, partial safety fac-
tors are used to consider the uncertainties in load, resistances and model uncertainties,
[297].
The first proposals about probabilistic-based safety concepts were found in the 1930s
in Germany and in the Soviet Union. The development of probabilistic safety concepts, in
general, experienced a strong impulse during and after World War II, not only in the
[email protected]
field of structures but also in the field of aeronautics. The Joint Committee of Structural

71
Chapter 4 Safety and uncertainties

deterministic probabilistic methods

methods
historical methods reliability fully probabilitic
empirical methods methods methods

calibration calibration calibration

semi-probabilistic
methods
method C
design using
method A partial safety factors method B

Figure 4.1: Calibration of the partial safety factors according to E UROCODE [71].

Safety (JCSS) introduced the probabilistic model code [121] on the basis of the proba-
bilistic safety concept of structures. The probabilistic model code includes a detailed
introduction to model the load and resistance parameters as random variables in order
to calculate the probability of failure of structures.
In E UROCODE [71] the safety of structures is defined as the capability of structures to
resist loads. Due to the fact that no building can resist all theoretically possible loads, the
resistance has to reach only a sufficient level [71]. Only by using a quantitative measure
can one offer a basis on decision on whether a structure is reliably or not. The reliability
is interpreted as a probability of failure not occurring, which is explained in detail later
in this chapter.
The design process including reliability is called reliability based design (RBD) [286].
Herein, the uncertain components of a system are simulated as random variables within
the reliability framework to evaluate the probability of failure. The E UROCODE, as well
other design codes [121, 127], probabilistic methods are used with the reliability based
design concept calibration of partial safety factors.
Proske [297] offers in figure 4.2 an overview of most of the safety concepts in struc-
tural engineering. Starting from basic empirical rules, Proske [297] adds concepts with
increasing complexity: As pointed out above, the global and semi-probabilistic safety
concepts are less complex than simplified and exact probability safety concepts. As a
next step, the reliability index and the probability of failure of a system are compared to
the a target reliability and probability of failure. Probabilistic methods like First Order
Reliability Method or Monte Carlo Method can be used for this, as described in the E U -
ROCODE [71]. Relatively new concepts like fuzzy-probabilistic procedures [252] and risk
based design concepts are also included in figure 4.2. One can deduce form this figure
that different safety concepts can cover different aspects. Proske [297] states that only
[email protected]
complex safety concepts fulfil the basic requirements indicated by basic human rights

72
 4.1 Safety in engineering

for safety and legal restrictions to save life.

Additional to the reliability-based safety concepts, there are the codes for structures
known as risk performance based design and risk based design [168, 286, 295]. There ex-
ist several different viewpoints with respect to the term risk [295]. Generally spoken, risk
is part of man’s judgement, when negative impact of dangerous activities is balanced by
their profits; such judgements are subject to bias and in many cases they contradict sta-
tistical facts, but they are the basis of acceptance or rejection of risks. Risk acceptance
is often formulated in an FN criterion, as shown in figure 4.3. Within the framework
of quantitative risk assessment, FN-curves show the frequency of occurrence (F) of an
event in relation to the number of lost lives (N). FN-curves are mainly based on the
discipline, in which the viewpoint was created and the requirements of that discipline.
These viewpoints have been summarized by Renn [306] into insurance-statistic based
view, toxicological-epidemiological view, engineering-technical view, economical view,
psychological view, social-theoretical view and cultural-theoretical view. The first group
is represented by the statistical-mathematical formulation and will be partly focused on
in this chapter. This formulation is based on the general formula:

R=C ·P (4.4)

where R is the risk, C is the negative consequence measure (damage or disadvantage)

and P is the probability of the occurence of C. Proske [295] points out that all compo-
nents can have different units, of course. For example, the negative consequences can be
given in monetary units, in time required, loss of space, loss of humans, loss of creatures,
loss of energy, loss of environment and so on.
The terms safety and risk have strong relations to the terms hazard and danger. The term
hazard implies the occurrence of a condition or phenomenon, which threatens disasters
to anthropogenic spheres of interest in a defined space and time [133]. In general, a
hazard is a natural, technical or social process or event that is a potential damage source.
There exist many further definitions of hazards; however, these will not be discussed
here.
In general, a hazard might be completely independent from the activities of humans,
for example an avalanche or a debris flow. However, if people are at the location of
the process, then these people might be in danger. Within this definition, Proske [295]
derives danger as a situation that yields without unhindered development to damage.
On one hand, danger can be seen as the opposite of safety, where no resources have to
be spent. Luhmann [225] stated that risks and hazards are opposites in terms of human
contribution. According to Proske [295], risks require possible human actions, whereas
hazards are independent from human actions. If that is true, however, then more tech-
nologies or more human resources simply means more risks due to the increased volume
of human actions and decisions.
Vulnerability is a term that permits an extension of the classical risk definition by only
two terms. Instead, further properties can be incorporated into the term vulnerability,
which itself is then part of the risk definition:
[email protected] R = f (p, A0 , ν0 , p0 ) (4.5)

73
74
human right “safety”

government duty “safety”

[email protected]
risk-based safety concept
Chapter 4 Safety and uncertainties

possibility-based safety concept

~ fuzzy-probability safety concept

~ exact probability safety concept

simplified probability safety concept

semi-probability concept

empirical safety concept with global safety factor

rules rules
Rules
empirical rules

Figure 4.2: Safety concepts in the context of structural engineering from Proske [295].
 4.1 Safety in engineering

Table 4.3: Target reliabilities according to E UROCODE [71].

Reliability consequences for loss of human reliability index βa

classes life, economic, social and βa for βa for
environmental consequences Ta = 1 year Ta = 50 years
RC 3 - high high 5.2 4.3
RC 2 - normal medium 4.7 3.8
RC 1 - low low 4.2 3.3

where R is the risk, p the probability of this event, A0 the value of an object, ν0 the
vulnerability of the object during the event and p0 the probability of exposure of the
object during the event, e.g. flood or landslide amongst other natural hazards [133].
Equation 4.5 can be further extended to several event scenarios and objects. However,
the term vulnerability remains to be defined. Again, as with many other terms, the
variety of definitions of vulnerability is virtually unmanageable according to [133, 295].
Deeper discussion and examples can be found amongst others in [133, 295].
At the basis of risk based design, Baecher & Christian [23] summarized the annual failure
probability and expected losses for a variety of common civil facilities and other large
structures or projects in figure 4.3. Baecher & Christian [23] and Meyerhof [251] are the
only two sources of FN-curves in geotechnical engineering according to the knowledge
in geotechnical engineering. In figure 4.4 Meyerhof [251] shows the global safety factor
in comparison to the lifetime probability of stability failure using different coefficients of
variability (COV). On top of this, observed and theoretical probabilities of failure in soil
engineering is compared to lifetime fatality risk per person. From figures 4.3 and 4.4 one
can derive that there are different reliability levels for different structures.
Table 4.3 shows the classification of target reliability levels provided in the E UROCODE
[71]. Reliability indices are given for two reference periods T ( 1 and 50 years) but with-
out any explicit link to the design working life Td . The reliability index β is a scalar
measure of safety equivalent to the probability of failure pf , but measured on another
scale. A detailed description of the reliability index is given in section 4.3.1.
The values are based on calibration and optimization and reflect results from several
studies. It is noted that similar β values as in table 4.3 are given in other national and
international guidelines. Examples of buildings and civil engineering works for RC 3
are bridges and public buildings, for RC 2 residential and office buildings and for RC 1
agricultural buildings and greenhouses.
In addition to this, reliability indices with respect to consequences and to relative cost
of safety measures are presented in JCSS [121] and ISO 2394 [183] offers more detail. The
bigger the reliability index β becomes, the more unreliable is the occurrence of failure,
which will be covered in section 4.3.1.
Calgaro [66] states that different criteria may be taken into account when choosing a
target reliability index. He emphasis that the combination of economic, risk acceptance,
psychological and legal criteria have to be taken into account.
The Life Quality Index (LQI) is a recently developed concept for determining accept-
[email protected]
ability of decisions involving life safety risks in engineering. It provides a rational basis

75
Chapter 4 Safety and uncertainties

0
10
marginally accepted
accepted
annual probability of failure

-1 mine
10 pit
slopes merchant shipping
geisers
10
-2 mobile
MOBIL
drill rigs
DRILL RIGS

foundations super
-3
fixed tankers
10 FIXED
drill rigs
canvey refi
neries
-4
10 super
dams tankers
-5 estim
10 ated
US
commercial d am
-6 aviation s
10
-2 -1 0 1 2 3 4
10 10 10 10 10 10 10 lives lost

Figure 4.3: Chart showing average annual risks posed by a variety of traditional civil fa-
cilities and other large structures or projects proposed by Baecher & Christian
[23].

lifetime fatality
global factor of safety risk per person
1.0 1.5 2.0 2.5 3.0
10
0
human life
lifetime probability of failure

natural disasters
high

COV mining
theor
y = 30 %
10
-2
FO motor vehicles
medium
CO

E CO ships, fires
V
Vt

R the
= 20 railways
he

-4 o ry
or

10 %
y

aircraft
=1

low
0%

gas piplines
-6
nuclear reactors
10
E FO , R FL

E = earthworks, FL = foundations on land, FO = offshore foundations,

R = earth retaining structures, COVtheory = theoretical COV in %
observed , theory

Figure 4.4: Lifetime probabilities of stabilitiy failures and comparative human risks from
Meyerhof [251].
[email protected]

76
 4.2 Basics of uncertainty quantification

Table 4.4: Reliability based design: different levels of accuracy from Honjo et al. [168].

level basic variables reliability

III random variables & probability density function probability of failure
II random variables & mean & standard deviation reliability index
I deterministic variables partial safety factors

for establishing target reliabilities for civil engineering systems [218, 299, 361]. The LQI
is a socio-economic utility function that depends on the wealth and life expectancy of a
society. Any decision that increases the value of the LQI is deemed acceptable. This in-
crease can be due to an increase in life expectancy (reduction of fatalities) or an increase
in societal wealth (reduced use of resources). In this way the LQI establishes a relation
between the resources invested in improving the safety of an engineering facility and
potential fatalities and injuries that are avoided by the investment. Hence it provides a
means to quantify the optimal trade-off between safety and cost [299, 361].

4.2 Basics of uncertainty quantification

Uncertainty quantification is forming the framework to focus on the effects of variability
within computational mechanics, as visualized in figure 4.5. In step A, a mechanical
model is set up together with assessment criteria for the behaviour of the system. This
step gathers all the ingredients used for a classical deterministic analysis of the physical
system to be analysed. In the next step B, the quantification of sources of uncertainty
is performed and random variables or random fields are used for the representation of
the different sources of uncertainties of the system. Within the uncertainty propagation
in step C, the response of the system is (with respect to the random input variables and
fields) evaluated, enclosing the uncertainty of the system. Numerous methods exist to
carry out this task as described in section 4.3.
Sudret [364] states that uncertainty propagation methods usually provide information
on the respective impact of the random input parameters on the response randomness.
A sensitivity analysis helps to identify the main sources of the response randomness.
Moreover, Sudret [364] point out that this sensitivity analysis may sometimes be the
unique goal of a probabilistic study.
Amongst others [57, 89], Sudret [364] reports that this representation can be done in
several levels of accuracy. Honjo et al. [168] as well as Phoon [286] show different levels
of accuracy summarized in table 4.4; different methods and approaches provide a more
detailed insight into the reliability of the system and form the key requirement for reli-
ability based design. In the basic case, one can use deterministic variables and partial
safety factors to simulate random variables of a geotechnical problem. This can be im-
proved by taking the mean value and the standard deviation of the random variables
into account. A result of this analysis technique is the reliability index. On top of this
[email protected]
is the simulation of random variables using the full probability density function. Via

77
Chapter 4 Safety and uncertainties

Figure 4.5: Steps in Uncertainty Quantification from Sudret [364].

this approach, it is possible to evaluate the probability of failure more precisely because
more information is available in comparison to the other levels of RBD and uncertainty
quantification.

4.2.1 Definition of the limit state function

When considering models of mechanical systems M, random variables X usually de-
scribe the randomness in the geometry, material properties and loading. They can also
represent model uncertainties [364]. This set also includes the variables used in the dis-
cretisation of random fields, if any. The model of the system yields a vector of response
quantities Y = M(X). In a mechanical context, these response quantities are e.g. dis-
placements, strain or stress components, or quantities computed from the latter.
The mechanical system is supposed to fail when some requirements of safety or ser-
viceability are not fulfilled. In the case of complex systems, one can set up different fail-
ure modes to capture the system behaviour in a precise way as described by Huber et al.
[175]. For each failure mode, a failure criterion is set up. It is mathematically represented
by a limit state function g(X, M(X), X′ ). The limit state function may depend on input
parameters X, response quantities M(X) that are obtained from the model and possibly
additional random variables and parameters gathered in X′ . For the sake of simplicity,
the sole notation X is used in the sequel to refer to all random variables involved in the
analysis. Let M be the size of X.
Conventionally, the limit state function g is formulated for the realisations x of the
random variable X in such a way that:
Ds = {x : g(x) > 0} is the safe domain in the space of parameters; (4.6)
[email protected]
D = {x : g(x) ≤ 0} is the failure domain. (4.7)
f

78
 4.2 Basics of uncertainty quantification

The set of points {x : g(x) = 0} defines the limit state surface [364, 366]. Denoting by
fX (x) the joint probability density function of a random vector X, the probability of
failure pf of the system is:
Z
pf = Prob [g (X1 , X2 , . . . , Xn ) ≤ 0] = fX (x)dx (4.8)
Df

Sudret [366] emphasis that in all but academic cases, this integral cannot be computed
analytically. Indeed, the failure domain depends on response quantities (e.g. displace-
ments, strains, stresses, etc.), which are usually computed by simplified equations [44,
57, 176, 286] or by means of numerical methods e.g. the finite element method [57, 176,
364, 366]. In other words, the failure domain is implicitly defined as a function of X.
Thus, numerical methods have to be employed to evaluate the probability of a failure.

4.2.2 Random number generation

The basis of the simulation of correlated random variables is the generation of random
numbers following a uniform distribution in an interval from zero to one. According
to Fenton & Griffiths [127], the most popular random number generators are the lin-
ear congruential generators (LCG). LCG generate independent, identically distributed
numbers with a very long periodicity fast by not requiring too much memory. More-
over, LCG have the ability to reproduce a given stream of random numbers exactly, as
emphasised by different researchers [85, 127, 266, 294].
The generation of non-uniform random numbers can be done via inverse transform,
convolution and acceptance-rejection as described by Phoon [286] or Fenton & Griffiths
[127] amongst others.

4.2.3 Generation of correlated random numbers

Widely used random number generators are optimized for producing sequences of num-
bers which appear to be uncorrelated. Hence the simulation of correlated random vari-
ables requires suitable transformations. The details of the transformation depend on
the joint probability density function of these variables. If the joint density function is
based on the Nataf model then the marginal density function fXi (xi ), i = 1, . . . n and the
correlation coefficients ρij , i, j = 1, . . . n have to be known. The simulation can then be
performed in a loop for k = 1, . . . m using these steps:

• Generate one sample u(k) of a vector of n uncorrelated standardized Gaussian ran-

dom variables.

• Transform the sample into correlated standard Gaussian space by means of

v(k) = L u(k) (4.9)

Here, L is the lower triangular matrix, which results from a Cholesky decomposi-
[email protected]
tion of the matrix of correlation coefficients ρ.

79
Chapter 4 Safety and uncertainties

• Transform each variable separately into non-Gaussian space using

h  i
(k) (k)
xi = FX−1i Φ vi (4.10)

4.2.4 Generation of random fields

Unconditional random fields: Various random field generators exist, which fulfil the
basic assumptions of homogeneity, stationarity, ergodicity and second order stationar-
ity, [82]. Deutsch & Journel [97] categorize random field generators into three differ-
ent categories: frequency domain simulation algorithms, random fractal simulation and
marked point processes. Chiles & Delfiner [79] add here the Markov Chain algorithms.
Deutsch & Journel [97] point out that frequency domain simulation algorithms are fre-
quently used in mining, hydrogeology and petroleum applications. Examples of these
algorithms are the Moving Average method (MA), Fast Fourier Transform (FFT), Turn-
ing band simulation (TB), Local Average Subdivision (LAS), Cholesky decomposition
(LU) and Sequential Gaussian Simulation (SGSIM).
The Sequential Gaussian Simulation method (SGSIM) is the most straightforward al-
gorithm for generating realizations of a multivariate Gaussian field in a sequential way
according to [79, 97], which is explained in detail in appendix E.

Conditional random fields: All the above mentioned approaches simulate uncondi-
tional random fields, which are not taking data points into account from e.g. measure-
ments, pre-knowledge, etc. Therefore, unconditional random fields are not spatially
consistent in the presence of data points. By studying the explanation of the SGSIM
in appendix E, one can derive the scheme of the direct conditional simulation within this
sequential scheme. In figures E.4 and E.5 in appendix E an overview of the main charac-
teristics of various random field simulators are shown.
In addition to this, there are also other approaches like the Kriging approach or simu-
lated annealing, which can be used for conditioning random fields [79, 97].
Li & Der Kiureghian [219] state that many applications in civil engineering call for a
combination of continuum mechanics and representation of uncertain media as random
fields. Therefore, it is necessary to map a random field onto a grid of a Finite Element
mesh. Several methods for discretisation of random fields have been proposed in the past
amongst others by [219, 245, 364, 366] in the framework of stochastic Finite Elements (see
section 4.3.3). These include the midpoint method (MP), the spatial averaging method
(SA), weighted integral method, the shape function method (SF) and the series expan-
sion method (SE), the Karhunen Loeve expansion (KL), the orthogonal series expansion
and the expansion optimal linear exstimation method (EOLE).
The simplest method of discretisation of a random field within the domain Ω is the
midpoint method. In this method, the field within the domain Ωe of an element is de-
scribed by a single random variable representing the value of the field at a central point
[email protected]
of the element, e.g. the centroid xc , as shown in figure 4.6. The field value within the

80
 4.3 Computation of failure probabilities

element is assumed to be a constant i.e.

vb(x) = v(xc ) ; x ∈ Ωe (4.11)

A realization of the field so defined is a stepwise function within discontinuities along

the element boundaries.
The spatial averaging method proposed by Vanmarcke & Grigoriu [387] describes the
field within each element in terms of the spatial average of the field over the element.
R
v(x) dΩ
vb(x) = ΩeR = vbe ; x ∈ Ωe (4.12)
Ωe
dΩ

The average values vbe now form the vector v. A realization of the field so defined is also
a stepwise function with discontinuities along the element boundaries. Li & Kiureghian
[219] state that the variance of the spatial average variable over an element is smaller
than the local variance of the random field, in general. Moreover, this method can map
a random field on structured as well as on non-structured finite element meshes.
The other above mentioned discretisation methods offer a continuous function of the
discretized field, but ask for a more complicated mathematical background as explained
in detail in literature [219, 245, 364, 366].

4.3 Computation of failure probabilities

Due to the vast amount of different methods available in this discipline existing proba-
bilistic methods can be categorised in various ways from different point of views . A pos-
sible classification of probabilistic methods for uncertainty quantification is portrayed in
figure 4.7, which includes the findings from different publications [57, 176, 191, 262].
Herein, it is differentiated between non-probabilistic and probabilistic methods. The
non-probabilistic approaches include interval analysis [258, 298], fuzzy approaches [107,

xi

We
xc

x
[email protected]
Figure 4.6: Random field discretisation after Li & Kiureghian [219].

81
Chapter 4 Safety and uncertainties

278], grey number theory [296], imprecise probability method based on p-box repre-
sentation [128, 415] and random set approaches [279, 290, 332], which are summarized
briefly in [262]. The probabilistic approaches aim for a computation of the probability of
failure, which is faster than the computationally time consuming Monte Carlo (MC) sam-
pling approach. Here it should be realised that each alternative to the robust MC method
implies some loss of accuracy. Therefore, the MC approach is used for verification and
calibration of these approaches. The Bayesian approach in uncertainty quantification is
described in various publications [23, 416] as well as the standard reliability methods
(e.g. FOSM, FORM, SORM) in [57, 326] and iterative random point sampling methods
in [57, 326, 327]. Prefixed point sampling methods like Taylor series, finite difference
methods or the Point Estimate method can be found in recent publications [13, 374].
Fenton has worked in various publications [127] and different applications in geotech-
nical engineering on the simulation of spatial variability using random fields within the

non-deterministic approaches

probabilistic methods non-probabilistic methods

interval analysis
Bayesian approach Standard reliability
methods
Random Finite fuzzy apporaches
Element Method FOSM
Iterative random grey number
Stochastic Finite FORM point sampling theory
Element Method
SORM directional samling,
Hasover-Lind method; imprecise
response surface methods adaptive sampling, probability
MCMC
random set
interpolation regression prefixed point approach
methods methods
sampling methods
Kriging
radial basis Taylor series finite
polynomial chaos
funciton difference method
expansion method

Support vector machine

point estimate
artificial neural network
method

Figure 4.7: Non-deterministic approaches for uncertainty quantification modified from

[email protected]
[57, 176, 191, 262].

82
 4.3 Computation of failure probabilities

Gaussian space
xj’

xc’
pdf
aj’
ac’

friction angle j’ [°]

(xc’*|xj’*) b

xj’ mj’
g>0

tan
(c’*|j’*)

ge
g>0 g<0

nt
g<0
b g<0
xc’ mc’
g<0 cohesion c’ [kN/m²]

Figure 4.8: First order reliability method in Honjo [168] and Sudret [364].

the Random Finite Element Method.

Within this section, the principles of the First Order Reliability Method and Monte
Carlo methods, together with the basics of the stochastic finite element method and the
response surface method, are explained in order to help the reader to understand in
depth the results of the next chapters.

4.3.1 First Order Reliability Method

The First Order Reliability Method (FORM) [152] is based on a description of the reli-
ability problem in standard Gaussian space N (µ = 0, σ = 1). In figure 4.8 the basics
of FORM are visualized as presented in Huber et al. [177]. The two random variables
Z = [c′ , ϕ′ ] are transformed by the q-q transformation from the physical space into Gaus-
sian variables ξc′ and ξϕ′ with N (µ = 0, σ = 1)

Yi = Φ−1 [FZi (Zi )]; i = 1 · · · n (4.13)

In the case of correlated variables Y the transformation from correlated Gaussian space
to standard Gaussian space can be done by means of

Ξ = L−1 Y (4.14)

in which L is calculated from the Cholesky-decomposition of CZZ .

CZZ = L LT (4.15)

Then a linearisation of the limit state function is performed in Gaussian space (ξc′ , ξϕ′ ).
The expansion point ξ ∗ is chosen so as to maximize the pdf within the failure domain Df .
Geometrically, this coincides with the point in the failure domain, having the minimum
[email protected]
distance β from the origin. From a safety engineering point of view, the point x∗ =

83
Chapter 4 Safety and uncertainties

[c′∗ , ϕ′∗ ] corresponding to ξ ∗ = [ξc′∗ , ξϕ′∗ ] is called the design point or most probable point
(MPP).
From the geometrical interpretation of the expansion point ξ ∗ in standard Gaussian
space it becomes quite clear that the calculation of the design point can be reduced to an
optimization problem.
 
∗ 1 T
ξ = argmin ξ ξ ; subjected to: g [z(ξ)] = 0 (4.16)
2

This leads to the Lagrange-function

1
L = ξ T ξ + λ g(ξ) → Min. (4.17)
2
Standard optimization procedures can be utilized to solve for the location of ξ like
e.g. Rackwitz-Fiessler algorithm, particle swarm algorithm amongst other methods de-
scribed in e.g. [57].
The equation of this hyperplane may be cast as:

β−αξ =0 (4.18)

where the unit vector α = ξ/β is also normal to the limit state surface at the design point
ξ∗ :
∇g ( T −1 (ξ ∗ ) )
α=− (4.19)
k∇g ( T −1 (ξ ∗ ) )k
Herein, T −1 is the q-q transformation from the Gaussian space into the physical space
shown in figure 4.8. The vector α describes the contributions of the random variables ξi
to the probability of failure pf .
A linear approximation of the failure surface at the design point will be accurate if
the failure function is linear or weakly non-linear (relatively flat). For heavily non-linear
failure functions, the FORM methods may not always be adequate to find a reasonably
correct failure probability. In such cases, a better approximation of the failure surface at
the design point is required. For this purpose, a second order (parabolic) failure surface
is fitted to the non-linear failure function at the design point [53, 93, 131, 376] to give
the Second Order Reliability Methods (SORM). It is a relatively complicated process and
computationally more time consuming as well. A detailed description of this method
can be found in Bucher [57] amongst others [53, 131, 191, 286, 366].

4.3.2 Monte Carlo simulation

The definition of the failure probability pf as given in equation 4.8 can be written as an
expected value, in which Ig (x1 . . . xn ) = 1 if g(x1 . . . xn ) ≤ 1 and Ig (·) else.
Z ∞Z ∞ Z ∞
pf =
[email protected] ... Ig (z1 . . . zn ) fz1 ...zn (z1 . . . zn ) dz1 . . . dzn (4.20)
−∞ −∞ −∞

84
 4.3 Computation of failure probabilities

Monte Carlo Method: In order to determine pf in principle all available statistical

methods for estimation of expected values are applicable. If m independent samples
z(k) of the random vector Z are available then the estimator
m
1 X

pf = Ig x(k) (4.21)
m k=1

yields a consistent and unbiased estimate for pf .

The problem associated with this approach is this: For small values of pf and small
values of m the confidence of the estimate is very low. The variance σp2f of the estimate
pf can be determined from
r
2 pf p2f pf pf
σp f = − ≈ → σp f = (4.22)
m m m m
The required number m of simulations is independent of the dimension n of the problem.

Importance sampling: Bucher [57] states that in order to reduce the standard deviation
σp2f of the estimator to the order of magnitude of the probability of failure itself m must
be in the range of m = 1/pf . For values of pf in the range of 10−6 this cannot be achieved
if each evaluation of the limit state function requires a complex analysis. Alternatively,
strategies are employed, which increase the "hit-rate" by artificially producing more sam-
ples in the failure domain than should occur according to the distribution functions. One
way to approach this solution is the introduction of a positive weighting function hY (z)
which can be interpreted as a density function of a random vector Y. Samples are taken
according to hY (z). The probability of failure is then estimated from
m  
1 X fY (z) fY (x)
pf = Ig (z) = E Ig (z) (4.23)
m k=1 hY (z) hY (z)

From the estimation procedure it can be seen that the variance of the estimator pf be-
comes  
2 1 fY (x)2
σp f = E Ig (x) (4.24)
m hY (x)2
A useful choice of hY should be based on minimizing σp2f . Ideally, the weighting function
should reduce the sampling error to zero. However, this cannot be achieved in reality
since such a function must have the property
(
1
f (z) for g(z) ≤ 0,
pf Z
hY (z) = (4.25)
0 for g(z) > 0

This property requires the knowledge of pf , which - of course - is unknown. Special

updating procedures such as adaptive sampling [57] can help to alleviate this problem.
Bucher [57] recommends to use the importance sampling concept in conjunction with
[email protected]
the FORM approach. Based on the previous FORM analysis, it may be attempted to

85
Chapter 4 Safety and uncertainties

obtain a general importance sampling concept. The efficiency of this concept depends
on the geometrical shape of the limit state function. In particular, limit state functions
with high curvatures or almost circular shapes cannot be covered very well.
Apart form importance sampling there are several other techniques like line sampling,
directional sampling, adaptive sampling, asymptotic sampling, line sampling or subset
simulation, which are explained in detail in several sources [57, 326, 364].

4.3.3 Stochastic Finite Element Method

The Stochastic Finite Element Method (SFEM) was introduced by Ghanem & Spanos
[138] and is an extension of the classical deterministic approach of the solution of stochas-
tic (static and dynamic) problems [326, 358].
Basically spoken, SFEM involves finite elements whose properties and boundary con-
ditions are random. From a mathematical point of view, SFEM is a powerful tool for
the solution of stochastic partial differential equations (SPDEs) and it has been treated
as such in numerous publications e.g. [57, 138, 286, 326, 358, 364, 366]. It has been
successfully applied to a wide variety of problems (e.g. solid, structural and fluid me-
chanics, heat transfer, geotechnical engineering) as described in Stefanou [358], in which
the stochastic medium is represented by random fields if appropriate. The result of the
SFEM is a so called polynomial chaos expansion, which represents the stochastic system
in a simplified way; in other disciplines this can be related with synonyms like surrogate,
meta-models or response surfaces.

Intrusive SFEM: Within SFEM, different authors [286, 364, 366] distinguish between
the so called intrusive and the non-intrusive approach.
Within the intrusive SFEM, named after the pioneering work of Ghanem & Spanos
[138], the aim is to represent the complete response PDF in an intrinsic way. The imple-
mentation of the intrusive SFEM has to be carried out for each class of problem. Herein,
the stiffness matrix as well as the boundary conditions consist of a mean (deterministic)
part and of stochastic parts, which can be solved by using various methods such as the
"weighted integral method", the "Neumann series expansion" method or the "Taylor se-
ries expansion" method, as described in detail in [245, 286, 366]. The response of the sys-
tem (which, after proper discretisation of the problem, is a random vector of unknown
joint probability density function) is expanded onto a particular basis of the space of
random vectors of finite variance called "polynomial chaos".
There are two main variants of SFEM in the literature: i) the perturbation approach,
which is based on a Taylor series expansion of the response vector [203] and ii) the spec-
tral stochastic finite element method (SSFEM) [138], where each response quantity is
represented using a series of random Hermite polynomials. A detailed description can
be found in [336, 358, 366]. Amongst others, Sett & Jeremic [336] also applied the SSFEM
framework to highly non-linear and dynamical geomechanical problems.
Sudret & Der Kiureghian [367] state as an overall conclusion that SSFEM has limited
applicability to reliability problems involving small failure probabilities. The polynomial
[email protected]
chaos expansion provides a global fit to each response quantity, which may be good in

86
 4.3 Computation of failure probabilities

the central region of the respective distribution, but poor in the tail regions. Since small
probability events are influenced by the tail regions of these probability distributions,
accurate results from SSFEM cannot be expected for such problems. This limitation is
more severe for problems involving random fields with short correlation lengths or large
coefficients of variation.

Non-intrusive SFEM: In order to apply this approach in a similar way to a wider range
of mechanical problems, the so called non-intrusive SFEM has been developed [286, 364].
Within the non-intrusive SFEM, the scalar response quantities S of the system e.g.
nodal displacements, strain or stress components are directly expanded onto the poly-
nomial chaos, which is truncated after the P term.
∞
X P
X −1
S = h(X) = Sj Ψj = Sj Ψj (4.26)
j=0 j=0

The big advantage of this approach is the fact that - basically - it is just the post-processing
of simulation results. Therefore, any FEM-code can be used to calculate the system re-
sponse in contrast to the intrusive SFEM [286, 364].
Sudret [364] proposes two methods to compute the coefficients in this expansion from
a series of deterministic finite element analyses, namely the projection method and the
regression method.
The PROJECTION METHOD is based on the orthogonality of the polynomial chaos [3,
286]. By premulitplying equation 4.26 with Ψi and taking the expectation of both mem-
bers, it becomes: "∞ #
X
E [S Ψ] ≈ E Sj Ψi Ψj (4.27)
j=0

Due to the orthogonality of the basis E [Ψi Ψj ] for any i 6= j, one can reformulate the
following equation:
E [S Ψj ]
S =  2 (4.28)
E Ψj
In equation 4.28 the denominator is known analytically, as derived in Appendix D, and
the numerator may be cast as a multi-dimensional integral:
Z
E [SΨj ] = h( X(ξ) ) Ψj (ξ) ϕm (ξ) dξ (4.29)
RM

where ϕm is the M -dimensional mulit-normal PDF, and where the dependency of S in ξ

through the iso-probabilistic transform of the input parameters X(ξ) has been given for
the sake of clarity. This integral can be computed by crude Monte Carlo simulation [57,
286]. However, the number of samples required in this case should be large enough to
obtain a sufficient accuracy. Sudret [364] states that in the case of using a computationally
demanding model for the evaluation of the system response, this approach is practically
[email protected]
not applicable. Alternatively, the Gaussian quadrature scheme can be used to evaluate

87
Chapter 4 Safety and uncertainties

the integral [43, 244] as a weighted summation of the integrand evaluated at selected
points (the so-called integration points).
The Smolyak sparse grids are also quoted in literature [3, 286] as a promising approach
for computing the integral of equation 4.29.
The REGRESSION METHOD is another approach for computing the response expansion
coefficients. It is the regression of the exact solution S with respect to the polynomial
chaos basis Ψi (ξ). The scalar response quantity S consists of a residual ǫ (a zero mean
random variable) and unknown coefficients S. e
p−1
X
S = h(X) = Sj Ψj + ǫ (4.30)
j=0

The minimization of the variance of the residual with respect to the unknown coefficients
leads to equation 4.31 by using a set of Q regression points in the standard normal space
ξ i and their isoprobabilistic transform xi
Q
( p−1
)2
1 X X
e = Argmin
S i
h(x ) − i
Sj Ψj (ξ ) (4.31)
Q i=1 j=0

Sudret [364] solves this minimization problem in the following way: Denoting by Ψ the
matrix whose coefficients are given by Ψij = Ψj (ξ i ), i = 1, . . . , Q; j = 0, . . . , p − 1 and
by Sex the vector containing the exact response valuse computed by the model Sex =
h(xi ), i = 1, . . . , Q, the solution to equation 4.31 reads:

−1 T
S = ΨT Ψ Ψ Sex (4.32)

This approach is comparable to the so called response surface method used in many
domains of natural sciences and engineering. Within this context, the set of x1 , . . . , xQ
is the so called experimental design. In equation 4.32, ΨT · Ψ is the information matrix.
Sudret [364] shows that an efficient design can be built from the roots of the Hermite
polynomials as follows:
• If p denotes the maximal degree of the polynomials in the truncated PC expansion,
then the p + 1 roots of the Hermite polynomial of degree p + 1 (denoted by Hep+1 )
are computed, say r1 , ..., rp+1 .
• From this set, M -tuplets are built using all possible combinations of the roots: rk =
(ri1 , . . . , riM ), 1 ≤ i1 ≤ . . . ≤ iM ≤ p + 1, k = 1, . . . , (p + 1)M .
• The Q points in the experimental design ξ1 , . . . , ξQ are selected among the rj by re-
taining those which are closest to the origin of the space, i.e. those with the smallest
norm, or equivalently those leading to the largest values of the PDF ϕM (ξ j ).
To choose the size of Q of the experimental design, the following empirical rule was
proposed by Berveiller et al. [44] based on a large number of numerical experiments.
[email protected] Q = (M − 1) P (4.33)

88
 4.3 Computation of failure probabilities

Herein, P is the number of unknown coefficients defined by the following equation com-
bining the PCE order M and the degree of the Hermite polynomial p.
 
M +p (M + p)!
P = = (4.34)
p M ! p!

Representation of the response PDF: Once the coefficients S of the PC expansion of a

response quantity are computed, the polynomial approximation can be simulated using
Monte Carlo simulation as shown in [286]. A sample of standard normal random vectors
ξ (1) . . . . , ξ (n) is generated.
Then the PDF can be plotted using a histogram representation [286, 366]. From equa-
tion 4.26 the mean and the variance σS2 , the skewness δS and kurtosis κS of the approxi-
mated response S are given by:
E[S] = S0 (4.35)
P
X −1
 
σS2 ≡ Var[S] = Sj2 E Ψ2j (4.36)
j=1

P −1 P −1 P −1
1  3
 1 XXX
δS ≡ 3 E (S − E[S]) = 3 E[Ψi Ψj Ψk ] Si Sj Sk (4.37)
σS σS i=1 j=1 k=1
P −1 P −1 P −1 P −1
1  4
 1 XXXX
κS ≡ E (S − E[S]) = E[Ψi Ψj Ψk Ψl ] Si Sj Sk Sl (4.38)
σS4 σS3 4 i=1 j=1 k=1 l=1

where the expectation of Ψ2j is given in Appendix D. Various authors [286, 364, 366] de-
scribe that the moments of higher order are obtained in a similar manner.

Reliability analysis of the SFEM results: This metamodel can be used to construct the
response PDF of the system as well as to compute the reliability of the observed sys-
tem, which is approximated by a polynomial chaos expansion. Surprisingly, this link
between structural reliability and the non-intrusive SFEM based on PC expansions is
relatively new [43, 364, 366]. The PC expansion can be used as a meta-model within the
framework of non deterministic approaches for uncertainty quantification, which offers
the engineer a first insight into the reliability of a system. One has to keep in mind that
the non-intrusive SFEM also has limited accuracy in evaluating small probabilities of
failure due to the approximation via PC expansions.

Assessment of the polynomial chaos approximation: It has been shown in the pre-
vious section that polynomial chaos (PC) approximations of the mathematical model
can be obtained using non-intrusive techniques, namely the projection approach or the
regression approach. Both methods provide a stochastic response surface whose perfor-
[email protected]
mance has to be assessed [47, 365]. Blatman & Sudret [47] point out that, in terms of

89
Chapter 4 Safety and uncertainties

statistical learning theory (see e.g. [388] ), the discrepancy between the model response
and the metamodel is measured by means of a risk functional, for instance the commonly
used mean-square error. Such a quantity depends on the PDF of the response, which is
unknown in our context.
The generalized error is defined in equation 4.39:
h i Z  2
 2
I MX,p
b = E (M(x) − Mxb,p ) = M(x) − MX,pb fx (x) dx (4.39)


Xb = x(1) , . . . , x(N ) T is the experimental design, the corresponding model evaluations



are Yb = y (1) = M x(1) , . . . , y (N ) = M x(N ) T and M b are the resulting PC approx-
X,p
imations. The notion of generalization error is a basic concept of statistical learning the-
ory as presented in [388]. Computing I[MX,p b ] requires a perfect knowledge of the model
function M, which is not the case in the context of geotechnical engineering in general
since the model is usually not analytical but numerical.
In literature [47, 388] it is proposed to compute the following empirical error or training
error in order to estimate equation 4.39.

h i 1 X
N 2
(i) (i)
IX MX,p
b = M(x ) − M b
X,p (x ) (4.40)
N i=1

Of common use is the related determination coefficient R2 which reads

h i
h i IX MX,p b
R2 MX,pb =1− (4.41)
Vd ar[Y ]

where !2
N N
d 1 X (i) 1 X (i)
V ar[Y ] = M(x ) − y
N − 1 i=1 N i=1
However, the use of R2 statistics might be misleading for comparing two different re-
gression base meta models since it automatically increases with the number of P basis
polynomials; furthermore, it is highly biased since it tends to R2 = 1 as P increases.
Blatman & Sudret [47] report that R2 generally underestimates the generalization error.
Therefore, the adjusted determination coefficient Radj
2
is recommended.
h i N −1  h i
2 2
Radj MX,p
b =1− 1 − R MX,p
b (4.42)
N −P −1
The Radj
2
statistic is penalized as P increases. Baltman & Sudret [47] report that Radj
2
still
often overpredicts the true approximation accuracy.
The cross-validation technique consists of dividing the data sample into two subsam-
ples. A metamodel is built from one subsample, i.e.the training set, and its performance
is assessed by comparing its predictions to the other subset, i.e.the test
 set. Let MX\i
[email protected]
be the metamodel that has been built from the experimental design X\ x(i) , e.g when

90
 4.3 Computation of failure probabilities

removing the ith observation from the training set X. The predicted residual is defined
as the difference between the model evaluation at x(i) and its prediction based on MX\i



∆(i) = M x(i) − MX\i x(i) (4.43)

The generalization error is then estimated by the mean predicted residual sum of squares
(PRESS), i.e. the following empirical mean square predicted residual, [47].

h i N
1 X (i)
2
∗
IX MX,p
b = ∆ (4.44)
N i=1

The corresponding determination coefficient with analogy to its empirical counterpart R2

is denoted by Q2 , [47]:
h i
h i ∗
IX MX,pb
Q2 MX,p b =1− P  PN (i) 2 (4.45)
1 (i) 1
N −1 =1 N M(x ) − N i=1 y

Blatman & Sudret [47] derive IX

∗
in equation 4.46 where hi is the ith diagonal term of the

−1
matrix Ψ Ψ Ψ
T
Ψ :
T

h i N

!2
1 X M x(i) − Mxb,p x(i)
Ix∗ MX,p
b = (4.46)
N i=1 1 − hi

The above presented metrics for assessing the accuracy of the PC approximation are also
quoted amongst other means by Field & Grigoriu [130].

4.3.4 Response surface methods

According to Bucher [57], response surface models are more or less simple mathemat-
ical models, which are designed to describe the possible experimental outcome (e.g.,
the structural response in terms of displacements, stresses, etc.) of a more or less com-
plex structural system as a function of variable factors (e.g., loads or system conditions).
Obviously, the chosen response surface model should give the best possible fit to any
collected data. In general, we can distinguish two different types of response surface
models:

• regression models (e.g. polynomials of varying degree or non-linear functions such

as exponentials or Hermite polynomials) [57, 286],

• interpolation models (e.g. polyhedra, radial basis functions) [57],

• artificial neural networks, support vector machines [57, 58] and

[email protected]
• Kriging and radial basis functions [57, 339].

91
Chapter 4 Safety and uncertainties

In most applications it is quite likely that the exact response function will not be known.
Therefore, it has to be replaced by a sufficiently versatile function, which will express
the relation between the response and the input variables satisfactorily.
Depending on the selected response surface model, support points have to be chosen
to estimate the unknown parameters of the response surface in a sufficient way. A set
of samples of the basic variables is generated for this purpose. In general, this is done
by applying predefined schemes, so called DESIGN OF EXPERIMENTS. Bucher [57] rec-
ommends that it is most helpful to setup an experimental scheme in a dimensionless
space. Within this, Bucher [57] describes saturated designs, which provide a number of
support points that just suffice to represent a certain class of response functions exactly,
and redundant designs, which provide more support points than required to define the
response surface.

[email protected]

92
 4.4 Sensitivity analysis

4.4 Sensitivity analysis

In chapter 5 and 6 the sensitivities analyses are used to quantify the relative importance
of relative importance of each input parameter of system, as described in [286, 317, 364].
Methods of sensitivity analysis are usually classified into two categories:

• L OCAL SENSITIVITY ANALYSIS investigates the local impact of input parameters on

the model. Sudret [364] points out that local SA is based on the computation of the
gradient of the response with respect to its parameters around a nominal value.

• G LOBAL SENSITIVITY ANALYSIS aims for the quantification of the output uncer-
tainty due to the uncertainty in the input parameters, which are taken singly or in
combination with others, [364].

Saltelli et al. [317] group the different techniques in SA into regression-based methods
and variance-based methods. Within the regression-based methods, the standardized re-
gression coefficients, Pearson correlation coefficients, partial correlation coefficients, and
standardized partial rank correlation coefficients are used to describe the correlation be-
tween inut and output. Sudret [364] points out that in the case of general non linear non
monotonic models, these approaches fail to produce satisfactory sensitivity measures.
The variance-based methods aim at decomposing the variance of the output as a sum
of the contributions of each input variables or combinations thereof. They are sometimes
called ANOVA techniques for "ANalysis Of VAriance", [364]. The correlation ratios in
McKay [249], the Fourier amplitude sensitivity test indices [319] and the Sobol indices
[318, 347] enter this category.
For the sake of completeness, an extensive overview on additional methods of local
and global sensitivity approaches can be found in Cacuci et al. [63] including screening
methods, non-parametric methods, variance based methods and density based methods.

4.4.1 Local sensitivity

As shown in section 4.3.1, FORM leads to the computation of a linearised limit state
function whose equation may be cast as:

gFORM (ξ) = β − α · ξ (4.47)

Herein, β is the reliability index and α is the unit vector to the design point. Sudret [364]
considers this linearised limit state function to be a margin function, which quantifies the
distance between a realization of the transformed input random vector and the failure
surface. Its variance straightforwardly reads:
M
X
Var [gFORM (ξ)] = αi2 = 1 (4.48)
i=1

Thus, the coefficients {α2 , i = 1 . . . , M }, which are also called FORM importance factors
i
[email protected]
by Ditlevsen & Madsen [105], correspond to the portion of the variance of the linearised

93
Chapter 4 Safety and uncertainties

margin, which is due to each ξi . When the input random variables X are independent,
there is a one-to-one mapping between Xi and ξi , i = 1 · · · M . Thus, αi2 is interpreted
as the importance of the i-th input parameter in the failure event, [364]. When the in-
put random variables are correlated, other measures of importance should be used as
explained in Haukaas & der Kiureghian [153].

4.4.2 Global sensitivity

The computation of Sobol’ indices is traditionally carried out by Monte Carlo simulation
as reported by [317, 318], which may be computationally unaffordable in the case of time
consuming models. In the context of non-intrusive SFEM, Sudret [364, 365] has shown
that Sobol’ indices can be derived analytically from the coefficients of the polynomial
chaos expansion of the response S, once the lattes have been compute by the projection
or regression approach. For instance, the first order Sobol’ indices, which quantify what
fraction of the response variance is due to each input variable :

VarXi [ E [S|Xi ] ]
δi = (4.49)
Var [S]

can be evaluated form the coefficients of the PC expansions in equation 4.26 as follows:
X
δiP C = Sα2 E [Ψα ] /σS2 (4.50)
α∈Ii

Herein, σS2 is the variance of the model response computed form the PC coefficients in
equation 4.36 and the summation set

Ii = {α : αi > 0 , αj6=i = 0} (4.51)

Higher order Sobol’ indices, which correspond to interactions of the input parameters,
can also be computed using this approach as described in Sudret [365] in detail.
By virtue of the knowledge of SA, engineers can rank the input variables by the
amount of their contributions to the output, and thus take measures accordingly to im-
prove the performance of the model, which is a core task in engineering.

4.5 Synopsis
This chapter summarises the concepts of safety and uncertainty. Besides this, the basics
of uncertainty quantification are explained in detail: The mechanical system is repre-
sented via a limit state function and its variables are represented via random variables
and/or random fields. Different methods to compute the failure probability of the me-
chanical system are discussed and followed up by the description of sensitivity analyses.
Global and local sensitivity analyses quantify the importance of each input parameter
within the scheme of uncertainty quantification.
[email protected]

94
 Chapter 5

Selected case studies in uncertainty

quantification using random properties

5.1 Introduction

Possibilities and limitations of probabilistic methods in civil engineering have been stud-
ied by Elishakoff [116]. Within a survey Elishakoff asked engineers and scientists on the
applicability of probabilistic methods in their fields all over the world. He concluded
that the engineering community can be divided into enthusiastic supporters of proba-
bilistic approaches in engineering and sceptic opponents. The criticism of the opponents
is mainly based on the complex mathematical and conceptual theory, which is necessary
to apply probabilistic design in engineering .
On top of this, it is stated in various other publications [23, 168, 286] that the broader
geotechnical engineering community is unfamiliar with reliability methods, particularly
pertaining to computational details, practical usefulness, pitfalls, and probabilistic char-
acterization of input parameters. In addition to this, the British Health and Safety Execu-
tive authority published the results of a survey on the application of reliability methods
for offshore engineering in a technical report [410]. Many of the respondents thought
that probabilistic techniques would be a welcome addition to their methods of analysis,
but that the approach was yet to be widely applied. Besides this, it is pointed out that
industry has concerns with the relatively small numbers of qualified and experienced
companies and individuals who can carry out designs and, more importantly, audit and
verify these designs.
Therefore, it is necessary to set up case studies to broaden and disseminate reliability
applications beyond the researcher and enlighten the background of probabilistic design
approaches as proposed by Phoon [286] amongst others. The key objectives of such
case studies are education by examples, demonstration of usefulness of reliability based
design in practical applications, development of user-friendly tools and evaluation of
reliability methods as proposed by the Geotechnical Safety Network (GeoSNet) [360].
Within this chapter, typical geotechnical case studies dealing with tunnelling and foot-
ing analysis are investigated within the framework of uncertainty quantification. Start-
ing from semi-analytical limit state equations (LSE), the reader is introduced to more
[email protected]
complex LSE, which are derived from two and three dimensional FEM models.

95
Chapter 5 Selected case studies in uncertainty quantification using random properties

5.2 Tunnel lining

According to prognoses of the European Commission, the growth in traffic between
Member States is expected to double by 2020. To meet the challenges connected with the
increased requirements for efficient traffic infrastructure the use of underground space
often constitutes an efficient and environmentally friendly solution. But there can be
also one essential disadvantage in building tunnels because, especially in an urban en-
vironment, large settlements due to tunnelling can cause tremendous consequences as
reported in [174, 178]. In the last couple of years years innovative tender methods ask
for the consideration of uncertainties in order to contribute usefully to risk management
in tunnelling.
The use of global safety factors offers one way to face this problem of dealing with un-
certainties, which is refined by partial safety factors later on. However, both approaches
miss a mathematical framework, which precisely allocates safety margins to stress and
resistance shares of the mechanical system.
This case study introduces the framework of uncertainty quantification by using an-
alytical formulas of a typical two dimensional problem. This includes a description of
the mechanical problem and of the stochastic variables, as well as a sensitivity study
followed up by an interpretation of the results.

5.2.1 Mechanical description of the problem

A key point in tunnel design is the design of the lining as reported in well known and
recent publications [108, 173, 177]. Möller [253] points out that the different phases of
the tunnel construction, namely excavation and support phases, have to be taken into
account to simulate the internal forces in the lining accurately. The earlier and simplified
way of Ahrens et al. [5] and Erdmann [120] is used to estimate the internal forces in the
tunnel lining. These analytical solutions of a two dimensional problem offer an insight
into the complex interaction between the soil and tunnel. Moreover, these equations
have been used to validate complex three dimensional numerical approaches by other
authors e.g. [253, 278, 332–334, 374].
Figure 5.1 presents an analytical continuum model, which was solved by Ahrens et
al. [5] and later adapted by Erdmann [120]. This analytical solution is based on the
following simplifications and assumptions: The tunnel lining is circular with radius R.
The tunnel longitudinal axis is parallel to the ground surface, inducing plane strain con-
ditions. The thickness of the lining d is constant. The lining is deforming without any
side contraction, i.e. Poisson’s ratio νconcrete is equal to zero. It is assumed that there is a
uniform initial stress field, with σh = K0 · σv , where K0 is the coefficient of lateral earth
pressure and σv = γsoil · H. Here γsoil is the soil unit weight and H is the tunnel depth as
indicated in Figure 5.1.
The lining is installed before tunnel excavation and is assumed to be rough with full
bonding. Both the lining and ground behave linear-elastically. Within this approach,
second order theory is neglected. Erdmann [120] supplemented the findings of Ahrens
[email protected]
et al. [5] to obtain equations (5.1) - (5.2) for a lining with full bonding, where ν is the Pois-

96
 5.2 Tunnel lining

sv sv

sh sh H+R

Econcrete , d

q R
Esoil
gsoil
n
K0
sh sh

sv sv

Figure 5.1: System of tunnel lining from Erdmann [120].

son’s ratio of the soil, and Econcrete F is the normal stiffness and Econcrete I is the flexural
rigidity of the lining respectively. Esoil is the elasticity modulus of the soil.
 
1 + K0 β β 1 − K0
N = γsoil H R/ 1 + + + γsoil H R n2 cos (2θ) (5.1)
2 1+ν α 2
1 − K0 2
M = γsoil H R m2 cos (2θ) (5.2)
2
Esoil R3 Esoil R
α= β= (5.3)
Econcrete I Econcrete F
α β
1 + 12(1+ν) + 4(1+ν)
n2 = α(3−2ν) β(5−6ν) αβ
(5.4)
1 + 12(3−4ν)(1+ν) + 4(3−4ν)(1+ν) + 12(3−4ν)(1+ν) 2

β
1+ 2(1+ν)
m2 = α(3−2ν) β(5−6ν) αβ
(5.5)
2+ 6(3−4ν)(1+ν)
+ 2(3−4ν)(1+ν)
+ 6(3−4ν)(1+ν)2

The system can be described with the limit state equation 5.6. This limit state equation
combines the approaches of Erdmann [120] and Sudret [364]. Sudret [364] considers
the interaction of the internal moment and normal force (see figure 5.1) via equation
(5.6). Herein, ultimate internal forces are Nult and Mult , which are dependent on the
compressive strength of the concrete fconcrete , as described by Sudret [364].
(d · Nult )2 + Mult
2
g (N, M ) = −1 (5.6)
[email protected] (d · N (θ))2 + M (θ)2

97
Chapter 5 Selected case studies in uncertainty quantification using random properties

Table 5.1: Properties of parametric studies on the tunnel lining.

parametric study 1 parametric study 2

property COV distribution COV distribution
ν = 0.30 15 − 40% lognormal 15 − 40% lognormal
K0 = 2.5 1 − 10% lognormal 1 − 10% lognormal
γsoil = 21 kN/m3 15 − 45% lognormal 15 − 45% lognormal
Esoil = 27 MN/m2 15 − 45% lognormal 15 − 45% lognormal
H = 30 m - deterministic 5% lognormal
R=5m - deterministic 1% lognormal
d = 0.35 m - deterministic 1% lognormal
Econcrete = 1, 500 MN/m2 - deterministic 10% lognormal
fconcrete = 45 MN/m2 - deterministic 6% lognormal

5.2.2 Stochastic variables and parametric studies

On the basis of a literature study [23, 282, 283, 286], the stochastic variables are listed in
table 5.1 for the following case studies.

Parametric study 1: At first, the influence of soil variability is investigated in paramet-

ric study 1. Therefore, the soil stiffness Esoil , Poisson’s ratio ν, the soil unit weight γsoil
and the K0 value are treated as random variables (table 5.1).
A combination of two reliability techniques is used for the uncertainty quantification
of this system. The evaluation of the reliability index β and of the probability of failure pf
is done by the combination of First-Order-Reliability Method (FORM) and Importance
sampling (IS). After the calculation of the design point using FORM, the IS algorithm is
employed to investigate the probability of failure more accurately.
The expected influence of soil variability can be clearly seen in figure 5.2: A higher
degree of soil variability leads to a lower degree of reliability of the system. In figure 5.2
(a), the reliability index β decreases with an increasing level of variability of the random
variables and vice versa for the probability of failure in figure 5.2 (b).
However, on the basis of these results one cannot be sure if the uncertainty of the
system is represented properly.

Parametric study 2: In order to quantify the effects of a fully random system, all vari-
ables are introduced with their random properties (table 5.1) in parametric study 2.
Herein, the coefficients of variation are taken from literature [23, 217, 286, 364, 399].
Looking at the results shown in figure 5.2, one can clearly observe that the most influ-
ential parameters have been modelled by stochastic variables: the results of both para-
metric studies are almost similar, but one cannot quantify the influence of each random
[email protected]
variable to the system behaviour.

98
 5.2 Tunnel lining

(a) 0 (b)
10

probability of failure pf
14
-10
12 10
reliability index b

10 -20
10
8 parametric study 1
-30
10
6 parametric study 2
-40
4 10
20 30 40 COV in % 20 30 40 COV in %
Esoil Esoil

20 30 40 COVg’ in % 20 30 40 COVg’ in %
15 25 35 COVn in % 15 25 35 COVn in %
2 4 6 COVK0 in % 2 4 6 COVK0 in %

Figure 5.2: Influence of the variability in parametric studies 1 and 2 on the tunnel lining.

5.2.3 Sensitivity analysis

A local and global sensitivity analysis has been carried out to enlighten the influence of
each random variable on the system response.
As mentioned in section 4.3.1, the local sensitivity analysis is a by-product of FORM. The
values of the importance vectors is shown in figure 5.4 (left). The negative importance
values αi indicate resistance parameters and vice versa. One can clearly see that the
thickness d and the soil stiffness Esoil , the soil weight γ’, the Poisson’s ratio ν, the over-
burden H and the compressive strength of concrete fconcrete have the largest influence on
the system behaviour.
The global sensitivity analysis is defined as the ratio between the variance of the inves-
tigated variable and variance of the system response as described in section 4.4.2. A
Polynomial chaos expansion (PCE) is used to approximate the system response due to
the uncertain random input parameters in table 5.1 (parametric study 2). The estimation
of PCE-accuracy is shown in figure 5.3 using different approaches. Herein, the Sobol’
indices δiP C are analytically calculated from a PCE of the limit state equation as derived
in section 4.4.2. The PCE approximations become impractical in the presence of a large
number of variables (n > 6) as also reported by Phoon [286]. This is impractical due to
the quite long computation time compared to the local sensitivity analysis. Figure 5.3
shows the accuracy of the approximated PCE, which is measured by different types of
error (see section 4.3.3). It can be clearly seen in figure 5.3 evaluated accuracy of the PCE
is increasing with the expansion order of the PCE. In addition to the empirical error also
the mean predicted residual sum of squares (PRESS) in figure 5.3 (a) offers similar results for
the increasing accuracy of the PCE with an increasing expansion order. It can be clearly
[email protected]
seen in figure 5.3 (b) that the accuracy of the PCE fitting is expressed using the deter-

99
 Chapter 5 Selected case studies in uncertainty quantification using random properties

(a) (b)

determination coefficient
15 19
x 10 x 10
empirical error

2 4 1.000

PRESS
1.5 3
0.995 det. coeff.
1 2
empirical error adj. det. coeff.
0.5 1 0.990
PRESS Q2 coefficient
0 0.985
0 1 2 3 4 1 2 3 4
expansion order M of the PCE expansion order M of the PCE

Figure 5.3: Accuracy estimation of the PCE in for the global sensitivity analysis.

(a) global sensitivity diPC (b)

0.8 0.7
local sensitivity ai

0.6
0.6
0.5
0.4
0.4
0.2
0.3
0.0 0.2
-0.2 0.1
-0.4 0.0
d r EconcreteEsoil g’ n K0 H fc d r EconcreteEsoil g’ n K0 H fc

Figure 5.4: Results of the local (a) and global (b) sensitivity analysis in parametric study
2 of a tunnel lining in a soil with COVϕ’ = 20% and a COVc’ = 10%.

mination coefficient, the adjusted determination coefficient and the Q2 coefficient. One
can derive from this that these measures quantify the PCE-fitting in a similar way: The
expansion order M = 3 can be taken for the evaluation of the global sensitivity factors in
figure 5.4.
By looking at the results of the global and local sensitivity analysis shown in figure
5.4 one can see comparable results for both techniques: Esoil , γsoil , H and fconcrete can
be identified as most influential variables on the system behaviour from the results of
both sensitivity measures. One would also expect the K0 value to play a major role in
this context, but both sensitivity measures are very small. This can be deduced to the
low coefficient of variation in table 5.1. In case of a bigger COV, the sensitivity of the K0
value is expected to increase.
The differences can be deduced to the afore mentioned definitions of these sensitivity
measures. Besides this, it has to be pointed out that negative local sensitivity measures
αi indicate a resistance parameter of the investigated system and vice versa for load
[email protected]
parameters.

100
 5.2 Tunnel lining

5.2.4 Conclusions
This first case study is showing exemplarily the application of the uncertainty quantifi-
cation framework in tunnelling. An analytical solution of a continuum problem in tun-
nelling is adapted to quantify the effects of soil variability as well as the performance of a
local and global sensitivity study of the input parameters. The proposed concept allows
one to investigate effects of soil variability precisely without using design approaches
described in the E UROCODE 7 [329].
The soil behaviour is modelled on basis of a linear elastic, perfectly plastic soil model.
Therefore, the results of this case study cannot be fully transferred into applied engi-
neering. But the presented studies offer a probabilistic description of a soil structure
interaction problem in tunnelling. It can be concluded that the reliability based design
methodology can be applied to any geotechnical problem, which can be described via a
limit state equation. This limit state equation can be a closed form solution or an equa-
tion, which is derived from FEM simulations.
On basis of these results, further investigations on the interaction between soil and
tunnel using numerical methods in combination with advanced constitutive models
would offer additional insights into this problem.

[email protected]

101
Chapter 5 Selected case studies in uncertainty quantification using random properties

5.3 Bearing capacity of vertically loaded strip footings

One of the core competences in foundation engineering is the design of footings. Fenton
& Griffiths [127] state that the design of a foundation involves the consideration of sev-
eral limit states, which can be separated into two groups: serviceability, which generally
translate into a maximum settlement or differential settlement, and ultimate limit states.
The latter are concerned with the maximum load, which can be placed on the footing
just prior to a bearing capacity failure.
This chapter contributes to the uncertainty quantification of strip footings. The effects
of uncertain soil strength properties has been and is a topic of various publications e.g.
Baecher & Christian [23], Breysse [54, 170], Cherubini [76], Phoon [285, 286], Russelli
[312], Subra [351], and Peschl [279] amongst others. Also the effects of spatially corre-
lated soil properties on the bearing capacity of footings has been investigated by Fenton
& Griffiths [127], Popescu et al. [289] and Huber et al. [172] amongst others.
Within this case study, the effects of different constitutive failure criteria are investi-
gated to contribute to the uncertainty of modelling soil behaviour. The presented case
studies show the effects of different constitutive failure criteria and point out the impor-
tance for the bearing capacity problem of strip footings.

5.3.1 Mechanical description of the system

The German standard DIN 4017 [101] is formulating the state-of-the-art and provides
equation 5.7 for the calculation of the bearing capacity of a vertically loaded strip footing.

qf = Nc · c′ + q0 · Nq + γsoil · b · Nb (5.7)
(
2+π for ϕ′ = 0
Nc = (5.8)
(Nq − 1)/ tan ϕ′ for ϕ′ 6= 0
 
π tan ϕ′ 1 + sin ϕ′
Nq = e Nb = 2 (Nq − 1) tan ϕ′ (5.9)
1 − sin ϕ′
Herein, Nc is the cohesion stability number, c′ the cohesion, Nq the depth stability num-
ber, q0 the surface load, ϕ′ is the effective friction angle, Nb the width stability number
and γsoil the soil unit weight.
The stability number Nc was derived analytically by Prandtl [293]. In 1920, Prandtl
succeeded in finding a solution for the problem of a strip load on a half plane that is
both statically and kinematically admissible. The material beneath the strip load can
be subdivided into three zones, as pictures in figure 5.5: (I) a wedge shapes zone, in
which the major principal stresses are vertical, called active Rankine zone, (II) a radial
shear zone, called Prandtl zone and (III) a passive Rankine zone, [394]. Based on Mohr’s
stress theory and using Airy’s stress function, Prandtl obtained an analytical expression
for the ultimate bearing capacity of a weightless soil.
Veruijt [394] reports that Buisman extended Prandtl’s theory by superimposing over-
[email protected]
burden pressure q0 and the unit weight of soil γsoil , which was further extended with

102
 5.3 Bearing capacity of vertically loaded strip footings

experimental and theoretical investigations by Keverling, Caquot, Terzaghi and Brinch-

Hansen amongst others. The theoretical investigations are focusing challenges: the an-
alytical expression of Prandtl’s ultimate bearing capacity is based on the assumption of
a weightless soil, which simplifies the system of differential equations as reported by
Perau [277]. Perau [277] collected different approaches on solving this issue numerically.
The results from limit analysis approach [232], kinematic approaches [398] or the meth-
ods of characteristics [345, 348] are offering a scatter of stability number Nb , which are
compared to the formula proposed in DIN 4017 [101] in figure 5.6.
Potts & Zdravkovic [292] derived the stability numbers from FEM calculations using
a linear elastic, ideal plastic model. The findings of their numerical study shows good
agreement with the results of equations 5.7 and 5.9.
On basis of this, parametric 2D-FEM studies are carried out to evaluate the stabil-
ity numbers Nb and Nc for different constitutive failure criteria namely Mohr-Coulomb

q0 q q0

I III I
II II j’ c’
gsoil = 0 kN/m³

Figure 5.5: Failure mechanism of the classic bearing capacity theory from Prandtl [293].

(a) (b)
Nb / Nb, DIN 4017
200 % q
180 %
b gsoil
160 %
j’ c’
140 %
120 % Weißmann Löffler
Vollweiden
DIN 4017
100 %
TGL Pregl
80 % Steenfelt Smoltczyk
60 %
10 15 20 25 30 35 40 45 50
friction angle j’

Figure 5.6: (a) Variability of the theoretical results of the width stability number Nb from
Perau [277] and
[email protected]
(b) geometry of the rigid strip footing.

103
Chapter 5 Selected case studies in uncertainty quantification using random properties

(MC) [243], Matsuoka-Nakai (MN) [243] and Lade-Duncan (LADE)[213]. Herein, the
contact between the rigid footing and the contact with the soil is assumed as rough with
full bounding. These three constitutive failure criteria are shown in figure 5.7. It can be
seen that the MC, MN and LADE criteria are identical for triaxial compression, whereas
MC and MN are identical for triaxial extension. It is clearly shown in the description in
Appendix F that for the MN and LADE criteria no additional input variables are needed.
The results in figure 5.8 (a) show that the width stability number Nb is slightly influ-
enced by the different constitutive failure criteria; the influence of MC, MN and LADE
on the cohesion stability number Nc can be clearly seen in figure 5.8 (b). It can be con-
cluded that the MC-criteria is a conservative lower bound for the estimation of the ulti-
mate shear strength of soil, as reported by others e.g. Schad [321]. Moreover, it has to be
pointed out that the advanced criteria MN and LADE have a larger friction angle under
plane strain conditions.
The bearing capacity problem can be described with the limit state equation 5.10. This
limit state equation compares the actual footing pressure q with the bearing capacity qf .

g (c, ϕ′ ) = q − qf (5.10)

5.3.2 Stochastic variables and parametric study

Parametric study 1: In this parametric study the effects of three different levels of soil
variability on the probability of failure are investigated. The variability of the friction
angle and of the cohesion are summed up in table 5.2. The reliability of the strip footing
is evaluated using the limit state equation 5.10 and the combination of the First-Order-
Reliability-Method (FORM) and Importance sampling (IS) within the FERUM [49] li-
braries.
Fragility curves are used to study the effects of increasing the variability as shown
in figure 5.9. Amongst others, Schultz et al. [328] describe that fragility curves show
how the reliability of a structure changes over the range of loading conditions to which

s1’

Mohr-Coulomb
Matsuoka-Nakai
Lade-Duncan

s2’ s3’

Figure 5.7: Different failure criteria of Mohr-Coulomb, Matsuoka-Nakai and Lade-

[email protected]
Duncan.

104
 5.3 Bearing capacity of vertically loaded strip footings

(a) (b)

cohesion stability number Nc

width stability number Nb

50 100 Mohr-Coloumb
Matsuoka-Nakai
40 80 Lade-Duncan

30 60

20 40

10 20

0 0
0 10 20 30 40 0 10 20 30 40
effective friction angle j’ effective friction angle j’

Figure 5.8: Stability numbers Nb (a) and Nc (b) for different constitutive failure criteria.

that structure might be exposed. This offers a more general approach for the proba-
bilistic description of a system by incorporating deterministic input parameters into a
probabilistic framework. The deterministic input in this parametric study is the footing
pressure. Furthermore, Ellingwood et al. [117] show that the fragility curves of build-
ings follow a lognormal cumulative distribution function in general. This is a benefit
because the fragility curve represented by a lognormal distribution function is continu-
ously defined over the entire range of loads instead of discrete points, [117]. Therefore,
lognormal cumulative distribution functions are fitted to the calculated failure proba-
bilities as shown in figure 5.9. Furthermore, it can be clearly seen that the larger COV
values of the cohesion and the friction angle cause a higher probability of failure.
In addition to this, the effects of different constitutive failure criteria are shown in

Table 5.2: Stochastic variables of the silty sand in two parametric studies.

Parametric study 1 Parametric study 2

µ COV µ COV
2 2
c ′
10 kN/m 10 - 160 % 10 kN/m 10 - 160 %
ϕ′ 25◦ 5 - 80 % 25◦ 5 - 80 %
ψ 0◦
deterministic 0◦
deterministic
γsoil 20 kN/m3 deterministic 20 kN/m3 20 %
b 5m deterministic 5m 10 %
d 0.8 m deterministic 0.8 m 10 %
50 - 10,000 kN/m2 deterministic
[email protected]
q 1,000 kN/m2 20 - 80 %

105
Chapter 5 Selected case studies in uncertainty quantification using random properties

COVj’ = 5 % COVc’ = 10 %
0.0 1.0
probability of non-failure 1-pf

0.1 0.9
COVj’ = 20 %
probability of failure pf

0.2 0.8
COVc’ = 40 %
0.4 0.7

0.5 0.6 COVj’ = 80 %

COVc’ = 160 %
0.5 0.5

0.6 0.4

0.7 0.3 MC
0.8 0.2 MN
LADE
0.9 0.1

1.0 4
0 1 2 3 4 5 6 7 8 9 x 10
footing pressure q in kN/m²

Figure 5.9: Fragility curves for different constitutive failure criteria and different COVs
of the cohesion c′ and of the friction angel ϕ’.

figure 5.9. The MC, MN and the LADE criteria have a significant influence on the prob-
ability of failure. For all investigated levels of soil variability, the MC criterion offers a
higher probability of failure in comparison to the MN and LADE criteria. From these
fragility curves it can be concluded that the MC criterion is the most conservative one in
this context.

Parametric study 2: The concept of the fragility curves offers a good insight into the
effects of soil variability. But this concept is based on a deterministic footing pressure.
Therefore, this approach is extended in this second parametric study to a fully proba-
bilistic analysis, which considers all input variables as random. Besides this, different
levels of variability of the load q are investigated. The details of theses lognormal distri-
butions are indicated in table 5.2.
It can be clearly deduced from figure 5.10 that the uncertainty and reliability of this
system are related: the higher the degree of uncertainty is, the higher is the probability of
failure. Again, the probability of failure for different levels of variability is evaluated by
a combination of FORM and IS within the FERUM libraries [49]. In the same fashion is
the influence of the variability of the load q. A high load variability is causing an higher
[email protected]
probability of failure. This effect is more evident for low levels of soil strength properties

106
 5.3 Bearing capacity of vertically loaded strip footings

(a) 0 (b)
10

probability of failure pf
14
-10
12 10
reliability index b

10 -20
10
8 parametric study 1
-30
10
6 parametric study 2
-40
4 10
20 30 40 COV in % 20 30 40 COV in %
E soil Esoil

20 30 40 COVg in %
soil
20 30 40 COVg in %
soil

15 25 35 COVn in % 15 25 35 COVn in %
2 4 6 COVK0 in % 2 4 6 COVK0 in %

Figure 5.10: Influence of the COV of the cohesion c′ , of the friction angel ϕ’ and of the
footing pressure on the probability of the failure by using the MC-criterion
in parametric study 2.

than for highly variable soils. One can conclude that the influence of the soil variability
is higher than the variability of the load. The results in figure 5.10 are evaluated for the
MC criterion.

5.3.3 Sensitivity analysis

Parametric study 1: Within parametric study 1, the local sensitivities are evaluated for
the range of the footing pressure from q = 500 kN/m2 to q = 3, 500 kN/m2 in figure
5.11. Within this, the coefficient of variation for the cohesion and the friction angle is
kept constant.
It can be seen that influence of the cohesion is decreasing with increasing footing pres-
sure and vice versa in the case of friction angle. Moreover, it is interesting to see that
the local sensitivity measures of MC, MN and LADE are the same relative to each other.
This can be related to the stress state and is explained in appendix F.

Parametric study 2: The local and global sensitivities are shown in figure 5.12 within
parametric study 2. Both measures of sensitivity show the comparative importance of
the investigated random input variables. In the left part of figure 5.12 the results of
the local sensitivity analysis are shown. The negative local sensitivity indicates that the
width of the footing B, the unit weight of soil γsoil and the soil strength c′ and ϕ′ are resis-
tance parameters, whereas the footing pressure q is a load parameter. From the measures
[email protected]
of local and global sensitivity it can be concluded that the soil strength parameters have

107
Chapter 5 Selected case studies in uncertainty quantification using random properties

(a) (b)
0 0

local sensitivity of j’
local sensitivity of c’

-0.2 -0.2

-0.4 -0.4

-0.6 -0.6
MC -0.8
-0.8 MN
LADE
-1.0 -1
500 1500 2500 3500 500 1500 2500 3500
footing pressure footing pressure

Figure 5.11: Local sensitivities of the cohesion c′ and the friction angle ϕ′ with respect to
the deterministic footing pressure in parametric study 1.

(a) (b)
M=5
local sensitivity ai

global sensitivity diPC

0.8 MC
0.0
0.6 MN
-0.2
LADE
-0.4 0.4

-0.6 0.2
-0.8 0.0
B g’ j’ c’ q B g’ j’ c’ q

Figure 5.12: Local (a) and global (b) sensitivity factors of the full probabilistic analysis in
parametric study 2 for COVϕ′ = 20%, COVc′ = 40% and COVq = 20%.

the biggest importance on the probability of failure. This implies that additional infor-
mation on cohesion and friction angle have more impact than for other parameters e.g.
the unit weight of soil γsoil .

5.3.4 Conclusion
The bearing capacity problem is investigated by means of 2D FEM calculations and prob-
abilistic methods within this case study. The results clearly indicate that the choice of the
constitutive failure criterion within a linear elastic, perfectly plastic constitutive model
has a significant impact on the deterministic bearing capacity as well as on the failure
probability of strip footings. Therefore, this source of model uncertainty has to be con-
sidered within an uncertainty quantification. From a deterministic as well as from a
probabilistic point of view, the Mohr-Coulomb criterion is the most conservative choice
[email protected]
in comparison to Matsuoka-Nakai and the Lade-Duncan. This is investigated by means

108
 5.3 Bearing capacity of vertically loaded strip footings

of fragility curves, which allow to visualize the influence of the different constitutive
failure criteria over a wide range of vertical footing pressures.
The load and resistance parameters of the studied problem are clearly identified by a
local sensitivity analysis and confirmed by the results of the global sensitivity analysis.
For sake of completeness, it has to be mentioned that the effects for a cross-correlation
ρc′ ,ϕ′ of the cohesion and the angle of friction are not investigated within the presented
parametric studies. This has been already been investigated in various publications e.g.
by Baecher & Christian [23] and Phoon [286] amongst others. Fenton [127] reports that
the probability of failure decreases with an increase of the negative correlation coefficient
ρc′ ,ϕ′ and vice versa.
It can be concluded that the presented findings are extending the state of the art from
a mechanical and a probabilistic point of view. The uncertainty of modelling the soil
behaviour should be investigated in further studies to derive tools to quantify this model
uncertainty in applied engineering.

[email protected]

109
Chapter 5 Selected case studies in uncertainty quantification using random properties

5.4 Tunnel face stability

During the construction of shallow tunnels by means of earth pressure balance machines,
the face stability is an important issue. To minimize settlements at the ground surface
and to prevent an uncontrolled collapse of the soil in the tunnel, a support pressure must
be maintained. The estimation of the required support pressures, e.g. for earth pressure
balance (EPB) shields to ensure the stability of the face, has been the topic of research
until the present day e.g. [12, 196, 216, 256, 311, 350, 392].
Within this case study the framework of uncertainty quantification is applied to the
face stability problem. Due to the absence of an analytical model, the 3D mechanical
system is modelled by the Finite Element Method (FEM) using different constitutive
failure criteria. Formulae for the collapse pressure of the tunnel face are derived from
these results, which offer the possibility to investigate the model uncertainty within the
probabilistic framework.

5.4.1 Mechanical description of the problem

Different researchers have worked on this problem using various approaches. Horn
[169] was the first using a wedge based model to calculate the minimum face pressure,
which is necessary to guarantee the stability of the system as shown in figure 5.13. This
approach was later improved by Anagnostou [12].
Leca & Dormieux [216] present an upper bound solution for the face stability of shal-
low tunnels by using a kinematic approach. This upper bound solution involves three
solutions based on consideration of three mechanisms, which are derived form the mo-
tion of rigid conical blocks; specifically, two active and one passive collapse mechanism.
According to Mollon et al. [256], the passive blow-out mode of failure does not occur for
the cases currently encountered in practice. Therefore, this collapse mechanism is not
investigated within this contribution.
Vermeer et al. [392] and Ruse [311] followed a different approach in evaluating the
stability of a tunnel heading. In contrast to Leca & Dormieux [216], no assumptions on
the shape of the collapse mechanism are made. By using the Finite Element Method
(FEM) and the Mohr-Coulomb (MC) failure criterion, a formula was derived for the
failure pressure of the tunnel face in equation 5.11. Herein c′ is the effective cohesion,

qs

H g’, c’, j’

D qt

[email protected]
Figure 5.13: Geometry of the tunnel.

110
 5.4 Tunnel face stability

(a) (b)
qt0

support pressure qt
control point A A
qt
s’h = K0 s’v

qcollapse

displacement u of the control point A

Figure 5.14: Flow area at collapse (a) and typical pressure displacement curve (b).

ϕ′ the effective friction angle, γ ′ the unit weight of the soil and D the diameter of the
circular tunnel (figure 5.13).

qcollapse = −c′ Nc + γ ′ D Nγ (5.11)

1
Nc′ ,M C = Nγ ′ ,M C = 1
9 tan ϕ′
− 0.05 (5.12)
tan ϕ′
When the ratio of the overburden and the diameter of the tunnel D are bigger than H/D
= 1.5, the overburden and the load q at the surface have no influence on the failure
pressure qcollapse , nor on the stiffness, the dilatation angle or the Poisson’s ratio of the
soil, as reported by Vermeer et al. [392]. Additional 3D FEM studies on the influence of
different constitutive failure criteria are conducted by the author.
As a symmetrical tunnel is considered, the collapse-load calculations are based on only
half a circular tunnel, which is cut lengthwise along the tunnel axis. Figure 5.14 (a) shows
a typical finite element mesh as used for the calculations. The ground is represented by
10-noded tetrahedral volume elements. The boundary conditions of the finite element
calculations are as follows: the ground surface is free to displace, the side surfaces have
roller boundaries and the base is fixed.
It is assumed that the initial stresses follow a geostatic stress distribution according
to the rule σh′ = K0 · σv′ , where σh′ is the horizontal effective stress and σv′ is the vertical
effective stress; K0 is the coefficient of lateral earth pressure. Ruse [311] found that the
K0 -value influences the magnitude of the displacements, but not the pressure at failure.
The first stage of the calculations is to remove the volume elements inside the tunnel
and to activate the shell elements of the lining. This does not disturb the equilibrium
as equivalent pressures are applied on the inside of the entire tunnel. To get full equiv-
alence between the initial supporting pressure and the initial geostatic stress field, the
pressure distribution is not constant but increases with depth. This is obviously signif-
icant for very shallow tunnels, but a nearly constant pressure occurs for deep tunnels.
[email protected]
The minimum amount of pressure needed to support the tunnel is then determined by

111
 Chapter 5 Selected case studies in uncertainty quantification using random properties

a stepwise reduction of the supporting pressure. A typical pressure-displacement curve

is shown in Figure 5.14 (b) , where qt is the support pressure at the level of the tunnel
axis and u is the displacement of the corresponding control point A at the tunnel face. It
has to be emphasised that this control point has to be chosen within the collapsing body;
otherwise the load-displacement curve in Figure 5.14 (b) will come to an almost sudden
end and the curve then cannot be used to conclude that failure has been reached.
The results of these parametric FEM studies using the Matsuoka-Nakai [243] (MN)
and the Lade-Duncan [213] (LADE) criteria are shown in figure 5.15. It can be seen
in figure 5.15 and derived from the equations below that the failure pressure qcollapse
is significantly reduced by introducing different constitutive failure criteria. One can
conclude form the experimental results presented in literature e.g. Kirsch [196] that there
is quite a wide range of possible values for Nc and Nγ in equation 5.11. The results of
the MN and LADE criteria in equations below are derived from an ordinary least square
fitting.

1
Nc,M N = Nγ,M N = 1
67 tan ϕ′
− 0.01 (5.13)
2 tan ϕ′

1
Nc,LADE = Nγ,LADE = 1
79 tan ϕ′
− 0.012 (5.14)
3 tan ϕ′

The system can be described with the limit state equation 5.15. This limit state equa-
tion compares the actual face pressure qt with the failure pressure qcollapse , as proposed

(a) (b)
soil weight stability number Ng

Anagnostou & Mohr-Coloumb

cohesion stability number Nc

5.0 0.6 Kovari Matsuoka-Nakai

Anagnostou &
Kovari Lade-Duncan
4.0
0.4
3.0
Krause
2.0 Leca &
0.2 Dormieux experimental
results in Kirsch
1.0

0.0 0.0
15 20 25 30 35 40 15 20 25 30 35 40
effective friction angle j’ effective friction angle j’

Figure 5.15: Variability of the stability numbers Nc (a) and Nγ (b) for different consti-
tutive failure criteria (Mohr-Coulomb, Matsuoka-Nakai, Lade-Duncan) in
[email protected]
comparison to experimental results from literature [196].

112
 5.4 Tunnel face stability

by Mollon et al. [255].

g (c′ , ϕ′ ) = qt − qcollapse (5.15)
For g = 0 one obtains a failure cure in the c′ − ϕ′ plane. These failure curves are also
refereed to as limit state surfaces. The limit state surfaces for different face pressures are
shown in figure 5.17 (a) and emphasise the key message in figure 5.16: the higher the
face pressure qt , the more unlikely is the collapse of the system.

5.4.2 Stochastic variables

For the stochastic soil properties, Phoon & Kulhawy [282] stated that, from an exhaus-
tive study of cone penetration test and triaxial test results, the coefficient of variation
(COV = σc′ /µc′ ) of cohesion c′ could vary from 10% to 55%; Cherubini [77] recommended
COVc′ = 12 − 45% for stiff clays and a higher limit of 80% for very soft clays. For the
COVϕ′ of the friction angle ϕ′ , Phoon & Kulhawy [282] proposed COVϕ′ = 5 − 15%.
On top of this, the correlation between cohesion and friction angle was found by
Cherubini [75] to be ρc′ ,ϕ′ = −60%, while Lumb [226], Phoon & Kulhawy [283], Wolff
[407], Yucemen et al. [414] and Speedie [354], amongst others, reported correlations be-
tween ρc′ ,ϕ′ = 0% and ρc′ ,ϕ′ = −70%.
The values used in this section for the statistical moments of the shear strength param-
eters belong to the intervals proposed by the above researchers. The stochastic variables
used in this section are chosen on the basis of a literature and are shown in table 5.3 and
5.4.

5.4.3 Parametric studies

Parametric study 1: The aim of the presented parametric studies is to quantify the im-
pact of soil variability for the tunnel heading problem using different constitutive failure
criteria within a probabilistic framework. The limit state equations have been imple-
mented into the F ERUM libaries [49].
In accordance with section 5.3, figure 5.16 can also be interpreted like a quasi-fragility
curve presented in section 5.3. It can be approximated by 1−Φ; Φ is the lognormal cumu-
lative distribution function. The deterministic face pressure qt is increased and plotted

Table 5.3: Properties of parametric study 1 on the tunnel heading using the Mohr-
Coulomb failure criterion.

property COV distribution

c′ = 2.5 kN/m 2 10 − 60 % lognormal
ϕ′ = 35◦ 5 − 30 % lognormal
γ ′ = 18 kN/m3 - deterministic
D = 10 m - deterministic
qt = 0 − 160 kN/m 2 - deterministic
[email protected]

113
Chapter 5 Selected case studies in uncertainty quantification using random properties

1.0
0.9

deterministic qcollapse
0.8 COVc’ = 10 % , COVj’ = 5 %
proebability of failure pf

0.7
COVc’ = 50 % , COVj’ = 25 %
COVc’ = 100 % , COVj’ = 50 %
0.6
0.5
0.4
0.3
0.2
0.1

0 qcollapse 50 100
face pressure qt [kN/m²]

Figure 5.16: Influence of the COV of the cohesion c′ and of the friction angle ϕ′ on face
pressure versus probability of failure using the Mohr-Coulomb criterion.

against the probability of failure pf . It can be seen in figure 5.16 that pf decreases with
an increasing deterministic face pressure qt and vice versa in the case of the reliability
index. It can also be concluded that, the higher the variability of ϕ′ and c′ , the higher is
the probability of failure pf . The deterministic value qcollapse shown in figure 5.16 is based
on the mean value of the cohesion and the friction angle.
The effects of positive and negative correlation between the cohesion c′ and the fric-
tion angle ϕ′ are shown in figure 5.17 (b). The ranges of ρc′ ,ϕ′ are chosen according to the
results of the literature study summarized in section 5.4.2. A positive correlation ρc′ ,ϕ′ in-
creases slightly the probability of failure. The influence of a negative correlation is more
obvious. In this parametric study it is found to be conservative to neglect a negative
correlation ρc′ ,ϕ′ , because the probability of failure is higher for ρc′ ,ϕ′ = 0.

Parametric study 2: In additional investigations, all parameters have been treated as

random variables in order to investigate the consequences of variability of the soil prop-
erties (ϕ′ , c′ , γ ′ ), of the geometry (represented by the tunnel diameter D) and of the
construction process (represented by the face pressure qt ). The stochastic properties are
defined in table 5.4 and have been used for a FORM analysis within the reliability as-
sessment.
If all variables are treated as random, the probability of failure is much higher than
in the case of parametric study 1. In that case, only the variability of the strength pa-
rameters was considered. Therefore, the probability of failure for a given face pressure
of qt = 50 kN/m2 (figure 5.16) is much lower than in the previous parametric study.
Additional calculations were carried out to get a more detailed insight into this problem.
[email protected]
For this purpose, the influence of the soil variability (c′ , ϕ′ ) and of the tunnelling pro-

114
 5.4 Tunnel face stability

friction angel j’
25° (a)
20° qt = 40 kN/m²
15° qt = 70 kN/m²
10° qt = 100 kN/m²
0 1 2 3 4 5 qt = 130 kN/m²
cohesion c’ [kN/m²]
probability of failure pf

0 (b)
10

-50
10

-100
10
-60 -40 -20 0 20 40 60
correlation coefficient rc’,j’

Figure 5.17: Limit state surfaces for different levels of the face pressure qt in (a) and for
probability of failure versus correlation coefficient between cohesion c′ and
friction angle ϕ′ in (b) for the MC- criterion.

cess represented by the face pressure qt is evaluated in additional calculations. The re-
sults of this study are shown in figure 5.18 (a). Herein, the soil variability is increased
and compared to different levels of variability of the face pressure for a given face pres-
sure qt = 50 kN/m2 . Again, it can clearly be seen that a high level of variability of the
variables causes a high probability of failure. Moreover it can be seen that the system
behaviour of highly random properties (COVc′ > 60%, COVϕ′ > 30%) are very similar.
The influence of different constitutive failure criteria is shown in figure 5.18 (b). Herein,
the probabilities of failure differs significantly for MC, MN and LADE. Again, the MC-
criterion offers conservative results in comparison to the more realisitic MN criterion,
whereas the results of the LADE criterion are even lower.

Table 5.4: Properties of parametric study 2 on the tunnel heading using the Mohr-
Coulomb failure criterion.

property COV distribution

c′ = 2.5 kN/m 2 20% lognormal
ϕ′ = 35◦ 10% lognormal
γ ′ = 18 kN/m 3 10% lognormal
D = 10 m 10% lognormal
qt = 50 kN/m 2 10% lognormal
[email protected]

115
 Chapter 5 Selected case studies in uncertainty quantification using random properties

(a) 0
(b)
10
probability of failure pf

0 MC

probability of failure pf
10
-2
10

-4
COVqt = 20 % 10 MN
10
-4 COVqt = 40 %
COVqt = 60 % 10
-8

COVqt = 80 %
COVqt = 100 % -12 LADE
-5 10
10
20 30 40 50 60 70
coefficient of variation COVc’ [%] 10
-16

20 40 60 80 COVc’ [%]
10 15 20 25 30 35 10 20 30 40 50
coefficient of variation COVj’ [%] coefficient of variation COVj’ [%]

Figure 5.18: (a) Variation of the COVs for the cohesion c′ , the friction angle ϕ′ and
the tunnel face pressure qt together with the corresponding probabilities
of failure pf .
(b) Variation of the coefficients of variation for the cohesion c′ and the
friction angle ϕ′ for a given coefficient of variation COVqt = 10%
the tunnel face pressure qt , together with the corresponding probabilities
of failure pf .

(a) (b)
global sensitivity diPC

0.2
0.8
0.0
local sensitivity ai

-0.2 0.6
-0.4 MC βMC = 5.80
0.4
-0.6 MN βMN = 10.06
-0.8 LADE βLADE = 21.90 0.2

-1.0 0
c’ j’ D g’ qt c’ j’ D g’ qt

Figure 5.19: Reliability indices βi and local sensitivity αi (a) as well as the global sensitiv-
ity values δiP C values (b) for parametric study 2 on the tunnel face stability
using different constitutive failure criteria MC, MN, LADE.

[email protected]

116
 5.4 Tunnel face stability

5.4.4 Sensitivity analysis

The local and global sensitivity of each random variable are shown in figure 5.19. The
load and resistance variables can be clearly identified from the local sensitivity measures
in figure 5.19 (a). The diameter D and the soil weight γ’ show a positive local sensitivity
αi . Therefore, they are load parameters of the limit state equation, whereas the cohesion
c’, the friction angle ϕ’ and the face pressure qt are resistance parameters. One can clearly
derive from both local and global sensitivity analyses that the cohesion c′ and the friction
angle ϕ′ have the largest influence on the reliability index, which is in agreement with
comparable studies by Mollon et al. [255].
Comparing the different constitutive failure criteria MC, MN and LADE, one can see
the same qualitative results of the sensitivity analyses. All three constitutive failure cri-
teria have show a nearly the same local and global sensitivity measures. This implies
that the contribution to the system behaviour is nearly the same for all investigated con-
stitutive failure criteria.

5.4.5 Conclusions
Within two parametric studies, the influence of soil variability, geometry and construc-
tion process on the probability of failure is studied. The limit state equation is derived
from 3D FEM studies using linear elastic, perfectly plastic constitutive models, which
are using the Mohr-Coulomb, the Matsuoka-Nakai and the Lade-Duncan failure criteria.
The assumption of uncorrelated shear strength parameters is found to be conservative
(i.e. it gives a greater probability of failure) in comparison to that of negatively correlated
parameters. The results of the parametric studies suggest that the strength parameters
of the soil and the tunnel face pressure have the major influence on the probability of
failure in comparison to the soil unit weight and tunnel diameter.
As a consequence of this, the influence of the different constitutive failure criteria is
investigated. One can clearly see that the well known Mohr-Coulomb criterion offers
very conservative results in relation to the more realistic Matsuoka-Nakai criterion and
to optimistic Lade-Duncan criterion. As pointed out in the appendix F, there are dif-
ferences in the definition of the failure surface, but the number of variables needed for
these criteria are the same. Therefore one can say that the complexity of these models
are similar.
It can be concluded from these results that the choice of the soil model and the consti-
tutive constitutive failure criteria has significant influence on the reliability of the tunnel
face. Additional investigation should focus on the model error and its considerations via
e.g. partial safety factors in applied engineering.

[email protected]

117
Chapter 5 Selected case studies in uncertainty quantification using random properties

5.5 Synopsis of the case studies

Three different case studies are quantifying the effects of soil variability on the perfor-
mance of geotechnical structures in the ultimate limite state. Herein, the soil variability
is considered via random variables following lognormal distributions.
This extends the state of the art given in E UROCODE 7 [119]: Soil variability is de-
scribed via probability distribution within the framework of uncertainty quantification
instead of constant safety factors.

Tunnel lining: The analytical solution of a continuum mechanical problem is used to

describe the soil-structure interaction of a tunnel lining with soil. Herein, the behaviour
of soil and concrete are approximated by purely elastic material models. The level of
complexity is increased from 4 to 9 random variables within two parametric studies. It
can be concluded from these results that only the driving parameters have to be con-
sidered as random properties. This is proven through the local and global sensitivity
analyses. The identification of the key parameters of a mechanical system by means of
sensitivity analysis is very important for practical issues because a low number of ran-
dom variables implies less computation time.

Bearing capacity of vertically loaded strip footings: Within this case study on the
ultimate limit state of a footing, the soil behaviour is modelled via linear elastic, per-
fectly plastic models, which follow different constitutive failure criteria; namely Mohr-
Coulomb, Masuoka-Nakai and Lade-Duncan. These approaches are used within para-
metric 2D FEM studies, which are used to derive a semi-analytical limit state equations.
Fragility curves are used as deterministic tools to quantify deterministically the stochas-
tic system. Moreover, a local and global sensitivity analysis is conducted to investigate
the contributions of soil strength parameters and other random input variables to the
probability of failure.

Tunnel face stability: The ultimate limit state of the tunnel heading is investigated
by using the Mohr-Coulomb, the Matsuoka-Nakai and the Lade-Duncan failure criteria,
which are used within a parametric, 3D FEM study. A semi semi-analytical limit state
equation is derived from these results, which is used in the subsequent reliability analy-
sis. Similar to the results of the case study on the bearing capacity of strip footings, one
can also identify the Mohr-Coulomb criterion as conservative approach for these two
problems in comparison to the Masuoka-Nakai and the Lade-Duncan criteria.

[email protected]

118
 Chapter 6
Selected case studies in uncertainty
quantification including spatial variability
It is stated in the E UROCODE 7 [119] that “characteristic values of soil and rock properties shall
take account of the variabilities of the property values”. Controversially, although statistical
methods are suggested as a possible way forward, there exists little guidance as to how
this should be achieved. This suggests a need to take soil variability into account within
a sound mathematical framework for geotechnical design.
The application of random fields, which represent the spatial variability of soil proper-
ties, is shown in different case studies in geotechnical engineering. The state-of-science
in uncertainty quantification is applied in the the estimation of tunnelling induced settle-
ments, slope stability and serviceability of footings. This chapter furthers the idea which
are presented in the three case studies in the last chapter.
Moreover, these applications also extend the common approaches with respect to mul-
tiple scales of soil variability within the framework of uncertainty quantification. Fi-
nally, the uncertainty of geology represented by macro-scale variability is investigated
by means of modern geostatistical simulation methods. This offers an insight into the
effects of geological uncertainty.

6.1 Estimation of tunnelling induced settlements

This case study investigates the consequences of soil variability in connection with tun-
nelling induced surface settlements. These investigations are performed for a single
layered and a two-layered soil involving different scales of spatial variability.
Soil variability can be simulated by the Random Finite Element Method, which is ex-
plained within this case study. The influencing factors on the probability of damage due
to differential settlements are identified and compared to each other. The results of this
contribution help to understand the influence of soil variability.
Schmidt [324] and Peck [274] were the first to show that the transverse settlement
trough, taking place after construction of a tunnel, in many cases can be well described
by the Gaussian function. Among others, Kolymbas [205] and Verruijt [393] derived
the settlement curve analytically for the case of a homogeneous subsoil. This practical
approach, as well as numerical methods such as the Finite Element Method (FEM) [253],
are often based on the assumption of a homogeneous soil. O’Reilly & New [267] and
Mair & Taylor [233] offered empirical formulae derived from case studies to take layered
[email protected]
soils into account.

119
Chapter 6 Selected case studies in uncertainty quantification including spatial variability

6.1.1 Mechanical description of the problem

Within this section, a 2D FEM mesh, shown in figure 6.1, is used to calculate the surface
settlements. This Plaxis 2D [8] FEM model consists of 2,326 15-noded elements, which
does not influence the stochastic soil properties due to the small grid size [174, 178].
Table 6.1 summarises the material parameters for the linear elastic, perfectly plastic soil
model and the linear elastic material model for the concrete lining.
The conventional tunnelling excavation is simulated via the so-called STRESS REDUC -
TION METHOD [253]. In this 2D method the 3D excavation problem is captured through a
stress relaxation factor. Using this, the stress relaxation of the ground due to the delayed
installation of the shotcrete lining and the load sharing between soil and lining are nicely
addressed. A full faced excavation with a stress relaxation of 35 % is chosen according to
Möller [253]. The stiffness of the building is not taken into account in this evaluation of
surface settlements. As described by several publications [205, 267, 274], the settlement
trough follows a Gaussian function.
As pictured in figure 6.1, the limit state due to differential settlements is investigated.
A limit state function g is defined to quantify the consequences of spatial variability.
This function describes the difference between the rotation α, which is evaluated by the
Random Finite Element Method (RFEM) and an ultimate rotation αultimate , that is

g (α) = αultimate − α (6.1)

The ultimate rotation αultimate = 1/500 due to differential settlements is taken from the
DIN 4019 [102] to avoid cracks in masonry. This RFEM approach calculates the green-
field settlements and does not consider soil structure interaction. Moreover, this proce-
dure does not take into account the different convex and concave parts of the surface
settlements as described by Netzel [263].

6.1.2 Stochastic variables

Within this section, the effects of spatial variability of the soil is investigated. There-
fore, random fields are used to represent spatially variable stiffness of the soil. The
RFEM is schematically pictured in figure 6.1 (a). The random fields are mapped onto
the FEM mesh via the spatial averaging approach and used to evaluate the system re-
sponse, which varies between the random fields.
Within RFEM, one has to pay attention to the generation of the random fields and to
the coarseness of the random field mesh. It can be deduced from section 4.2.3 that the
random field mesh has to be finer than the FEM mesh and one has to keep in mind that
the coarseness of the random field is also dependent on the simulated correlation length.
The libraries of GSLIB [97] are used for this. The finer random fields have been
mapped onto the coarser FEM mesh via the spatial averaging approach described in
section 4.2.4. In this study the random field mesh is 10 times finer than the smallest el-
ement of the FEM mesh. This averaging over every single FEM element enables one to
[email protected]
use non-structured FEM meshes.

120
 6.1 Estimation of tunnelling induced settlements

Table 6.1: Material properties for the parametric study on tunnelling settlement in an
single-layered subsoil.

Soil (linear-elastic, perfectly-plastic soil model using the MC criterion)

soil unit weight γsoil ’ = 20 kN/m3
friction angle ϕ’ = 20 ◦
cohesion c’ = 40 kN/m 2
Poisson’s ratio ν = 0.35
coefficient of horizontal earth pressure K0 = 0.34
modulus of elasticity lognormally distributed
µE = 60 MN/m2
COVE = 10 − 75 %
θh /D = 0.5-5
θh / θv = 1 − 50
exponential correlation function
Shotcrete (linear elastic material model)
thickness of the lininig d = 0.35 m
concrete unit weight γc = 25 kN/m3
modulus of elasticity E = 7, 500 MN/m2
Poisson’s ratio ν = 0.20

In figure 6.1 the probability of damage is evaluated based on a modified Monte-Carlo

approach. Although the Monte-Carlo approach is the most robust and accurate reliabil-
ity method, it is very time consuming to reach good accuracy of the probability of failure,
especially in the presence of small probabilities as mentioned in section 4.3.2. Therefore,
the author fits a normal distribution to the system response of 300 random field realisa-
tions as proposed by Huber et al. [174] (see figure 6.1 b and c). This approach speeds up
the evaluation of the influences of the variation of geometrical and stochastic properties
relative to each other. Moreover, the convergence of this modified Monte-Carlo approach
is checked by looking at the mean system behaviour after 300 RFEM calculations. It is
found that that the mean value and standard deviation of the differential surface settle-
ments do not change significantly after 300 random field realisations, [174]. The results
are approximations, but nevertheless a basis for the investigations of influences of the
above mentioned properties in comparison to each other. For this reason, this approach
is used for the evaluation of the results in this chapter 6.1.

[email protected]

121
Chapter 6 Selected case studies in uncertainty quantification including spatial variability

6.1.3 Parametric study on tunnelling settlement in a single-layered

subsoil
Within this case study, different geometrical boundaries, such as the width of the build-
ing B, the location of the building relative to the tunnel axis and the overburden H of the
tunnel are varied. On top of this also the effects of coefficient of variation of the stiffness
COVE , the vertical correlation length θv , the horizontal correlation length θh and the ra-
tio of horizontal and vertical correlation lengths θh /θv are investigated. All the presented
results are symmetrical with respect to the tunnel axis. This is related to the symmetrical
distribution of the principal stresses due to the simulation of the tunnel excavation.
Tunnelling creates a settlement trough that is centred directly above the tunnel. As
mentioned in section 6.1.1, this settlement trough is following a Gaussian distribution
function. As a consequence, the probability of damage is highest for a building above
the steepest slope of this trough. In figure 6.1 the variable L stands for distance from
midpoint of the building.
This is investigated in figure 6.2 (a). Different locations and different widths of a build-
ing are investigated. It can be concluded from this that larger buildings are less endan-
gered than smaller ones, which is related to the geometry of the building in relation to
the tunnel diameter D. However, one has to keep in mind that the building is assumed
as a rigid block, otherwise the probability of damage for large buildings is expected to
be significantly bigger.
In addition to the location and width of the building, the effects of the stochastic soil
properties on the probability of damage pdamage are investigated in figure 6.2 (b,c,d). By
comparing figures 6.2 (b) and (c) it can be clearly seen that the effects of the variability of
the stiffness COVE are larger than the effects of the spatial variability of the stiffness. This
low influence of the spatial correlation structure can attributed to the construction pro-
cess of the tunnelling. The changes of the stresses field due to the tunnelling excavation
process are predominant in relation to the variations caused by the spatial correlation of
the stiffness.
In contrast to this are the results in 6.2 (d). It is shown that the probability of dam-
age pdamage is strongly influenced by introducing an anisotropy of the spatial correlation
θh /θv . In the case of modelling the stiffness of the subsoil via isotropically random fields
(θh /θv = 1), the probability of damage is far higher than in case of highly anisotropically
random fields.
In addition to this, the influence of anisotropy θh /θv is shown in figure 6.3. The max-
imal probabilities of damage pdamage are plotted with the ratios of anisotropy θh /θv in
figure 6.3. The isotropic correlation structure θh /θv = 1.0 causes the biggest probabil-
ity of damage. The larger the horizontal correlation length θh becomes, the lower is the
probability of damage.
By looking into geostatistic textbooks [79], one finds some hint to explain this: Large
ratios of anisotropy θh /θv can be interpreted as soil layering. In this extreme case, the hor-
izontal correlation length is θh is assumed to be endless whereas the vertical correlation
length is shorter, which leads to a nearly one-dimensional distribution of soil properties.
[email protected]
In figure 6.3 it is shown that this leads to a lower probability of damage than in the case

122
 6.1 Estimation of tunnelling induced settlements

(a) Generation of Mapping the fields Performing a FEM calculation

random fields onto the FEM mesh B
P1 P2
H
D
4D L = (P1+P2)/2

10 D
density function
probability

sZ sZ
X Y

pdamage mZ
Z = aultimate-a
Evaluation of the probability of damage Evaluation of the system response

fitted normal distribution of g

(mZ=0.00176;sZ =0.00076)
empirical distribution of g
Cumulative distribution

100 % 600
denstiy function
Probability
function

400
50 %
200

0% 0
0 0.002 0.004 0 0.002 0.004
(b) Limit state function g Limit state function g (c)

Figure 6.1: (a) Evaluation scheme of the probability of damage pdamage due to
differential settlements,
(b) cumulative distribution function and
(c) histogram with fitted probability density function of the normally
distributed limit state function g, based on an underlying random field
(B/D = 2, H/D = 1, µ = 60 MN/m2 , COV = 50 %, θh = θv = 2D).
[email protected]

123
Chapter 6 Selected case studies in uncertainty quantification including spatial variability

L/D
0 1 2 3
0
(a)
B / D = 1.0 qv / D = qv / D = 2
log(pdamage)

B / D = 2.0 H/D=1
-2 B / D = 3.0 COVE = 50 %
B / D = 4.0

-4

L/D
0 1 2 3
0
(b)
log(pdamage)

-2 COVE = 10 % qv / D = qv / D = 2
COVE = 50 % H/D=1
COVE = 75 % B/D=2
-4

L/D
0 1 2 3
(c ) 0
log(pdamage)

q / D = 0.5 = qv / D = qv / D
-2
q/D=1 H/D=1
q/D=2 B/D=2
-4 q/D=5 COVE = 50 %

L coordinate / diameter D
0 1 2 3
(d) 0
log(pdamage)

qh / qv = 1 qh / D= 2
-10 qh / qv = 5 H/D=1
qh / qv = 10 B/D=2
-20 qh / qv = 50 COVE = 50 %

Figure 6.2: (a) Influence of the width B of the building,

(b) influence of the coefficient of variation,
(c) correlation lengths and
(d) influence of the anisotropy of the correlation structure on the surface
settlements due to tunnel excavation.
[email protected]

124
 6.1 Estimation of tunnelling induced settlements

maximum probability of
0
10

damage max pdamage

H/D = 0.40
-5
10 COVE = 50 %
B = 2D
-10
10

-15
10

0 5 10 50 100
ratio of anisotropy qh / qv

Figure 6.3: Effects of the ratio of anisotropy θh /θv on the maximum pdamage .

of an isotropically random field simulation.

6.1.4 Parametric study on tunnelling settlement in a two-layered soil

The results of the parametric study in a single-layered subsoil show clearly that the in-
fluence of spatial anisotropy is tremendously high in comparison to the the variation of
geometrical conditions or (isotropic) stochastic properties. Therefore, the consequences
of anisotropy are investigated in the sequel on a more generous basis. The extreme case
of anisotropy is that the horizontal correlation length becomes nearly infinity. Then all
points are practically correlated in the horizontal but not in the vertical direction. This
macro-scale correlation can also be interpreted as soil layering.
Starting from this idea, a concept was set up to simulate soil layering or, geostatisti-
cally speaking: to simulate spatial variability at two scales: the meso-scale variability
inside both soil layers is represented by two random fields following the properties in-
dicated in table 6.2. The soil layers are separated by a one dimensional random field,
which is conditioned to soil investigations. This one-dimensional random field is rep-
resenting a macro-scale variability in this parametric study. The soil investigations are
indicating the vertical location of the boundary between the upper and lower layer.
Figure 6.4 shows the tunnel and the location of three boreholes (B1, B2, B3). The ratio
of overburden was chosen with H/D = 1 to skip the influences of shallow tunnels on the
probability of damage.
At first, the influence of the number of boreholes is investigated. This can be inter-
preted as different levels of knowledge of the soil layering, as shown in figure 6.4. The
results are shown in figure 6.5 (a). In case of one borehole B1, the reliability of the system
is smaller than in case of 2 (B1+B2) or three (B1+B2+B3) soil investigations.
Apart from this, the effects of the soil layer boundaries are investigated. For this rea-
son, the correlation length of the one dimensional random field θboundary is increased,
while keeping the spatial variability of the upper and lower layers constant (table 6.2).
[email protected]
The increase of the reliability of the system is shown in figure 6.5(b). Herein, a random

125
Chapter 6 Selected case studies in uncertainty quantification including spatial variability

L = (P1+P2)/2
D D
B
B3 B1 B2 P1 P2

4D

10 D

Figure 6.4: FEM mesh and location of the boreholes (B1,B2 B3).

Table 6.2: Material properties for the parametric study on tunnelling settlement in a lay-
ered soil.

Soil (linear-elastic, perfectly-plastic material model)

upper layer lower layer
3
γ = 20 kN/m
′
21 kN/m2
ϕ′ = 20 ◦ 30◦
c′ = 30 kN/m2 20 kN/m2
ν = 0.35 0.30
K0 = 0.34 0.34
2
E-modulus µ = 60 MN/m µ = 300 MN/m2
COV = 50 % COV = 50 %
θh / D = 1 θh / D = 1
θh / θv = 1-50 θh / θv = 1-50
exponential correlation function
Shotcrete (linear-elastic material model)
d = 0.35 m
γ shotcrete = 25 kN/m3
E = 7,500 MN/m2
ν = 0.20
[email protected]

126
 6.1 Estimation of tunnelling induced settlements

(a) B/D=2 (b)

H/D = 1
4 1 borehole (B1) B/D=2 2.0 qv / D = 1
10

scaling factor b
reliability index b

2 boreholes (B1+B2) H /D = 1 qh / D = 1
q /D=2 1.5
10
3
3 boreholes (B1+B2+B3) boundary 1 borehole
qv / D = 1
2
qh / D = 1 1.0
10

10
1
0.5

0
10
0 0.5 1 1.5 2 2.5 0 1.0 2.0 3.0 4.0 5.0
L/D qboundary / D

(c)
1.0 B/D=2
scaling factor b

0.8 H/D = 1
qboundary / D= 2 D
0.6 1 borehole
0.4
0.2
0.0
1 2 10 20 50
anisotropy qh/qv

Figure 6.5: Results of parametric study on tunnelling in a layered soil.

process is conditioned to one borehole to simulate the boundary between the isotropic,
homogeneous random fields of the upper and lower layers. Note that these random
fields are unconditional and not conditioned to the borehole data. The effects of in-
creasing values of θboundary are quantified via β̄. This ratio β̄ quantifies the effects of
different values of θboundary , which is normalized by the reliability index β with respect
to θboundary /D = 2 using one borehole for conditioning. It is clearly indicated in figure
6.5(b) that the boundary between both layers has a significant influence on the probabil-
ity of damage, which is shown via the reliability index β. It can be concluded that a long
correlation length of the stochastic boundary implies a larger reliability against damage
of the building.
Another issue of interest is the effects of the anisotropy of the spatial correlations in
the upper and lower layers. For this reason the spatial anisotropy θh /θv of the upper and
lower layers are simultaneously stepwise increased. The boundary correlation length
of θboundary /D = 2 is kept constant. The results in figure 6.5 (c) show that, the larger
the horizontal correlation lengths are, the bigger is the pdamage and the smaller is the
[email protected]
reliability index β. Comparing the results of the figure 6.5 (b) and (c), one can see that

127
Chapter 6 Selected case studies in uncertainty quantification including spatial variability

the macro-scale variability of the boundary between the soil layers has more effect than
the spatial anisotropy θh /θv within the layers.

6.1.5 Conclusions
This case study investigates the effects of spatial soil variability on a complex soil-structure
interaction problem. The surface settlements are used for a simplified analysis of differ-
ential settlements of a rigid building, which are introduced by shallow tunnelling. These
surface settlements are calculated on the basis of the linear elastic, perfectly plastic Mohr-
Coulomb model and they are used to estimate the probability of damage of a building.
Within this, the interaction between subsoil and the weightless and rigid building is in-
vestigated.
Although these assumptions simplify this complex soil-structure interaction problem,
the effects of the spatial variability of the soil is quantified via the probability of dam-
age by varying geometrical properties (location of the building L, width of the building
B) and stochastic soil properties (coefficient of variation COVE , horizontal correlation
length θh , vertical correlation length θv and the ration θh /θv ). Within this, the single lay-
ered subsoil is assumed to have spatially correlated properties.
In addition to this, the effects of spatial variability of the stiffness at different scales
are investigated. This is done in a simplified way: The medium-scale spatial variability
of the stiffness is represented by random fields. The large scale spatial soil variability is
considered by introducing two different layers. Herein, a stochastic process is separating
the two layers, which are represented by random fields. At first, the stochastic process
is conditioned to one borehole. It is found that additional boreholes for conditioning
the stochastic process lower the probability of damage. Moreover, it is shown that a
large correlation length of the stochastic process also lowers the probability of damage.
Similar effects are observed for a large ratio of anisotropy θh /θv .
The presented case studies indicate the effects of spatial soil variability by using sim-
plified approaches for estimating the probability of damage. Additional investigations
on the soil-structure interaction should be carried out using advanced constitutive soil
models. These studies would contribute quantitatively to the presented results.
Above all, it has to be pointed out that spatial soil variability is three dimensional
and it cannot be fully described by two dimensional investigations. Therefore, only 3D
RFEM studies can fully evaluate the impacts of the spatial variability of soil properties
at different scales.
However, the uncertainty quantification, used in the case studies in chapter 5, is not
fully performed. Although random fields are employed to represent the spatial soil vari-
ability and to evaluate the response of the mechanical model, the sensitivity analyses are
not performed due to the absence of adequate methods presented in literature. This asks
for the development of new approaches for sensitivity analyses in order to quantify the
contribution of spatially correlated variables to the system response.

[email protected]

128
 6.2 Slope stability

6.2 Slope stability

The slope stability problem is a very general geotechnical problem. Not only in urban,
but also in rural areas slope stability problems can occur because slope stability is con-
cerned with the stability of natural slopes, excavations, embankments, dams, road cuts,
mining pits or landfills, [86].
The main objectives of slope stability analysis are finding endangered areas, investiga-
tion of potential failure mechanisms, determination of the slope sensitivity to different
triggering mechanisms, designing of optimal slopes with regard to safety, reliability and
economics, designing possible remedial measures, e.g. barriers and stabilization, [355].
Due to the importance of the slope stability problem, lots of scientists have been deal-
ing with this issue from a deterministic as well from a probabilistic perspective. The
probabilistic analysis of slope stability has been in the focus of science since more than
40 years e.g. [114, 371, 383, 407, 414] and is still a topic of ongoing research [157, 158, 254,
266, 355, 391]. These investigations are focusing on the effects of uncertain soil parame-
ters including spatial variability as well as time dependent seepage forces.
This case study progresses this line of research. The influence of soil variability and
spatial variability on a single-layered soil slope is investigated in two dimensional slope
stability calculations. These calculations are compared to each other using different ap-
proaches including different correlation functions describing spatial variable soil proper-
ties and different random field generators. In addition to this, the effects of the anisotropy
of spatial correlation on the slope reliability are investigated. Also a two-layered slope is
investigated incorporating different scales of spatial variability. The sensitivity analysis
of this problem helps to understand the contributions of the different sources of uncer-
tainty and the different scales of variability to the failure probability of the systems.

strength reduction factor 3H 2H 3H

1.0 1.4
safety
global

0.3
1.5 H

0.4
δmax

(a) (b)

Figure 6.6: (a) Strength reduction factor versus maximum settlement of a 2D

homogeneous slope (c′ = 50 kN/m2 , ϕ′ = 0) and
(b) FEM mesh of the investigated slopes with 8-noded,
[email protected]
quadrilateral elements.

129
Chapter 6 Selected case studies in uncertainty quantification including spatial variability

6.2.1 Mechanical description of the problem

Conventional slope stability analysis: Amongst others, Craig [86] reports that the sta-
bility of a slope is usually assessed using limit equilibrium methods. Stability analysis
using the limit equilibrium approach involves solving the equilibrium problem by as-
suming force and/or moment equilibrium.
Over the years, many limit equilibrium methods for slope stability analysis have been
developed and applied in practice, including the ordinary method of slices of Fellenius,
Bishop’s modified method, force equilibrium methods, Janbu’s generalised procedure
of slices, Morgenstern and Price’s method and Spencer’s method. These methods have
been used for slope stability charts, which are useful for a preliminary analysis and a
quick estimation of the stability of slopes [86]. However, in practice, detailed slope sta-
bility analysis is usually performed using a computer program and most of the available
computer programs are based on the limit equilibrium approach.
In the conventional limit equilibrium approach, the stability of a slope is measured by
the factor of safety (FOS), which is defined as the ratio between the shear strength of the
soil to the shear stress required for limit equilibrium.

Finite Element Method for slope stability analysis: The finite element method (FEM)
is a powerful technique for slope stability analysis. Herein, the strength parameters
of the soil are reduced until the collapse of the system, which is called the strength-
reduction method. The strength reduction method is illustrated by the strength reduc-
tion factor versus maximum settlement plot in figure 6.6 (a) for the FEM mesh shown in
figure 6.6 (b). Although the FEM has been commonly used in the deformation analysis
of embankments and other geotechnical problems, it is still not widely used for the sta-
bility analysis of slopes as compared with the conventional limit equilibrium methods
[86, 147]. This can be deduced to the simplicity of the latter approach and to the available
computer programs usually providing a quick and accurate estimation of the FOS of a
slope.
In contrast, the FEM involves more complex theory and it usually requires more time
for developing model parameters, performing the computer analyses and interpreting
the results as described in detail in Smith & Griffiths [342].
Despite this, the FEM as in several advantages over the conventional limit equilibrium
methods, as stated by Griffiths & Lane [147]:

• No assumption is required in advance with respect to the shape and location of the
slip surface. Therefore, the failure finds its way through weak zones of the soil.

• There is no need to make assumptions about internal forces, which appear to be

one of the major sources of inaccuracy for some limit equilibrium methods. The
finite element method preserves global equilibrium until "failure" is reached.

• The FEM solution provides information about deformations at pre-failure stress

[email protected]
levels if realistic soil stiffness parameters are used.

130
 6.2 Slope stability

• The FEM is able to provide information on progressive failure up to and including

overall shear failure.

Smith & Griffiths [342] state that the FOS computed by the FEM is in good agreement
with that calculated by limit equilibrium methods, which is proven by various studies.

6.2.2 Parametric studies of single-layered soil slopes

Influence of soil variability on slope reliability: The aim of this section is to investi-
gate and quantify the effects of the coefficient of variation and the spatial variability of
soil properties on the estimated probability of failure pf of a slope. For this reason, nu-
merical studies are conducted using a modified version of the FEM programme of Smith
& Griffiths [342] within the framework of MATLAB and of the FERUM libraries [49].
Within this section, the author does not follow methods like the First-Order-Second-
Moment approach or Point-Estimate-Methods linked with limit equilibrium methods as
published by [114, 265] amongst others, but the Monte Carlo approach with numerical
methods as proposed by Fenton & Griffiths [127] and Hicks & Spencer [158] amongst
others.
The First-Order-Reliability-Method (FORM) is used to evaluate the reliability of the
slope shown in figure 6.6 (b). The cohesion c’ is considered as random variable as in-
dicated in table 6.3 without taking spatial variability into account. The effects of the
increase of soil variability on the probability of failure pf and on the reliability index β is
shown in figure 6.7. The probability of failure pf is significantly higher for large COVc’ s
than for low COVc’ s and vice versa in case of the reliability index β. This simple study
does not take the spatially variability of the cohesion into account.

Influence of spatial variability on slope reliability: Therefore, the Random Finite Ele-
ment Method (RFEM) approach [127] is used to quantify the effects of spatial variability.
Within this RFEM study, random fields represent the spatial variability of a purely cohe-
sive soil, which is represented by an idealised linear-elastic, perfectly-plastic soil model.

Table 6.3: Lognormal distributed input parameters of the parametric studies.

Soil properties
soil unit weight γ’ = 21 kN/m3
friction angle ϕ = 0◦
cohesion c’ lognormal distribution
µ = 50 kN/m2
COVc’ = 10-100 % exp. corr. function
Θ = θ/H = 1-10
angel of dilatancy ψ = 0◦
Poisson’s ratio ν = 0.3
[email protected]
Geometry height H = 10 m

131
Chapter 6 Selected case studies in uncertainty quantification including spatial variability

probability of failure pf
2.5 pf 10
reliability index b

2.0
-1
10
1.5
1.0 -2
10
0.5
b
-3
0.0 10
10 20 30 40 50 60 70 80 90
COVc’ in %

Figure 6.7: Effect of varying COVc on the reliability index β and on the probability of fail-
ure pf using lognormally distributed random variables for µc = 50 kN/m2 .

On the basis of the Monte-Carlo approach in chapter 4.3.2, each random field realisation
is represents a possible spatial distribution of higher and lower values, which influence
the system response.
Following the concepts of Fenton & Griffiths [127], the COV and the correlation length
of an isotropic random field of a cohesive slope are investigated using two different
random field generators, namely Sequential Gaussian Simulation Method (SGSIM) and
Sequential Indicator Simulation Method (SISIM). SGSIM and SISIM are basically similar
in the way of random field generation, but differ in the representation of spatial corre-
lation, as explained in appendix E. In the case of SISIM random fields, also the extreme
high and low values have a spatial correlation in contrast to SGSIM random fields. This
can be considered in the SISIM approach via different indicator correlation lengths. A
more detailed explanation of the differences and similarities of these approaches can be
found in Appendix E. The results of the case studies on the evaluation of the spatial vari-
ability of soil properties in chapter 3 clearly indicated that the investigated measurement
data cannot be fully described by a mean value, a standard deviation and only one single
spatial correlation function. Therefore, the SISIM approach is employed to investigate
the effects of this.
The libraries of GSLIB [97] are used for the generation of random fields, which are used
as input for the modified Monte-Carlo approach inside the RFEM framework, described
in section 6.1.3. After checking the convergence of the mean and standard deviation of
the REFM results, a normal distribution function is fitted to the simulation results via
the best-fit criterion in order to estimate the pf . Via this simplification it is possible to
investigate also small failure probabilities within reasonable computation times.
In the figure 6.8 (b) the effect of the normalized isotropic correlation length Θ = θ/H
on reliability index β for COVc = 20% can be seen. The reliability index of small normal-
ized correlation lengths is significantly higher than for large correlation lengths. Similar
results are reported Fenton & Griffiths [127].
[email protected]
Moreover, it is interesting to compare the outcomes of the different random field gen-

132
 6.2 Slope stability

(a) (b)

SGSIM
reliability index b

reliability index b
30 30
SISIM

20 20

10 10

0 0
0 20 40 60 80 100 0 2 4 6 8 10
COVc’ in % isotropic correlation length Θ = q/H

Figure 6.8: (a) Effects of varying COVc on the probability of failure using different
random field generators with a Θ = θ/H = 2 and random variables
for µc = 50 kN/m2 ;
(b) effects of varying Θ = θ/H on the probability of failure for different
random field generators in comparison to random variables with
a COVc’ = 20%.

erators SGSIM and SISIM. As pointed out before, the SGSIM and LAS approach are used
in well known publications e.g. [127, 156, 158] due to their simplicity in describing a ran-
dom field by the mean value, a standard deviation and one correlation function. Com-
paring the results in figure 6.8, one can deduce that the influence of the spatial correlation
structure simulated via SGSIM and SISIM is offering comparable results. Although the
differences between the two approaches are relatively small, the SISIM approach offers
slightly lower reliability indices for the same spatial correlation lengths; the SGSIM ran-
dom fields only differ in terms of the indicator correlation length, as described in detail
in appendix E.4.
Additional conclusions can be drawn from figure 6.8 (a). Herein, the COVc has been
increased while keeping the isotropic correlation length constant. This offers a clear in-
sight into the effect of increasing COVc , which is similiar to the presented uncertainty
quantification of a single-layered soil slope using random variables instead of random
fields, which represent soil variability. The bigger the COVc of the soil the more unreli-
able is the slope.
By comparing figures 6.7 and 6.8, one can deduce that the consideration of spatial vari-
ability results in a higher slope reliability. In the case of an infinite correlation length, all
elements of a random field are fully correlated; therefore, the results of the RFEM and
reliability methods are assumed to converge. If there is a very small spatial correlation
of the cohesive soil, every single element of the random field is theoretically indepen-
dent. As a consequence, there will be no variation of the system response between each
realisation of a random field. In this case the slope stability problem is just influenced by
[email protected]
the mean value and the reliability index is infinity.

133
Chapter 6 Selected case studies in uncertainty quantification including spatial variability
Cumulative probability of FOS

1
lognormal distribution
m = 50 kN/m² , COV = 20 %
0.8 Q = q / H = 2.0

0.6
exponential correlation function
-7
0.4 pf = 4.51 10
Gaussian correlation function
-9
0.2 pf = 1.85 10

spherical correlation function

0 pf = 8.51 10
-11
1.3 1.4 1.5 1.6 1.7
Factor of safety (FOS)

Figure 6.9: Influence of the spherical, exponential and Gaussian correlation function on
the probability of failure of a slope.

Influence of the correlation function: An additional parametric study is carried out

within the RFEM framework to focus on the influence of the spatial correlation function.
For this reason, the spherical, exponential and Gaussian correlation functions are used
within the Sequential Gaussian Method to generate 300 isotropic random field realisa-
tions as input for the modified Monte-Carlo approach. It can be seen in figure 6.9 that
the resulting pf values differ a lot and vary from pf = 4.51 · 10−7 to 8.51 · 10−11 . The
exponential correlation function shows the highest pf in comparison to the other func-
tions. Therefore, this exponential correlation function is used within this chapter for the
generation of random fields.

Sensitivity analysis of the reliability of a single layered soil slope: Within this pre-
sented investigation, the influence of the stochastic properties is shown. One can clearly
see the effect of an increasing or decreasing spatial correlation and COVc’ . The transfer
of these findings to a practical situation is rather difficult, because at one site there will
not be a wide range of correlation lengths or COVs as shown in figure 6.8. As shown
in the results of previous chapter 3, one will get a probability density function (pdf) of
the correlation length as a result of the Bayesian Model Averaging approach of measure-
ment data, results presented in literature, correlation lengths derived from experiments
in comparable soils and expert knowledge. Linking this with the results of the para-
metric studies shown in figure 6.8, it is possible to calculate the most probable system
behaviour due to the probability density function of the correlation length. This idea
offers the calculation of one single value of pf of the single-layered soil slope. Also the
surrogate model of polynomial-chaos-expansion (PCE) polynomial can be employed to
approximate the system response, which offers as a by product the basis for the global
[email protected]
sensitivity analysis.

134
 6.2 Slope stability

Table 6.4: Lognormal distributed input parameters of the parametric study.

mean value COV

cohesion c′ mean value µc′ 50 kN/m2 25 %
coefficient of variation COVc′ 50 % 10 %
horizontal correlation length Θhor = θhor /H 0.35 10 %
vertical correlation length Θver = θver /H 7.00 10 %

For this reason, the PCE approach is used within the concept of the Random Finite
Element Method (RFEM). This PCE mimics the system behaviour and represents the
mechanical system in a mean way as shown in appendix G. As outlined in section 4.4.2,
the global sensitivity indices can be calculated analytically by the presented formulae,
which quantify the sensitivity of the input parameters on pf .
Within this, the sensitivity of the uncertain variables for a 2D reliability analysis of the
slope is shown in figure 6.16. The mean value, the coefficient of variation and the correla-
tion lengths Θhor and Θver are considered as stochastic variables as indicated in table 6.4.
The system behaviour is approximated by a PCE at the different PCE evaluation points.
300 random field realisations are used for the computation of a mean system behaviour
at each PCE evaluation point. The results are approximated by a normal distribution
function, which allows one to extrapolate also small failure probabilities. Although this
approach is speeding up the evaluation of the system response significantly, there is still
a big computational effort for this study.
The fitting of the PCE to the system response is carried out with the approaches de-
scribed in section 4.3.3 and shown in figure 6.10. It can be derived from figures 6.10 (a)
and (b) that the expansion order M = 3 is enough to represent the system. By looking
at the determination coefficient Q2 , one can see that the fitting of the PCE is accurate
to almost 100 % for the expansion order M = 3. Besides this, it can be clearly derived
from these results that the determination coefficient, the adjusted determination coeffi-
cient and the Q2 determination coefficient quantify the fitting of the PCE-approximation
clearer in comparison to the empirical error or PRESS. Therefore, the author recommends
to used the determination coefficient and the Q2 determination coefficient in this context.
One can clearly deduce form the global sensitivity measures in figure 6.15 (c) that the
mean and COV of the cohesive soil are - as expected - the most influencing variables;
the vertical correlation length is far less important than the horizontal one. From this,
one can conclude that the vertical correlation length has a lower influence than the more
dominating horizontal correlation length. As a consequence, more effort should be put
into the investigation of the horizontal correlation length.

Effects of anisotropy of the spatial correlation on the probability of failure: As pointed

out in chapter 2.4.1, the spatial variability of natural soils is anisotropic. Hicks & Spencer
[158] point out that man-made deposits have a ratio of the horizontal and vertical corre-
lation length of 5 < θhor /θver < 25, which is smaller than the ratio of natural soils with
θhor /θver > 25 as reported by Hicks & Samy[157].
[email protected]
Hicks & Spencer [158] identified in extensive three dimensional RFEM studies three

135
Chapter 6 Selected case studies in uncertainty quantification including spatial variability

(a) (b)

PRESS
empirical error

determination coefficient
-4
x 10
PRESS 1.000
empirical error

0.01
1 0.999 det. coeff.
adj. det. coeff.
Q2 coeff.
0 0.000 0.998
1 2 3 1 2 3
expansion order M expansion order M
(c)
98 %
100
global sensitivity
index di in [%]

50

0.21 % ~0 % 1.75 %
0
mc COVc Qver Qhor

Figure 6.10: Different measures for the accuracy of the PCE to the system response of
a 2D slope stability analysis (a,b) and (c) global sensitivity measures of the
input variables.

different three failure modes, as illustrated by the typical deformed meshes and contours
of horizontal (out-of-face) displacement shown in figure 6.11. These modes depend on
the value of θhor relative to slope geometry, as defined by the slope height (HS ) and length
(L), and are summarised as follows [158, 266]:
• Mode 1: For θhor <HS there is little opportunity for failure to develop through semi-
continuous weaker zones. Hence, failure goes through weak and strong zones
alike, there is considerable averaging of property values over the failure surface,
and the slope fails along its entire length. This case is analogous to a conventional
2D analysis based on the mean value, [266].
• Mode 2: For HS < θhor < L/2, there is a tendency for failure to propagate through
semi-continuous weaker zones, leading to discrete 3D failures and a decrease in
reliability as the slope length increases. Hicks & Spencer [158] showed how proba-
bilistic theory could be used to predict the reliability of longer slopes based on the
detailed 3D stochastic analysis of shorter slopes.
• Mode 3: For θ
hor > L/2, the failure mechanism reverts to along the slope length. Al-
[email protected]
though it is similar in appearance to Mode 1, it is a fundamentally different mecha-

136
 6.2 Slope stability

Mode 1
100 %

80 %
probability of failure pf

Mode 2
60 %

Mode 1
40 % Mode 2
Mode 3
Mode 3
20 %

0%
0.8 1.0 1.2 1.4 1.6 1.8
Factor of safety (FOS)

Figure 6.11: Different failure modes derived in Hicks & Spencer [158].

nism. In this case, failure propagates along weaker layers and there is a wide range
of possible solutions that depend on the locations of these layers. The solution for
this mode is analogous to a 2D stochastic analysis.

Effects of c − ϕ correlation on the probability of failure: The results discussed so far

are based on the assumption of purely cohesive material, which can be represented by
a random variable or a random field. The next step is modelling the soil as a cohesive-
frictional material, as mentioned by Chock [81]. Moreover, Chock [81] presents a study
on the influence of the cross correlation ρc′ ,ϕ′ between the effective cohesion and the
effective friction angle on the probability of failure of slopes. The results of [81] indicate
that, for all cases of COV and correlation lengths, negative correlation correlation ρc′ ,ϕ′
leads to a lower estimate of the probability of failure, while a positive correlation ρc′ ,ϕ′
leads to a higher estimate of the probability of failure.

[email protected]

137
 Chapter 6 Selected case studies in uncertainty quantification including spatial variability

(a) (b)
3H Ls 3H 3H Ls 3H

Hs

HB 1.5 H HB

Figure 6.12: Geometry of the layered slope including one borehole using random vari-
ables (a) and random fields (b).

6.2.3 Parametric studies of two-layered soil slopes

In the previous section, the influence of soil variability on a single-layered soil slope is
investigated. This concept is extended to a more general basis. In the following, the
slope soil profiles consists of two soil layers with different properties.
The uncertainty of the layer boundary and geometry of the slope is investigated using
a the framework of reliability based design (RBD).

Uncertainty quantification using random variables At first, the uncertainties of the

strength properties are taken into account for the uncertainty quantification in case study
I.A as indicated in table 6.5. In addition to this, the depth of the lower layer is considered
as a lognormal distributed variable in case study I.B. Moreover, the boundary between
the upper and lower layer is assumed to be a horizontal line in case study I.C, as shown
in figure 6.12 (a).
In figure 6.13 the results of C ASE STUDY A are shown. One can clearly see in figure 6.13
(a) the effects of taking different random variables into account. It is clear that the system
in C ASE STUDY I.A is more reliable than in C ASE STUDY I.B, due to the uncertainty in the
geometry (HS , LS ). By considering HB as an additional random variable in C ASE STUDY
I.C, the resulting reliability is even smaller. From these results it can be deduced that,
the more uncertain variables that are introduced into a reliability analysis, the lower is
the system reliability. However, it can also be possible that additional correlation be-
tween the random variables leads to different results. Furthermore, it is shown that an
uncertain geometry of this slope problem introduces a significant uncertainty into the
stability.

Uncertainty quantification including spatial variability of soil properties at different

scales: But the consideration of the soil layer boundary as a simple horizontal line
might not capture reality. The layering can be interpreted as large scale spatial variabil-
ity, as indicated in section 2.4.1. Chiles & Delfiner [79] state that geological uncertainty
can be considered via so called categorical variables. They [79] describe the spatial dis-
tribution of soil types via so called categorical variables, which can be described by mul-
[email protected]
tivariate distributions. These concepts can only be used in the presence of detailed soil

138
 6.2 Slope stability

Table 6.5: Lognormal distributed input parameters of the C ASE STUDIES I, II and III.

upper layer lower layer layer boundary

µ COV µ COV σ/H Θboundary
C ASE STUDY I.A
cohesion c′ 10 kN/m2 10 % 11 kN/m2 20 % 0 ∞
friction angle ϕ′ 25◦ 10 % 33◦ 20 %
slope height HS 20 m deterministic
slope length LS 15 m deterministic
depth of the lower layer HB 10 m deterministic
C ASE STUDY I.B
cohesion c′ 10 kN/m2 10 % 11 kN/m2 20 % 0 ∞
friction angle ϕ′ 25◦ 10 % 33◦ 20 %
slope height HS 20 m 10 %
slope length LS 15 m 10 %
depth of the lower layer HB 10 m deterministic
C ASE STUDY I.C
cohesion c′ 10 kN/m2 10 % 11 kN/m2 20 % 0 ∞
friction angle ϕ′ 25◦ 10 % 33◦ 20 %
slope height HS 20 m 10 %
slope length LS 15 m 10 %
depth of the lower layer HB 10 m 1%
C ASE STUDY II
cohesion c′ 10 kN/m2 10% 11 kN/m2 20 % 0.5 0.10 to ∞
friction angle ϕ′ 25◦ 10 % 33◦ 20 %
Θhor = θhor /H = Θver 1.1 1.1
slope height HS 20 m deterministic
slope length LS 15 m deterministic
depth of the lower layer HB 10 m deterministic
C ASE STUDY III
cohesion c′ 10 kN/m2 10% 11 kN/m2 20 % 0.1 to 1
friction angle ϕ′ 25◦ 10 % 33◦ 20 %
Θhor = Θver 1.0 10 % 2.0 20 %
Θboundary 10 10%
slope height HS 20 m deterministic
slope length LS 15 m deterministic
depth of the lower layer HB 10 m deterministic

[email protected]

139
Chapter 6 Selected case studies in uncertainty quantification including spatial variability

(a) (b)
4 random variables 0.6

global sensitivity dPC

5 random variables
0.5
6 7 random variables
reliability index b

0.4
5
4 0.3
3 0.2
2
0.1
1
0 0.0
I.A I.B I.C c’ j’ c’ j’ HB LS HS
CASE STUDIES upper layer lower layer geometry
CASE STUDY I.A
CASE STUDY I.B
CASE STUDY I.C

Figure 6.13: C ASE STUDY I: Reliability evaluation of a two layered slope stability problem
(a) and the corresponding global sensitivity measures (b).

investigations like in reservoir engineering, which are usually not present in geotechni-
cal engineering.
Therefore, a simplified approach is set up. The combination of macro- and meso-scale
variability of soil properties is captured via a two step procedure: at first, the borderline
between the upper and lower layer is simulated via a one-dimensional random field,
which is conditioned to the soil investigation as shown in figure 6.12 (b). The SGSIM-
random fields of the strength properties of the upper and lower layer are generated and
mapped on the FEM mesh via the local averaging approach. The stochastic properties
are summarized table 6.5.
C ASE STUDY II is conducted to simulate the soil layer boundary as a simple horizontal
line and alternatively as a one dimensional random process, which is conditioned to a
soil investigation at the head of the slope as indicated in figure 6.12 (b). The properties
of both layers are shown in table 6.5. This illustrative example is showing the influence
of spatial variability at two scales; Figure 6.14 shows the scaling factor β̄ with respect to
the correlation length Θboundary = θboundary /H and the variance σboundary
2
of the stochastic
process, which is representing the boundary between the lower and upper layer. The
normalized reliability index β̄ is scaled by the reliability index β of C ASE STUDY I.A.
One can clearly see that the reliability is decreasing with a bigger variance σboundary
2
and a
longer correlation length Θboundary . The influence of the stochastic process between both
layers is vanishing in the case of an infinite correlation length Θboundary . The probability
of failure is significantly higher in the presence of spatial variability of both layers. It can
[email protected]
be concluded that the consideration of spatial variability is an important step towards

140
 6.2 Slope stability

200
2
σ boundary= 0.2
reliability index b in %
scaling factor of the
190 2
σ boundary= 1.0
180

170

160

150
0 20 40 60 80 100
correlation length Θboundary

Figure 6.14: C ASE STUDY II: Normalized reliability index of different RFEM approaches
for a two layered slope with respect to a varying Θboundary = θboundary /H.

realistic modelling of the presented problem.

Sensitivity analysis of a layered soil slope: The global sensitivity is investigated by

combining the RFEM approach with the PCE approximation of the system response. By
employing the PCE-approach, it is possible to derive the global sensitivity measures δiP C
analytically. The results of C ASE STUDY II are extended in the sense that also the uncer-
tainty of the boundary correlation length Θboundary is taken into account. The stochastic
soil strength parameters of both layers as well as the stochastic boundary between the
layers are listed for C ASE STUDY III in table 6.5.
The accuracy of PCE-approximation is estimated for different PCE expansion orders
using different determination coefficients in figure 6.15. The calculated determination
coefficients indicated are approximately 100 %, which implies that the PCE representa-
tion for M = 5 cannot be significantly improved by introducing a higher expansion order.
Therefore, the expansion order M = 5 is used for deriving the global sensitivity mea-
sures δiP C . In figure 6.16 one can see that the sensitivity for the boundary correlation
length Θboundary is more influential than the influence of the isotropic correlation length
of the upper layer. Figure 6.16 shows also the sensitivities δij between the investigated
properties. However, the other investigate sensitivities are below 10 % and therefore one
can neglect these sensitivity measures.

6.2.4 Conclusions
This case study investigates reliability of single- and a two-layered soil slopes. Herein, a
linear elastic perfectly plastic constitutive model on basis of the Mohr-Coulomb criterion
is used in a 2D FEM model.
At first, the effects of spatial soil variability are quantified for single layered soil slopes,
[email protected]
which are compared to slopes with random and not spatially correlated soil properties.

141
Chapter 6 Selected case studies in uncertainty quantification including spatial variability
determination coefficient

1.00

0.98
determination coefficient
0.96 adjusted determination coefficient
Q2 coefficient
0.94

1 2 3 4 5
expansion order M of the PCE

Figure 6.15: C ASE STUDY III: Different determination coefficients for the PCE accuracy.
glolbal sensitivity of the parameters

0.5 1 ... isotropic correlation length of the upper layer

2 ... isotropic correlation length of the lower layer
0.4
3 ... correlation length of the layer boundary

0.3

0.2

0.1

0 PC PC PC PC PC PC PC
d3 d1 d13 d2 d23 d12 d123

Figure 6.16: C ASE STUDY III: Global sensitivity measures δiP C for the isotropic correlation
length of the upper layer (1) and the lower layer (2) as well as for the layer
boundary (3) for the PCE expansion order M = 5.

[email protected]

142
 6.2 Slope stability

Within the investigation of the effects of different correlation function on the reliability
of single layered soil slopes, the exponential correlation function is identified as the most
conservative one. Moreover, it is shown that different complex correlation structures of
spatial variability, which are simulated by SGSIM and SISM algorithms, do not have a
major effect on the calculated probability of failure.
The influences of mean value, coefficient of variation, correlation function and random
field generators are investigated within the RFEM framework. Besides this, the effects
of the mean value, the coefficient of variation and the anisotropy of the horizontal and
vertical correlation length are quantified within a novel sensitivity analysis within the
RFEM framework. It is shown that the mean valued, the coefficient of variation and
the horizontal correlation length have the biggest influence on the probability of failure,
whereas the contribution of the vertical correlation length is very small.
The effects of spatial variability at different scales is in the focus of additional paramet-
ric studies on two-layered soil slopes. Again the importance of spatial soil variability is
highlighted by comparing the effects of random properties and spatially correlated ran-
dom properties. Within this study the effects of an uncertain geometry are investigated.
It is shown that the slope geometry and the soil layering have a bigger influence in com-
parison to the stochastic soil properties. The effects of spatial soil variability at different
scales are quantified with respect to random soil properties. The spatial variability at the
large scale and the spatial variability of the upper layer are contributing the most to the
probability of failure of the investigated two layered soil slope.
The presented studies are fundamental investigations on the effects of soil variabil-
ity. These studies can be enriched by additional investigations using 3D slope stability
analyses, which would contribute to the quantification of the effects of spatial variability
for slope analyses. In addition to this, the investigation of the uncertain slope geometry
would offer additional insight into 3D slope stability analyses.
The presented findings can be further extended by additional 3D slope stability anal-
yses, which would represent more exactly the spatial variability of the subsoil.

[email protected]

143
Chapter 6 Selected case studies in uncertainty quantification including spatial variability

6.3 Risk based characterisation of an urban site

When planning new surface or underground infrastructures, it is vital to anticipate what
geological conditions are likely to be encountered even before taking any specific survey
or investigations works, as stated by Raspa et al [301]. An improved assessment of ge-
ometry and mechanical properties of underground layers would minimize project risks,
from an economical point of view as well as from a technological point of view.
Faber & Stewart [122] define risk is a random event that may possibly occur and, if
it did occur, would have a negative impact on the goals of the project. Thus, a risk is
composed of three elements: the scenario, its probability of occurrence; and the size of
its impact, if it did occur. Geological risk is defined by De Marsily et al. [88] as any
geological conditions, which represents a potential threat to health, safety or welfare of
people. Public administrations, civil protection agencies and research institutions are
continuously involved in managing environmental hazards and planning future devel-
opments of urban areas, which both require in-depth knowledge of the subsoil as stated
by Raspa et al. [301] amongst others [127, 149, 273].
The aim of this case study is to show a procedure of risk based site characterisation,
which is employed to quantify the effects of macro-scale soil variability. Herein, the
damage estimation of buildings is performed, which is triggered by the uncertainty of
geology. This damage estimation of rigid buildings is performed by the calculation of
differential settlements of buildings using an advanced constitutive model. The uncer-
tainty of geology is simulated via the Pluri-Gaussian simulation approach, which is cap-
turing the uncertainty of the domain boundaries of three spatially distributed soil types.
Moreover, this approach considers soil investigations, stochastic soil properties and ex-
pert judgement within a sound mathematical framework of categorical random field
simulation.

6.3.1 Site description

Geological description of the subsoil: The investigated site is located in the urban area
of Rome. Additional information on the location of the site and its geological settings
are given in the Appendix H. The model of the subsoil of Roma consists of the integra-
tion and analysis of the main available geological and geotechnical data: stratigraphy,
lithology and texture, physical and mechanical properties, and hydrogeology. Informa-
tion from more than 6,000 boreholes and measured stratigraphic logs, geological maps,
and in-situ tests were homogenized, classified, and archived in a database thus far, as
described in Raspa et al. [301]. Geological information retrieved from the database was
interpreted and encoded to reconstruct the stratigraphic framework of the Tevere Valley.
Attention was specifically focused on the upper Pleistocene–Holocene alluvial deposits.
The investigated site is a part of a complex geological system. This fluvio-deltaic area
consists of complex channels and floodplain areas, which changed over time. Due to
this, the complexity of the subsoil is almost impossible to describe via deterministic ap-
proaches properly. Therefore, geostatistical simulation approaches offer a proper way to
[email protected]
model the subsoil-uncertainty.

144
 6.3 Risk based characterisation of an urban site

Table 6.6: Mean values and standard deviations of the lognormally distributed measure-
ment and literature data.

c′ in kN/m2 ϕ′ in degree E in MN/m2

µ σ µ σ µ σ
measurements
I gravel, medium and coarse sand 0.0 0.0 39.4 3.2 42.7 11.7
II silty fine sand, sandy silt 9.9 8.3 35.3 3.2 16.3 6.0
III silty clay, clayey silt, organic clay 14.6 11.6 23.8 3.2 3.3 1.6
IV bedrock 75.9 2.9 23.5 1.0 61.0 0.0
literature data
I gravel, medium and coarse sand 0.0 0.0 31.5 1.8 45.0 12.5
II silty fine sand, sandy silt 2.5 1.3 36.0 2.0 27.5 6.3
III silty clay, clayey silt, organic clay 10.0 2.5 25.0 2.5 7.5 1.8
IV bedrock 76.0 3.8 24.0 1.2 610.0 30.5
data combinded via the Bayesian approach
I gravel, medium and coarse sand 0.01 0.0 33.3 1.5 43.9 8.4
II silty fine sand, sandy silt 2.7 1.2 35.8 1.7 21.8 4.4
III silty clay, clayey silt, organic clay 10.2 2.4 24.6 2.0 51.6 1.2
IV bedrock 75.9 2.1 23.7 0.8 61.0 0.1

Information from more than 2,000 boreholes penetrating these deposits was used by
Marconi [238] to model their lithological and textural associations in a conceptual geo-
logical model. This model has been used as the basis for the geostatistical simulations.
For this case study, only geotechnical boreholes with geotechnical information penetrat-
ing these deposits were selected to derive geomechanical properties of the subsoil. This
set includes 283 boreholes and 719 samples.
Five main soil types are derived from this large dataset and summarized in table 6.6,
which have been used in the latter to simulate the soil behaviour. S OIL TYPE I consists
of gravel, medium and coarse sand, S OIL TYPE II of silty fine sand, and sandy silt, S OIL
TYPE III of silty clay, clayey silt and organic clay and S OIL TYPE IV of bedrock. Table
6.7 provides the number of measurement data for the strength and stiffness properties
of the different soil types. The author combined the measurement data with literature
[346] data using the Bayesian approach. Via this, the description of the variability of the
subsoil is enriched by the combination of measurements and expert judgement.

Simulation of the geological uncertainty. The uncertainty of the geology is seen as un-
certainty of different soil types, which includes the uncertainty of the domain boundaries
of the spatially distributed soil types. This is done via the Pluri-Gaussian Simulation ap-
proach, which is simulating spatially distributed categorical variables. The principle of
the Pluri-Gaussian Simulation is to simulate one or several continuous Gaussian fields
and to truncate them in order to produce a categorical variable. Armstrong et al. [14]
[email protected]
describe the concept of the Pluri-Gaussian simulation approach in depth and illustrate

145
Chapter 6 Selected case studies in uncertainty quantification including spatial variability

Table 6.7: Number of measurements inside each soil type.

soil types c′ ϕ′ E
number of measurement samples
I gravel, medium and coarse sand 8 8 4
II silty fine sand, sandy silt 40 40 3
III silty clay, clayey silt, organic clay 86 86 92
IV bedrock 2 2 1

the basic idea, as shown in figure 6.17: Two Gaussian random fields (figure 6.17 a,b)
are used to describe the main characteristics of the soil, e.g. anisotropies. The different
soil types are generated by truncating several Gaussian random fields (figure 6.17 a,b).
This provides flexibility to handle complex transitions and anisotropy among the soil
types. The truncation rule is called the lithotype rule (figure 6.17 c) because it synthe-
sizes transitions between the soil types. One can easily deduce from this that the choice
of a lithotype rule is a major step of the methodology. In this case study, the lithotype
rule was derived from borehole logs providing a good basis, on which soil types can
and cannot be in contact. These proportions are not constant over the domain, but vary
vertically and laterally because of the existence of trends in the geological processes as
shown in Appendix H. The mathematical theory of the Pluri-Gaussian [215, 241] meth-
ods is described in detail in by Marconi [238] amongst others [14, 118]. This includes the
conditioning of the unconditional Pluri-Gaussian random field to borehole data via the
Gibbs sampler.
In Appendix H, the details of the Pluri-Gaussian Simulations are listed, which have
been provided by Marconi [238]: The plan view of the investigated area, the spatial
statistics of the soil types, the lithotype rules and illustrative illustrations of one realisa-
tion of the Pluri-Gaussian Simulations are shown in Appendix H.

6.3.2 Calculation of fragility curves

As mentioned in section 5.3, fragility curves are useful tools to quantify the uncertainty
of the soil: the deterministic load on a footing is stepwise increased, while the probability
of damage is evaluated. The probability of damage pdamage is defined as the probability
of exceeding a defined level of allowable settlements αultimate . These fragility curves are
used as the basis for the risk based site characterisation as described in the next section
6.3.3.
The area of this case study is shown in Appendix H in figure H.4 and is subdivided
into 35 areas, which are used as input for the 2D RFEM calculations. The fragility curves
are calculated form the RFEM results.

Mechanical description of the problem: Figure 6.19 shows the 2D FEM mesh of a
rigid building (B/W = 30/15 m), which is used for the evaluation of the load - displace-
[email protected]
ment curves. Within the stepwise construction of the rigid building (B=30 m) differential

146
 6.3 Risk based characterisation of an urban site

(a) Y1 Y2 (b)

Y2
(c) (d)
(2)
(1)
(3)
Y1

Figure 6.17: Principles of the Pluri-Gaussian Simulation approach, [14]:

Two uncorrelated Gaussian random fields Y1 (a) and Y2 (b) with different
anisotropies are combined via the lithotype rule (c); the red square high-
lights the way the lithotype rule is used to construct the Pluri-Gaussian ran-
dom field simulation (d).

antropic backfill
Soil type I
Soil type II
Soil type III
Soil type IV

Figure 6.18: Part of the realisation of the Pluri-Gaussian random field (200/200/75m) in-
cluding all soil types.

[email protected]

147
Chapter 6 Selected case studies in uncertainty quantification including spatial variability

4.5 B B 4.5 B
q
P1 P2

3B
Figure 6.19: 2D FEM model used in this case study.

settlements between points P1 and P2 are observed, which are related to the spatial dis-
tribution of the 5 different soil types, as shown in table 6.6. Herein, the contact between
the rigid building and the subsoil is assumed to be rigid with full bounding.
The soil behaviour is simulated by a small strain, double hardening model, which was
developed by Benz [39]. This advanced, non-linear soil model is able to take the small
strain behaviour as well as the non-linear stress-strain relationship of soil into account.
The input variables of this soil model are given in the Appendix H in table H.1.

Stochastic soil properties: Although there is a large number of borehole data, it is quite
difficult to estimate the strength and stiffness parameters of the different soil types: Most
of these boreholes aimed for a geological description and only a relative small number
of boreholes focused on geotechnical characterisation as indicated in table 6.7. There-
fore, the measurement data are combined with expert knowledge from the Geotechnical
Handbook [346] via the Bayesian approach as proposed by Ching et al. [80]. The results
of this combination are given in table 6.6. One can clearly see that the combined values
of the cohesion, friction angle and E-modulus show a smaller standard deviation. These
lognormal distributed values are used for the random number generation of the strength
and stiffness parameters of each soil type. These random variables, which are based on
the mean values and standard deviations in table 6.6, are used in combination with the
300 Pluri-Gaussian random field realisations for the description of the geological hetero-
geneity. Each Pluri-Gaussian random field realisation has a length of 620 m, a width of
436 m and an overall-depth of 75 m with a horizontal discretisation of ∆X = ∆Y = 5 m
and a vertical discretisation of ∆Z = 0.05 m. These Pluri-Gaussian random fields were
generated by Marconi [238].
Due to the scarcity of data describing the mechanical behaviour of anthropic backfill
[email protected]
material, this material is not considered in the subsequent investigations.

148
 6.3 Risk based characterisation of an urban site

Random Finite Element approach & fragility curves: As mentioned above, the 3D
Pluri-Gaussian random fields are used for the description of the geological uncertainty.
A modified RFEM approach is used to quantify the effects of geological uncertainty.
Within the preprocessing phase of this case study, a simple mapping procedure has been
used to map the 3D random field onto the 2D FEM mesh. The Pluri-Gaussian random
field is divided into blocks (5B × 10B × 75 m). The transformation from 3D to 2D is
carried out by averaging the soil properties in the third direction. Then the finer 2D
random field is averaged over the coarser FEM mesh.
Within the processing phase, the differential settlements of the rigid building due to a
load up to q = 2, 000 kN/m2 are evaluated by using the RFEM framework.
The boundaries of the mechanical model are chosen so as to minimize the influence on
the settlement prediction as indicated in figure 6.19. This follows the recommendations
for numerical modelling in geotechnical engineering in [291, 292].
As the first step within the postprocessing phase, the fragility curves are evaluated in the
middle of the building, which is shifted over the entire field. The fragility curves describe
the probability of exceeding the ultimate differential settlements due to a load, which is
increased to q = 2, 000 kN/m2 . As shown in literature e.g. [286, 289], this fragility curve
can be approximated by a cumulative lognormal distribution function, which offers the
possibility of a continuous definition of fragility due to differential settlements. For this
purpose the probabilities of exceeding the ultimate differential settlement (αultimate =
1/300, 1/500, 1/600 and 1/1, 000) for the different load levels are evaluated. Due to the
fact that especially for small load levels the probability of exceeding αultimate is very
small, the system response was approximated via a normal distribution function. Via
this, it is possible to estimate the probability of exceeding αultimate . One has to keep in
mind that the absolute numbers of these probabilities are just approximations and offer
a good basis to compare two different locations of the building.

6.3.3 Evaluation of the ultimate loads according to the state-of-the-art

The E UROCODE defines the probability of exceeding of the serviceability limit state (SLS)
with pexceeding = 1.91 · 10−3 . This probability pexceeding is employed to estimate the ulti-
mate load qSLS by using the fragility curves. In figure 6.20 the fragility curve for ultimate
differential settlements αultimate = 1/1, 000 is shown. The lognormal cumulative distri-
bution function is fitted to the values via least square fitting, which have been derived
from the RFEM calculations. This lognormal cumulative distribution function is used
for the calculation of the load qSLS , which does not exceed the probability of the SLS for
differential settlements αultimate = 1/1, 000. One can extrapolate from this that for some
regions, which have a very low probability of exceeding the ultimate differential set-
tlements αultimate between the loads q = 0 kN/m2 to 2, 000 kN/m2 , the load qSLS might
result in a higher value of the footing pressure than in the performed calculations. There-
fore, these values are just valid for qualitative comparison of different positions and their
susceptibility towards the probability of occurrence of differential settlements.
One has to keep this in mind while studying figure 6.21 and figure 6.22. These fig-
[email protected]
ures show the boreholes and the 35 locations of the midpoints of the building, which

149
Chapter 6 Selected case studies in uncertainty quantification including spatial variability

diff. settlement limit aulitmate=1/1,000

-3
10
probability of exceeding 3.5
3.0
2.5
-3
pexceedance = 1.91·10
2.0
1.5
1.0 qSLS = 1,711 kN/m²
0.5
0.0
0 400 800 1,200 1,600 2,000
footing pressure q in kN/m²

Figure 6.20: Sample fragility curve for ultimate differential settlements

αultimate = 1/1, 000 and the fitted lognormal cumulative distribution
function.

are indicated by black points. The orange squares indicate the location of the boreholes
in the figures 6.21 and 6.22. The ultimate loads qSLS , which are calculated for four dif-
ferential settlement limits αultimate = 1/300, 1/500, 1/600, 1/1, 000. These ultimate loads
qSLS are plotted as grey-shaded regions in the background of figures 6.21 and 6.22. The
grey-shaded regions between the points are interpolated using the Ordinary Kriging al-
gorithm. By looking at the figures 6.21 and 6.22, one can clearly identify areas, which are
less and more endangered to differential settlements under a specific load.

6.3.4 Conclusion
The presented case study presents a methodology for risk based site characterisation.
This approach uses geostatistical simulation techniques and FEM modelling to derive
the ultimate load over a site. Via the presented methodology, the geological uncertainty
is combined with stochastic, geotechnical soil properties, which are based on measure-
ment data and on expert judgement. The uncertainty of the complex geological condi-
tions is quantified within a mathematical framework. By using RFEM in combination
with fragility curves, these results are translated into into maps, which offer another in-
sight into the effects of soil heterogeneity and the resulting risk, which is a novelty at the
interface between engineering geology, geostatistics and geotechnical engineering. A
by-product of this methodology of risk based site characterisation is the quantification
of the effects of macro-scale soil variability.
By using the Pluri-Gaussian Simulation algorithm one has to be aware that modelling
geological uncertainty by using different Gaussian random fields is a simplification. As
pointed out in chapter 2, a Gaussian random field is a simple description of spatial vari-
[email protected]
ability using a mean value, a standard deviation and one single covariance function.

150
 6.3 Risk based characterisation of an urban site

(a)

300
relative y-coordinate in meter

250

200
midpoints of
150 the building
boreholes
100

50

0 qSLS =
4,500 kN/m²
0 50 100 150 200 250
4,000 kN/m²
relative x-coordinate in meter
(b) 3,500 kN/m²

300 3,000 kN/m²

relative y-coordinate in meter

2,500 kN/m²
250
2,000 kN/m²
200
1,500 kN/m²
150 1,000 kN/m²

100 500 kN/m²

50

0 50 100 150 200 250

relative x-coordinate in meter

Figure 6.21: Map of allowable loads pSLS on a footing with αulitmate = 1/300 (a) and
αulitmate = 1/500 (b).
[email protected]

151
Chapter 6 Selected case studies in uncertainty quantification including spatial variability

(a)

300
relative y-coordinate in meter

250

200
midpoints of
150 the building
boreholes
100

50

0 qSLS =
4,500 kN/m²
0 50 100 150 200 250
4,000 kN/m²
relative x-coordinate in meter
(b) 3,500 kN/m²

300 3,000 kN/m²

relative y-coordinate in meter

2,500 kN/m²
250
2,000 kN/m²
200
1,500 kN/m²
150 1,000 kN/m²

100 500 kN/m²

50

0 50 100 150 200 250

relative x-coordinate in meter

Figure 6.22: Map of allowable loads pSLS on a footing with αulitmate = 1/600 (a) and
αulitmate = 1/1, 000 (b).
[email protected]

152
 6.3 Risk based characterisation of an urban site

Moreover, one has to keep in mind that the lithotype is a subjective interpretation of ge-
ological conditions, although the lithotype rule is based on the analysis of a large set of
boreholes.
However, the geological uncertainty can also be simulated by other simulation ap-
proaches in this context, which are able to simulate categorical variables described in
literature [305].
Besides this, the Pluri-Gaussian simulation algorithm needs a large number of bore-
holes and laboratory tests to perform well. The stochastic properties of the Gaussian
random fields and the definition of the lithotype rule needs a quite large database of
boreholes. Otherwise, this simulation algorithm cannot be recommended for similar
investigations in the case of a small dataset.
The presented case study on the risk based characterisation of an urban site extends
the state of research presented in different contributions. Amongst others [24, 172, 312],
Fenton [127] focused on the effects of meso-scale soil variability and neglected the macro-
scale variability. On the basis of these results, Chen et al. [73] combine the micro- and
the meso-scale variability via a multi-scale random field approach, which is applied to
the ultimate limit state and to the serviceability limit state of strip footings.
However, the studied contributions [24, 127, 172, 312] do not take into account the
available prior information like literature knowledge or soil investigations, a concep-
tual geological model or subjective engineering judgement. Therefore, the presented
case study offers a promising scheme of evaluating the effects of geological, macro-scale
uncertainty in the presence of a large borehole database. Moreover, the presented ap-
proach offers a realistic interpretation of spatially variable site conditions via the adopted
fragility curve approach, which allows a realistic estimation of the risk of differential set-
tlements. The results of this are transformed back to a limiting load, which is obeying
the probability for SLS given in the E UROCODE 7. Via this, it is possible to identify
endangered areas of a building site and this procedure quantifies the consequences of
geological uncertainty.

[email protected]

153
Chapter 6 Selected case studies in uncertainty quantification including spatial variability

6.4 Synopsis of the case studies

The presented case studies contribute to the description of uncertain performance of
geotechnical structures in the ultimate and serviceability limit states. The effects of meso-
and macro-scale variations of soil properties are investigated within three different case
studies of geotechnical problems. These investigations show a possible way to combine
two different scales of variability.

Tunnelling induced settlements: The RFEM approach is applied to the evaluation of

tunnelling induced settlements. The effects of stochastic properties and the spatial vari-
ability of a single layered soil on the surface settlements is investigated. This approach
is extended to a two layered system, which captures different scales of variability. The
macro-scale variability is captured via a random process, which is conditioned to bore-
holes. The meso-scale variability inside the soil layers is simulated via random fields.

Slope stability: The reliability of single-layered soil slopes is investigated by means

of RFEM. Within this, different random field generators are investigated and global
sensitivity measures are used to quantify the influence of uncertain input parameters
(µc , COVc , Θver , Θhor ). The effects of multi-scale soil variability are considered via a sim-
plified approach, similar to chapter 6.1. The effects of soil variability become clearer by
comparing the results to a simplified approach using random variables and neglecting
spatial variability. Within this, the uncertainty of the slope geometry is quantified in
comparison to soil variability. Global sensitivity measures quantify the influence of the
multi-scale spatial variability.

Risk based characterisation of an urban site: This case study investigates the effects of
geological uncertainty. The different soil types are simulated as categorical variables by
means of the Pluri-Gaussian simulation approach. This approach considers soil investi-
gations and geostatistical subsoil characteristics as well as engineering judgement. These
quantifications of macro-scale soil variability are used for the calculation of the footing
serviceability. The footing serviceability is investigated by differential settlements and
means of fragility curves. These fragility curves are the basis for risk maps, which vi-
sualize the effects of geological uncertainty and identifies endangered area. This study
clearly quantifies the influence of macro scale soil variability and indicates clearly the
need to consider this.

[email protected]

154
 Chapter 7
Summary and conclusions

7.1 Potential and limitations of probabilistic methods

Geotechnical design is traditionally based on deterministic analysis using global or par-
tial safety factors to take soil variability into account. These traditional approaches are
mainly based on experience. Therefore, there is a need to evaluate the safety margins in
the geotechnical design within a mathematical framework by using probabilistic meth-
ods to evaluate the probability of failure. Fluctuations of loads, variability of material
properties, uncertainty in analysis models etc. all contribute to a failure probability [389].
The reliability, defined as the complement of the failure probability, is a rational measure
of safety. Reliability methods deal with the uncertain nature of loads, resistance etc. and
lead to assessment of the reliability. Reliability methods are based on established anal-
ysis models for the structure in conjunction with available information about loads and
resistances and their associated uncertainties. The analysis models are usually imper-
fect, and the information about loads and resistances is usually incomplete. Therefore
the reliability as assessed by reliability methods is generally not a purely physical prop-
erty of the structure in its environment of loads. It is rather a nominal measure of the
safety of the structure, given a certain analysis model and a certain amount and quality
of information.
Correspondingly, also the estimated failure probability is dependent on the analysis
model and the level of information, and it can therefore usually not be interpreted as the
frequency of occurrence of failure for that particular type of structure [389].
Measuring the safety of a structure by its reliability makes the reliability a useful de-
cision parameter. Fulfilment of a requirement to the reliability is then necessary in order
to ensure a sufficient safety level in design. Such a requirement can either be derived
by a utility optimisation in a decision analysis, or by requiring that the safety level as
resulting from the design by a reliability analysis shall be the same as the safety level
resulting from current deterministic design practice.
Probabilistic analyses are powerful and versatile tools for investigating the influence
of uncertainties on a given geotechnical problem. However, one has always to keep
in mind that the probabilistic analyses are done on the basis of the given input, but
not from an inherent understanding of the statistics and physics of the problem. Thus,
probabilistic analyses may be most beneficially used for enhancing the understanding
of a physical problem that has already been identified. Preferably, the problem should
even be largely understood aside of the probabilistic analysis itself.
[email protected]
The relative influence of the most important factors (e.g. variance or spatial corre-

155
Chapter 7 Summary and conclusions

lation) may be studied via parametric studies, and this will bring confidence to the
predicted behaviour in a given design situation. The way to conduct an uncertainty
quantification in geotechnical engineering is shown in case studies onto tunnel lining,
tunnel face stability, settlements introduced by tunnelling, slope stability, bearing capac-
ity of vertically loaded strip footings and serviceability of footings. In these case studies
semi-analytically defined and numerically simulated limit states are investigated with
respect to their behaviour due to random variables and random fields. The contribution
of the input variables is analysed by means of local and global sensitivity analyses. The
presented case studies will show the potential of the framework of uncertainty quantifi-
cation.
It is also very important to be aware of the limitations that lie in a probabilistic analy-
sis. The presented approaches in chapter 2 allow to consider pre-knowledge from expert
judgement, literature and/or available field tests, which can enrich a stochastic site char-
acterisation. However, there is a need to conduct a bigger soil investigation campaign
compared to the state-of-the-art because more input data are needed to derive statistical
distributions and parameters describing spatial variability.
Besides this, another difficulty in using probabilistic concepts in applied engineer-
ing is the difficult statistical background, which is needed to understand the results of
probabilistic analyses. Within chapter 3, a large number of field tests is analysed and
contributes to the database on stochastic soil properties, which is assembled by the re-
sults of a large literature review in appendix A. The basics of safety and uncertainty
are summarized in chapter 4 and shall help to develop an understanding of uncertainty
quantification and to interpret results of probabilistic analyses presented in chapters 5
and 6. These case studies show the application of uncertainty quantification and shall
guide the reader to a comprehensive understanding of the presented approaches. These
different case studies in tunnelling and foundation engineering show the effects of soil
variability and spatial variability at different scales. Moreover, sensitivity analyses are
carried out to investigate the contribution of each random parameter to the probability
of failure.

7.2 How well do we need to know soil variability?

Practitioners may ask how well do we have to know soil variability in order to make
proper predictions. Generally spoken: The more we know about the subsoil, the better
the predictions will be. The subjective engineering judgement shall be enriched by using
the framework of uncertainty quantification to consider soil variability. The framework
of uncertainty quantification offers a sound mathematical framework for a rigorous con-
sideration of errors and uncertainties within a geotechnical design process. As shown
in the case studies in the chapters 5 and 6, different limit state formulation ranging from
semi-analytical to FEM defined can be incorporated depending on the level of accuracy.
Moreover, also different uncertainties can be considered via (cross-correlated) random
variables or random fields. But this framework needs statistical input parameters, which
can be derived from soil samples of a site investigation. With a site investigation strategy,
[email protected]
soil sampling strategies are a key-point in detecting soil variability within site character-

156
 7.3 What is the best way to evaluate the effects of soil variability?

isation. Different sampling concepts are offered in literature [257] e.g. simple random
pattern, stratified random pattern or cluster sampling, and are used in groundwater
ecology and geostatistics to guide the choice of additional soil samples. These concepts
mainly focus on the properties of different homogeneous soil layers, which can be evalu-
ated according to the statistical methods described in guidelines e.g. [195, 390]. It can be
deduced from the case studies in chapter 6 that the spatial variability of soil properties
has significant influence on the behaviour of geotechnical structures. Therefore, it is nec-
essary to quantify theses effects in a rational manner. Not only literature data but also
field investigations have to be conducted to investigate the the spatial correlation of soil
properties. Consequently, the sampling schemes for detecting spatial variability at dif-
ferent scales have to be used from geostatistics [79, 248] or soil science [314] to evaluate
the these properties in an economic and cost effective way.

7.3 What is the best way to evaluate the effects of soil

variability?
The quantification of the effects of soil variability is probably one of the most important
issues in geotechnical design. By using probabilistic methods, the recommendation for
the most reliable way to calculate the probability of failure is an easy task. Focusing
purely on the evaluation of the calculation of the probability of failure, it is the Monte-
Carlo method. The Monte-Carlo method is a robust method for the evaluation of the
probability of failure, although it is also the most time consuming one in comparison to
e.g. FORM. As elaborated in section 4.3, different authors published various methods to
evaluate the probability of failure apart from the Monte Carlo methods to overcome this
drawback.
While reading the case studies in chapters 5 and 6, one can clearly see not only the
effects of representing soil variability using random variables and random fields. Sensi-
tivity analyses quantify the contribution of each uncertain variable and help the engineer
to simulate effects of the important random variables, which helps to simplify the cal-
culation. As a consequence, the evaluation of the probability of failure is speeded up by
taking just the important variables into account.
Within this context, it has to be stressed that another important issue is the complexity
of the limit state equation. The limit state equation is describing the system behaviour
via a close form solution, an empirical equation, a numerical simulation model or a sur-
rogate model. The best way to simulate the system response is the use of a numerical
model in 3D together with an advanced constitutive model. But this cannot fulfil the
requirements for a fast evaluation of the probability of failure. Therefore, the engineer
has to make compromises to keep the mechanical simulation model as simple and as ac-
curate as possible for a realistic description of the soil structure interaction as indicated
by Potts et al. [291]. Within this, the approach of using surrogates or meta-models offers
a promising way in this context.
There is no single answer to the question: which scales of variability do we have to
[email protected]
consider in geotechnical design? Generally spoken, the very small scales of spatial vari-

157
Chapter 7 Summary and conclusions

ability do not play a major role. As shown in chapter 6, micro-scale variations (e.g. at
the grain-size level) have minor effects on the overall failure probability of a structure,
whereas large geological spatial variations (e.g. soil layers) can have significant effects
as shown in the case studies. One can deduce from the presented case studies that the
effects are significant for the combination of meso- and macro-scale spatial soil variabil-
ity. The presented engineering approach is taking these two scales of spatial variability
into account: The soil layer boundary is represented by a conditional, stochastic process,
which is separating layers with spatial variable properties. By using the Polynomial-
Chaos-Expansion (PCE) in connection with the Random Finite Element approach for the
representation of the system response, it is possible to consider not only a single corre-
lation length, but a lognormal distribution of correlation lengths. Moreover, the global
sensitivity measures of the uncertain variables can be derived from the PCE analytically.
This offers a possible way to quantify the effects of multi-scale spatial variability of soil
properties.

7.4 Recommendations for research

The following recommendations for future research in science, education and applied
engineering can be drawn from this research.

Science: Additional studies should be conducted to gain more knowledge and experi-
ence in the of stochastic quantification of soil properties. The presented literature study
and the results of the case studies analysing a large set of CPT tests offer a good basis
for this. This would form a good starting point for the formulation of a guideline for the
evaluation of spatial variability and heterogeneity comparable to recommendations like
[11, 17, 67, 95]. The need for cost effective sampling schemes is closely linked with this.
Apart from the presented Methods of Moments and Maximum Likelihood approaches
in chapter 2, the author recommends inverse methods like in geostatistics or earthquake
engineering to investigate the effects of soil variability. These methods are based on
Bayesian approaches [164] or sequential approaches[171, 304] amongst others.
Moreover, the quality of the random field representation within the framework of un-
certainty quantification and reliability based design is an important task. The forecast
quality in meteorology [309] might be a staring point for further developments in this
context .
Apart from this, the partial safety factors in the EC7, which are mainly based on expe-
rience, should be enriched and extended with the results from additional investigations.
This would lead to a more precise separation between soil uncertainties and human re-
lated errors.
Another interesting part in the context of economic design is the optimization of a
geotechnical structure including uncertainties. Reliability based design optimization
includes different failure modes of complex structures incorporating the variability of
[email protected]
loads and resistance forces in a proper way.

158
 7.4 Recommendations for research

Education: In order to introduce probabilistic concepts in applied sciences, it is nec-

essary to make engineers familiar with statistics and concepts of reliability. Therefore,
it would be necessary to teach students in these fields and to teach them the linguis-
tic aspects to be able to speak and exchange with experts from applied mathematics or
physics. Mathematics and physics have a longer tradition in using the concepts of statis-
tics in their fields.

Applied engineering: Guidelines and worked examples would be the best arguments
to convince applied engineers to used reliability based design approaches. This has al-
ready been started by different groups like GeoSNET [360] or JCSS [399] amongst others
[262, 280]. It has to be pointed out that user friendly software tools would support this,
which would allow one to estimate the reliability of complex systems in a fast and effi-
cient way by using robust and efficient algorithms.

[email protected]

159
[email protected]
 Bibliography
[1] M. Aboufirassi and M.A. Marino. Cokriging of aquifer transmissivities from field
measurements of transmissivity and specific capacity. Mathematical Geology, 16(1):
19–35, 1984.

[2] M. Abramovich and I.A. Stegun. Handbook of Mathematical Functions, Graphs and
Mathematical Tables. National Bureau of Standards Applied Mathematics, Wash-
ington, 1964.

[3] M. Abramovich and I.A. Stegun. Handbook of Mathematical Functions with Formulas,
Graphs and Mathematical Tables. Dover, 1974.

[4] A. AghaKouchak, A. Bárdossy, and E. Habib. Conditional simulation of remotely

sensed rainfall data using a non-gaussian v-transformed copula. Advances in Water
Resources, 33(6):624–634, 2010.

[5] H. Ahrens, E. Lindner, and K.H. Lux. Zur Dimensionierung von Tunnelausbauten
nach den Empfehlungen zur Berechnung von Tunneln im Lockergestein (1980).
Die Bautechnik, 8:260–311, 1982.

[6] H. Akaike. Information theory and an extension of the maximum likelihood prin-
ciple. In Proceedings of Second International Symposium on Information Theory, vol-
ume 1, pages 267–281. Springer Verlag, 1973.

[7] H. Akaike. A Bayesian analysis of the minium AIC procedure. Ann. Inst. Statist.
Math., 30:9–14, 1978.

[8] R. Al-Khoury, K.J. Bakker, P.G. Bonnier, H.J. Burd, G. Soltys, and P.A. Vermeer.
PLAXIS 2D Version 9. R.B.J. Brinkgreve and W. Broere and Waterman, 2008.

[9] D. Alber and W. Reitmeier. Beschreibung der räumlichen Streuungen von Bodenken-
nwerten mit Hilfe der Zeitreihenanalyse, volume 7. Technische Universität München,
1986.

[10] E.E. Alonso and R.J. Krizek. Stochastic formulation of soil properties. In Proc. 2nd
Int. Conf. on Applications of Statistics & Probability in Soil and Structural Engineering,
Aachen, volume 2, pages 9–32, 1975.

[11] American Society of Civil Engineers. Standard Guideline for the Geostatistical Esti-
mation and Block-averaging of Homogeneous and Isotropic Saturated Hydraulic Conduc-
tivity. American Society of Civil Engineers, Environmental and Water Resources
[email protected]
Institute, 2010.

161
Bibliography

[12] G. Anagnostou and K. Kovari. Face stability conditions with earth-pressure-

balanced shields. Tunnelling and Underground Space Technology, 11(2):165–173, 1996.

[13] A.H.S. Ang and W.H. Tang. Probability concepts in engineering planning and design,
vol. 2: Decision, risk, and reliability. JOHN WILEY & SONS, INC., 1984.

[14] M. Armstrong, A. Galli, H. Beucher, G. Loc’h, D. Renard, B. Doligez, R. Eschard,

and F. Geffroy. Plurigaussian simulations in geosciences. Springer, 2011.

[15] A. Asaoka and D.A. Grivas. Spatial variability of the undrained strength of clays.
Journal of the geotechnical engineering division, 108(5):743–756, 1982.

[16] D. Assimaki, A. Pecker, R. Popescu, and J. Prevost. Effects of spatial variability of

soil properties on surface ground motion. Journal of earthquake engineering, 7:1–44,
2003.

[17] ASTM D 5549-94. Standard Guide for the Contents of Geostatistical Site Investiga-
tion Report, 2001.

[18] R. Azzouz and C. Bacconnet. Optimal approach to the planning of a network of in

situ soil test. In Proc. 6th Int. Conf. on Applications of Statistics and Probability in Soil
and Structural Engineering, pages pp. 654–659, Mexico City, 1991.

[19] R. Azzouz, C. Bacconnet, and J.-C. Faugeras. Analyse geostatistique dune cam-
pagne de reconnaissance au penetrometre statique. In Proc. 5th Int. Conf. on Appli-
cations of Statistics and Probability in Soil and Structural Engineering, pages 821– 828,
Vancouver, 1987.

[20] G.L. Babu and S.M. Dasaka. The effect of spatial correlation of cone tip resistance
on the bearing capacity of shallow foundations. Geotechnical and Geological Engi-
neering, 26(1):37–46, 2008.

[21] G.B. Baecher. Site Exploration: A probabilistic approach. PhD thesis, Massachusetts
Instiute of Technology, 1972.

[22] G.B. Baecher. Analyzing exploration strategies. In C.H. Dowding, editor, Site Char-
acterization and Exploration, New York, 1978. American Society of Civil Engineers.

[23] G.B. Baecher and J.T. Christian. Reliability and statistics in geotechnical engineering.
John Wiley & Sons Inc., 2003.

[24] G.B. Baecher and T.S. Ingra. Stochastic FEM in settlement predictions. Journal of
the Geotechnical Engineering Division, 107(4):449–463, 1981.

[25] G.B. Baecher, M. Chan, T.S. Ingra, T. Lee, and L.R. Nucci. Geotechnical reliability
of offshore gravity platforms. Technical Report Report MITSG 80-20, Sea Grant
[email protected]
College Program, Mass. Inst, of Tech, 1980.

162
 Bibliography

[26] G.B. Baecher, J.D. Ditmars, D.E. Edgar, and C.H. Dowding. Radioactive waste iso-
lation in salt: A method for evaluating the effectiveness of site characterization
measures. Technical report, Report ANL/EESTM-342, Argonne National Labora-
tory, Ill, 1988.

[27] S. Bakhtiari. Stochstiac Finite Element Slope Stability Analysis. PhD thesis, Manch-
ester Centre for Civil and Construction Engineering, University of Manchester,
England, 2011.

[28] A.A. Bakr, L.W. Gelhar, A.L. Gutjahr, and J.R. MacMillan. Stochastic analysis of
spatial variability in subsurface flows. 1. comparison of one-and three-dimensional
flows. Water Resources Research, 14(2):263–271, 1978.

[29] A. Bárdossy. Copula-based geostatistical models for groundwater quality param-

eters. Water Resources Research, 42(11):11416–11428, 2006.

[30] A. Bárdossy and Z.W. Kundzewicz. Geostatistical methods for detection of outliers
in groundwater quality spatial fields. Journal of Hydrology, 115(1-4):343–359, 1990.

[31] A. Bárdossy, U. Haberlandt, and J. Grimm-Strele. Interpolation of groundwater

quality parameters using additional information. In A. Soares, editor, Proceedings of
geoENV I-Geostatistics for Environmental Applications. Kluver Academic Publishers,
1997.

[32] R. J. Barnes. Bounding the required sample size for geological site characterization.
Mathematical Geology, 20(5):477–490, 1988.

[33] C. Bellehumeur and P. Legendre. Multiscale sources of variation in ecological vari-

ables: modeling spatial dispersion, elaborating sampling designs. Landscape Ecol-
ogy, 13(1):15–25, 1998.

[34] J.S. Bendat and A.G. Piersol. Random data analysis and measurement procedures, vol-
ume 11. IOP Publishing, 2000.

[35] J.P. Benkendorfer, C.V. Deutsch, P.D. LaCroix, L.H. Landis, Y.A. Al-Askar, A.A.
Al-AbdulKarim, and J. Cole. Integrated Reservoir Modelling of a Major Arabian
Carbonate Reservoir. Middle east oil show, 298(96):1–26, 1995.

[36] R. J. Bennett. Spatial Time Series. Pion Ltd., London, 1979.

[37] C. Benson. Predicting excursions beyond regulatory thresholds of hydraulic con-

ductivity using quality control measurements. In Proceedings of the First Canadian
Conference on Environmental Geotechnics, pages 14–17, 1991.

[38] T. Benz. Small-strain stiffness of soils and its numerical consequences. PhD thesis,
University of Stuttgart, Institute of Geotechnical Engineering, 2007.

[39] T. Benz. Small strain double hardening model. Technical report, NTNU, Trond-
[email protected]
heim, 2012.

163
Bibliography

[40] D. T. Bergado and K. T. How. A Probabilistic Spatial-Temporal Modelling of Em-

bankment Settlement on Soft Bangkok Cly: A Case of Bangna-Bangpakong High-
way, Thailand. In Proc. of 6th Int. Conf. on Applications of Statistics and Probability in
Soil and Structural Engineering, pages pp. 808–815, 1991.

[41] D.T. Bergado and L.R. Anderson. Stochastic analysis of pore pressure uncertainty
for the probabilistic assessment of the safety of earth slopes. Soils and Foundations,
25(2):87–105, 1985.

[42] J.O. Berger, V. De Oliveira, and B. Sansó. Objective Bayesian analysis of spatially
correlated data. Journal of the American Statistical Association, 96(456):1361–1374,
2001.

[43] M. Berveiller, B. Sudret, and M. Lemaire. Presentation of two methods for com-
puting the response coefficients in stochastic finite element analysis. In Proceedings
of 9th ASCE specialty Conference on Probabilistic Mechanics and Structural Reliability,
Albuquerque, USA, 2004.

[44] M. Berveiller, B. Sudret, and M. Lemaire. Stochastic finite element: a non-intrusive

approach by regression. Revue Européenne de Mécanique Numérique-Volume, 15(1-2):
81–92, 2006.

[45] P. Biçer and P.T. Özener. Vertical variability of the undrained shear strength of
golden horn clay. Marine Georesources and Geotechnology, 27(4):309–321, 2009.

[46] A. A. Binsariti. Statistical analyses and stochastic modeling of the Cortaro aquifer in
Southern Arizona. PhD thesis, University of Arizona, 1980.

[47] G. Blatman and B. Sudret. An adaptive algorithm to build up sparse polynomial

chaos expansions for stochastic finite element analysis. Probabilistic Engineering
Mechanics, 25(2):183–197, 2010.

[48] R. Borja, editor. Multiscale and multiphysics processes in geomechanics. Springer, 2011.

[49] J.-M. Bourinet, C. Mattrand, and V. Dubourg. A review of recent features and
improvements added to FERUM software. In Proceedings of 10th International Con-
ference on Structural Safety and Reliability (ICOSSAR’09), Osaka, Japan, 2009.

[50] D.S. Bowles and H.H. Ko, editors. Probabilistic Characterization of Soil Properties:
Bridge Between Theory and Practice. ASCE, 1984.

[51] G.E.P. Box and G.M. Jenkins, editors. Time Series Analysis: Forecasting and Control.
Holden-Day, San Fransisco, 1970.

[52] H. Bozdogan. Model selection and akaike’s information criterium: The general
tehory and its analytical extensions. Psychometrika, 52:345–370, 1987.

[53] K. Breitung. Asymptotic approximations for multi-normal integrals. J. Engrg.

[email protected]
Mech. Div., ASCE, 110(3):357–366, 1984.

164
 Bibliography

[54] D. Breysse, H. Niandou, S. Elachachi, and L. Houy. A generic approach to soil-

structure interaction considering the effects of soil heterogeneity. Geotechnique, 55
(2):143–150, 2005.

[55] P.I. Brooker, J.P. Winchester, and A.C. Adams. A geostatistical study of soil data
from an irrigated vineyard near Waikerie, South Australia. Environment Interna-
tional, 21(5):699–704, 1995.

[56] D.J. Brus and J.J. De Gruijter. Random sampling or geostatistical modelling?
choosing between design-based and model-based sampling strategies for soil
(with discussion). Geoderma, 80(1-2):1–44, 1997.

[57] C. Bucher. Computational Analysis of Randomness in Structural Mechanics. Taylor &

Francis, 2009.

[58] C. Bucher and T. Most. A comparison of approximate response functions in struc-

tural reliability analysis. Probabilistic Engineering Mechanics, 23(2-3):154–163, 2008.

[59] P.A. Burrough. Multiscale sources of spatial variation in soil. I - The application
of fractal concepts to nested levels of soil variation. Journal of Soil Science, 34(3):
577–597, 1983.

[60] P.A. Burrough. Multiscale sources of spatial variation in soil. III - The application
of fractal concepts to nested levels of soil variation. Journal of Soil Science, 34(3):
578–599, 1983.

[61] P.A. Burrough. Multiscale sources of spatial variation in soil. II - A non-Brownian

fractal model and its application in soil survey. Journal of Soil Science, 34(3):599–
620, 1983.

[62] E. Byers and D.B. Stephens. Statistical and stochastic analyses of hydraulic con-
ductivity and particle-size in a fluvial sand. Soil Science Society of America Journal,
47(6):1072, 1983.

[63] D.G. Cacuci, M. Ionescu-Bujor, and I.M. Navon. Sensitivity and Uncertainty Analy-
sis: Applications to large-scale systems, volume 2. CRC, 2005.

[64] J. Caers. Modeling Uncertainty in the Earth Sciences. Wiley, 2011.

[65] F. Cafaro and C. Cherubini. Large sample spacing in evaluation of vertical strength
variability of clayey soil. Journal of geotechnical and geoenvironmental engineering,
128:558, 2002.

[66] J.A. Calgaro. Safety philosophy of eurocodes. In Proceedings of the 3rd International
Symposium on Geotechnical Safety and Risk, 2011.

[67] E.O.F. Calle, G.A.M Kruse, M.T. van der Meer, W.R. Halter, and B.M. Effing.
Schematiseren Geotechnische Faalmechanismen bij Dijken - SBW Faalmechanis-
[email protected]
men TR Grondonderzoek, (in Dutch). Technical report, Deltares, 2009.

165
Bibliography

[68] R.G. Campanella, D.S. Wickremesinghe, and P.K. Robertson. Statistical treatment
of cone penetration test data. In Proceedings of the 5th International Conference on
Reliability and Risk Analysis in Civil Engineering, 1987.

[69] S.F. Carle and G.E. Fogg. Transition probability based indicator geostatistics. Math-
ematical Geology, 28(4):453–476, 1996.

[70] R. Caspeele. Proabilistic evaluation of conformity control and the use of the Bayesian up-
dating techniques in the framework of safety analyses of concerte structures. PhD thesis,
University Gent, 2009.

[71] CEN. Eurocode: Basis of structural design. European standard, EN 1990: 2002,
2002.

[72] R. Chaplow. The geology and hydrogeology of sellafield: an overview. Quarterly

Journal of Engineering Geology and Hydrogeology, 29(Supplement 1):1–12, 1996.

[73] Q. Chen, A. Seifried, J.E. Andrade, and J.W. Baker. Characterization of random
fields and their impact on the mechanics of geosystems at multiple scales. Inter-
national Journal for Numerical and Analytical Methods in Geomechanics, 36(2):140–165,
2010.

[74] Q. Cheng. Multifractality and spatial statistics. Computers & Geosciences, 25:949–
961, 1999.

[75] C. Cherubini. Data and considerations on the variability of geotechnical properties

of soils. In Proceedings of the international conference on safety and reliability, ESREL,
volume 97, pages 1583–1591, 1997.

[76] C. Cherubini. Reliability evaluation of shallow foundation bearing capacity on c’–

ϕ’ soils. Canadian Geotechnical Journal, 37(1):264–269, 2000.

[77] C. Cherubini, I. Giasi, and L. Rethati. The coeffcient of variation of some geotech-
nical parameters. In Proceedings of Probabilistic Methods in Geotechnical Gengineering,
1993.

[78] P. Chiasson and Y.J. Wang. Spatial variability of sensitive Champlain sea clay and
an application of stochastic slope stability analysis of a cut. In Proceedings of the 2nd
International Workshop on Characterisation and Engineering Properties of Natural Soils.
Singapore, November, 2006.

[79] J.P. Chiles and P. Delfiner. Geostatistics–modeling spatial uncertainty. Wiley-

Interscience, 1999.

[80] J. Ching, J.R. Chen, J.Y. Yeh, and K.K. Phoon. Updating uncertainties in friction
angles of clean sands. Journal of Geotechnical and Geoenvironmental Engineering, 138
[email protected]
(2):217–229, 2011.

166
 Bibliography

[81] Y.H. Chok. Modelling the effects of soil variability and vegetation on the stability of
natural slopes. PhD thesis, University of Adelaide, School of Civil, Environmental
and Mining Engineering, 2008.

[82] G. Christakos. Random field models in earth sciences. Academic Press, 1992.

[83] J. Chu, W. Hsču, H. Zhu, and A.G. Journel. The Amoco case study. Report 4,
Stanford Center for Reservoir Forecasting, 1991.

[84] W.G. Cochran. Sampling techniques. Wiley, London, 3rd edition, 1977.

[85] P.D. Coddington. Random number generators for parallel computers. The NHSE
Review, 2:1–26, 1997.

[86] R.F. Craig. Soil mechanics. Taylor & Francis, 2004.

[87] N. Cressie. Fitting variogram models by weighted least squares. Mathematical

Geology, 17(5):563–586, 1985.

[88] G. De Marsily, F. Delay, J. Gonçalvès, P. Renard, V. Teles, and S. Violette. Dealing

with spatial heterogeneity. Hydrogeology Journal, 13(1):161–183, 2005.

[89] E. De Rocquigny, N. Devictor, and S. Tarantola. Uncertainty in industrial practice.

Wiley Online Library, 2008.

[90] D.J. DeGroot and G.B. Baecher. Estimating autocovariance of in-situ soil proper-
ties. Journal of Geotechnical Engineering, 119(1):147–166, 1993.

[91] J.P. Delhomme. Spatial variability and uncertainty in groundwater flow parame-
ters: a geostatistical approach. Water Resources Research, 15(2):269–280, 1979.

[92] F.A. Delory, A.M. Crawford, and M.E.M. Gibson. Measurements on a tunnel lining
in very dense till. Canadian Geotechnical Journal, 16(1):190–199, 1979.

[93] A. Der Kiureghian, H-Z. Lin, and S.-J. Hwang. Second-order reliability approxi-
mations. J. Engrg. Mech. Div., ASCE, 113(8):1208–1225, 1987.

[94] A.J. Desbarats and R. Dimitrakopoulos. Geostatistical simulation of regionalized

pore-size distributions using min/max autocorrelation factors. Mathematical Geol-
ogy, 32(8):919–942, 2000.

[95] C. Deutsch, M. Monteiro, S. Zanon, and O. Leuangthong. Procedures and guide-

lines for assessing and reporting uncertainty in geostatistical reservoir modeling.
Technical report, Technical report, Centre for Computational Geostatistics, Univer-
sity of Alberta, Edmonton, AB, 2002.

[96] C.V. Deutsch. Geostatistical Reservoir Modeling. Oxford University Press, New York,
[email protected]
USA, 2002.

167
Bibliography

[97] C.V. Deutsch and A.G. Journel. GSLIB - Geostastical Software Library and Users’s
Guide. Oxford University Press, 1998.
[98] J.L. Devary and P.G. Doctor. Pore velocity estimation uncertainties. Water Resources
Research, 18(4):1157–1164, 1982.
[99] J. Diaz-Padilla and E.H. Vanmarcke. Settlement of Structures on Shallow Founda-
tions: A Probabilistic Analysis. Dept. of Civil Engineering, Massachusetts Institute
of Technology, 1974.
[100] C.R. Dietrich and M.R. Osborne. Estimation of covariance parameters in kriging
via restricted maximum likelihood. Mathematical Geology, 23(1):119–135, 1991.
[101] DIN 4017. Baugrund-Berechnung des Grundbruchwiderstands von Flachgrün-
dungen, 2006.
[102] DIN 4019. Baugrund-Setzungsberechnungen, 2001.
[103] DIN 4094-1. Subsoil - field investigations - part 1: Cone penetration tests, 06 2002.
[104] DIN 4094-5. Geotechnical field investigations - part 5: Borehole deformation tests.,
2001.
[105] O. Ditlevsen and H.O. Madsen. Structural reliability methods, volume 178. Wiley
Chichester, 1996.
[106] D.N. Veritas. Recommended practice report dnv-rp-c207: Statistical representation
of soil data. Technical report, D.N. Veritas, 1997.
[107] G.R. Dodagoudar and G. Venkatachalam. Reliability analysis of slopes using fuzzy
sets theory. Computers and Geotechnics, 27(2):101–115, 2000.
[108] H. Duddeck. Empfehlungen zur Berechnung von Tunneln im Lockergestein. Die
Bautechnik, 10:349–356, 1980.
[109] A.S. Dyminski, J. Howie, D. Shuttle, and A. Kormann. Soil variability study for
embankment design of port of navegantes, brazil. In Proceedings of GeoCongress
2006: Geotechnical Engineering in the Information Technology Age, pages 1–6. ASCE,
2006.
[110] M.D. Ecker and A.E. Gelfand. Bayesian variogram modeling for an isotropic spa-
tial process. Journal of Agricultural, Biological, and Environmental Statistics, 3:347–
369, 1997.
[111] B. Efron. Bootstrap methods: another look at the jackknife. The annals of Statistics,
7(1):1–26, 1979.
[112] J. Eggleston and S. Rojstaczer. Identification of large-scale hydraulic conductiv-
ity trends and the influence of trends on contaminant transport. Water Resources
[email protected]
Research, 34(9):2155–2168, 1998.

168
 Bibliography

[113] M. El Gonnouni, Y. Riou, and PY Hicher. Geostatistical method for analysing

soil displacement from underground urban construction. Geotechnique, 55:171–182,
2005.

[114] H. El-Ramly, N.R. Morgenstern, and D.M. Cruden. Probabilistic slope stability
analysis for practice. Canadian Geotechnical Journal, 39(3):665–683, 2002.

[115] A.M.M. Elfeki. Stochastic characterization of geological heterogeneity and its impact
on groundwater contaminant transport. PhD thesis, TU Delft, Delft University of
Technolog, 1996.

[116] I. Elishakoff. Possible limitations of probabilistic methods in engineering. Appl.

Mech. Rev., 1(53):19–36, 2000.

[117] B.R. Ellingwood, O.C. Celik, and K. Kinali. Fragility assessment of building struc-
tural systems in mid-america. Earthquake Engineering & Structural Dynamics, 36
(13):1935–1952, 2007.

[118] X. Emery. Properties and limitations of sequential indicator simulation. Stochastic

Environmental Research and Risk Assessment, 18(6):414–424, 2004.

[119] EN 1997-1:2004. Eurocode 7 - Geotechnical design–Part 1: General rules, 2004.

[120] J. Erdmann. Vergleich ebener und Entwicklung räumlicher Berechnungsverfahren für

Tunnel. PhD thesis, Technische Universität Carolo-Wilhelmina zu Braunschweig,
1983.

[121] M. Faber and T. Vrouwenvelder. Probabilistic model code. Technical re-

port, Joint Committee of Structural Safety, www.jcss.ethz.ch, May 2001. URL
www.jcss.ethz.ch.

[122] M.H. Faber and M.G. Stewart. Risk assessment for civil engineering facilities: crit-
ical overview and discussion. Reliability engineering & system safety, 80(2):173–184,
2003.

[123] M. N. Fardis. Probabilistic liquefaction of sands during earthquakes. Technical

Report R79-14, Dept. of Civil Engrg., MIT, 1979.

[124] M.A. Fardis and D. Veneziano. Estimation of SPT-N and relative density. Journal
of the Geotechnical Eengineering Division, 107(10):1345–1359, 1981.

[125] G.A. Fenton. Random field modeling of CPT data. Journal of Geotechnical and Geoen-
vironmental Engineering, 125(6):486, 1999.

[126] G.A. Fenton. Estimation for stochastic soil models. Journal of Geotechnical and
Geoenvironmental Engineering, 125(6):470, 1999.

[127] G.A. Fenton and D.V. Griffiths. Risk assessment in geotechnical engineering. John
[email protected]
Wiley & Sons, 2008.

169
Bibliography

[128] S. Ferson. Constructing probability boxes and Dempster-Shafer structures. Sandia Na-
tional Laboratories, 2002.

[129] Federal Highway Administration (FHWA). Federal highway ad-

ministration’s geotechnical research program, 06 2012. URL
http://www.fhwa.dot.gov/publications/publicroads/.

[130] R.V. Field and M. Grigoriu. On the accuracy of the polynomial chaos approxima-
tion. Probabilistic Engineering Mechanics, 19(1):65–80, 2004.

[131] B. Fiessler, H.-J. Neumann, and R. Rackwitz. Quadratic limit states in structural
reliability. J. Engrg. Mech. Div., ASCE, 105(4):661–676, 1979.

[132] G.E. Fogg, F.J. Lucia, and R.K. Senger. Stochastic simulation of interwell-scale
heterogeneity for improved prediction of sweep efficiency in a carbonate reservoir.
In NIPER/DOE Second International Reservoir Characterization Technical Conference,
Dallas, TX, pages 355–381, 1989.

[133] S. Fuchs. Probabilities and uncertainties in natural hazard risk assessment. In

D. Proske, M. Mehdianpour, and L. Gucma, editors, Proceedings of 4th International
Probabilistic Symposium, pages 189–204, Berlin, 2006.

[134] C. Gascuel-Odoux and P. Boivin. Variability of variograms and spatial estimates

due to soil sampling: a case study. Geoderma, 62(1-3):165–182, 1994.

[135] L.W. Gelhar, C. Welty, and K.R. Rehfeldt. A critical review of data on field-scale
dispersion in aquifers. Water Resources Research, 28(7):1955–1974, 1992.

[136] A. Gelman, J.B. Carlin, H.S. Stern, and D.B. Rubin. Bayesian data analysis. CRC
press, 2004.

[137] M.G. Genton. Highly robust variogram estimation. Mathematical Geology, 30(2):
213–221, 1998.

[138] R.G. Ghanem and P.D. Spanos. Stochastic finite elements: a spectral approach.
Springer, 1991.

[139] A. Glockner, A. Igelbüscher, C. Moormann, H. Jud, K. Lindner, and H. Kun-

zmann. Bau eines neuen Stahlwerks auf weichen Sedimentböden naha Rio D
Janeiro geotechnische Herausforderungen in ungewöhnlichen Dimensionen. In
Vorträge der Baugrundtagung 2008, pages 199–206, 2007.

[140] D.I. Goggin. Patterns of Permeability in Eolian Deposits: Page Sandstone (Juras-
sic), Northeastern Arizona, SPE, U. of Texas ll. A. Chandler, U. of Texas. SPE reprint
series, 2(27):46, 1988.

[141] D.J. Goggin, M.A. Chandler, G. Kocurek, and L. Lake. Permeability transects of eo-
lian sands and their use in generating random permeability fields. SPE Formation
[email protected]
Evaluation, 7(1):7–16, 1992.

170
 Bibliography

[142] J. Goldsworthy. Reducing the Risk of Foundation Failures by Improving the Effective-
ness of Geotechnical Investigations. PhD thesis, School of Civili and environmental
Engineering, University of Adelaide, 2006.

[143] J.J. Gómez-Hernández and X.H. Wen. Multigaussian models: The danger of par-
simony. Statistical Methods & Applications, 4(2):167–181, 1995.

[144] J.J. Gómez-Hernández and X.H. Wen. To be or not to be multi-Gaussian? A reflec-

tion on stochastic hydrogeology. Advances in Water Resources, 21(1):47–61, 1998.

[145] P. Goovaerts. Geostatistics for natural resources evaluation. Oxford University Press,
USA, 1997.

[146] P.A. Graniero and V.B. Robinson. A real–time adaptive sampling method for field
mapping in patchy, heterogeneous environments. Transactions in GIS, 7(1):31–53,
2003.

[147] D.V. Griffiths and P.A. Lane. Slope stability analysis by finite elements. Geotech-
nique, 49(3):387–403, 1999.

[148] G. Gudehus. Sicherheitsnachweise für Grundbauwerke. Geotechnik, 10(1):4–34,

1987.

[149] R. Hack, B. Orlic, S. Ozmutlu, S. Zhu, and N. Rengers. Three and more dimen-
sional modelling in geo-engineering. Bulletin of Engineering Geology and the Envi-
ronment, 65(2):143–153, 2006.

[150] I.S. Halim and W.H. Tang. Bayesian method for characterization of geological
anomaly. In Proceedings of Uncertainty Modeling and Analysis, pages 585–590. IEEE,
1990.

[151] C.P. Haslauer, J. Li, and A. Bárdossy. Application of copulas in geostatistics. In

Proceedings of geoENV VII–Geostatistics for Environmental Applications, pages 395–
404. Springer, 2010.

[152] A.M. Hasofer and N.C. Lind. Exact and invariant second-moment code format.
Journal of the Engineering Mechanics Division, 100(1):111–121, 1974.

[153] T. Haukaas and A. Der Kiureghian. Strategies for finding the design point in non-
linear finite element reliability analysis. Probabilistic Engineering Mechanics, 21(2):
133 – 147, 2006. ISSN 0266-8920.

[154] Y.A. Hegazy, P.W. Mayne, and S. Rouhani. Geostatistical assessement of spatial
variability in piezocone tests. In Proceedings of Uncertainty in the geological environ-
ment: from theory to practice, New York, 1996. ASCE.

[155] K.M. Hess. Use of a borehole flowmeter to determine spatial heterogeneity of

hydraulic conductivity and macrodispersion in a sand and gravel aquifer, cape
[email protected]
cod, massachusetts. In Proceedings of New Field Techniques for Quantizing the Physical

171
Bibliography

and Chemical Properties of Heterogeneous Aquifers, National Water Well Association,

Dublin, OH, pages 497–508, 1989.

[156] M.A. Hicks and C. Onisiphorou. Stochastic evaluation of static liquefaction in a

predominantly dilative sand fill. Geotechnique, 55:123–133, 2005.

[157] M.A. Hicks and K. Samy. Influence of anisotropic spatial variability on slope re-
liability. In Proceedings of the 8th International Symposium on Numerical Models in
Geomechanics, Rome, Italy, pages 535–9, 2002.

[158] M.A. Hicks and W.A. Spencer. Influence of heterogeneity on the reliability and
failure of a long 3D slope. Computers and Geotechnics, 37(7):948–955, 2010.

[159] C. Hilldale. A probabilistic approach to estimating differential settlement of foot-

ings on sand. Master’s thesis, Massachusetts Institute of Technology, 1971.

[160] K. Hoeg and W.H. Tang. Probabilistic considerations in the foundation engineer-
ing for offshore structures. In Proceedings of 2nd International Conference on Struc-
tural Safety and Reliability, ICOSSAR, Munich, pages 267–269, 1977.

[161] R.J. Hoeksema and P.K. Kitanidis. Analysis of the spatial structure of properties of
selected aquifers. Water Resources Research, 21(4):563–572, 1985.

[162] J.A. Hoeting, D. Madigan, A.E. Raftery, and C. T. Volinsky. Bayesian Model Aver-
aging: a tutorial. Statistical Science, 14(14):382–417, 1999.

[163] W. Hof. Zum Begriff Sicherheit. Beton- und Stahlbetonbau, 86(12):286–289, 1991.

[164] A. Hommels, M. Huber, F. Molenkamp, and P.A.: Vermeer. Inverse modelling

including spatial variability applied to the construction of a road embankment. In
T. Benz, editor, Proceedings of the 7th European Conference on Numerical Methods in
Geotechnical Engineering, Trondheim, 2010 (NUMGE), pages 369–374, 2010.

[165] Y. Honjo. A new identification method for autocorrelation function of geotech-

nical data: the exteded bayesian metho. In R. Poettler, H. Klapperich, and H.F.
Schweiger, editors, Proceedings of International Conference on Probabilistic in Geotech-
nics: Technical and Economic Risk Estimation, pages 71–78, 2002.

[166] Y. Honjo. Challenges in geotechnical reliability based design. In N. Vogt, B. Schup-

pener, D. Straub, and G. Bräu, editors, Proceedings of ISGSR: Geotechnical Safety and
Risk, pages 11–28, 2011.

[167] Y. Honjo and K. Kuroda. A new look at fluctuating geotechnical data for reliability
design. Soils and Foundations, 31(1):110–120, 1991.

[168] Y. Honjo, T. Hara, M. Suzuki, M. Shirato, T.C. Le Kieu, and Y. Kikuchi. Code
calibration in reliability based design level i verification format for geotechnical
structures. In Proceedings of 2nd International Symposium on Geotechnical Safety and
[email protected]
Risk, pages 435–452, 2009.

172
 Bibliography

[169] N. Horn. Horizontaler Erddruck auf senkrechte Abschlussflächen von Tunnel-

röhren. In Proceedings of Landeskonferenz der ungarischen Tiefbauindustrie (German
translation by STUVA Düsseldorf), pages 7–16, 1961.

[170] L. Houy, D. Breysse, and A. Dents. Influence of soil heterogeneity on load redis-
tribution and settlement of a hyperstatic three-support frame. Geotechnique, 55(2):
163–170, 2005.

[171] L.Y. Hu, G. Blanc, and B. Noetinger. Gradual deformation and iterative calibration
of sequential stochastic simulations. Mathematical Geology, 33(4):475–489, 2001.

[172] M. Huber, A. Möllmann, A. Bárdossy, and P.A. Vermeer. Contributions to proba-

bilistic soil modelling. In H. Vrijling P.H.A.J.M. van Gelder, D. Proske, editor, 7th
International Probabilistic Symposium, pages 519–530, 2009.

[173] M. Huber, P.A. Vermeer, and A. Bardossy. Untersuchungen zur Variabilität des
Untergrundes. In H. Schad and C. Vogt, editors, Proceedings of Bauen in Boden und
Fels, 2009.

[174] M. Huber, M.A. Hicks, P.A. Vermeer, and C. Moormann. Probabilistic calcula-
tion of differential settlement due to tunnelling. In L. Gucma, P. van Gelder, and
D. Proske, editors, Proceedings of 8th International Probabilistic Symposium, 2010.

[175] M. Huber, A. Möllmann, and P.A. Vermeer. PC-River–reliability analysis of em-

bankment stability. Georisk, 5(2):132–142, 2011.

[176] M. Huber, M.R.S. Sadaghiani, C. Moormann, and K.J. Witt. Beitrag zur räumlichen
Beschreibung der Heterogenität des Untergrundes. In J. Tiedemann, editor, Pro-
ceedings of 18. Tagung für Ingenieurgeologie, 2011.

[177] M. Huber, P.A. Vermeer, M.A. Hicks, and C. Moormann. Reliability based design
in tunnelling. In Ch. Brenguer, editor, Proceedings of the European Safety and Relia-
bility Conference: Advances in Safety, Reliability and Risk Management, Troyes, France,
pages 308–317, 2011.

[178] M. Huber, P.A. Vermeer, and C. Moormann. Evaluation and implications of soil
variability in tunneling. In M.H. Faber, editor, Proceedings of Applications of Statistics
and Probability in Civil Engineering, Proceedings of the 11th International Conference
on Applications of Statistics and Probability in Civil Engineering (ICASP11), Zürich,
Switzerland, pages 2785–2791, 2011.

[179] P. Hufschmied. Estimation of three-dimensional statistically anisotropic hydraulic

conductivity field by means of single well pumping tests combined with flowme-
ter measurements. Hydrogeologie, 2:163–174, 1986.

[180] J. Iqbal, J.A. Thomasson, J.N. Jenkins, P.R. Owens, and F.D. Whisler. Spatial vari-
ability analysis of soil physical properties of alluvial soils. Soil Science Society of
[email protected]
America Journal, 69:1338–1350, 2005.

173
Bibliography

[181] E.H. Isaaks. Indicator simulation: application to the simulation of a high grade
uranium mineralization. In Geostatistics for Natural resources characterization; P II.
D. Reidel, Dordrecht, pages 1057–1069, 1984.
[182] E.H. Isaaks and R.M. Srivastava. Applied geostatistics, volume 2. Oxford University
Press New York, 1989.
[183] ISO 2394. General principles on the reliability of structures, 1998.
[184] M. B. Jacksa. The influence of spatial variability on the geotechnical design properties of
a stiff, overconsolidated clay. PhD thesis, The University of Adelaide, 1995.
[185] M.B. Jaksa, P.I. Brooker, and W.S. Kaggwa. Modelling the spatial variability of
the undrained shear strength of clay soils using geostatistics. In Proceedings of 5th
International Geostatistics Congress, pages 1284–1295, 1997.
[186] M.B. Jaksa, W.S. Kaggwa, and P.I. Brooker. Experimental evaluation of the scale of
fluctuation of a stiff clay. In Proceedings of 8th International Confernece on the Appli-
cation of Statistics and Probability, volume 1, pages 415–422. Sydney, AA Balkema,
Rotterdam, 1999.
[187] M.B. Jaksa, W.S. Kaggwa, G.A. Fenton, and H.G. Poulos. A framework for quanti-
fying the reliability of geotechnical investigations. In Proceedings of 9th International
Conference on the Application of Statistics and Probability in Civil Engineering, pages
1285–1291, 2003.
[188] A.G. Journel. Resampling from stochastic simulations. Environmental and Ecological
Statistics, 1(1):63–91, 1994.
[189] A.G. Journel and F. Alabert. Non-gaussian data expansion in the earth sciences.
Terra Nova, 1(2):123–134, 1989.
[190] A.G. Journel and C.J. Huijbregts. Mining geostatics. Academic Press, London, 1978.
[191] H. Karadeniz and T. Vrouwenvelder. SAFERELNET : Task 5.1: Overview reliability
methods. Technical report, TU Delft, The Netherlands, October 2003.
[192] J.M. Keaveny, F. Nadim, and S. Lacasse. Autocorrelation functions for offshore
geotechnical data. In Proceedings of the 5th International Conference on Structural
Safety and Reliability (ICOSSAR), pages 263–270, 1994.
[193] R. Kerry and M. A. Oliver. Determining the effect of asymmetric data on the vari-
ogram II - Outliers. Computers & Geosciences, 33(10):1233–1260, 2007.
[194] R. Kerry and M.A. Oliver. Determining the effect of asymmetric data on the vari-
ogram I - Underlying asymmetry. Computers & Geosciences, 33(10):1212–1232, 2007.
[195] L. Kirkup and R.B. Frenkel. An Introduction to Uncertainty in Measurement: Using the
GUM (Guide to the Expression of Uncertainty in Measurement). Cambridge University
[email protected]
Press, 2006.

174
 Bibliography

[196] A. Kirsch. On the face stability of shallow tunnels in sand. PhD thesis, Universtität
Innsbruck, 2009.

[197] P.K. Kitanidis. Statistical estimation of polynomial generalized covariance func-

tions and hydrologic applications. Water Resources Research, 19(4):909–921, 1983.

[198] P.K. Kitanidis. Parameter uncertainty in estimation of spatial functions: Bayesian

analysis. Water Resources Research WRERAO Vol. 22, 22(4):499–507, 1986.

[199] P.K. Kitanidis. Parametric estimation of covariances of regionalized variables. Jour-

nal of the American Water Resources Association, 23(4):557–567, 1987.

[200] P.K. Kitanidis and R.W. Lane. Maximum likelihood parameter estimation of hy-
drologic spatial processes by the Gauss-Newton method. Journal of hydrology, 79
(1-2):53–71, 1985.

[201] M. Kittridge, L. Lake, F. Lucia, and G. Fogg. Outcrop/subsurface comparisons of

heterogeneity in the San Andres Formation. SPE Formation Evaluation, 5(3):233–
240, 1990.

[202] H. Klammler, M. McVay, D. Horhota, and P. Lai. Influence of spatially variable

side friction on single drilled shaft resistance and lrfd resistance factors. Journal of
Geotechnical and Geoenvironmental Engineering, 136:1114, 2010.

[203] M. Kleiber and T.D. Hien. The Stochastic Finite Element Method. John Wiley & Sons
Inc, 1992.

[204] C.E. Koltermann and S.M. Gorelick. Heterogeneity in sedimentary deposits: A

review of structure-imitating, process-imitating and descriptive approaches. Water
Resources Research, 32(9):2617–2658, 1996.

[205] D. Kolymbas. Tunnelling and tunnel mechanics: a rational approach to tunnelling.

Springer Verlag, 2005.

[206] A.S.D.A.C.M. Kormann and J.H.D. Shuttle. Variability Analysis of SPT and CPT
data for a Reliability-based Embankment Design of a Southern Brazilian Port Site.
In Proceedings of XIII Brazilian Conference on Soil Mechanics and Geotechnical Engi-
neering. Curitiba, Brazil, 2006.

[207] H. Kreuter. Ingenieurgeologische Aspekte geostatistischer Methoden. PhD thesis, Tech-

nische Universität Karlsruhe,Institute of Geotechnicla Engineering, 1996.

[208] P. Kulatilake and A. Ghosh. An investigation into accuracy of spatial variation

estimation using static cone penetrometer data. In Proceedings of the 1st International
Conference on Penetration Testing, Orlando, Fla, pages 815–821, 1988.

[209] P.H.S.W. Kulatilake and J.G. Um. Spatial variation of cone tip resistance for the
clay site at texas university. Geotechnical and Geological Engineering, 21(2):149–165,
[email protected]
2003.

175
Bibliography

[210] H. Kupfersberger and C.V. Deutsch. Methodology for integrating analog geologic
data in 3-d variogram modeling. AAPG bulletin, 83(8):1262–1278, 1999.
[211] S. Lacasse and J.Y. de Lamballerie. Statistical treatment of CPT data. In Proceedings
of the International Symposium on Cone Penetration Testing, pages 4–5, 1995.
[212] S.M. Lacasse, C.C. Ladd, and A.K. Barsvary. Undrained behavior of embankments
on new liskeard varved clay. Canadian Geotechnical Journal, 14(3):367–388, 1977.
[213] P.V. Lade and J.M. Duncan. Elastoplastic stress-strain theory for cohesionless soil.
Journal of the Geotechnical Engineering Division, 101(10):1037–1053, 1975.
[214] R.M. Lark. Estimating variograms of soil properties by the method-of-moments
and maximum likelihood. European Journal of Soil Science, 51(4):717–728, 2000.
[215] G. Le Loc’h, H. Beucher, A. Galli, B. Doligez, HERESIM Group, et al. Improve-
ment in the truncated gaussian method: Combining several gaussian functions.
In ECMOR IV, Proceedings of the 4th European Conference on the Mathematics of Oil
Recovery, R oslash ros, Norway, 1994.
[216] E. Leca and L. Dormieux. Upper and lower bound solutions for the face stability
of shallow circular tunnels in frictional material. Geotechnique, 40(4):581–606, 1990.
[217] M. Lemaire, A. Chateauneuf, and J.C. Mitteau. Structural reliability. Wiley Online
Library, 2009.
[218] A. Lentz. Acceptability of civil engineering decisions involving human consequences.
PhD thesis, Technischen Universität München, München, 2007.
[219] C.C. Li and A. der Kiureghian. Optimal discretization of random fields. Journal of
Engineering Mechanics, 119:1136, 1993.
[220] J. Li, A. Bárdossy, L. Guenni, and M. Liu. A copula based observation network
design approach. Environmental Modelling & Software, 26(11):1349–1357, 2011.
[221] K.S. Li and I.K. Lee. The assessment of geotechnical safety. In Selected Topics in
Geotechnical Engineering-Lumb Volume, 1991.
[222] C. N. Liu and C.-H. Chen. Spatial correlation structures of CPT data in a liquefac-
tion site. Engineering Geology, 111(1-4):43 – 50, 2010.
[223] C.N. Liu and C.H. Chen. Estimating spatial correlation structures based on CPT
data. Georisk: Assessment and Management of Risk for Engineered Systems and Geohaz-
ards, 4(2):99–108, 2010.
[224] K. Loague and G.A. Gander. R-5 revisited: Spatial variability of infiltration on a
small rangeland catchment. Water Resources Research, 26(5):957–971, 1990.
[225] N. Luhmann. Die Moral des Risikos und das Risiko der Moral. In Proceedings of
[email protected]
Risiko und Gesellschaft, pages 327–338. Westdeutscher Verlag, Opladen, 1997.

176
 Bibliography

[226] P. Lumb. Safety factors and the probability distribution of soil strength. Canadian
Geotechnical Journal, 7(3):225–242, 1970.

[227] P. Lumb. Application of statistics in soil mechanics. In Soil Mechanics: New Hori-
zons, pages 44–112, 1974.

[228] P. Lumb. Spatial variability of soil properties. In Proceedings of the 2nd International
Conference on Application of Statistics and Probability to Soil and Structural Engineering,
Aachen, volume 2, pages 397–421, 1975.

[229] P. Lumb and J.K. Holt. The undrained shear strength of a soft marine clay from
Hong Kong. Geotechnique, 18(1):25–36, 1968.

[230] H.J. Lund, H. Ates, E. Kasap, and R.W. Tillman. Comparison of Single and Multi-
Facies Variograms of Newcastle Sandstone: Measures for the Distribution of Bar-
riers to Flow. In Low Permeability Reservoirs Symposium, pages 507–522, 1995.

[231] R.J. Luxmoore, B.P. Spalding, and I.M. Munro. Areal variation and chemical modi-
fication of weathered shale infiltration characteristics. Soil Science Society of America
Journal, 45(4):687, 1981.

[232] A.V. Lyamin, R. Salgado, S.W. Sloan, and M. Prezzi. Two-and three-dimensional
bearing capacity of footings in sand. Géotechnique, 57(8):647–662, 2007.

[233] R.J. Mair and R.N. Taylor. Theme lecture: Bored tunneling in the urban envi-
ronment. In Proceedings of the 14th International Conference on Soil Mechanics and
Foundation Engineering, pages 2353–2385, 1997.

[234] H.B. Mann. Nonparametric tests against trend. Econometrica: Journal of the Econo-
metric Society, 2:245–259, 1945.

[235] A. Marache, J. Dubost, D. Breysse, A. Denis, and S. Dominique. Understanding

subsurface geological and geotechnical complexity at various scales in urban soils
using a 3D model. Georisk: Assessment and Management of Risk for Engineered Sys-
tems and Geohazards, 1(1):1–14, 2009.

[236] B.P. Marchant and R.M. Lark. Estimating variogram uncertainty. Mathematical
Geology, 36(8):867–898, 2004.

[237] B.P. Marchant and R.M. Lark. Adaptive sampling and reconnaissance surveys for
geostatistical mapping of the soil. European Journal of Soil Science, 57(6):831–845,
2006.

[238] F. Marconi. Ricostruzione con metodi geostatistici di corpi alluvionali sepolti per la valu-
tazione delle pericolosita geologiche in aree urbane. PhD thesis, Universita degli studi
di Roma "Sapienza", Facolta di Ingegneria, Dipartimento Ingegneria Chimica Ma-
[email protected]
teriali Ambiente, 2013.

177
Bibliography

[239] G. Matheron. The theory of regionalized variables and its applications, volume 5. École
national supérieure des mines, 1971.

[240] G. Matheron and F. Blondel. Traité de géostatistique appliquée. Editions Technip,

1962.

[241] G. Matheron, H. Beucher, A. Galli, D. Guerillot, and C. Ravenne. Conditional

simulation of the geometry of fluvio-deltaic reservoirs. In SPE Annual Technical
Conference and Exhibition, 1987.

[242] M. Matsuo. Reliability in embankment design. Technical Report R76-33., Depart-

ment of Civil Engineering, Massachusetts Institute of Technology, Cambridge, Res.
Rept. R, 1976.

[243] H. Matsuoka and T. Nakai. Relationship among Tresca, Mises, Mohr-Coulomb

and Matsuoka-Nakai failure criteria. Soils and Foundations, 25(4):123–128, 1985.

[244] H.G. Matthies and A. Keese. Galerkin methods for linear and nonlinear elliptic
stochastic partial differential equations. Computer Methods in Applied Mechanics
and Engineering, 194(12):1295–1331, 2005.

[245] H.G. Matthies, C. Brenner, C. Bucher, and C. Guedes Soares. Uncertainties in

probabilistic numerical analysis of structurs and solids – stochastic finite elements.
Structural Safety, 19(3):283–335, 1997.

[246] B. Matti. Geological heterogeneity in landslides: characterization and flow modelling.

PhD thesis, Ecole Polytechnique Federale de Lausanne, 2008.

[247] A.B. McBratney and R. Webster. Choosing functions for semi-variograms of soil
properties and fitting them to sampling estimates. European Journal of Soil Science,
37(4):617–639, 1986.

[248] A.B. McBratney, R. Webster, and T.M. Burgess. The design of optimal sampling
schemes for local estimation and mapping of of regionalized variables: Theory
and method. Computers & Geosciences, 7(4):331–334, 1981.

[249] M.D. McKay. Evaluating prediction uncertainty. Technical report, Office of Nu-
clear Regulatory Research. Division of Systems Technology and Los Alamos Na-
tional Laboratory, 1995.

[250] S.A. McKenna. Probabilistic Approach to Site Characterization: MIU Site, Tono
Region, Japan. Technical report, Sandia National Labs., Albuquerque, NM (US);
Sandia National Labs., Livermore, CA (US), 2001.

[251] G. G. Meyerhof. Development of geotechnical limit state design. Canadian Geotech-

nical Journal, 32(1):128–136, 1995.

[252] B. Möller, M. Beer, W. Graf, A. Hoffmann, and J.-U. Sickert. Modellierung von
[email protected]
Unschärfe im Ingenieurbau. Bauinformatik Journal, 3(11):697–708, 2000.

178
 Bibliography

[253] S.C. Möller. Tunnel induced settlements and structural forces in linings. PhD thesis,
Institute of Geotechnical Engineering, University of Stuttgart, 2006.

[254] A.F.D. Möllmann. Probabilistische Untersuchung von Hochwasserschutzdeichen mit

analytischen Verfahren und der Finite-Elemente-Methode. PhD thesis, University of
Stuttgart, Institute of Geotechnical Engineering, 2009.

[255] G. Mollon, D. Dias, and A.H. Soubra. Probabilistic analysis of circular tunnels in
homogeneous soil using response surface methodology. Journal of Geotechnical and
Geoenvironmental Engineering, 135:1314–1325, 2009.

[256] G. Mollon, D. Dias, and A.H. Soubra. Face stability analysis of circular tunnels
driven by a pressurized shield. Journal of Geotechnical and Geoenvironmental Engi-
neering, 136:215–229, 2010.

[257] Monitor Environmental Consultants Ltd. Secondary model procedure for the de-
velopment of appropriate soil sampling strategies for land contamination, 2012.
URL www.environment-agency.gov.uk.

[258] R.L. Muhanna, H. Zhang, and R.L. Mullen. Interval finite elements as a basis for
generalized models of uncertainty in engineering mechanics. Reliable Computing,
13(2):173–194, 2007.

[259] D.J. Mulla. Estimating spatial patterns in water content, matric suction, and hy-
draulic conductivity. Soil Sciences Society of America Journal, 52(6):1547–1553, 1988.

[260] D.S. Murty and G.L.S. Babu. Effect of spatial correlation structure and transforma-
tion model on the design of shallow strip footing. Geomechanics and Geoengineering,
5(2):91–97, 2010.

[261] J. Murzewski. Sicherheit der Baukonstruktionen. VEB Verlag für Bauwesen, Berlin,
1974.

[262] A. Nasekhian. Application of non-probabilistic and probabilistic concepts in finite ele-

ment analysis fo tunneling. PhD thesis, TU Graz, Institute of Soil Mechanics, 2011.

[263] H.D. Netzel. Building response due to ground movements. PhD thesis, Delft University
of Technology, 2011.

[264] I.T. Ng and C.Y. Zhou. Uncertainty-based optimization of site characterization

using cpt. In Proceedings of 2nd International Symposium on Cone Penetration Testing,
Huntington Beach, CA, USA, May 2010 2010.

[265] V.U. Nguyen and R.N. Chowdhury. Simulation for risk analysis with correlated
variables. Geotechnique, 35(1):47–58, 1985.

[266] J.D. Nuttall. Parallel Implementation and Application of the Random Finite Element
[email protected]
Method. PhD thesis, The University of Manchester, Manchester, UK, 2011.

179
Bibliography

[267] M. P. O’Reilly and B. M. New. Settlements above tunnels in the UK-their magni-
tude and prediction. In Proceedings of Tunnelling, 82, pages 173–181, 1982.

[268] J. Ortiz and C.V. Deutsch. Calculation of uncertainty in the variogram. Mathemat-
ical Geology, 34(2):169–183, 2002.

[269] J. Ouellet, D.E. Gill, and M. Soulié. Geostatistical approach to the study of induced
damage around underground rock excavations. Canadian Geotechnical Journal, 24
(3):384–391, 1987.

[270] E. Pardo-Iguzquiza. Maximum likelihood estimation of spatial covariance param-

eters. Mathematical Geology, 30(1):95–108, 1998.

[271] E. Pardo-Igúzquiza. Varfit: a fortran-77 program for fitting variogram models by

weighted least squares. Computers & Geosciences, 25(3):251–261, 1999.

[272] E. Pardo-Iguzquiza and P. Dowd. Variance–covariance matrix of the experimental

variogram: Assessing variogram uncertainty. Mathematical Geology, 33(4):397–419,
2001.

[273] R.L. Parsons and J.D. Frost. Evaluating site investigation quality using GIS and
geostatistics. Journal of Geotechnical and Geoenvironmental Engineering, 128(6):451–
461, 2002.

[274] R.B. Peck. Deep excavations and tunnelling in soft ground. In 7th Int. Conf. on Soil
Mech. and Found. Engrg., Mexico, pages 225–290. Sociedad Mexican de Mecanica de
Suelos, 1969.

[275] Pacific Earthquake Engineering Research Center (PEER). Strong motion database,
06 2012. URL http://peer.berkeley.edu.

[276] B. Peintinger and R. Rackwitz. Berichte zur Zuverlaessigkeitstheorie der Bauw-

erke: Numerical uncertainty analysis of slopes. Technical report, Technische Uni-
versität München, 1980.

[277] E.W. Perau. Ein systematischer Ansatz zur Berechnung des Grundbruchwiderstands von
Fundamenten. PhD thesis, Universität Gesamthochschule Essen, 1995.

[278] G. Peschl and H. Schweiger. Reliability analysis in geotechnics with finite

elements-comparison of probabilistic, stochastic and fuzzy set methods. In Pro-
ceedings of Symposium on Imprecise Probabilities and their Applications ISIPTA’03, 2003.

[279] G.M. Peschl. Reliability Analysis in Geotechnics with the Random Set Element Method.
PhD thesis, TU Graz, Institute of Soil Mechanics, 2004.

[280] G.M. Peschl. Reliability analyses in geotechnics with the random set finite element
[email protected]
method. PhD thesis, TU Graz, Institute of Soil Mechanics, 2011.

180
 Bibliography

[281] A.N. Pettitt and A.B. McBratney. Sampling designs for estimating spatial variance
components. Journal of the Royal Statistical Society. Series C (Applied Statistics), 42(1):
185–209, 1993.

[282] K.-K. Phoon and F.H. Kulhawy. Evaluation of geotechnical property variabaility.
Canadian Geotechnical Journal, 36:625–639, 1999.

[283] K.-K. Phoon and F.H. Kulhawy. Characterization of geotechnical variability. Cana-
dian Geotechnical Journal, 36:612–624, 1999.

[284] K.-K Phoon, S.-T. Quek, and P. An. Identification of statistically homogeneous oil
layers using modified Bartlett stastistics. Journal of Geotechnical and Geoenvironmen-
tal Engieneering, 120(7):649–659, 2003.

[285] K.K. Phoon. Reliability-based design of foundations for transmission line structures.
PhD thesis, Cornell University, 1995.

[286] K.K. Phoon. Reliability-Based Design in Geotechnical Engineering - Computations and

Applications. Taylor & Francis, 2008.

[287] J.S. Piersol. Random data analysis and measurement procedures. Measurement
Science and Technology, 11:1825, 2000.

[288] R. Popescu. Stochastic variability of soil properties: data analysis, digital simulation,
effects on system behavior. PhD thesis, Princeton University, 1995.

[289] R. Popescu, G. Deodatis, and A. Nobahar. Effects of random heterogeneity of soil

properties on bearing capacity. Probabilistic Engineering Mechanics, 20(4):324 – 341,
2005.

[290] R. Pöttler and H.F. Schweiger. Anwendung der Random-Set-Methode (RSM) für
Standsicherheitsuntersuchungen im Tunnelbau. Bautechnik, 83(11):754–759, 2006.

[291] D. Potts, K. Axelsson, L. Grande, H.F. Schweiger, and M. Long. Guidelines for the
use of advanced numerical analysis. Thomas Telford Services Limited, 2002.

[292] D.M. Potts and L. Zdravković. Finite element analysis in geotechnical engineering:
application, volume 2. Thomas Telford Services Limited, 2001.

[293] L. Prandtl. Über die Härte plastischer Körper. Göttingen Nachr. Math. Phys. Kl, 12:
74–85, 1920.

[294] W.H. Press, S.A. Teukolsky, W.T. Vettrling, and B.P. Flannery. Numerical Recipes in
FORTRAN - The Art of Scientific Computing, volume second edition. Cambridge
University Press, second edition edition, 1992.

[295] D. Proske. Catalogue of Risks - Natural, Technical, Social and Health Risks. Springer-
[email protected]
Verlag Berlin, 2008.

181
Bibliography

[296] D. Proske. Debris flow impact uncertainty modeling with Grey Numbers. In
L. Gucma, P. van Gelder, and D. Proske, editors, Proceedings of 8th International
Probabilistic Symposium, pages 285–294, 2010.

[297] D. Proske, P. Lieberwirth, and P. van Gelder. Sicherheitsbeurteilung historischer Stein-

bogenbrücken. Dirk Proske Verlag, Wien Dresden, 2006.

[298] Z.P. Qiu and I. Elishakoff. Antioptimization of structures with large uncertain-but-
non-random parameters via interval analysis. Computer methods in Applied Mechan-
ics and Engineering, 152(3):361–372, 1998.

[299] R. Rackwitz. Optimization and risk acceptability based on the life quality index.
Structural Safety, 24(2-4):297–331, 2002.

[300] A. Raftery, F. Balabdaoui, T. Gneiting, and M. Polakowski. Using Bayesian Model

Averaging to calibrate froecast ensembles. Technical Report 440, Department of
statistics, Unviersity of Washington, December 2004.

[301] G. Raspa, M. Moscatelli, F. Stigliano, A. Patera, F. Marconi, D. Folle, R. Vallone,

M. Mancini, G.P. Cavinato, S. Milli, et al. Geotechnical characterization of the up-
per Pleistocene-Holocene alluvial deposits of Roma (Italy) by means of multivari-
ate geostatistics: Cross-validation results. Engineering Geology, 101(3-4):251–268,
2008.

[302] K.R. Rehfeldt. Prediction of macrodispersivity in heterogeneous aquifers. PhD thesis,

Massachusetts Institute of Technology, Dept. of Civil Engineering, 1988.

[303] K.R. Rehfeldt, J.M. Boggs, and L.W. Gelhar. Field study of dispersion in a heteroge-
neous aquifer 3. geostatistical analysis of hydraulic conductivity. Water Resources
Research, 28(12):3309–3324, 1992.

[304] N. Remy, A. Boucher, and J. Wu. Applied geostatistics with SGeMS: A user’s guide.
Cambridge University Press, 2008.

[305] N. Remy, Alexandre. Boucher, and Jianbing. Wu. Applied geostatistics with SGeMS.
Cambridge University Press, 2009.

[306] O. Renn. Concepts of risk: a classification. In S. Krimsky and D. Golding, editors,

Proceedings of Social Theories of Risk, pages 53–79. Praeger, London, 1992.

[307] M. Riva, L. Guadagnini, A. Guadagnini, T. Ptak, and E. Martac. Probabilistic study

of well capture zones distribution at the Lauswiesen field site. Journal of Contami-
nant Hydrology, 88(1-2):92–118, 2006.

[308] P.K. Robertson. Soil classification using the cone penetration test. Canadian Geotech-
nical Journal, 27(1):151–158, 1990.

[309] P.J. Roebber. Visualizing multiple measures of forecast quality. Weather and Fore-
[email protected]
casting, 24(2):601–608, 2009.

182
 Bibliography

[310] K.O. Ronold and P. Bjerager. Model uncertainty representation in geotechnical

reliability analyses. Journal of Geotechnical Engineering, 118:363, 1992.

[311] N.M. Ruse. Räumliche Betrachtung der Standsicherheit der Ortsbrust beim Tunnelvor-
trieb. PhD thesis, Institute of Geotechnical Engineering, University of Stuttgart,
2004.

[312] C. Russelli. Probabilistic methods applied to the bearing capacity problem. PhD thesis,
Institute of Geotechnical Engineering, University of Stuttgart, 2008.

[313] D. Russo. A geostatistical approach to solute transport in heterogeneous fields and

its applications to salinity management. Water Resources Research, 20(9):1260–1270,
1984.

[314] D. Russo. Design of an optimal sampling network for estimating the variogram.
Soil Science Society of America Journal, 48(4):708–716, 1984.

[315] D. Russo and E. Bresler. Soil hydraulic properties as stochastic processes: I. an

analysis of field spatial variability. Soil Science Society of America Journal, 45(4):682,
1981.

[316] L. Sachs. Angewandte Statistik. Springer, 2002.

[317] A. Saltelli. Sensitivity analysis in practice: a guide to assessing scientific models. John
Wiley & Sons Inc, 2004.

[318] A. Saltelli and I.M. Sobol. About the use of rank transformation in sensitivity
analysis of model output. Reliability Engineering & System Safety, 50(3):225–239,
1995.

[319] A. Saltelli, S. Tarantola, and K.P.S. Chan. A quantitative model-independent

method for global sensitivity analysis of model output. Technometrics, 41:39–56,
1999.

[320] P. Samui and T.G. Sitharam. Site characterization model using artificial neural
network and kriging. International Journal of Geomechanics, 10:171–180, 2010.

[321] H. Schad. Zeit-und geschwindigkeitsabhängiges Materialverhalten in der Geotechnik:

Experimentelle Erfassung und numerische Analyse. PhD thesis, Institute of Geotech-
nical Engineering, University of Stuttgart, 1992.

[322] H. Schad and G. Teutsch. Geostatistical analysis of hydraulic conductivity related

data based on core samples from a heterogeneous fluvial aquifer. Statistical Meth-
ods and Applications, 4(1):125–140, 1995.

[323] M.T. Schafmeister and A. Pekdeger. Spatial structure of hydraulic conductivity in

various porous media-problems and experiences. Geostatistics Tróia’92, 1:733–743,
[email protected]
1993.

183
Bibliography

[324] B. Schmidt. Settlements and ground movements associated with tunneling in soil. PhD
thesis, University of Illinois., 1969.

[325] J. Schneider. Sicherheit und Zuverlässigkeit im Bauwesen: Grundwissen für Ingenieure.

VDF, Hochschul-Verl. an der ETH, 1996.

[326] G.I. Schueller. A state-of-the-art report on computational stochastic mechanics.

Probabilistic Engineering Mechanics, 12(4):197–321, 1997.

[327] G.I. Schueller, H.J. Pradlwarter, and P.S. Koutsourelakis. A comparative study of
reliability estimation procedures for high dimensions. In Proceedings of 16th ASCE
Engineering Mechanics Conference, pages 233 – –243, 2003.

[328] M.T. Schultz, B.P. Gouldby, J.D. Simm, and J.L. Wibowo. Beyond the factor of
safety: developing fragility curves to characterize system reliability. Technical re-
port, DTIC Document, 2010.

[329] B. Schuppener. Eurocode 7 geotechnical design. part 1: General rules and its latest
developments. Georisk, 4(1):32–42, 2010.

[330] B. Schuppener. Nachweiskonzepte und Sicherheit in der Geotechnik. In C. Boley,

editor, Handbuch Geotechnik, pages 225–258. Vieweg+Teubner Verlag, 2012. ISBN
978-3-8348-8272-1.

[331] T. Schweckendiek and E.O.F. Calle. A factor of safety for geotechnical charac-
terization. In Proceedings of Seventeenth Southeast Asian Geotechnical Conference
(17SEAGC)–Geo-Engineering for Natural Hazard Mitigation and Sustainable Develop-
ment, volume 2, pages 227–230, 2010.

[332] H.F. Schweiger and G.M. Peschl. Reliability analysis in geotechnics with the ran-
dom set finite element method. Computers and Geotechnics, 32(6):422–435, 2005.

[333] H.F. Schweiger, R. Thurner, and R. Pöttler. Reliability analysis in geotechnics with
deterministic finite elements. International Journal of Geomechanics, 1(4):389–413,
2001.

[334] H.F. Schweiger, G.M. Peschl, and R. Pöttler. Application of the random set finite
element method for analysing tunnel excavation. Georisk, 1(1):43–56, 2007.

[335] R.K. Senger, F.J. Lucia, C. Kerans, M. Ferris, and C. E. Fogg. Geostatistical/Geo-
logical Permeability Characterization of Carbonate Ramp Deposits in Sand Andres
Outcrop, Algerita Escarpment, New Mexico . In Permian Basin Oil and Gas Recovery
Conference, Midland, Texas, pages 287–301. Society of Petroleum Engineers, 1992.

[336] K. Sett, B. Jeremić, and M.L. Kavvas. Probabilistic elasto-plasticity: solution and
verification in 1D. Acta Geotechnica, 2(3):211–220, 2007.

[337] M.A. Shahin, M.B. Jaksa, and H.R. Maier. Artificial neural network applications in
[email protected]
geotechnical engineering. Australian Geomechanics, 36(1):49–62, 2001.

184
 Bibliography

[338] J.E. Shigley and C.R. Mischke. Mechanical Engineering Design. McGraw Hill, New
York, 2001.

[339] T.W. Simpson, T.M. Mauery, J.J. Korte, and F. Mistree. Kriging models for global
approximation in simulation-based multidisciplinary design optimization. AIAA
journal, 39(12):2233–2241, 2001.

[340] J.B. Sisson and P.J. Wierenga. Spatial variability of steady-state infiltration rates as
a stochastic process. Soil Science Society of America Journal, 45(4):699, 1981.

[341] T.G. Sitharam, P. Samui, and P. Anbazhagan. Spatial variability of rock depth in
bangalore using geostatistical, neural network and support vector machine mod-
els. Geotechnical and Geological Engineering, 26(5):503–517, 2008.

[342] I.M. Smith and D.V. Griffiths. Programming the finite element method. John Wiley &
Sons, Inc., 1998.

[343] J.L. Smith and R.A. Freeze. A stochastic analysis of steady-state groundwater flow in a
bounded domain. 1978., 1978.

[344] L. Smith. Spatial variability of flow parameters in a stratified sand. Mathematical

Geology, 13(1):1–21, 1981.

[345] H.U. Smoltczyk. Ermittlung eingeschränkt plastischer Verformungen im Sand unter

Flachfundamenten. W. Ernst, 1960.

[346] U. Smoltczyk. Geotechnical Engineering Handbook - Volumen I: Fundamentals. Wiley-

VCH, 2003.

[347] I.M. Sobol. Global sensitivity indices for nonlinear mathematical models and their
monte carlo estimates. Mathematics and computers in simulation, 55(1-3):271–280,
2001.

[348] V.V. Sokolovski and J.K. Kushner. Statics of granular media. Journal of Applied
Mechanics, 33:239, 1966.

[349] P. Soos. Die Rolle des Baugrundes bei der Anwendung der neuen Sicherheitsthe-
orie im Grundbau. Geotechnik, 19:82–91, 1990.

[350] A.H. Soubra. Kinematical approach to the face stability analysis of shallow circular
tunnels. In Proceedings of 8th International Symposium on Plasticity, pages 443–445,
2000.

[351] A.H. Soubra, D.S. Youssef Abdel Massih, and K. Kalfa. Bearing capacity of foun-
dations resting on a spatially random soil. In Geocongress008, 2008.

[352] M. Soulié. Geostatistical applications in geotechnics. Geostatistics for natural re-

[email protected]
sources characterization, Pt, 2:703–730, 1984.

185
Bibliography

[353] M. Soulié, P. Montes, and V. Silvestri. Modelling spatial variability of soil parame-
ters. Canadian Geotechnical Journal, 27(5):617–630, 1990.

[354] M.G. Speedie. Selection of design value from shear test results. In Proceedings of the
2nd Australia-New Zealand Conference on Soil Mechanics and Foundation Engineering,
pages 107–109, Wellington, 1965.

[355] W.A. Spencer. Parallel stochastic and finite element modelling of clay slope stability in
3D. PhD thesis, University of Manchester, UK, 2007.

[356] M.J. Spry, F.H. Kulhawy, and M.D. Grigoriu. Reliability-based foundation design
for transmission line structures: Geotechnical site characterization strategy. Tech-
nical Report EL-5507, Electric power research institute, Palo Alto, California, 1988.

[357] F.I. Stalkup. Permeability variation in a sandstone barrier island-tidal channel-

tidal delta complex, ferron sandstone (lower cretaceous), central utah. In Proceed-
ings of SPE Annual Technical Conference and Exhibition, 1986.

[358] G. Stefanou. The stochastic finite element method: Past, present and future. Com-
puter Methods in Applied Mechanics and Engineering, 198:1031 – 1051, 2009. ISSN
0045-7825.

[359] A. Stein and J.W. van Groenigen. Constrained optimization of spatial sampling us-
ing continuous simulated annealing. Journal of Environmental Quality, 27(5):1078–
1086, 1998.

[360] D. Straub. Reliability benchmarks in geotechnical engineering, May 2012. URL

http://geosnet.geoengineer.org/.

[361] D. Straub, A. Lentz, I. Papaioannou, and R. Rackwitz. Life quality index for as-
sessing risk acceptance in geotechnical engineering. In Proceedings of the 3rd Inter-
national Symposium on Geotechnical Safety and Risk, pages 37–46, 2011.

[362] R. Suchomel and D. Mašín. Spatial variability of soil parameters in an analysis of

a strip footing using hypoplastic model. In Proceedings of 7th European conference
on numerical methods in geomechanics, Trondheim, Norway, Taylor & Francis Group,
London, pages 383–388, 2010.

[363] E.A. Sudicky. A natural gradient experiment on solute transport in a sand aquifer:
Spatial variability of hydraulic conductivity and its role in the dispersion process.
Water Resources Research, 22(13):2069–2082, 1986.

[364] B. Sudret. Uncertainty propagation and sensitivity analysis in mechanical models–

Contributions to structural reliability and stochastic spectral methods. Habilitation a
diriger des recherches, Université Blaise Pascal, Clermont-Ferrand, France, 2007.

[365] B. Sudret. Global sensitivity analysis using polynomial chaos expansions. Reliabil-
[email protected]
ity Engineering & System Safety, 93(7):964–979, 2008.

186
 Bibliography

[366] B. Sudret and A. Der Kiureghian. Stochastic finite element methods and reliability:
A state-of-the-art report. Technical report, Department of Civil & Cnvironmental
Cngineering, University of California, Berkeley, 2000.

[367] B. Sudret and A. Der Kiureghian. Comparison of finite element reliability methods.
Probabilistic Engineering Mechanics, 17(4):337–348, 2002.

[368] W.H. Tang. Probabilistic evaluation of penetration resistances. Journal of the

Geotechnical Engineering Division, 105(10):1173–1191, 1979.

[369] W.H. Tang. Principles of probabilistic characterizations of soil properties. In Prob-

abilistic Characterization of Soil Properties - Bridge Between Theory and Practice, pages
74–89. ASCE, 1984.

[370] W.H. Tang. Updating anomaly statistics-single anomaly case. Structural safety, 4
(2):151–163, 1986.

[371] J.M. Tantalla, J.H. Prevost, and G. Deodatis. Spatial variability of soil properties
in slope stability analysis: fragility curve generation. In Proceedings of ICOSSAR,
volume 1, pages 17–21, 2001.

[372] D.W. Taylor. Fundamentals of soil mechanics. Soil Science, 66(2):161, 1948.

[373] K. Terzaghi and R.B. Peck. Soil mechanics in engineering practice. Wiley, 1948.

[374] R. Thurner. Probabilistische Untersuchungen in der Geotechnik mittels deterministischer

Finite Elemente-Methode. PhD thesis, TU Graz, Institute of Soil Mechanics, 2001.

[375] A. Tillmann, A. Englert, Z. Nyari, I. Fejes, J. Vanderborght, and H. Vereecken.

Characterization of subsoil heterogeneity, estimation of grain size distribution and
hydraulic conductivity at the krauthausen test site using cone penetration test.
Journal of Contaminant Hydrology, 95(1-2):57–75, 2008.

[376] L. Tvedt. Distribution of quadratic forms in normal space– approximation to struc-

tural reliability. Journal of Engineering Mechanics Division, ASCE, 116(6):1183–1197,
1990.

[377] K. Tyler, A. Henriquez, and T. Svanes. Modeling heterogeneities in fluvial do-

mains: A review of the influence on production profiles. Stochastic Modeling and
Geostatistics, AAPG Comput. Appl. Geol, 3:77–89, 1994.

[378] Brussels Universite Catholique de Louvain. Ofda/cred international disaster

database, 2012. URL www.emdat.be.

[379] K. Unlu, D.R. Nielsen, and W. Biggar. Stochastic analysis of state unsaturated flow.
one dimensional monte carlo simulations and comparisons with spectral pertur-
bation analysis and field observations. Water Resources Research, 26(1):2207–2218,
[email protected]
1990.

187
Bibliography

[380] U.S. Geological Survey (USGS). Earthquake hazards program, 06 2012. URL
http://earthquake.usgs.gov/.

[381] M. Uzielli, G. Vannucchi, and K.K. Phoon. Random field characterisation of stress-
normalised cone penetration testing parameters. Geotechnique, 55(1):3–20, 2005.

[382] P.H.A.J.M. van Gelder. Stasticial methods fro the risk-based design of civil structures.
PhD thesis, Delft University of Technology, 2000.

[383] E.H. Vanmarcke. Reliability of earth slopes. Journal of Geotechnical Engineering, 103:
1247–1265, 1977.

[384] E.H. Vanmarcke. Probabilistic characterization of soil profiles. In Site Characteriza-

tion & Exploration, pages 199–219. ASCE, 1978.

[385] E.H. Vanmarcke. Random fields: analysis and synthesis. MIT Press, 1984.

[386] E.H. Vanmarcke and N.F. Fuleihan. Probabilistic prediction of levee settlements.
In Proceedings of the 2nd International Conference on Applications of Statistics and Prob-
ability in Soil and Structural Engineering, Aachen, volume 2, pages 175–190, 1975.

[387] E.H. Vanmarcke and M. Grigoriu. Stochastic finite element analysis of simple
beams. Journal of Engineering Mechanics, 109:1203, 1983.

[388] V.N. Vapnik. The nature of statistical learning theory. Springer-Verlag New York Inc.,
2000.

[389] Det Norske Veritas. Structural reliability analysis of marine structures. D.N.V., 1992.

[390] Det Norske Veritas. Statistical representation of soil data. Technical report, D.N.V.,
1997.

[391] P. A. Vermeer, B. Westrich, A. Möllmann, U. Merkel, and M. Huber. PC-River -

Zuverlässigkeitsanalys und Risikoabschätzung für den Hochwasserschutz unter
integrierter Berücksichtigung geotechnischer, hydrologischer und hydraulischer
Einflussgrössen. Technical report, University of Stuttgart, Institute of Geotechnical
Engineering, 2011.

[392] P.A. Vermeer, N. Ruse, and T. Marcher. Tunnel heading stability in drained ground.
Felsbau, 20(6):8–18, 2002.

[393] A. Verruijt and J.R. Booker. Surface settlements due to deformation of a tunnel in
an elastic half plane. Geotechnique, 46(4):753–756, 1996.

[394] A. Verruijt and S. van Baars. Soil mechanics. VSSD, 2007.

[395] M.F. Versteeg. External safety policy in the netherlands: an approach to risk man-
[email protected]
agement. Journal of hazardous materials, 17(2):215–222, 1988.

188
 Bibliography

[396] G. Vessia, C. Cherubini, J. Pieczyńska, and W. Puła. Application of random finite

element method to bearing capacity design of strip footing. Journal of Geoengineer-
ing, 4(3):103–112, 2009.

[397] J.P. Visodic. Design Stress Factors, volume 55. ASME International, New York, 1948.

[398] U.H. Vollenweider. Beitrag zur Behandlung bodenmechanischer Grenzwertprobleme

nach der Plastizitätstheorie. PhD thesis, Institut für Baustatik der ETH Zürich, 1969.

[399] T. Vrouwenvelder. The JCSS probabilistic model code. Structural Safety, 19(3):245–
251, 1997.

[400] T. Vrouwenvelder and E.O.F. Calle. Measuring spatial correlation of soil proper-
ties. HERON, 48(4):2003, 2003.

[401] H. Wackernagel. Multivariate geostatistics: an introduction with applications. Springer

Verlag, 2003.

[402] R. Webster and M.A. Oliver. Sample adequately to estimate variograms of soil
properties. Journal of soil science, 43(1):177–192, 1992.

[403] R. Webster and M.A. Oliver. Geostatistics for environmental scientists. John Wiley &
Sons Inc., 2007.

[404] D. Wickremesinghe and R.G. Campanella. Scale of fluctuation as a descriptor of

soil variability. In Proceedings of Conference on Probabilistic Methods in Geotech. Engrg,
pages 233–239, 1993.

[405] D.S. Wickremesinghe. Stasticial characterisation of soil profile using in situ tests. PhD
thesis, University of British Columbia, 1989.

[406] T. Wolff. Ein methodischer Bemessungsansatz zur Abschätzung des Tragverhaltens

von Pfahlgründungen in weichem kalkigem Sedimentgestein. PhD thesis, Bauhaus-
Universität Weimar, 2009.

[407] T.F. Wolff. Analysis and design of embankment dam slopes: a probabilistic approach. PhD
thesis, Purdue University, Lafayette, Ind., 1985.

[408] S.Y. Wong. Stochastic Characterization and Reliability of Saturated Soils. PhD thesis,
Ph. D. Thesis, Manchester Centre for Civil and Construction Engineering, Univer-
sity of Manchester, England, 2004.

[409] A.D. Woodbury and EA Sudicky. The geostatistical characteristics of the Borden
aquifer. Water Resources Research, 27(4):533–546, 1991.

[410] WS Atkins Consultants Ltd . Risk implications in site characterisation and analysis
for offshore engineering and design. Technical report, Health and Safety Execu-
[email protected]
tive, 2004.

189
Bibliography

[411] T.H. Wu. Uncertainty, safety, and decision in soil engineering. Journal of the Geotech-
nical Engineering Division, 100(3):329–348, 1974.

[412] T.H. Wu and A. El-Jandali. Use of time series in geotechnical data analysis. Geotech.
Test. J, 8(4):151–158, 1985.

[413] J. R. Ximenez de Embun and M. R. Romana. The formulation of soil model from
penetrometer probabilistic information - application to differential settlement pre-
diction. In Proc. 4th Int. Conf. on Applications of Statistics and Probability in Soil and
Structural Engineering, pages 1601–1613, Pitagora Editrice, 1983.

[414] M.S. Yucemen, W.H. Tang, and A.H.S. Ang. A probabilistic study of safety and
design of earth slopes. Technical report, Engineering Experiment Station, College
of Engineering, University of Illinois at Urbana-Champaign, 1973.

[415] H. Zhang, R.L. Mullen, and R.L. Muhanna. Interval Monte Carlo methods for
structural reliability. Structural Safety, 32(3):183–190, 2010.

[416] J. Zhang, L.M. Zhang, and W.H. Tang. Bayesian framework for characterizing
geotechnical model uncertainty. Journal of geotechnical and geoenvironmental engi-
neering, 135:932, 2009.

[417] Z.J. Zhang and M.T. Tumay. Statistical to fuzzy approach toward CPT soil classi-
fication. Journal of Geotechnical and Geoenvironmental Engieneering, 125(3):179–186,
1999.

[email protected]

190
 Appendix A
Database on stochastic soil properties

A.1 Database on the variability of rock

[email protected]

191
 Table A.1: Database on the variability of rock

192
property soil type θver [m] θhor [m] applied spatial model reference
permeability outcrop of grainstone and dolomit 0.1 0.8 [201]
permeability sand and gravel aquifer 0.5 20 [179]
permeability sandstone aquifer 1 [28]
permeability subsurface of dolomitized packstone and sandstone 2.6 2.7 [201]
permeability eolian sandstone outcrop 3 8 [140]
permeability limestone aquifer 3,500 [91]

[email protected]
permeability limestone aquifer 6,300 [91]
permeability chalk 7,500 [91]
permeability sandstone aquifer 17,000 [91]
permeability sandstone aquifer 45,000 [161]
SPT soft rock, hard soils 4 spherical [202]
SPT soft rock, hard soils 5 spherical [202]
Appendix A Database on stochastic soil properties
 A.2 Database on the variability of frictional soil

A.2 Database on the variability of frictional soil

[email protected]

193
194
Table A.2: Database on the variability of frictional soils.

property soil type θver [m] θhor [m] applied spatial model reference
ϕ’ + c’ sand 5.1 242 [362]
buldk density medium gravel 0.2 exponential [375]
buldk density coarse sand, fine gravel 0.3 exponential [375]
buldk density medium gravel 0.4 exponential [375]
CPT sand 0 ∼2 exponential [10]

[email protected]
CPT alluvial deposits 0.1 ∼ 2.6 exponential [413]
CPT sand 0.1 ∼2 exponential [222]
CPT sand 0.3 ∼ 1.2 3.5 ∼ 15 Campanella [408]
CPT sand 0.3 LAS [405]
CPT sand 0.4 ∼ 2.4 LAS [27]
CPT sand 0.5 5.4 LAS [288]
CPT sand 1 12.1 LAS [288]
CPT clean sand 1 [208]
CPT soft glacial clay, North sea 1 [412]
Appendix A Database on stochastic soil properties

CPT soft glacial clay, North sea 1 [412]

CPT sand and gravel 1.2 [352]
CPT fine silty sand, North sea 1.4 [412]
CPT clean sand 1.6 [208]
CPT sand 2 20 exponential [353]
CPT sand 2.6 16 Gaussian [16]
CPT sandy gravel 5 95 Gaussian [207]
CPT coastal sand 5 [25]
CPT sand 28 LAS [405]
CPT sand 32.4 LAS [405]
CPT prairie soil 8 [224]
CPT dense sand 9.6 [211]
d10 to d90 sand 6.1 exponential [123]
Continued on next page
 Table A.2 – continued from previous page
property soil type θver [m] θhor [m] applied spatial model reference
d10 to d90 sand 49 exponential [123]
d50 Pleistocene Quadra sand 1 [344]
d50 Pleistocene Quadra sand 6.5 [344]
d50 sand 3.3 [344]
density sand 3 [269]

[email protected]
index property sand 0.2 ∼2 [148]
index property sandy clay 0.4 ∼ 1.1 1.2 ∼ 2.5 [349]
index property poorly sorted gravel 0.7 15 [307]
index property sand 0.9 exponential [109]
index property sand 1 ∼3 5 ∼ 15 [276]
permeability sand 3 25 exponential [303]
permeability fluvial sand 0.1 3 [62]
permeability braided river environment 0.3 10 [322]
permeability medium gravel 0.3 ∼ 0.9 2 ∼ 11.4 exponential [375]
permeability medium gravel 0.4 1.8 exponential [375]
permeability medium gravel 0.4 0.3 exponential [375]
permeability sand aquifer 0.5 21.5 [409]
permeability sandy gravel 0.6 1.2 [377]
permeability coarse sand, fine gravel 0.6 5 exponential [375]
permeability sandy soil 0.9 9 [335]
permeability fluvial sand 1.5 13 [302]
permeability sand 1.7 [344]
permeability sand 3.7 exponential [123]
permeability sand 3.7 exponential [124]
permeability strongly weathered rock 3.8 600 [132]
permeability sandstone 4.5 1.6 [141]
permeability heterogeneous alluvial aquifer 4.8 38.4 [303]
permeability sand 6 900 [94]
Continued on next page

195
A.2 Database on the variability of frictional soil
 Table A.2 – continued from previous page

196
property soil type θver [m] θhor [m] applied spatial model reference
permeability sand 7.2 5.4 [357]
permeability sand 12 ∼ 16 40 [379]
permeability sand 12 1,200 [83]
permeability limestone 42 6,300 [35]
permeability sand » 40 [123]
permeability compacted clay 0.5 ∼2 [37]
permeability weathered shale subsoil 2 [231]

[email protected]
permeability Pleistocene Quadra sand 3.3 [344]
permeability sand 6 exponential [92]
permeability strony layer, medium gravel 9.5 exponential [375]
permeability sand and gravel aquifer 12.5 [112]
permeability mediterranean soil 20 [315]
permeability sand 40 ∼ 280 [159]
permeability and 67 exponential [123]
permeability sand 67 exponential [124]
permeability salt dome 1,500 [26]
Appendix A Database on stochastic soil properties

permeability chalk 10,000 [88]

porosity gravelly sand 0 ∼ 2.2 8.5 ∼ 12 exponential [10]
sand fraction poorly sorted gravel 0.7 15 [307]
sand fraction sandy silt 50 ∼ 60 exponential, spherical [113]
sand fraction sandy gravel 100 ∼ 120 exponential, spherical [113]
SPT clean sand 0 ∼4 exponential [10]
SPT sand fill 0.3 exponential [228]
SPT dune sand 20 [25]
water content coarse sand, fine gravel 0.2 exponential [375]
water content medium gravel 0.2 exponential [375]
water content medium gravel 0.4 exponential [375]
 A.3 Database on the variability of cohesive soil

A.3 Database on the variability of cohesive soil

[email protected]

197
198
Table A.3: Database on the variability of cohesive soils.

property soil type θver [m] θhor [m] applied spatial model reference
CPT organic, soft clay 0.1 [206]
CPT organic clay 0.1 exp [109]
CPT clay 0.1 ∼ 0.7 exp + cos [68]
CPT clay, silty sand 0.1 ∼ 0.6 exp [381]
CPT deltatic soils 0.1 ∼ 0.4 [404]

[email protected]
CPT clay 0.2 ∼ 0.3 exp [209]
CPT organic, soft clay 0.2 [206]
CPT organic clay 0.2 exp [109]
CPT clay, silty sand 0.2 exp [381]
CPT clay 0.2 ∼ 0.7 exp [396]
CPT clay 0.2 ∼ 0.4 0.2 ∼ 0.7 [65]
CPT clay 0.2 ∼ 1.7 0.5 ∼ 3 [9]
CPT silty clay 0.2 exp [221]
CPT clay, silty sand 0.2 ∼ 0.5 exp [381]
Appendix A Database on stochastic soil properties

CPT deltatic soils 0.2 ∼ 0.3 [404]

CPT silt to silty clay 0.3 ∼1 exp [381]
CPT organic, soft clay 0.3 [206]
CPT organic clay 0.3 ∼ 0.4 exp [109]
CPT sandy silt 0.4 exp [381]
CPT deltatic soils 0.4 [404]
CPT clay 0.4 [156]
CPT clay, silty sand 0.4 ∼1 exp [381]
CPT clay 0.5 ∼5 20 ∼ 200 [125]
CPT silty sand,clay 0.6 ∼ 0.8 exp [381]
CPT sandy clay 0.6 [19]
CPT clay 0.7 [156]
CPT silty clay 0.7 exp [356]
Continued on next page
 Table A.3 – continued from previous page
property soil type θver [m] θhor [m] applied spatial model reference
CPT clay 0.7 [156]
CPT clay mixed with silt 0.8 exp [381]
CPT clean sand, silty sand 1 ∼ 1.1 exp [381]
CPT silty clay 1 5 ∼ 12 [211]
CPT silty clays 1 ∼4 exp [40]

[email protected]
CPT varved clay 1 [90]
CPT clay 1.2 ∼ 5.6 exp [45]
CPT marine clay 1.5 exp [242]
CPT clay 2 5 exp [20]
CPT sea clay 2 40 [78]
CPT soft clay 2 40 [167]
CPT sensitive clay 2 [78]
CPT sensitive clay 2 [78]
CPT sensitive clay 2 [78]
CPT sensitive clay 2 [78]
CPT sandy clay 2.5 ∼ 4.5 20 ∼ 30 exp [207]
CPT clay 4 24 exp [16]
CPT varved clay 5 exp [383]
CPT alluvial clay 6.5 [167]
CPT clay 8.6 8.6 [411]
CPT varved clay 46 exp [212]
CPT marine clay 55 35 ∼ 60 exp [15]
CPT marine clay 6 exp [227]
CPT clay 0.2 [260]
CPT silty clay 0.2 ∼ 1.9 exp [223]
CPT sandy clay 25 [18]
CPT seabed deposits 30 ∼ 60 [160]
CPT seabed deposits 53 exp [368]
Continued on next page

199
A.3 Database on the variability of cohesive soil
 Table A.3 – continued from previous page

200
property soil type θver [m] θhor [m] applied spatial model reference
CPT alluvial clay 200 [167]
CPT silty clay 212 ∼ 245 exp [264]
CPT clay 278 ∼ 286 exp [264]
CPT clay 0.6 ∼ 1.8 exp [186]
CPT silty clay 1 [211]
CPT sensitive clay 2 [78]
CPT clay 2 exp [310]

[email protected]
CPT clay 10 289 exp [185]
CPT deep glacial sands 20 ∼ 35 exp [400]
CPT clay 9.6 [211]
CPT North sea soil 13.9 [192]
CPT marine clay 400 exp [229]
CPT North sea clay 30 [368]
CPT North sea soil 37.5 [192]
CPT (friction ratio) clay 0 ∼ 0.6 exp [10]
CPT (friction ratio) clay to silty clay 0.1 exp [381]
Appendix A Database on stochastic soil properties

∼ 0.5
CPT (friction ratio) clay, silty sand 0.1 ∼ 0.5 exp [381]
CPT (friction ratio) clean sand, silty sand 0.2 ∼ 0.6 exp [381]
CPT (friction ratio) clay, silty sand 0.2 ∼ 0.3 exp [381]
CPT (friction ratio) claen sand, silty sand 0.3 ∼ 0.6 exp [381]
CPT (friction ratio) clay, silty sand 0.4 exp [381]
CPT (friction ratio) silty sand to sandy silt 0.4 exp [381]
CPT (friction ratio) claen sand,silty sand 0.5 ∼ 0.6 exp [381]
CPT (friction ratio) silty sand to sandy silt 0.6 exp [381]
CU (triaxial) offshore soil 1 [192]
CU (triaxial) offshore soil 3 [192]
density compacted clay 5 [25]
density compacted clay 4 ∼5 [25]
Continued on next page
 Table A.3 – continued from previous page
property soil type θver [m] θhor [m] applied spatial model reference
density clay 10 ∼ 15 [9]
Index Properties marine clay 1.6 exp [227]
liquid limit soft silty loam 0 ∼ 3.6 18 ∼ 22 exp [10]
liquid limit clay »0 [227]
liquid limit clay »0 [228]

[email protected]
modulus of elasticity clay 2 ∼5 [186]
modulus of elasticity clay 6 298.5 spherical [185]
modulus of elasticity clay 400 [235]
permeability glacial Lacustirne sand 0.1 3 [363]
permeability glacial outwash sand 0.3 5 [155]
permeability clay 0.4 5 [343]
permeability fine sands 1.5 14 [323]
permeability clay, fine sands 5.4 3 [230]
permeability silty clay 0.1 [340]
permeability fluvial soil 7.6 [135]
permeability alluvial soil 15 [180]
permeability alluvial aquifer 150 [91]
permeability gravelly loamy sand 500 [313]
permeability alluvial clay 800 [46]
permeability alluvial aquifer 820 [98]
permeability alluvial aquifer 1,800 [91]
permeability clay 4,000 [1]
ϕ’ clay 800 [301]
preconsolidation pressure soft, silty clay 0.6 exp [99]
preconsolidation pressure soft, silty clay 180 exp [99]
sand fraction soft silty loam 0 ∼ 5.4 19.6 ∼ 23 exp + cos [10]
sand fraction clay, silty sand 0.3 exp [381]
sand fraction silty sand,clay 0.6 exp [381]
Continued on next page

201
A.3 Database on the variability of cohesive soil
 Table A.3 – continued from previous page

202
property soil type θver [m] θhor [m] applied spatial model reference
sand fraction clay, silty sand 1 exp [381]
silt fraction soft silty loam 0 ∼ 5.4 19.5 ∼ 23 exp + cos [10]
silt fraction soft silty loam 0 ∼ 3.8 exp [10]
UCS Mexico City clay 0 ∼ 0.9 exp [10]
unit weight soft silty loam 0 ∼3 17 ∼ 22 exp + cos [10]
unit weight soft clay 1.2 [386]
void ratio soft silty loam 16.8 exp + cos [10]

[email protected]
0 ∼ 4.5 ∼ 22
Void ratio soft, silty clay 3 exp [99]
VST very soft clay 1 22 [41]
VST soft clay 1.2 [15]
VST sensitive clay 2 [78]
VST soft clay 2.4 exp [15]
VST sensitive clay 3 30 [353]
VST marine clay 3 [353]
VST organic soft clay 3.1 [352]
VST soft clay 6.2 exp [15]
Appendix A Database on stochastic soil properties

VST soft clay 0 ∼ 0.5 exp + cos [10]

VST marine clay 30 [353]
water content sandy clay 40 ∼ 70 spherical [259]
water content soft organic silty clay 3 exp [99]
water content clay 0 ∼ 2.5 exp [10]
water content mexican clay 0 ∼ 1.6 exp [10]
water content soft, silty loam 0 ∼5 exp + cos [10]
water content clay 5 ∼ 10 [9]
water content sandy clay 400 [55]
 A.3 Database on the variability of cohesive soil

Figure A.2 shows the frequency of the used correlation functions and techniques used
in the different entries of the database entries. One can clearly deduce from figure A.2
that nearly 50% of the literature sources in the database only offered the experimental
variogram or engineering judgement, but not a theoretical covariance function. Many of
the studied publications fitted an exponential correlation function to the experimental
values.

83% cohesive soils

frictional soils

59% 59% rock

60 %

30% 30% 31% 28%

30 %
17%
0%
0%
CPT permeability others

Figure A.1: Statistics of the available properties in the literature database.

frictional soils cohesive soils

29 %

48%

45 %
4% 47%
7% 4% 1%

exponential LAS
Gaussian, spherical, bilinear without

Figure A.2: Frequency of the correlation functions in the database for frictonal and cohe-
[email protected]
sive soils.

203
Appendix A Database on stochastic soil properties

pdf

4000 (a)
2000

0
0 10 20 30 40 50
vertical correlation length qver

pdf pdf pdf

4000 (b) 10 (c) (d)

50
2000 5

0 0 0
0 5 10 20 30 50
qver qver qver

Figure A.3: Histogram of vertical correlation lengths in cohesive soils for all scales (a),
for small(b), medium (c) and large scales (d).

[email protected]

204
 A.3 Database on the variability of cohesive soil

pdf
1500 (a)
1000

500

0 50 100 150 200 250 300 350 400

horizontal correlation length qver

pdf (b) pdf (c)

600 150

400 100

200 50

0 20 40 60 80 100 200 300 400

qhor qhor

Figure A.4: Histogram of horizontal correlation lengths in cohesive soils for all scales (a),
medium (b) and large scales (c).

[email protected]

205
Appendix A Database on stochastic soil properties

pdf
1500
(a)
1000

500

0
0 5 10 15 20 25 30 35
vertical correlation length qver

pdf pdf
(b) (c)
400
50
200

0 0
0 5 25 30 35
qver qver

Figure A.5: Histogram of vertical correlation lengths in frictional soils for all scales (a),
medium (b) and large scales (c).

pdf
300
(a)
200
100

0 10 20 30 40 50 60 70 80 90
horizontal correlation length qhor

pdf pdf
80 (b) (c)
100
60
40
20
0
0 10 20 90
qhor qhor

Figure A.6: Histogram of horizontal correlation lengths in frictional soils for all scales
(a), medium (b) and large scales (c).
[email protected]

206
 Appendix B
Measurement data of CPT - databases

[email protected]

207
Appendix B Measurement data of CPT - databases

Friction Ratio Tip Resistance

Fs/Qt (%) Qt (MN/m²)
0.00 0.0 10.0 0.0 50.0

5.00

10.00
Depth
(m)
15.00

20.00

Figure B.1: Plan view and typical CPT profile of NGES site A LAMEDA.

[email protected]

208
 Friction Ratio Tip Resistance
Fs/Qt (%) Qt (MN/m²)
0.00 0.0 10.0 0.0 50.0

5.00

Depth
(m)

20.00

25.00

30.00

Figure B.2: Plan view and typical CPT profile of NGES site E VANSVILLE A REA.

[email protected]

209
Appendix B Measurement data of CPT - databases

Friction Ratio Tip Resistance

Fs/Qt (%) Qt (MN/m²)
0.00 0.0 10.0 0.0 50.0

Depth
(m)

10.00

20.00

Figure B.3: Plan view and typical CPT profile of NGES site L ANCESTER.

[email protected]

210
 Friction Ratio Tip Resistance
Fs/Qt (%) Qt (MN/m²)
0.00 0.0 10.0 0.0 50.0

5.00

Depth
(m)

20.00

25.00

30.00

Figure B.4: Plan view and typical CPT profile of NGES site S AN B ERNADINO C OUNTY.

[email protected]

211
Appendix B Measurement data of CPT - databases

Friction Ratio Tip Resistance

Fs/Qt (%) Qt (MN/m²)
0.00 0.0 10.0 0.0 50.0

Depth
(m)

10.00

20.00

Figure B.5: Plan view and typical CPT profile of NGES site S AN L UIS O BISPO C OUNTY.

[email protected]

212
 Friction Ratio Tip Resistance
Fs/Qt (%) Qt (MN/m²)
0.00 0.0 10.0 0.0 50.0

5.00

Depth
(m)

20.00

25.00

30.00

Figure B.6: Plan view and typical CPT profile of NGES site S ANTA C LARA C OUNTY.

[email protected]

213
Appendix B Measurement data of CPT - databases

Friction Ratio Tip Resistance

Fs/Qt (%) Qt (MN/m²)
0.00 0.0 10.0 0.0 50.0

5.00

Depth
(m)

20.00

25.00

30.00

Figure B.7: Plan view and typical CPT profile of NGES site S OLANO C OUNTY.

[email protected]

214
 Tip Resistance Sleeve Friction
q (MPa) fs (kPa)
T
0 10 20 30 40 50 60 0 1000 2000
0

6
Depth (m)

10

12

Figure B.8: Plan view and typical CPT profile of PEER site A NSSALL.

[email protected]

215
Appendix B Measurement data of CPT - databases

qc (MPa) fs (MPa)
0 10 20 30 40 0.00 0.10 0.20
0 0

1 1

2 2

3 3

4 4

5 5
Depth (m)

6 6

7 7

8 8

9 9

10 10

11 11

12 12

Figure B.9: Plan view and typical CPT profile of PEER site B ERKELEY.

[email protected]

216
 Appendix C
Basic definitions & statistical background

C.1 Moments of a distribution function

Every distribution function can be characterized by different moments listed below.

• First moment of distribution & mean value

Z ∞
µ= x f (x) dx = E[X] (C.1)
−∞

• Second moment of distribution & variance

Z ∞
2
 
σ = (x − µ)2 f (x) dx = E (X − µ)2 (C.2)
−∞

Coefficient of variation
σ
COV = (C.3)
µ
• Third moment of distribution & skewness
Z ∞  
1 3 µ3 E (X − µ)3
γ1 = 3 (x − µ) f (x) dx = 3 = (C.4)
σ −∞ σ σ3

• Fourth moment of distribution & excess kurtosis

Z ∞  
1 4 µ4 E (X − µ)4
γ2,excess = 4 (x − µ) f (x) dx − 3 = 4 − 3 = −3 (C.5)
σ −∞ σ σ4

[email protected]

217
Appendix C Basic definitions & statistical background

C.2 Distribution functions

Looking into statistical textbooks like Sachs [316] or Abramovich & Stegun [2], one can
find a vast number of different probability distribution function. According to Remy et
al. [304], a distribution function should account for all information available; it provides
all that is needed to quantify the uncertainty about the actual outcome of the variable x.
For example,
• probability intervals can be derived as
Z b
Prob {Z ∈ (a, b)} = F(b) − F(a) = f (z) dz. (C.6)
a

• quantile values can be derived such as the 0.10 quantile or the 1st decile:
q0.10 = F −1 (0.10) = z − outcome value such that Prob(Z ≤ q0.10 ) = 0.10 (C.7)

Phoon [286] as well as Baecher & Christian [23] point out the normal and the lognor-
mal distribution function as widespread probability distribution functions in geotechni-
cal engineering, which are explained in detail afterwards.

C.2.1 Normal distribution function

The normal distribution is a continuous probability distribution that has a bell-shaped
probability density function, also known as the Gaussian function.
   
2 1 (x − µ)2 1 x−µ
f (x; µ, σ ) = √ exp − = Φ (C.8)
σ 2π 2 σ2 σ σ
The location of the distribution function is controlled by the mean value µ; the shape
is defined by the variance σ 2 . A normal distribution with N (µ = 0, σ 2 = 1) is called
the standard normal. Function f(x) is unimodal and symmetric around the point x = µ,
which is at the same time the mode (peak value of the distribution function that occurs
most frequently), the median (middle value and location parameter separating the the
lower and upper half of the distribution) and the mean of the distribution. The inflection
points of the curve occur one standard deviation away from the mean (i.e. at x = µ − σ
and x = µ + σ). The n-th derivative is given by Φ(n) (x) = (−1)n Hn (x) Φ(x), where Hn is
the Hermite polynomial of order n.
The cumulative distribution function (CDF) describes probability of a random variable
falling in the interval (−∞, x]. The CDF of the standard normal distribution Φ can be
computed as an integral of the probability density function:
Z x    
1 2 1 (x − µ)2 1 x−µ
Φ(x) = √ f (x; µ, σ ) = √ exp − = Φ (C.9)
2 π −∞ σ 2π 2 σ2 σ σ
 
x−µ
F (x; µ, σ) = Φ (C.10)
[email protected] σ

218
 C.2 Distribution functions

Table C.1: Summary of the normal distribution

support x∈R
mean µ
median µ
mode µ
variance σ2
skewness 0
excess kurtosis 0

C.2.2 Lognormal distribution function

In probability theory, a log-normal distribution is a continuous probability distribution
of a random variable whose logarithm is normally distributed. If X is a random variable
with a normal distribution, then Y = exp(X) has a lognormal distribution; likewise, if Y
is log-normally distributed, then X = log(Y ) is normally distributed.
A variable might be modelled as log-normal if it can be thought of as the multiplica-
tive product of many independent random variables each of which is positive. In a
lognormal distribution X, the parameters denoted µ and σ, are the mean and standard
deviation, respectively, of the variable’ s natural logarithm (by definition, the variable’ s
logarithm is normally distributed), which means using Z as a standard normal variable.

X = eµ+σ Z (C.11)
On a non-logarithmised scale, µ and σ can be called the location parameter and the scale
parameter respectively. The probability density function of a lognormal distribution is
as shown in the following equation.
 
1 (ln x − µ)2
fX (x; µ, σ) = √ exp x>0 (C.12)
xσ 2π 2 σ2

The cumulative distribution function of a lognormal distribution is

 
ln x − µ
FX (x; µ, σ) = Φ (C.13)
σ

[email protected]

219
Appendix C Basic definitions & statistical background

Table C.2: Summary of the lognormal distribution

support x ∈ (0, +∞)

2
mean e µ+σ /2
median eµ
2
mode e µ−σ
2 2
variance e σ −√1 e 2µ+σ
2
skewness e σ + 2 e σ2 − 1
2 2 2
excess kurtosis e 4σ + 2 e 3σ + 3 e 2σ − 6

Table C.3: Summary of the multivariate normal distribution

support x ∈ µ + span(Σ) ⊆ Rk )
mean µ
median µ
mode µ
variance Σ

C.2.3 Multivariate normal distribution

The multivariate normal distribution of a k-dimensional random vector z = [Z1 , Z2 , ..., Zk ]
of the random variables Z can be written in the following notation:

z = Nk (µ, COVZZ ) (C.14)

with a k - dimensional mean vector

µ = [E[Z1 ], E[Z2 ], ..., E[Zk ]] (C.15)

and a k × k covariance matrix

COVZZ = [ COV [Zi , Zj ] ] , i = 1, 2, ..., k; j = 1, 2, ..., k (C.16)

The main diagonal gives the variance and the off-diagonals are symmetrical covariances.
The covariance matrix is not singular and definite. In case of a singular covariance ma-
trix, the corresponding distribution has no density.
The probability density function f (z) can be written in the following form.

f (z) = (2 π)k 2 |COVZZ |−1/2 exp[−1/2 (z − µ)T COVZZ (z − µ)] (C.17)

According to Sachs [316], no analytical expression exists for den cumulative density
function.

[email protected]

220
 C.2 Distribution functions

C.2.4 Estimation methods for probability distribution functions

In civil engineering practice many parameter estimation methods for probability distri-
bution functions are in circulation. According to van Gelder [382], well known methods
are for example:

• the method of moments,

• the method of maximum likelihood,

• the method of least squares,

• the method of Bayesian estimation,

• the method of probability weighted moments,

• the method of L-moments and

• the method of maximum entropy.

In Gelder [382] a comparison of the different methods together with applications of these
approaches are presented.

[email protected]

221
[email protected]
 Appendix D
Polynomial Chaos Expansion

D.1 Hermite polynomials

The Hermite polynomials Hen (ξ) are solution of the following differential equation:
y ′′ − x y ′ + n y = 0 n∈N (D.1)
The Hermite polynomials Hen (x) are defined in the following formula as stated in Abramovic
& Stegun [2].
He0 (x) = 1
(D.2)
Hen+1 (x) = x Hen (x) − n Hen−1 (x)
They are orthogonal with respect to the Gaussian probability measure:
Z ∞
Hem (x) Hen (x) ϕ(x) dx = n! δmn (D.3)
−∞
√
where ϕ(x) = 1/ 2π exp(−x2 /2) is the standard normal PDF. If ξ is a standard normal
random variable, the following relationship holds:
E[Hem (ξ) Hen (ξ)] = n! δmn (D.4)
The first three Hermite polynomials are

He1 (x) = x
He2 (x) = x2 − 1 (D.5)
He3 (x) = x3 − 3 x

D.2 Computation of the expectation of products of

bivariate Hermite polynomials
A multivariate Hermite polynomial is defined as the product of several univariate Her-
mite polynomials of different variables. For n variables, its expression is given by
Γi1,i2,...in (ξ1 , ξ2 , . . . , ξn ) = Hei1 (ξ1 ) · Hei2 (ξ2 ) · . . . · Hein (ξn ) (D.6)
In case of two variables, the expression of the bivariate Hermite polynomials is:
[email protected] Γi,j (ξ1 , ξ2 ) = Hei1 (ξ1 ) · Hei2 (ξ2 ) (D.7)

223
Appendix D Polynomial Chaos Expansion

For a simple use in mathematical formulas, the multivariate Hermite polynomials are
often renamed and sorted by using only one numerical index, for example:

Γi,j (ξ1 , ξ2 ) = ψi (D.8)

Each polynomial ψi of the basis of the PCE of two variables ξ1 and ξ2 can be entirely
defined by two indexes i1 and i2 such that

ψi = Γi,j (ξ1 , ξ2 ) = Hei (ξ1 ) · Hei (ξ2 ) (D.9)

With this notation, the following equations have been derived by Sudret & Kiureghian
[366].
E(ψi2 ) = i1 ! · i2 !
E(ψi · ψj · ψk ) = Di1,j1,k1 · Di2,j2,k2 (D.10)
E(ψi · ψj · ψk · ψl ) = Di1,j1,k1,l2 · Di2,j2,k2,l2
In these expressions, the D terms are obtained by:
 (

 i! j! (i + j + k)even
Ci,j,k = ((i+j−k)/2)! ((j+k−i)/2)! ((k+i−j)/2)!
k ∈ [|i − j| , i + j] (D.11)


0 otherwise

Di,j,k = Ci,j,k · k! (D.12)

X
Di,j,k,l = Di,j,q · Ck,l,q (D.13)
q≥0

[email protected]

224
 D.3 Two variables PCEs used for different values of the PCE order

D.3 Two variables PCEs used for different values of the

PCE order
M = expansion order of the PCE
PCEnum = number of the unknown PCE coefficients
m = number of the available collocation points

M Roots of the uni- Expression PCEs for different orders PCEnum m

variate Hermite
polynomials or-
der M+1
 √
2 0; ± 3 U2 = a0,0 · Γ0,0 + a1,0 · Γ1,0 (ξ1 ) + a0,1 · Γ0,1 (ξ2 ) + 6 9
a2,0 · Γ2,0 (ξ1 ) + a1,1 · Γ1,1 (ξ1 , ξ2 ) + a0,2 ·
Γ0,2 (ξ2 )
n p √ o
3 ± 3± 6 U3 = a0,0 · Γ0,0 + a1,0 · Γ1,0 (ξ1 ) + a0,1 · Γ0,1 (ξ2 ) + 10 16+1
a2,0 · Γ2,0 (ξ1 ) + a1,1 ·1,1 (ξ1 , ξ2 ) + a0,2 ·
Γ0,2 (ξ2 )+a3,0 ·Γ3,0 (ξ1 )+a2,1 ·2,1 (ξ1 , ξ2 )+
a1,2 · Γ1,2 (ξ1 , ξ2 ) + a0,3 · Γ0,3 (ξ2 )
n p √ o
4 0; ± 5± 10 U4 = a0,0 · Γ0,0 + a1,0 · Γ1,0 (ξ1 ) + a0,1 · 15 25
Γ0,1 (ξ2 )+a2,0 ·Γ2,0 (ξ1 )+a1,1 ·Γ1,1 (ξ1 , ξ2 )+
a0,2 · Γ0,2 (ξ2 ) + a3,0 · Γ3,0 (ξ1 ) + a2,1 ·
Γ2,1 (ξ1 , ξ2 ) + a1,2 · Γ1,2 (ξ1 , ξ2 ) + a0,3 ·
Γ0,3 (ξ2 )+a4,0 ·Γ4,0 (ξ1 )+a3,1 ·Γ3,1 (ξ1 , ξ2 )+
a2,2 · Γ2,2 (ξ1 , ξ2 ) + a1,3 · Γ1,3 (ξ1 , ξ2 ) + a0,4 ·
Γ0,4 (ξ2 )
 
 ±3.324257 ; 
5 ±3.889176 ; U5 = a0,0 · Γ0,0 + a1,0 · Γ1,0 (ξ1 ) + a0,1 · 21 36+1
 
±0.616707 Γ0,1 (ξ2 )+a2,0 ·Γ2,0 (ξ1 )+a1,1 ·Γ1,1 (ξ1 , ξ2 )+
a0,2 · Γ0,2 (ξ2 ) + a3,0 · Γ3,0 (ξ1 ) + a2,1 ·
Γ2,1 (ξ1 , ξ2 ) + a1,2 · Γ1,2 (ξ1 , ξ2 ) + a0,3 ·
Γ0,3 (ξ2 )+a4,0 ·Γ4,0 (ξ1 )+a3,1 ·Γ3,1 (ξ1 , ξ2 )+
a2,2 · Γ2,2 (ξ1 , ξ2 ) + a1,3 · Γ1,3 (ξ1 , ξ2 ) + a0,4 ·
Γ0,4 (ξ2 )+a5,0 ·Γ5,0 (ξ1 )+a4,1 ·Γ4,1 (ξ1 , ξ2 )+
a3,2 · Γ3,2 (ξ1 , ξ2 ) + a2,3 · Γ2,3 (ξ1 , ξ2 ) + a1,4 ·
[email protected] Γ1,4 (ξ1 , ξ2 ) + a0,5 · Γ0,5 (ξ2 )

225
[email protected]
 Appendix E
Sequential simulation algorithms
The aim of sequential simulation, as it was originally constructed, is to reproduce de-
sirable multivariate properties through the sequential use of conditional distributions
[79, 97]. Consider a set {Z(uj ), j = 1, . . . N } of N random variables
 l defined at N loca-
tions uj . The objective is to generate several joint realizations z (uj ), j = 1, . . . , N , l =
1, . . . , L conditional to the available data and to some structural model such as the vari-
ogram. It can be shown that the N - point multivariate distribution can be decomposed
into a set of N one-point conditional cumulative distribution function’s as

F (u1 , . . . , un ; z1 . . . , zN | (n)) = F (uN ; zN | (n + N − 1)) · (E.1)

·F (uN −1 ; zN −1 | (n + N − 2)) ·
·F (u2 ; z2 | (n + 1)) · F (u1 ; z1 | (n))

where F (uN ; zN | (n + N − 1)) is the conditional CDF of Z(uN ) given the set of n orig-
inal data values and the previous (N − 1) realizations z (l) (uj ) , j = 1, . . . , N − 1 [145].
This decomposition allows us to generate an image by sequentially visiting each node.
Sequential simulation, under a given multivariate distribution, amounts to read the de-
composition E.2 from left to right, i.e. the purpose of sequential simulation is to repro-
duce the properties of the given multivariate distribution. The simulation algorithm
proceeds as follows according to Goovaerts [145]:

• Perform a transformation if necessitated by the theory

• Define a random path visiting all nodes

• For each node ui , i = 1, . . . , N do

– model the conditional distribution F (ui ; z| (n + i − 1)) of Z (ui ), given the n
original data values and all i − 1 previously drawn values z (l) (uj ) , ui , j =
1, . . . , i − 1
– draw the simulated value z (l) (uj ) from F (ui ; z|(n + i − 1))

• end loop

• Perform a back-transform to identify the target histogram

The sequential simulation principle is independent of the algorithm or model to establish

[email protected]
the equation E.2 [97].

227
Appendix E Sequential simulation algorithms

Within the Sequential Gaussian Simulation algorithm, all conditional cumulative dis-
tribution F (·) are assumed Gaussian and their means and variances are given b a series
fo N simple kriging systems; in terms of the Sequential Indicator Simulation algorithm,
the conditional cumulative distribution are obtained directly by indicator kriging, as
shown below.

E.1 Sequential Gaussian Simulation algorithm

The most straight forward algorithm for generating realizations of a multivariate Gaus-
sian field is provided by the sequential principle. As described by various authors
[97, 403], each value is simulated sequentially according to its normal conditional cu-
mulative distribution (nccd) function, which must be determined at each location to be
simulated. The conditioning data comprise all the original data and all previously sim-
ulated values within the neighbourhood of the point being simulated. Sequential Gaus-
sian simulation (SGSIM) starts with the assumption that the kriging error is normally
distributed with mean 0 and variance σK 2
(x0 ), i.e. N (0, σK
2
(x0 )). In these circumstances
b
the probability distribution for the true values is N (Z(x0 ), σK 2
(x0 )); it is simply shifted
b
by Z(x0 ).
Deutsch & Journel [97] as well Webster & Oliver [403] list the following steps of SGSIM
algorithm:

1. Ensure that the data are approximately normal; transform to a standard normal
distribution if necessary.

2. Compute and model the variogram.

3. Specify the coordinates of the points at which you want to simulate. These will
usually be on a grid.

4. Determine the sequence in which the points, xj ; j = 1, 2, . . . will be visited for the
simulation. Choosing the points at random will maximize the diversity of different
realizations.

5. Simulate at each of these points as follows:

(a) Use simple kriging with the variogram model to obtain Z(x b i ) and σ 2 (xi .
K
b
(b) Draw a value at random from a normal distribution N (Z(xi ), σK (xi )).
2

(c) Insert this value into the grid at xi , and add it to the data.
(d) Proceed to the next node and simulate the value at this point in the grid.
(e) Repeat steps (a) to (c) until all of the nodes have been simulated.

6. Back-transform the simulated values if there is a need to.

The SGSIM algorithm is very fast and straightforward because the modelling of the
ccdf at each location u requires the solution of only a sling kriging system at that loca-
[email protected]
tion. The implicit assumption is that the spatial variability of the attribute values can be

228
 E.2 Kriging algorithm

fully characterized by sa single covariance function. In particular, this precludes mod-

elling patterns of spatial continuity specific to difference classes of values. Possibly more
critical, the maximum entropy property of the multi-Gaussian random field model does
not allow for any significant correlation of extreme values [96, 97, 143, 144, 189] and for
a given covariance maximizes their scattering in space, which is called destructuration
effect by Goovaerts [145] .

E.2 Kriging algorithm

As reported by Remy et al. [304], Kriging has been historically at the source of accep-
tance of geostatistics and remains a major data integration tool and is used in most geo-
statistical estimation and simulation algorithms. Kriging is a generic name adopted by
geostatisticians for a family of generalized least squares regression algorithms, [145].
All kriging estimators are but variants of the basis linear regression estimator Z ∗ (u) as
defined as
n(u)
X
∗
Z (u) − m(u) = λα (u) [Z(uα ) − m(uα )] (E.2)
α=1
where λα is the weight assigned to the location z(uα ) interpreted as a realization of the
random variable Z(u). The quantities m(u) and m(uα ) are the expected values of the
random variables Z(u) and Z(uα ). The number of data involved in the estimation as
well as their weights may change form on location to another. Goovaerts [145] reports
amongst others [79, 97], that in practice, only the n(u) closest to the location u being es-
timated are retained, the data within a given neighbourhood or window W (u) centered
on u.
The interpretation of the unknown value z(u) and data values z(uα ) as realizations
of the random variables Z(u) and Z(uα ) allows on to define the estimation error as a
random variable Z ∗ (u) − Z(u). All flavours of kriging share the same objective of min-
imizing the estimation or error variance σE2 (u) under the constraint of unbiasedness of
the estimator that is,
σE2 (u) = Var [Z ∗ (u) − Z(u)] (E.3)
is minimized unter the constraint that.
E [Z ∗ (u) − Z(u)] = 0 (E.4)
Goovaerts [145] distinguishes three kriging variants according to the model considered
for the rend m(u).
1. Simple kriging considers the mean m(u) to be known and constant through the
study area A
m(u) = m , known ∀ u ∈ A (E.5)
2. Ordinary kriging accounts for local fluctuations of the mean by limiting the domain
of stationarity of the mean to the local neighbourhood W (u)
[email protected] m(u′ ) = constant but known ∀ u′ ∈ W (u) (E.6)

229
Appendix E Sequential simulation algorithms

3. Kriging with a trend model considers that the unknown local mean m(u′ ) smoothly
varies with each local neighbourhood W (u), hence over the entire study area A.
The trend component is modelled as a linear combination of functions fk (u) of the
coordinates:
XK
′
m(u ) = ak (u′ ) fk (u′ ) (E.7)
k=0

withak (u) ≈ ak constant but unknown ∀ u′ ∈ W (u)

The coefficients m(u′ ) are unknown and deemed constant with each local neigh-
bourhood W (u). By convention, f0 (u′ ) = 1, hence the case K = 0 is equivalent to
ordinary kriging (constant but unknown mean a0 ).

E [Z ∗ (u) − Z(u)] = 0 (E.8)

E.3 Sequential Indicator Simulation algorithm

In earth sciences, connected strings of large or small values are common and are crit-
ical for many applications, as shown by various publications e.g. [97, 143, 144, 189].
Amongst others, Goovaerts [145] states that the multi-Gaussian random field model is
inappropriate, whenever the structural analysis or qualitative information indicates that
extreme values could be better correlated in space than medium values. Even in absence
of of information about connectivity of extreme values, the user must be aware that the
analytical simplicity of SGSIM is balancey by the risk of understanding the potential for
critical features, such as strings of small or large values.
The Sequential Indicator Simulation (SGSIM) algorithm is the most widely used non-
multi-Gaussian simulation technique. The indicator formalism introduced in section
2.4.2 is used to model the sequence of conditional cdfs from which simulated values are
drawn. Unlike SGSIM, the indicator approach allows one to account for class specific
patterns of spatial continuity through different indicator semivariogram models [145].
Consider first the simulation of a single continuous attribute z at N grid nodes u′j
conditional only to the z- data {z (uα ) , α = 1, . . . , n}. Sequential indicator simulation
proceeds as follows [145]:

• Discretize the range of variation of z into (K + 1) classes using K threshold values

zk . Then, transform each datum z(uα ) into a vector of hard indicator data, defined
as (
1 if z(uα ) ≤ zk k = 1, . . . , K
i (uα ; zk ) =
0 otherwise

• Define a random path visiting each node of the grid only once.
[email protected]
• At each node u′ :

230
 E.4 Comparison of SGSIM and SISIM algorithms

1. Determine the K ccdf values F (uα ; zk |(n)) using an indicator kriging algo-
rithm e.g. simple indicator kriging. The conditioning information consists of
indicator transform of neighbouring original z-data and previously simulated
z-values.
2. Correct for any order relation deviations resulting from negative kriging weights
[145]. Then build a complete ccdf model (u′ ; z|(n)) , ∀z.
3. Draw a simulated valued z (l) (u′ ) form that conditional cumulative distribu-
tion function.
4. Add the simulated value to the conditioning data set.
5. Proceed to the next node along the random path, and repeat steps 1 to 4.

Goovaerts [145] explains to repeat

 (l′ ) the entire procedure
′ with a different random path to
generate another realization z (uj , j = 1, . . . , N , l 6= l. Additional information on
′

the SGSIM algorithm can be found in various publications [79, 96, 97, 143–145, 189, 304].

E.4 Comparison of SGSIM and SISIM algorithms

Two different random field realisations are shown in figure E.1. Both random fields fol-
low the same isotropic correlation length, but they look different. One can deduce from
the description of these two algorithms the similarities and the differences. It is shown
in figure E.2 (a) that the variogram and in figure figure E.2 (b) the cumulative distribu-
tion function are nearly the same. But in terms of the indicator correlation lengths both
algorithms offer a different insight into their spatial correlation structure . As shown in
figure E.3, the indicator correlation lengths are different for the extreme values, whereas
for the median values they are similar. One can derive from this that the influence of
different algorithms for the simulation of spatial correlation results in two completely
different visualisations of spatial variability.

[email protected]

231
Appendix E Sequential simulation algorithms

-4
-2
0
2
4

qver

qhor

Figure E.1: Random field realisations of the SGSIM (a) and the SISIM (b) algorithm with
an isotropic, spatial correlation (θver = θhor ).

For the sake of completeness, figures E.4 and E.5 summarize different methods for the
simulation of random fields, which are taken from Chiles & Delfiner [79]

[email protected]

232
 E.4 Comparison of SGSIM and SISIM algorithms

g(t) g(t)
Cumulative probability

0.9 0.6

0.7
0.4
0.5
SGSIM algorithm
0.3 0.2

0.1 SISIM algorithm

0.0
-3 -1 1 3 0 40 80 120
lag t
data

Figure E.2: Cumulative probability distribution (a) and variograms (b) of the SGSIM and
SISIM random field in figure E.1.
cumulative distribution

-4 -2 0 2 4
modulus of elasticity [MN/m²] indicator correlation
length qind
SGSIM maximum indicator
SISIM correlation length qind,max

Figure E.3: Indicator correlation lengths for each threshold of the SGSIM and SISIM
algorithm.

[email protected]

233
Appendix E Sequential simulation algorithms

Figure E.4: Main characteristics of various random field simulation methods from Chiles
& Delfiner [79].

[email protected]

234
 E.4 Comparison of SGSIM and SISIM algorithms

Figure E.5: Main characteristics of various random field simulation methods from Chiles
& Delfiner [79].

[email protected]

235
[email protected]
 Appendix F
Failure criteria in constitutive modelling of soil
Benz [38] states, that the Mohr-Coulomb failure criterion for soils is one of the earliest
and most trusted failure criteria.
It is experimentally verified in triaxial compression and extension and is of striking
simplicity. However, the Mohr-Coulomb (MC) criterion is very conservative for inter-
mediate principal stress states between triaxial compression and extension. The Mohr-
Coulomb failure criterion in principal stress space is defined as:

f1 = |σ1 − σ2 | − (σ1 + σ2 ) sin ϕ − 2 c cos ϕ

f2 = |σ2 − σ3 | − (σ2 + σ3 ) sin ϕ − 2 c cos ϕ (F.1)
f3 = |σ3 − σ1 | − (σ3 + σ1 ) sin ϕ − 2 c cos ϕ

Matsuoka & Nakai [243] (MN) proposed a failure criterion that is in better agreement
with the experimental data, which is shown in figure 5.7 with the MC and the Lade -
Duncan criteria [213] (LADE). Matsuoka & Nakai [243] propose the concept of a Spatial
Mobilized Plane (SMP), which defines the plane of maximum spatial, averaged particle
mobilization in principal stress space. The SMP is geometrically constructed by deriving
the mobilized (Mohr-Coulomb) friction angles for each principal stress pair separately
(figure F.1, left) and sketching the respective mobilized planes in principal stress space
(figure F.1, right). Matsuoka & Nakai derive their failure criterion by limiting the aver-
aged ratio of spatial normal stress to averaged spatial shear stress on this plane. Their
failure stress ratio can be expressed as a simple function of the first, second, and third
stress invariant, I1 , I2 , and I3 as shown in equation F.3 and F.3. With the SMP concept, the
Matsuoka-Nakai criterion automatically retains the well established material strength of
the Mohr-Coulomb criterion in triaxial compression and extension. The likewise well-
known failure criterion by Lade & Duncan appears compared to the Mohr-Coulomb
criterion and the Matsuoka-Nakai criterion rather optimistic in plane strain conditions
and triaxial extension. Benz [38] points out that using bifurcation analysis, progressive
failures would most likely "correct" for the Lade criterion’s overly optimistic, ultimate
material strength estimate.
Both failure criteria, Matsuoka-Nakai and Lade, are functions of the first, second, and
third stress invariants, I1 , I2 , and I3 respectively:

I1 I2 9 − sin2 ϕ
fMN = − c1 = 0 with c1 = (F.2)
I3 −1 + sin2 ϕ
I13 (−3 + sin ϕ)3
[email protected]
fLade = − c2 = 0 with c2 = (F.3)
I3 (−1 + sin ϕ) (−1 + sin ϕ)2

237
Appendix F Failure criteria in constitutive modelling of soil

(a) (b)
t sI
s1
SMP
Mobilized
fmob23
Planes fmob13
fmob13 45°+ 2
fmob12
fmob12 s3
45°+ 2
sIII
s3 s2 s1 s
fmob23
45°+ 2
s2
sII

Figure F.1: The SMP concept of Matsuoka & Nakai.

(a) Three mobilized planes where the maximum shear stress to normal stress
ratio is reached for the respective principal stresses.
(b) SMP in principal stress space from Benz [38].

where

I1 = σij
1
I2 = (σij σij − σii σjj ) (F.4)
2
1
I3 = (σii σjj σkk + 2 σij σjk σki − 3 σij σji σkk )
6
In principal stress, the stress invariants simplify to:

I 1 = σ1 + σ2 + σ3
I2 = −σ1 σ2 − σ2 σ3 − σ3 σ1 (F.5)
I 3 = σ1 σ2 σ3

The constants c1 and c2 in equation F.3 and F.3 are defined so that both failure criteria are
identical to the Mohr-Coulomb criterion in triaxial compression.

[email protected]

238
 Appendix G
Application of non-intrusive SFEM to a
reliability problems
The non-intrusive Stochastic Finite Element Method (SFEM) is applied to a simple ex-
ample in order to visualize the steps described in section 4.3.3. The limit state eqution
5.11 is used for the comparison of the probability of failure between non-intrusive SFEM
and FORM, the statistical moments of the response and for the calculation of the Sobol
indices. Within this example the soil properties of the parametric study 1 in section 5.4
and table 5.3 and a COVϕ′ = 20% and COVc′ = 10% are used.
The deterministic results of the system response as a function of the random input
variables in the shown in figure G.1.
The probabilistic results of a parametric study are shown in the figure G.2. Generally
speaking, it can be deduced that the higher the face pressure, the higher the reliability of
the system.
FORM and non-intrusive SFEM represent this qualitatively in the same way. The in-
fluence of the cohesion c′ and of the friction angle ϕ′ remains constant while varying the
face pressure qt .
To enlighten the concept of non-intrusive SFEM, the collocation points for different
PCE orders are plotted in the Gaussian as well in the physical space (figure G.4). The
dependency of the statistical moments to the PCE order is shown in figure G.3. Together
with the results of figure G.5 and figure G.6, the reader can see the convergence of the
non-intrusive SFEM approximation. The mean value, the variance, the skewness and
also the kurtosis show only minor changes at the PCE order M = 5. This can also be
deduced from Figure G.5 : The PCE order M = 5 is acceptable for the probability of
failure pf . This is strengthened the graphs of the empirical error and the coefficient of
determination in figure G.3.
The difference of both approaches can be clearly seen in Figure 4. The lower the prob-
ability of failure, the bigger is the difference of FORM and CSRSM. This is due to the
approximation of the system response via a PCE within CSRSM, which becomes more
inaccurate with deceasing probability of failure. The dependency of the Sobol indices to
the PCE order M is rather low, which is enlightened in Figure G.7.
It can be concluded from the presented study that the non-intrusive SFEM cannot be
accurately applied to evaluate the small probability of failures, which was also reported
in literature e.g. Phoon [286]. This can be deduced to the concept approximation of the
stochastic system response using a high order polynomial, which does not represent the
small tails of the distribution of system response. But the concept of non-intrusive SFEM
[email protected]
is can be efficiently applied to evaluate the first and second statistical moments of the

239
Appendix G Application of non-intrusive SFEM to a reliability problems

system response as well as for the analytical evaluation of global sensitivities. These
benefits can be applied to any sort of uncertain input variable ranging from soil strength
properties, correlation lengths,... etc.
If this approach is applied to the system response assuming uncertain correlation
lengths, on as to keep in mind that the evaluation points are mean values and the fit-
ted PCE model will just approximate the real response surface; therefore, the sensitivity
factors derived from the PCE are not global factors but help to understand the impor-
tance between the input variables between each other.

system response
system response
qt [kN/m²]
40 30
35
30
25
25
20
15 20
10
5 15
0
45 10
fric 40 4.0 4.5 ²]
tion 35
ang 30 3.0 3.5
c’ [ kN/m
2.5 sion
el j
’ [° 25 1.5 2.0 cohe
]

pdf

pdf

Figure G.1: Input variables and system response of an PCE with order M = 5 and a face
[email protected]
pressure of qt = 40 kN/m2 .

240
 0
10

probability of failure pf
-5
10

-10
10
non-intrusive SFEM (M= 5)
-15 FORM
10
10 20 30 40 50
face pressure qt in kN/m²

Figure G.2: Comparison of non-intrusive SFEM and FORM in terms of the probability of
failure for a COVϕ′ = 5 % and COVϕ′ = 10 % .

-3
x 10
0.01
empirical error

PRESS

2
0.005
1

0 0
1 2 3 4 5 1 2 3 4 5
expansion order M of the PCE expansion order M of the PCE
determination coefficient

1
0.9995
0.999 determination coefficient
adjusted determination coefficient
0.9985 Q2
0.998
1 2 3 4 5
expansion order M of the PCE

Figure G.3: Acccuracy of the PCE fitting for a face pressure qt = 40 kN/m2 .
[email protected]

241
Appendix G Application of non-intrusive SFEM to a reliability problems

Gaussian space (ξj’ ξc’)

, Physical space (φ’, c’)
order M = 1 45
4
40
φ’ 35
j’ 0
30
-4
-4 -2 0 2 4 2 2.5 3 3.5
ξ c’ c’

order M = 2 45
4
40
ξ j’ 0 φ’ 35
30
-4
-4 -2 0 2 4 2 2.5 3 3.5
ξ c’ c

order M = 3 45
4
40
ξ j’ 0 φ’ 35
30
-4
-4 -2 0 2 4 2 2.5 3 3.5
ξ c’ c
c’

order M = 4 45
4
40
ξ j’ 0 φ’ 35
30
-4
-4 -2 0 2 4 2 2.5 3 3.5
ξ c’ c’

order M = 5 45
4
40
ξ j’ 0 φ’ 35
30
-4 25
-4 -2 0 2 4 1.5 2 2.5 3 3.5
ξ c’ c’

Figure G.4: Collocation points in Gaussian space and physical space for several PCE ex-
pansion orders M.

[email protected]

242
 2 0.4
Mean value

1.5

Variance
0.3

1 0.2

0.5 0.1

0 0
1 2 3 4 5 6 1 2 3 4 5 6
PCE order M PCE order M
1 2
Skewness

0.8 Kurtosis 1.5

0.6
1
0.4
0.2 0.5

0 0
1 2 3 4 5 6 1 2 3 4 5 6
PCE order M PCE order M

Figure G.5: Statistical moments of the approximated system response for a face pressure
qt = 40 kN/m2 .
proability of failure p f

-2
10

-4
10

-6
10

-8
10
0 2 4 6
PCE order M

Figure G.6: Probability of failure as a function of the PCE expansion order.

[email protected]

243
Appendix G Application of non-intrusive SFEM to a reliability problems

-3
0.951 0.052 6 x 10

Sobol index δφ’, c’

Sobol index δc’
Sobol index δφ’

0.95 0.051 5
0.949 4
0.05
0.948
0.049 3
0.947
0.048 2
0.946
0.945 0.047 1
0.944 0.046 0
1 2 3 4 5 6 1 2 3 4 5 6 1 2 3 4 5 6
PCE expansion order M PCE expansion order M PCE expansion order M

Figure G.7: Sobol indices for several PCE expansion orders for the face pressure qt = 40
kN/m 2 .

[email protected]

244
 Appendix H
Pluri-Gaussian Simulation Method

H.1 Introduction
Soil properties vary naturally through space as a result of the complex geological pro-
cesses through which soil evolve, [204]. Physical and chemical processes, including
structural deformation, deposition and diagenesis control the geometry and texture of
sedimentary deposits and create soil variability at different scales.
This multi-scale spatial variability of soil properties can be captured by geostatis-
tical simulation approaches. Koltermann & Gorelick [204] group the simulation ap-
proaches in three main categroies: structure imitating, process imitating and descriptive.
Structure-imitating methods rely on spatial statistics, probabilistic rules and determin-
istic constraints to depict geometric relations within aquifers and reservoirs. Process-
imitating methods solve governing equations to represent either the processes through
which soil form or the physics of subsurface fluid-flow and transport. Descriptive meth-
ods divide subsoil into zones by synthesizing measured soil properties and geologic
observations into a conceptual depositional model.
Structure-imitating methods are further subdivided into Gaussian and Non-Gaussian
methods. Gaussian-related methods produce images of a continuous variable with the
same mean, variance, covariance function and a Gaussian univariate distribution of val-
ues. The continuous distribution can be truncated or grouped into categories, [204].
Non-Gaussian algorithms include indicator-based methods, simulated annealing, Boolean
methods, and Markov chains. Boolean methods do not reproduce a covariance func-
tion but create a geologic image by random1y generating in space simple shapes com-
monly observed in sedimentary deposits. In contrast, indicator-based methods are based
on variogram models, simulated annealing solves an optimization formulation derived
from the problem of solidification of a solid from a hot liquid upon cooling, and Markov
chains are based on transition probabilities, [204]. Geologic information can be consid-
ered through training images that contain geologic features deemed important to the
investigated problem.
Geologic information can be considered through training images that contain geologic
features deemed important to subsurface fluid flow and transport. Kupfersberger &
Deutsch [210] report that training images can be geologic maps, cross sections, well logs
data, fence diagrams, outcrop maps, conditioning data grouped into zones or images
from quantitative depositional models. The form and parameters of a model of spatial
[email protected]
correlation must reflect the features observed in the training image, [79].

245
Appendix H Pluri-Gaussian Simulation Method

(a) (b) (c)

-0.6 0.5

Figure H.1: (a) Histogram of a standard normal distribution (i.e. N (µ = 0, σ = 1))

showing the two thresholds -0.6 and 0.5,
(b) simulated greytone image with N (µ = 0, σ = 1) and
(c) same image after being truncated at the thresholds -0.6 an 0.5.
Values below -0.6 have been shaded dark grey, those between -0.6 and 0.5
are coloured light grey while values above 0.5 are shown in white, [14].

H.2 Pluri-Gaussian simulation approach

The principle of truncated Gaussian simulation (TGS) was established almost 20 years
ago Chiles & Delfiner [79] and by Isaaks [181] and then developed further developed as
an efficient method for simulating spatial categorical variables that can be represented
by indicators.
The underlying idea in both truncated Gaussian and pluri-Gaussian simulations is
to set up one or more simulations of standard normal random functions in the area of
interest and to attribute the soil type depending on the simulated values at each point.
This is done by truncating. When only one Gaussian random field is used, the truncation
is effectively defined by the values of thresholds as shown in figure H.1. When two or
more Gaussian random fields are used, the situation is more complex. It is represented
graphically via the lithotype rule.

H.2.1 Geological constraints

The complex geological setting is pictured in figure H.2. The simplified geological sketch
map of the study area from Raspa et al.[301] shows togeather with the conceptual geo-
logical model the complex subsoil of the study area. This is even more emphasised by
figure H.3 (a). The boreholes are shown in figure H.3 (b) in a 3D plot and in plan view in
figure H.4. Moreover, the area of the case study is indicated in figure H.4.
Figure H.5 shows the the results of variogram analysis of the categorical variable soil
type, which includes the values 1,2,3 for the corresponding soil type. The variograms
are calculated for all directions on the horizontal plane with a mutual lag distance of 100
m and tolerance of 50 m. It allows to analyse the spatial continuity of the variable in all
[email protected]
directions as well as investigating the anisotropy. From the map in figure H.5 one can

246
 H.2 Pluri-Gaussian simulation approach

identify two directions (N50 and N150) with a very high anisotropy.
Experimental variograms in the horizontal plane coordinate system of three indicator
variables considered, calculated with a pitch of 100 m, 50 m tolerance step angular toler-
ance of 45◦ . The variograms are calculated in both directions N60 and N150 (direction of
maximum anisotropy) and exhibit a variability of the spatial distribution of soil types,
especially medium-large, non-stationarity of the phenomenon, made more apparent by
the study of proportions in figure H.8 (b), is partly masked by the fact of having areas
with mixed distributions very different.
In figure H.8 (a) the vertical proportion curves of an area 500 m x 500 m and map of
the vertical proportion curves of the soil types 1, 2 and 3 H.8 (b).
These proportion curves in figure H.8 are analysed to derive the lithotypes, which are
used in the pluri-Gaussian simulations shown in figure H.9.

[email protected]

247
Appendix H Pluri-Gaussian Simulation Method

(a)

(b)

Figure H.2: (a) Simplified geological sketch map of the study area from Raspa et al.[301].
Legend:
(1) upper Pleistocene - Holocene alluvial deposits,
(2) middle-upper Pleistocene volcanic bedrock,
(3) Plio-Pleistocene sedimentary bedrock,
(4) boreholes with geotechnical samples,
(5) boreholes with samples endowed with the full set of
geotechnical information and
(6) track of the geological cross section in figure H.3 (a).

(b) conceptual geological model of the Tevere valley by [238].

[email protected]

248
 H.2 Pluri-Gaussian simulation approach

(a) Tenever NE
SW 30
River
20

10
0 m above
seal level
geotechnical
-10
borehole
-20
antropic backfill
-30
Soil type I
-40
Soil type II
-50
Soil type III
-60
Soil type IV 0 500 m

(b)

Figure H.3: (a) Geological cross section of the recent alluvial deposits filling the
Tevere valley from [301].
(b) 3D plot of the selected boreholes with location of the geotechnical
samples (black points) for location of the boreholes in figure
[email protected]
H.2 from [301].

249
Appendix H Pluri-Gaussian Simulation Method

boreholes
geological soil
investigation

area of the case study

400 m x 600 m

Figure H.4: Plan view investigated area and area of the case study (400 m x 600 m) in
the city of Rome with locations of the boreholes and soil investigations from
[238].
[email protected]

250
 H.2 Pluri-Gaussian simulation approach

gind(t)

Figure H.5: Variogram map from Marconi [238].

[email protected]

251
Appendix H Pluri-Gaussian Simulation Method
gind,1(t)

experimental indicator variogram (N60)

experimental indicator variogram (N150)

t [m]
gind,2(t)
gind,1-2(t)

t [m] t [m]
gind,3(t)
gind,1-3(t)

gind,2-3(t)

t [m]

Figure H.6: Horizontal indicator variograms of the three different soil types from [238].

[email protected]

252
 H.2 Pluri-Gaussian simulation approach
gind,1(t)

t [m]
gind,1-2(t)

gind,2(t)

t [m] t [m]
gind,3(t)
gind,2-3(t)
gind,1-3(t)

t [m]

Figure H.7: Vertical indicator variograms of the three different soil types from [238].

[email protected]

253
Appendix H Pluri-Gaussian Simulation Method

Soil type I
Soil type II
Soil type III

20
10
0
-10
-20
-30
-40
-50
0 50 100

4 646 000

4 645 000

4 644 000

4 643 000

4 642 000

4 641 000

4 640 000
288 000 289 000 290 000 291 000 292 000

Figure H.8: Vertical proportion curve of an area 500 m x 500 m (a) and map of the vertical
proportion curves of the soil types 1, 2 and 3, [238].
[email protected]

254
 H.2 Pluri-Gaussian simulation approach

(a) (b) (c)

Y1 Y1 Y1

3 1 2 3 1 2

2 1 3

Y2 Y2 Y2

Figure H.9: Used lithotypes for the Pluri-Gaussian Simulation [238].

H.2.2 Pluri-Gaussian simulations

The Pluri-Gaussian simulation algorithm of [238] was used to simulate 300 random fields
using the mesh shown in figure H.10. A part of one realisation is shown in figure H.11
and H.12.

[email protected]

255
Appendix H Pluri-Gaussian Simulation Method

coarse grid (40 x 40 x 0.5 m)

fine grid (5 x 5 x 0.5 m)

0 1 2
km

Figure H.10: Plan view of the Pluri-Gaussian mesh from Marconi [238].

[email protected]

256
 H.2 Pluri-Gaussian simulation approach

(a)

(c)

(b)

Figure H.11: Part of the realisation of the pluri-Gaussian random field for soil type 1 (a),
soil type 2 (b).

[email protected]

257
Appendix H Pluri-Gaussian Simulation Method

(a)

(b )

(c)

Figure H.12: Part of the realisation of the pluriGaussian random field including soil type
3 (a), soil type 4 (b), antropogenetic top soil layer (c).

[email protected]

258
 Table H.1: Input parameters of the soil model.

Soil type I Soil type II Soil type III Soil type IV

◦
friction angel ϕ’ [ ] stochastic stochastic stochastic stochastic
cohesion c’ [kN/m2 ] stochastic stochastic stochastic stochastic
tension cut-off st [kN/m2 ] 0 0 0 0
dilatancy angel ψ [◦ ] 0 0 0 0
power m [-] 0.5000 0.80 0.89 0.40

[email protected]
reference pressure pref [kN/m2 ] 100 100 100 100
failure ratio Rf [-] 0.9 0.9 0.9 0.9
Poisson’s ratio ν [-] 0.33 0.37 0.37 0.37
reference Young’s modulus E 50 [kN/m2 ] stochastic stochastic stochastic stochastic
reference Young’s modulus for unloading-reloading E ur [kN/m2 ] 105, 000 12, 000 8, 000 300, 000
ratio of shear moduli G0 / Gur [-] 2.0 2.0 2.0 2.0
threshold shear strain γ 0.7 [-] 10-4 10-4 10-4 10-4

259
H.2 Pluri-Gaussian simulation approach
[email protected]
 Lebenslauf
30.09.1979 Geboren in Linz (Österreich)

1999-2004 Technische Universität Graz (Österreich)

Studium des Bauingenieurwesens (Dipl.-Ing.)

1999-2005 Technische Universität Graz (Österreich)

Studium des Wirtschaftsingenieurwesen-Bauwesen (Dipl.-Ing.)

2005 Technische Universität Graz, Institut für Wasserbau & Wasserwirtschaft

Forschungsstipendiat

2005-2007 FHCE - Floegl Hydro Consulting Engineering

Projektingenieur in den Bereichen Wasserbau, Geotechnik und
Ingenieurbau

2007-2013 Universität Stuttgart, Institut für Geotechnik

akademischer Mitarbeiter

[email protected]

261
[email protected]
 Mitteilungen des Instituts für Geotechnik
der Universität Stuttgart

Mitteilungen des Baugrundinstitutes Stuttgart

(Institut für Grundbau und Bodenmechanik)
der Universität Stuttgart
Hrsg.: Prof. Dr.-Ing. Dr. h.c. U. Smoltczyk
Nr. 01 Thamm, B. R. (1974) Anfangssetzungen und
Anfangsporenwasserüberdrücke
eines normalverdichteten wasser-
gesättigten Tones

Nr. 02 Gußmann, P. (1975) Einheitliche Berechnung von

Grundbruch und Böschungsbruch

Nr. 03 Feeser, V. (1975) Die Bedeutung des

Kalziumkarbonats für die
bodenphysikalischen
Eigenschaften vom Löß

Nr. 04 Du Thin, K. (1976) Standsicherheit von Böschungen:

Programm-Dokumentation

Nr. 05 Smoltczyk, U. (1976) Messungen an Schleusen in der

Pertschi, O. UDSSR. Schleusennorm der
Hilmer, K. UDSSR (SN 30365)

[email protected]

263
 Mitteilungen des Instituts für Geotechnik

Nr. 06 Hilmer, K. (1976) Erddruck auf

Schleusenkammerwände

Nr. 07 Laumans, Q. (1977) Verhalten einer ebenen, in Sand

eingespannten Wand bei nicht-
linearen Stoffeigenschaften des
Bodens

Nr. 08 Lächler, W. (1977) Beitrag zum Problem der Teil-

flächenpressung bei Beton am Bei-
spiel der Pfahlkopfanschlüsse

Nr. 09 Spotka, H. (1977) Einfluß der Bodenverdichtung

mittels Oberflächenrüttelgeräten
auf den Erd-druck einer
Stützwand bei Sand

Nr. 10 Schad, H. (1979) Nichtlineare Stoffgleichungen für

Böden und ihre Verwendung bei
der numerischen Analyse von
Grundbau-aufgaben

Nr. 11 Ulrich, G. (1980) Verschiebungs- und

kraftgesteuerte Platten-
druckversuche auf
konsolidierenden Böden
Gußmann, P. Zum Modellgesetz der
Konsolidation

Nr. 12 Salden, D. (1980) Der Einfluß der Sohlenform auf

die Traglast von Fundamenten
[email protected]

264
 Mitteilungen des Instituts für Geotechnik

Nr. 13 Seeger, H. (1980) Beitrag zur Ermittlung des

horizontalen Bettungsmoduls von
Böden durch Seitendruckversuche
im Bohrloch

Nr. 14 Schmidt, H.H. (1981) Beitrag zur Ermittlung des

Erddrucks auf Stützwände bei
nachgiebigem Baugrund

Nr. 15 Smoltczyk, U. (1981) Vorstudie über bauliche

Schweikert, O. Alternativen für
Durchgangsstraßen in Siedlungen

Nr. 16 Malcharek, K. (1981) Vergleich nationaler Richtlinien

Smoltczyk, U. für die Berechnung von
Fundamenten

Nr. 17 Gruhle, H.D. (1981) Das Verhalten des Baugrundes

unter Einwirkung vertikal
gezogener Ankerplatten als
räumliches Problem des
Erdwiderstandes

Nr. 18 Kobler, W. (1982) Untersuchungen über Böschungs-

und Grundbruch bei begrenzten
Lastflächen

Nr. 19 Lutz, W. (1983) Tragfähigkeit des geschlitzten

Baugrunds neben Linienlasten

[email protected]

265
 Mitteilungen des Instituts für Geotechnik

Nr. 20 Smoltczyk, U. (1983) Studienunterlagen

“Bodenmechanik und
Grundbau“; überarbeitete
Ausgabe 1993

Nr. 21 Schweikert, O. (1984) Der Einfluß des

Böschungswinkels auf die
Berechnung des aktiven
Erddrucks

Nr. 22 Vogt, N. (1984) Erdwiderstandsermittlung bei

monotonen und wiederholten
Wand-bewegungen in Sand

Nr. 23 Buchmaier, R. (1985) Zur Berechnung von Kon-

solidationsproblemen bei nicht-
linearem Stoffverhalten

Nr. 24 Schad, H. (1985) Möglichkeiten der Böschungs-

Smoltczyk, U. sicherung bei kleinen Baugruben

Schad, H. Sonderkonstruktionen der

Zoller, P. Böschungssicherung

Nr. 25 Gußmann, P. (1986) Die Methode der Kinematischen

Elemente

Nr. 26 Steinmann, B. (1985) Zum Verhalten bindiger Böden

bei monotoner einaxialer
Beanspruchung

Nr. 27 Lee, S.D. (1987) Untersuchungen zur

Standsicherheit von Schlitzen im
[email protected] Sand neben Einzelfundamenten

266
 Mitteilungen des Instituts für Geotechnik

Mitteilungen des
Instituts für Geotechnik der Universität Stuttgart
Hrsg.: Prof. Dr.-Ing. Dr. h.c. U. Smoltczyk
Nr. 28 Kolb, H. (1988) Ermittlung der Sohlreibung von
Gründungskörpern unter
horizontalem kinematischen
Zwang

Nr. 29 Ochmann, H. (1988) Ebene Grenzzustände von

Erdböschungen im
stochastischen
Sicherheitskonzept

Nr. 30 Breinlinger, F. (1989) Bodenmechanische

Stoffgleichungen bei großen
Deformationen sowie Be- und
Entlastungsvorgängen

Nr. 31 Smoltczyk, U. (1989) Beitrag zur Bemessung von

Breilinger, F. Tunneln in offener Bauweise
Schad, H.
Wittlinger, M.
Nr. 32 Gußmann, P. (1990) Beiträge zur Anwendung der
Schanz, T. KEM (Erddruck, Grundbuch,
Smoltczyk, U. Standsicher-heit von
Willand, E. Böschungen)

Nr. 33 Gruhle, H.D. (1990) Der räumliche Erdwiderstand

vor überwiegend horizontal
belasteten Ankerplatten

[email protected]

267
 Mitteilungen des Instituts für Geotechnik

Nr. 34 Henne, J. (1995) Zur Bewehrung von verformten

Bodenschichten durch Einsatz
zugfester Geokunststoffe

Nr. 35 Wittlinger, M. (1994) Ebene

Verformungsuntersuchungen zur
We-ckung des Erdwiderstandes
bindiger Böden

Nr. 36 Schad, H. (1992) Zeit- und geschwindigkeits-

abhängiges Materialverhalten in
der Geotechnik – Experimentelle
Erfassung und numerische
Analyse

Nr. 37 Belz, I. (1992) Zur Ermittlung dynamischer

Boden-kennwerte in situ aus der
Systemantwort des Erregers

Nr. 38 Ma, J. (1994) Untersuchungen zur

Standsicherheit der durch
Stützscheiben stabilisierten
Böschungen

Nr. 39 Smoltczyk, U. (1994) Sonderheft: 25 Jahre Lehre und

Forschung in der Geotechnik

Nr. 40 Rilling, B. (1994) Untersuchungen zur

Grenztragfähigkeit bindiger
Schüttstoffe am Beispiel von
Lößlehm

[email protected]

268
 Mitteilungen des Instituts für Geotechnik

Mitteilungen des
Instituts für Geotechnik der Universität Stuttgart
Hrsg.: Prof. Dr.-Ing. P.A. Vermeer
Nr. 41 Vermeer, P.A. (1996) Deponiebau und Geotechnik

Nr. 42 Vermeer, P.A. (1997) Baugruben in Locker- und Fest-

gestein

Nr. 43 Brinkmann, C. (1998) Untersuchungen zum Verhalten

von Dichtungsübergängen im
Staudammbau

Nr. 44 Fiechter-Scharr, (1998) Beeinflussung von Erdbaustoffen

I. durch Beimischen eines organo-
philen Bentonits

Nr. 45 Schanz, T. (1998) Zur Modellierung des

mechanischen Verhaltens von
Reibungsmaterialien

Nr. 46 Akinrogunde, (1999) Propagation of Cement Grout in

A.E. Rock Discontinuities Under
Injection Conditions

Nr. 47 Vogt-Breyer, C. (1999) Experimentelle und numerische

Untersuchungen zum
Tragverhalten und zur
Bemessung horizontaler
Schraubanker

Nr. 48 Vermeer, P.A. (1999) Neue Entwicklungen in der

[email protected] Geotechnik

269
 Mitteilungen des Instituts für Geotechnik

Nr. 49 Marcher, T. (2002) Resultate eines

Versuchsprogramms an Beaucaire
Mergel

Nr. 50 Marcher, T. (2003) Nichtlokale Modellierung der

Entfestigung dichter Sande und
steifer Tone

Nr. 51 Ruse, N.M. (2004) Räumliche Betrachtung der

Stand-sicherheit der Ortsbrust
beim Tunnelvortrieb

Nr. 52 Beutinger, P.H. (2005) Ein geotechnischer Beitrag zur

Standsicherheit mobiler Bau-
maschinen

Nr. 53 Wehnert, M. (2006) Ein Beitrag zur drainierten und

undrainierten Analyse in der
Geotechnik

Nr. 54 Möller, S. C. (2006) Tunnel induced settlements and

forces in linings

Nr. 55 Benz, T. (2007) Small-Strain Stiffness of Soils and

its Numerical Consequences

Nr. 56 Abed, A. (2008) Numerical Modeling of

Expansive Soil Behavior

[email protected]

270
 Mitteilungen des Instituts für Geotechnik

Nr. 57 Hintner, J. (2008) Analyse der Fundamentver-

schiebungen infolge vertikaler
und geneigter Belastung

Nr. 58 Russelli, C. (2008) Probabilistic Methods applied to

the Bearing Capacity Problem

Nr. 59 Peña Olarte, (2008) Influence of Particle Shape on the

A.A. Global Mechanical Response of
Granular Packings:
Micromechanical Investigation of
the Critical State in Soil
Mechanics

Nr. 60 Neher, H. P. (2008) Zeitabhängiges Materialverhalten

und Anisotropie von weichen
Böden – Theorie und Anwendung

Nr. 61 Vermeer, P.A. (2008) Von der Forschung zur Praxis:

Symposium zum 80. Geburtstag
von Prof. U. Smoltczyk

Nr. 62 Satibi, S. (2009) Numerical Analysis and Design

Criteria of Embankments on
Floating Piles

Nr. 63 Lächler, A. (2009) Bedeutung herstellungsbedingter

Einflüsse auf das Trag- und
Verformungsverhalten von
Schlitzwänden
[email protected]

271
 Mitteilungen des Instituts für Geotechnik

Nr. 64 Möllmann, A. (2009) Probabilistische Untersuchung

von Hochwasserschutzdeichen
mit analytischen Verfahren und
der Finite-Elemente-Methode

Mitteilungen des
Instituts für Geotechnik der Universität Stuttgart
Hrsg.: Prof. Dr.-Ing. habil. Ch. Moormann
Nr. 65 Moormann, Ch. (2011) 7. Stuttgarter Geotechnik-
Symposium

Nr. 66 Beuth, L. (2012) Formulation an Application of a

Quasi-Static Material Point
Method

Nr. 67 Moormann, Ch., (2012) Proceedings of the

Huber, M., 10th International Probabilistic
Proske, D. Workshop

Nr. 68 Schneider, M. (2013) Zur energetischen Nutzung von

Tunnelbauwerken – Messungen
und numerische Berechnungen
am Beispiel Fasanenhoftunnel

Nr. 69 Huber, M. (2013) Soil variability and its

consequences in geotechnical
engineering

[email protected]

272