Reservoir Characterization

Reservoir Characterization
Scrivener Publishing
100 Cummings Center, Suite 541J
Beverly, MA 01915-6106
Sustainable Energy Engineering
Series Editor: Fred Aminzadeh
Scope: This series’s mission is to publish peer-reviewed, original research seeking sustainable methods
of worldwide energy production, distribution, and utilization through engineering, scientific, and
technological advances in the areas of both fossil fuels and renewable energy. Energy issues cannot
be addressed in isolation, without attention to the economy and the environment. Thus, this series
introduces the “E cubed” concept, addressing sustainability with a three-pronged emphasis on energy,
economy and environment, publishing research in all of these areas, and their intersections, as they
apply to global energy sustainability. This unique multi-disciplinary theme allows the introduction
of new and cutting-edge processes and technologies across all areas of energy production, transportation,
and transmission, including fossil fuels and renewable energy and their intersection with the economy
and environment. Papers are invited on any individual topic or those which are interdisciplinary.
Publishers at Scrivener
Martin Scrivener ([email protected])
Phillip Carmical ([email protected])
Reservoir Characterization
Fundamentals and Applications
Fred Aminzadeh
This edition first published 2022 by John Wiley & Sons, Inc., 111 River Street, Hoboken, NJ 07030, USA
and Scrivener Publishing LLC, 100 Cummings Center, Suite 541J, Beverly, MA 01915, USA
© 2022 Scrivener Publishing LLC
For more information about Scrivener publications please visit www.scrivenerpublishing.com.
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, or other-
wise, except as permitted by law. Advice on how to obtain permission to reuse material from this title
is available at http://www.wiley.com/go/permissions.
Wiley Global Headquarters

111 River Street, Hoboken, NJ 07030, USA
For details of our global editorial offices, customer services, and more information about Wiley prod-
ucts visit us at www.wiley.com.
Limit of Liability/Disclaimer of Warranty

While the publisher and authors have used their best efforts in preparing this work, they make no rep
resentations or warranties with respect to the accuracy or completeness of the contents of this work and
specifically disclaim all warranties, including without limitation any implied warranties of merchant-
ability or fitness for a particular purpose. No warranty may be created or extended by sales representa
tives, written sales materials, or promotional statements for this work. The fact that an organization,
website, or product is referred to in this work as a citation and/or potential source of further informa
tion does not mean that the publisher and authors endorse the information or services the organiza
tion, website, or product may provide or recommendations it may make. This work is sold with the
understanding that the publisher is not engaged in rendering professional services. The advice and
strategies contained herein may not be suitable for your situation. You should consult with a specialist
where appropriate. Neither the publisher nor authors shall be liable for any loss of profit or any other
commercial damages, including but not limited to special, incidental, consequential, or other damages.
Further, readers should be aware that websites listed in this work may have changed or disappeared
between when this work was written and when it is read.
Library of Congress Cataloging-in-Publication Data
ISBN 9781119556213
Cover image: Geo/Rock Wall, 31647625 © Pzaxe | Dreamstime.com

Cover design by Kris Hackerott
Set in size of 11pt and Minion Pro by Manila Typesetting Company, Makati, Philippines
Printed in the USA
10 9 8 7 6 5 4 3 2 1
Contents
Foreword xix
Preface xxiii
Part 1: Introduction 1
1 Reservoir Characterization: Fundamental
and Applications - An Overview 3
Fred Aminzadeh
1.1 Introduction to Reservoir Characterization? 3
1.2 Data Requirements for Reservoir Characterization 5
1.3 SURE Challenge 7
1.4 Reservoir Characterization in the Exploration,
Development and Production Phases 10
1.4.1 Exploration Stage/Development Stage 10
1.4.2 Primary Production Stage 11
1.4.3 Secondary/Tertiary Production Stage 11
1.5 Dynamic Reservoir Characterization (DRC) 12
1.5.1 4D Seismic for DRC 13
1.5.2 Microseismic Data for DRC 14
1.6 More on Reservoir Characterization and Reservoir
Modeling for Reservoir Simulation 15
1.6.1 Rock Physics 16
1.6.2 Reservoir Modeling 17
1.7 Conclusion 20
References 20
v
vi Contents
Part 2: General Reservoir Characterization

and Anomaly Detection 23
2 A Comparison Between Estimated Shear Wave Velocity
and Elastic Modulus by Empirical Equations and that
of Laboratory Measurements at Reservoir Pressure Condition 25
Haleh Azizia, Hamid Reza Siahkoohi, Brian Evans,
Nasser Keshavarz Farajkhah and Ezatollah KazemZadeh
2.1 Introduction 26
2.2 Methodology 28
2.1.2 Estimating the Shear Wave Velocity 28
2.2.2 Estimating Geomechanical Parameters 31
2.3 Laboratory Set Up and Measurements 32
2.3.1 Laboratory Data Collection 34
2.4 Results and Discussion 35
2.5 Conclusions 41
2.6 Acknowledgment 43
References 43
3 Anomaly Detection within Homogenous Geologic Area 47
Simon Katz, Fred Aminzadeh, George Chilingar
and Leonid Khilyuk
3.1 Introduction 48
3.2 Anomaly Detection Methodology 49
3.3 Basic Anomaly Detection Classifiers 50
3.4 Prior and Posterior Characteristics of Anomaly
Detection Performance 52
3.5 ROC Curve Analysis 55
3.6 Optimization of Aggregated AD Classifier Using Part
of the Anomaly Identified by Universal Classifiers 58
3.7 Bootstrap Based Tests of Anomaly Type Hypothesis 61
3.8 Conclusion 64
References 65
4 Characterization of Carbonate Source-Derived
Hydrocarbons Using Advanced Geochemical Technologies 69
Hossein Alimi
4.1 Introduction 70
4.2 Samples and Analyses Performed 71
4.3 Results and Discussions 72
Contents vii
4.4 Summary and Conclusions 79

References 80
5 Strategies in High-Data-Rate MWD Mud Pulse Telemetry 81
Yinao Su, Limin Sheng, Lin Li, Hailong Bian, Rong Shi,
Xiaoying Zhuang and Wilson Chin
5.1 Summary 82
5.1.1 High Data Rates and Energy Sustainability 82
5.1.2 Introduction 83
5.1.3 MWD Telemetry Basics 85
5.1.4 New Telemetry Approach 87
5.2 New Technology Elements 88
5.2.1 Downhole Source and Signal Optimization 89
5.2.2 Surface Signal Processing and Noise Removal 92
5.2.3 Pressure, Torque and Erosion Computer Modeling 93
5.2.4 Wind Tunnel Analysis: Studying New Approaches 96
5.2.5 Example Test Results 108
5.3 Directional Wave Filtering 111
5.3.1 Background Remarks 111
5.3.2 Theory 112
5.3.3 Calculations 116
5.4 Conclusions 132
Acknowledgments 133
References 133
6 Detection of Geologic Anomalies with Monte Carlo
Clustering Assemblies 135
Simon Katz, Fred Aminzadeh, George Chilingar,
Leonid Khilyuk and Matin Lockpour
6.1 Introduction 135
6.2 Analysis of Inhomogeneity of the Training and Test
Sets and Instability of Clustering 136
6.3 Formation of Multiple Randomized Test Sets
and Construction of the Clustering Assemblies 138
6.4 Irregularity Index of Individual Clusters
in the Cluster Set 139
6.5 Anomaly Indexes of Individual Records
and Clustering Assemblies 141
6.6 Prior and Posterior True and False Discovery Rates
for Anomalous and Regular Records 142
viii Contents
6.7 Estimates of Prior False Discovery Rates

for Anomalous Cluster Sets, Clusters, and Individual
Records. Permeability Dataset 142
6.8 Posterior Analysis of Efficiency of Anomaly
Identification. High Permeability Anomaly 144
6.9 Identification of Records in the Gas Sand Dataset
as Anomalous, using Brine Sand Dataset as Data
with Regular Records 146
6.10 Notations 149
6.11 Conclusions 149
References 150
7 Dissimilarity Analysis of Petrophysical Parameters
as Gas-Sand Predictors 151
Simon Katz, George Chilingar, Fred Aminzadeh
and Leonid Khilyuk
7.2 Petrophysical Parameters for Gas-Sand Identification 152
7.3 Lithologic and Fluid Content Dissimilarities of Values
of Petrophysical Parameters 154
7.4 Parameter Ranking and Efficiency of Identification
of Gas-Sands 155
7.5 ROC Curve Analysis with Cross Validation 159
7.6 Ranking Parameters According to AUC Values 161
7.7 Classification with Multidimensional Parameters
as Gas Predictors 163
7.8 Conclusions 164
Definitions and Notations 166
References 166
8 Use of Type Curve for Analyzing Non-Newtonian Fluid
Flow Tests Distorted by Wellbore Storage Effects 169
Fahd Siddiqui and Mohamed Y. Soliman
8.2 Objective 173
8.3 Problem Analysis 173
8.3.1 Model Assumptions 174
8.3.2 Solution Without the Wellbore Storage Distortion 175
8.3.3 Wellbore Storage and Skin Effects 175
8.3.4 Solution by Mathematical Inspection 175
8.3.5 Solution Verification 176
Contents ix
8.4 Use of Finite Element 176

8.5 Analysis Methodology 177
8.5.1 Finding the n Value 177
8.5.2 Dimensionless Wellbore Storage 178
8.5.3 Use of Type Curves 178
8.5.4 Match Point 179
8.5.5 Uncertainty in Analysis 180
8.6 Test Data Examples 180
8.6.3 Analysis Recommendations 185
8.6.5 Analysis Recommendations 186
8.6.6 Match point 186
8.7 Conclusion 188
Nomenclature 188
References 189
Appendix A: Non-Linear Boundary Condition
and Laplace Transform 189
Appendix B: Type Curve Charts for Various Power
Law Indices 191
Part 3: Reservoir Permeability Detection 195

9 Permeability Prediction Using Machine Learning,
Exponential, Multiplicative, and Hybrid Models 197
Simon Katz, Fred Aminzadeh, George Chilingar
and M. Lackpour
9.2 Additive, Multiplicative, Exponential, and Hybrid
Permeability Models 198
9.3 Combination of Basis Function Expansion
and Exhaustive Search for Optimum Subset of Predictors 200
9.4 Outliers in the Forecasts Produced with Four
Permeability Models 201
9.5 Additive, Multiplicative, and Exponential
Committee Machines 203
9.6 Permeability Forecast with First Level
Committee Machines. Sandstone Dataset 206
9.7 Permeability Prediction with First Level
Committee Machines. Carbonate Reservoirs 210
x Contents
9.8 Analysis of Accuracy of Outlier Replacement

by The First and Second Level Committee Machines.
Sandstone Dataset 212
9.9 Conclusion 214
Notations and Definitions 215
References 216
10 Geological and Geophysical Criteria for Identifying Zones
of High Gas Permeability of Coals (Using the Example
of Kuzbass CBM Deposits) 217
A.G. Pogosyan
10.2 Physical Properties and External Load Conditions
on a Coal Reservoir 219
10.3 Basis for Evaluating Physical and Mechanical Coalbed
Properties in the Borehole Environment 225
Acknowledgement 228
References 229
11 Rock Permeability Forecasts Using Machine Learning
and Monte Carlo Committee Machines 231
Simon Katz, Fred Aminzadeh, Wennan Long,
George Chilingar and Matin Lackpour
11.2 Monte Carlo Cross Validation and Monte
Carlo Committee Machines 233
11.3 Performance of Extended MC Cross Validation
and Construction MC Committee Machines 236
11.4 Parameters of Distribution of the Number
of Individual Forecasts in Monte Carlo Cross Validation 237
11.5 Linear Regression Permeability Forecast
with Empirical Permeability Models 238
11.6 Accuracy of the Forecasts with Machine
Learning Methods 242
11.7 Analysis of Instability of the Forecast 244
11.8 Enhancement of Stability of the MC
Committee Machines Forecast Via Increase
of the Number of Individual Forecasts 246
Nomenclature 247
Contents xi
Appendix 1- Description of Permeability Models from

Different Fields 248
Appendix 2- A Brief Overview of Modular Networks or
Committee Machines 249
References 251
Part 4: Reserves Evaluation/Decision Making 253

12 The Gulf of Mexico Petroleum System – Foundation
for Science-Based Decision Making 255
Corinne Disenhof, MacKenzie Mark-Moser and Kelly Rose
Introduction 256
Basin Development and Geologic Overview 257
Petroleum System 259
Reservoir Geology 259
Hydrocarbons 261
Salt and Structure 262
Conclusions 263
Acknowledgments and Disclaimer 264
References 265
13 Forecast and Uncertainty Analysis of Production
Decline Trends with Bootstrap and Monte Carlo Modeling 269
Simon Katz, George Chilingar and Leonid Khilyuk
13.2 Simulated Decline Curves 271
13.3 Nonlinear Least Squares for Decline Curve
Approximation 273
13.4 New Method of Grid Search for Approximation
and Forecast of Decline Curves 273
13.5 Iterative Minimization of Least Squares
with Multiple Approximating Models 275
13.6 Grid Search Followed by Iterative Minimization
with Levenberg-Marquardt Algorithm 276
13.7 Two Methods for Aggregated Forecast and Analysis
of Forecast Uncertainty 277
13.8 Uncertainty Quantile Ranges Obtained Using Monte
Carlo and Bootstrap Methods 279
13.9 Monte Carlo Forecast and Analysis
13.10 Block Bootstrap Forecast and Analysis
xii Contents
13.11 Comparative Analysis of Results of Monte Carlo

and Bootstrap Simulations 285
References 288
14 Oil and Gas Company Production, Reserves, and Valuation 289
Mark J. Kaiser
14.2 Reserves 292
14.2.1 Proved Reserves 292
14.2.2 Proved Reserves Categories 292
14.2.3 Reserves Reporting 293
14.2.4 Probable and Possible Reserves 293
14.2.5 Contractual Differences 294
14.3 Production 294
14.4 Factors that Impact Company Value 295
14.4.1 Ownership 295
14.4.1.1 International Oil Companies 295
14.4.1.2 National Oil Companies 296
14.4.1.3 Government Sponsored Entities 296
14.4.1.4 Independents and Juniors 297
14.4.2 Degree of Integration 297
14.4.3 Product Mix 298
14.4.4 Commodity Price 298
14.4.5 Production Cost 299
14.4.6 Finding Cost 299
14.4.7 Assets 300
14.4.8 Capital Structure 300
14.4.9 Geologic Diversification 301
14.4.10 Geographic Diversification 301
14.4.11 Unobservable Factors 302
14.5 Summary Statistics 303
14.5.1 Sample 303
14.5.2 Variables 303
14.5.3 Data Source 305
14.5.4 International Oil Companies 305
14.5.5 Independents 308
14.6 Market Capitalization 309
14.6.1 Functional Specification 309
14.6.2 Expectations 309
14.7 International Oil Companies 310
Contents xiii
14.8 U.S. Independents 312

14.8.1 Large vs. Small Cap, Oil vs. Gas 312
14.8.2 Consolidated Small-Caps 314
14.8.3 Multinational vs. Domestic 314
14.8.4 Conventional vs. Unconventional 315
14.8.5 Production and Reserves 316
14.8.6 Regression Models 316
14.9 Private Companies 318
14.10 National Oil Companies of OPEC 320
14.11 Government Sponsored Enterprises
and Other International Companies 320
References 324
Part 5: Unconventional Reservoirs 337

15 An Analytical Thermal-Model for Optimization
of Gas-Drilling in Unconventional Tight-Sand Reservoirs 339
Boyun Guo, Gao Li and Jinze Song
15.2 Mathematical Model 341
15.3 Model Comparison 346
15.4 Sensitivity Analysis 348
15.5 Model Applications 349
Nomenclature 352
Acknowledgements 353
References 353
Appendix A: Steady Heat Transfer Solution for Fluid
Temperature in Counter-Current Flow 355
Assumptions 355
Governing Equation 355
Boundary Conditions 360
Solution 360
16 Development of an Analytical Model for Predicting the Fluid
Temperature Profile in Drilling Gas Hydrates Reservoirs 363
Liqun Shan, Boyun Guo and Xiao Cai
16.2 Mathematical Model 365
16.3 Case Study 373
xiv Contents
16.4 Sensitivity Analysis 374

Nomenclature 378
References 379
17 Distinguishing Between Brine-Saturated and Gas-Saturated
Shaly Formations with a Monte-Carlo Simulation
of Seismic Velocities 383
Simon Katz, George Chilingar and Leonid Khilyuk
17.2 Random Models for Seismic Velocities 385
17.3 Variability of Seismic Velocities Predicted
by Random Models 387
17.4 The Separability of (Vp, Vs) Clusters for Gas- and
Brine-Saturated Formations 388
17.5 Reliability Analysis of Identifying Gas-Filled Formations 389
17.5.1 Classification with K-Nearest Neighbor 391
17.5.2 Classification with Recursive Partitioning 392
17.5.3 Classification with Linear Discriminant Analysis 394
17.5.4 Comparison of the Three Classification Techniques 395
References 397
18 Shale Mechanical Properties Influence Factors Overview
and Experimental Investigation on Water Content Effects 399
Hui Li, Bitao Lai and Shuhua Lin
18.2 Influence Factors 400
18.2.1 Effective Pressure 401
18.2.2 Porosity 402
18.2.3 Water Content 403
18.2.4 Salt Solutions 405
18.2.5 Total Organic Carbon (TOC) 406
18.2.6 Clay Content 407
18.2.7 Bedding Plane Orientation 408
18.2.8 Mineralogy 411
18.2.9 Anisotropy 413
18.2.10 Temperature 413
18.3 Experimental Investigation of Water Saturation
Effects on Shale’s Mechanical Properties 414
Contents xv
18.3.1 Experiment Description 414

18.3.2 Results and Discussion 414
18.3.3 Error Analysis of Experiments 417
References 420
Part 6: Enhance Oil Recovery 427

19 A Numerical Investigation of Enhanced Oil Recovery
Using Hydrophilic Nanofuids 429
Yin Feng, Liyuan Cao and Erxiu Shi
19.2 Simulation Framework 432
19.2.1 Background 432
19.2.2 Two Essential Computational Components 433
19.2.2.1 Flow Model 433
19.2.2.2 Nanoparticle Transport
and Retention Model 435
19.3 Coupling of Mathematical Models 437
19.4 Verification Cases 439
19.4.1 Effect of Time Steps on the Performance
of the in House Simulator 439
19.4.2 Comparison with Eclipse 440
19.4.3 Comparison with Software MNM1D 442
19.5 Results 443
19.5.1 Continuous Injection 445
19.5.1.1 Effect of Injection Time on Oil
Recovery and Nanoparticle Adsorption 445
19.5.1.2 Effect of Injection Rate on Oil
19.5.2 Slug Injection 449
19.5.2.1 Effect of Injection Time on Oil
19.5.2.2 Effect of Slug Size on Oil Recovery
and Nanoparticle Adsorption 451
19.5.3 Water Postflush 452
19.5.3.1 Effect of Injection Time Length 452
19.5.3.2 Effect of Flow Rate Ratio Between
Water and Nanofuids on Oil
and Nanoparticle Recovery 452
xvi Contents
19.5.4 3D Model Showcase 455

19.6 Discussions 457
19.7 Conclusions and Future Work 459
References 461
20 3D Seismic-Assisted CO2-EOR Flow Simulation
for the Tensleep Formation at Teapot Dome, USA 463
Payam Kavousi Ghahfarokhi, Thomas H. Wilson
and Alan Lee Brown
20.1 Presentation Sequence 464
20.3 Geological Background 468
20.4 Discrete Fracture Network (DFN) 469
20.5 Petrophysical Modeling 473
20.6 PVT Analysis 473
20.7 Streamline Analysis 479
20.8 CO2-EOR 479
Acknowledgement 483
References 484
Part 7: New Advances in Reservoir

Characterization-Machine Learning Applications 487
21 Application of Machine Learning in Reservoir
Characterization 489
Fred Aminzadeh
21.1 Brief Introduction to Reservoir Characterization 489
21.2 Artificial Intelligence and Machine (Deep)
Learning Review 491
21.2.1 Support Vector Machines 492
21.2.2 Clustering (Unsupervised Classification) 492
21.2.3 Ensemble Methods 497
21.2.4 Artificial Neural Networks (ANN)-
Based Methods 498
21.3 Artificial Intelligence and Machine (Deep) Learning
Applications to Reservoir Characterization 502
21.3.1 3D Structural Model Development 503
21.3.2 Sedimentary Modeling 506
21.3.3 3D Petrophysical Modeling 508
21.3.4 Dynamic Modeling and Simulations 512
Contents xvii
21.4 Machine (Deep) Learning and Enhanced Oil

Recovery (EOR) 513
21.4.1 ANNs for EOR Performance
and Economics 514
21.4.2 ANNs for EOR Screening 516
21.5 Conclusion 517
Acknowledgement 518
References 518
Index 525
Foreword
What is reservoir characterization? As you will see from this book, this is
a very advanced topic so let’s break it down a bit and start form the basics.
What is a reservoir? This is ‘a place where something is kept in store’. And
what is characterization? That is ‘to describe the character or quality’ all
according to the Webster dictionary. So, we are arrived at: ‘describe the
character of something that’s kept in store’. It seems relatively benign and
easy but ‘the devil is in the details’ is perhaps the best way to get the readers
intrigued and immersed in this topic. So, we are left wondering what are
these details where the devil resides? And here starts the story…..
In fact, a better wording would be ‘Subsurface Reservoir Characterization’
or SRC. There have been on the order of thousands of studies in reser-
voir characterization over the life time of this field. As such, this topic has
evolved and matured with many learnings. As illustrated in this book,
there are now well established and tested workflows SRC and I’d like to go
over some aspects of these understandings and workflows.
First, it is key to understand that SRC is a continuously changing,
multi-discipline and multi-scale topic. For continuously changing a good
example would be the recent impact of say machine learning methods. I
have learned that if our data quality is good enough and there are physical
relationships between reservoir data and properties, machine learning can
be an excellent way to quickly uncover relationships in a multi-variable
universe. However, once again, even here, the devil is in the details….
Multi-discipline is a word we easily use but have difficulty implementing.
In many projects the geologist is tasked with building a static reservoir
model and then passing it on to the reservoir engineer to build a dynamic
model and history match production. However, it has been challenging
to form a loop versus a linear workflow or for the dynamic model to be
updated with new static information or cover a range of possible models
that fit the data…… As for multi-scale, the discipline involves integration
of data from a wide range of data, say, nanometer (electron microscope),
to centimeter (cutting and core samples), to decimeter (well log), to meter
xix
xx Foreword
(seismic) scale. Spatially most of these data are acquired within a small
portion of one or several wells and geophysical data gives the capability to
extrapolate away from the wells with lower resolution. Due to uncertainties
in the data, rapid variations in the subsurface, and sparse sampling multi-
scale integration can be a challenging task. There is a good discussion of
“SURE Challenge” in the book where the author addresses the above men-
tioned challenges of integration incolving multitude of data set with differ-
ent Scale, Unvertainty, Resoultion and Environemnet. It is suggested that
different AI and Data Analytics techniques may be best equipped to handle
the SURE Challenge.
The second component can be categorized into input data quality
(informally ‘garbage in, garbage out’). Any workflow that is lets say cutting
edge cant work without high quality input data. Further, it may cause mis-
interpretation that a workflow is ‘not’ a good workflow or appropriate sim-
ply because the input data was the culprit. The input data in fact starts from
data acquisition, then to data processing and finally to data interpretation
and integration. One of the pitfalls along the way is to simply obtain the
data as an interpreter and not be aware of lets say the ‘history’. An exam-
ple would be to apply amplitude based seismic analysis to data that non-
amplitude preserving processing was applied to (Automatic Gain Control
or AGC would be a simple example). However, the same could be happen-
ing with say well-log or production data. The good news is that over time
in every SRC related discipline data quality has been improving with not
only better tools but also more frequent data acquisition during the life of a
reservoir. Further, over time we have learned to build much better process-
ing tools that provide high quality data for the integration component. The
net result of this has improved our ability to conduct integrated studies and
quantitative products. One example of this near to my heart is joint seis-
mic inversion of PP reflected waves with PS (or converted) reflected waves
from a reservoir. We have seen that with improved acquisition and process-
ing, the joint PP/PS inversion can substantially improve pre-stack seismic
inversion providing a stable S-impedance as well as a P-impedance that can
provide valuable information such as formation properties, porosity, Total
Organic Carbon (TOC,) fluid types, and time-lapse reservoir pressure and
saturation changes over the life of the reservoir. Such improvements are
going on in all the subsurface disciplines thanks to modern acquisition and
more diverse data with higher quality.
This book is an excellent resource for beginners in SRC to get an over-
view of the topic and for expert to study most recent advances in their
own and related disciplines. The book covers a wide range of topics from
conventional to unconventional reservoirs, from geology to geophysics to
Foreword xxi
petroleum engineering, from laboratory measurements to field applica-

tions, from deterministic to statistical methods, from primary depletion
to EOR with CO2 injection, from static to dynamic SRC, as well as use of
AI for reservoir charcaterization. In the end, SRC requires best practices to
be implemented to be value generating. This books certainly provides the
necessary best practices.
Dr. Ali Tura

Professor of Geophysics
Co-director of Reservoir Characterization Project (RCP)
Colorado School of Mine,
Denver, Colorado
Preface
An important step in exploration development, monitoring, and man-

agement a reservoir as well optimizing production and planning for post
primary production decisions is reservoir characterization. Upon the
completion of the preliminary task of reservoir characterization, and as we
continue to produce from the reservoir or use different methods to stimu-
late it, many of its properties change. This requires updating the reservoir
model, bringing up the concept of dynamic reservoir characterization. To
achieve this goal, we incorporate the newly acquired petrophysical, seis-
mic, micro seismic and production data. The updated model would be
a better representative of the status of the reservoir. Both static reservoir
properties, such as porosity, permeability, and facies type; and dynamic
reservoir properties, such as pressure, fluid saturation, and temperature,
needs to be updated as more field data become available.
Among the reason for focusing on reservoir characterization is the fact
based on the estimates by experts, more than 95% of the world’s oil pro-
duction in the 21st century will come from existing fields. This will require
significant improvements in the current recovery rates of less than 50%
in most reservoirs. Improved secondary and tertiary recovery through
enhanced recovery of oil and gas require by better understanding and
monitoring of the reservoir will be an important element of the much-
needed increase in the recovery factor. Increased production will be made
possible only through effective dynamic reservoir characterization.
We need to recognize the fact that reservoir characterization is a mul-
tidisciplinary field. It attempts to describe petroleum deposits and the
nature of the rocks that contain hydrocarbons using a variety of data types.
Reservoir characterization relies on expertise from petroleum engineers,
geologists, geochemists, petrophysicists, and of course geophysicists, The
integration of information from these fields, with the aid of advanced data
analysis techniques as well as artificial intelligence (AI) based methods will
make our reservoir models more accurate and the updating process much
faster.
xxiii
xxiv Preface
This book will provide a comprehensive body of technical material on

different aspects of reservoir characterization. It is divided into 7 parts:
Part 1 is an introductory chapter covering the general concepts, of reser-
voir characterization. It includes an overview of what is meant by reservoir
characterization as it is applied in different stages of its life, from explora-
tion to post primary production stages. It also highlights the challenges of
data integration of different data types, the previously mentioned dynamic
reservoir characterization, and reservoir stimulation for enhanced oil (or
gas) recovery.
Part 2 deals with general issues on reservoir characterization and anom-
aly detection. It is comprised of 7 chapters on different related topics such
as: (1) Comparison between estimated shear wave velocity and elastic
modulus at in situ pressure condition (2) Anomaly detection (3) geochem-
ical analysis on characterization of carbonate source-derived hydrocar-
bons, (4) MWD mud pulse telemetry, (5) Use of Monte Carlo clustering
to detect geologic anomalies, (6) Gas-sand predictors using dissimilarity
analysis, and (7) Fluid flow tests distorted by wellbore storage effects. Part
3 is dedicated to reservoir permeability detection, being one of the most
important reservoir properties. What are covered here are three different
techniques. Two of them involves use of two different machine learning
techniques to predict permeability, namely, exponential/multiplicative and
Monte Carlo/committee machines. The other chapter discusses geoscience
criteria identifying high gas permeability zones.
One of the reasons for reservoir characterization is to assess the recov-
erable reserves in the reservoir. Part 4 addresses reserves evaluation and
decision-making issues. The first chapter of this part discusses foundation
for science-based decision making, using data from the Gulf of Mexico.
The next chapter in Part 4 investigates decline trends in a reservoir using
Bootstrap and Monte Carlo modeling. This Part concludes with a typical
production, reserves, and valuation method used in an oil and gas company.
Given the tremendous success with the development and production
from shale reservoirs over the last 2 decades, Part 5 is dedicated to the
unconventional reservoirs. The chapters in Part 5 include: (1) Optimization
of Gas-Drilling in Unconventional Tight-Sand Reservoirs, (2) Predicting
the Fluid Temperature Profile in Drilling Gas Hydrates Reservoirs, (3)
Distinguishing between brine and gas-saturated shaly formations, and (4)
Influence of shale mechanical properties on water content effects.
Part 6 is about enhanced oil recovery. It covers EOR with hydro-
philic nanofluids, as well as CO2-EOR Flow Simulation for the Tensleep
Formation using 3D seismic data. Part 7 is the concluding section, high-
lighting new advances in reservoir characterization. It discusses the recent
Preface xxv
application of machine learning in reservoir characterization. It also dis-

cusses the future trends in reservoir characterization and the impact of
data explosion associated with the real time reservoir monitoring and res-
ervoir surveillance. It also describes how the “Big Data” concepts and data
analytics techniques will play a role in the next generation reservoir char-
acterization technology developments.
It should be understood reservoir characterization is an evolving tech-
nology. It is our hope that this volume will be a meaningful addition to the
current body of literature and will help pave the way for further advances
on the subject matter in the future.
Fred Aminzadeh
Santa Barbara, California
September 22, 2021
Part 1
INTRODUCTION
1
Reservoir Characterization: Fundamental
and Applications - An Overview
Fred Aminzadeh *
FACT Inc., Santa Barbara, CA, USA
Abstract
This article provides a brief overview of reservoir characterization at different
stages of a field from exploration to development to production and post pri-
mary production. It demonstrates the challenges associated with integration of
different data types. It also shows how “Dynamic Reservoir Characterization”
can assist in monitoring of the field for various well stimulation processes such
as enhanced oil recovery as well as reservoir stimulation. Different sections of
this entry attempt to highlight different aspects of reservoir characterization,
as an exploration tool, development tool, production tool and monitoring tool.
As reservoirs age, different measures are taken to extend their productive life.
This includes different types of reservoir stimulation and enhanced oil (or gas)
recovery.
Keywords: Reservoir characterization, data integration challenges, 3D/4D

seismic, micro-seismic data, reservoir monitoring, dynamic reservoir
characterization, rock physics and enhanced oil recovery (EOR)
1.1 Introduction to Reservoir Characterization?

As discussed in Aminzadeh and Dasgupta [2], Reservoir Characterization is to
assess reservoir condition and its properties using the available data from
core/log data to seismic and production data. This is done to assist in delin-
eating or describing a reservoir. Reservoir characterization and modeling
have become increasingly important for optimizing field development.
*Email: [email protected]
Fred Aminzadeh (ed.) Reservoir Characterization: Fundamentals and Applications, (3–22) © 2022
Scrivener Publishing LLC
3
4 Reservoir Characterization
Reservoir valuation and producing from a field demands a realistic descrip-

tion of the reservoir, requiring an integrated reservoir characterization and
modeling. An integrated approach for reservoir modeling bridges the tradi-
tional disciplinary divides and tears down interdisciplinary barriers, lead-
ing to better handling of uncertainties and improvement of the reservoir
model for field development. Integrated reservoir management requires
better characterization of the reservoir and it is imperative to a successful
operation throughout the life of the reservoir. Dynamic reservoir character-
ization is to understand the changes in reservoir properties to monitor its
performance as we produce from reservoir and/or stimulate the reservoir
to enhance production. This is accomplished by the analysis of data from
combination of different sources, to extract additional information about
the in-situ conditions of the reservoir, including the formation tempera-
ture, pressure, and the properties of the oil, gas, and brine. Other reservoir
properties that can affect measured data are density, hydrocarbon viscos-
ity, stresses, and fractures. We start with the reservoir description process
that generates models of reservoir architecture, lithologies and facies. The
geometry of the flow units is established, physical rock properties such as
porosities and permeabilities of flow layers. Three properties are related to
the pore space: porosity--the fraction of the entire volume part occupied by
pores, cracks and fractures, internal surface: the magnitude of the surface of
pores as related to the rock mass pore volume and controls interface-effects
at the boundary grain - pore fluid, permeability: the ability to flow fluid
through rock pores. Porosity and specific internal surface are scalar proper-
ties, permeability is a tensor.
Figure 1.1 shows integration of reservoir structure or architecture and
reservoir detailed properties from calibration with well data for the reser-
voir model. Reservoir description is an iterative process and need.
Different aspects of Figure 1.1, from the input data to the process (well
data, seismic data, production data, etc.) will be discussed in Section 1.2
on the data requirements. The difficulties associated with the integration
of different data sources will be addressed in Section 1.3, under “SURE
Challenge”. In Section 1.4 we discuss different aspects of reservoir char-
acterizations fin different stages of reservoir life. The exploration and
development stage deal with preliminary determination of the reservoir
structural model, stratigraphic and facies models. This is followed by the
production phase with a focus on porosity, permeability and fluid satura-
tion, involving reservoir/flow simulation and history matching. The recov-
ery stages involve injection of water/CO2 or steam to increase production.
We discuss Dynamic Reservoir Characterization (DRC) in Section 1.5.
We note that 4D seismic and microseismic data play an important role
Reservoir Characterization: Fundamental and Applications 5
Structural Stratigraphic
model model
Reservoir grid Upscaling
Well and
seismic
data
Facies proportions Geological model: Flow

∆
model Facies, porosity, simulation
permeability
Geomodel updating
Integration
Production
of 4D seismic Integration Forecasts
data y of production
tor
his data asts
no ching o rec
t f
ma CCS ctio
n
h
wit Inje
Figure 1.1 Different components of reservoir characterization, from Fornel and Estublier
[5].
in geo-model updating monitoring production and the EOR/reservoir

stimulation process. Sections 1.6 goes into more details on rock physics
and reservoir modeling and how reservoir characterization can be used as
an input to reservoir simulation and help with enhanced oil recovery and
other well stimulation processes.
1.2 Data Requirements for Reservoir

Characterization
Well data provides vertically high-resolution model at the well location,
however, the distribution of well in a field are sparse. Combining well
information with geophysical and geological data allows the necessary
constraints for extrapolating high resolution well data beyond where they
are measured thus increasing the coverage. For every phase of the reser-
voir life cycle from discovery to development to operating to maturity and
well stimulation (enhanced oil recovery) phase, geophysical tools are used
to create reservoir model with the associated properties and update the
model based new data collected.
Integration of geophysical data with geologic data, and engineering

measurements improves our understanding of the reservoir, reduces
uncertainties and mitigates the risk. The detailed spatial coverage offered
are calibrated with analysis of well logs, pressure tests, cores, fracture
system, geologic depositional knowledge and other information from
appraisal wells. 3D seismic is the primary geophysical technique used
to create the original reservoir models. 4D seismic (time lapse data) and
other new measurements (micro-seismic, new log/pressure data and pro-
duction data help create updated (dynamic) reservoir model. In addition,
gravity, controlled source EM, borehole measurements such as vertical
seismic profiling-VSP, borehole gravimeter-BRGM, cross well seismic,
cross well EM are also used to build the original and updated reservoir
models.
The required information for the petroleum engineers and geologists
includes subsurface lithology, net pay, porosity, permeability, reservoir
fluid-fill, fluid contacts, reservoir pressure and stress regime. Geophysical
tools infer reservoir properties from the measured physical observations
by blending these with measurements made at the wells like well logs, well
tests and core analyses. During the field appraisal and development stages,
understanding of the reservoir matrix properties and fluid distribution
within the reservoir are of great importance.
------sand bodies---river channels---oil/gas traps------
crack-----fissure------fracture------fault-----------------------------------------
pore--lamina--bed/set--facies--compartment--paraseq.--deltas--basin--
1mm 1m 10m 100m 1km 10km
ultrasonic--sonic logging------------
cross-hole--vsp------------
high-resolution seismic-----
reflection seismic----------------
earthquake seismology
gravity/magnetic
remote sensing data
Figure 1.2 Wide range of physical scale for different data types associated with different
geological and reservoir features.
1.3 SURE Challenge

The ultimate goal is not only to identify and delineate hydrocarbon charged
in reservoirs, but also to quantitatively determine the volume and distri-
bution of oil and gas it contains and quantify the associated uncertainty.
No single measurement has the required response to achieve this. It is
therefore essential to integrate the various types of data to a common earth
model. This information includes seismic data, various types of well data,
and geologic concepts. The challenge is to integrate measurements that are
of different Scale, Uncertainty, Resolution, and Environment or the SURE
Challenge as was introduced by Aminzadeh [1] and further elaborated
at Aminzadeh and Dasgupta [2] and Aminzadeh [3]. The entire process
of exploration for reservoirs to its abandonment involves acquisition and
analysis of different types of data. These data types are associated with an
enormous range of scale as shown in Figure 1.2. This spans ultra-sonic
measurements of pores of the order of 1 millimeter to remote sensing mea-
surements of basins of over 10 Kilometers wide. Examples of many other
data measurements for many other objects that lie in between all the fea-
tures are shown in this figure.
Admittedly not all the data types are integrated at the same time.
Nevertheless, Scale and the wide range of differences for different data
types is one of the challenges in reservoir characterizations. To make the
matter more complicated is the fact that different data types are associated
Uncertainty/Coverage Pyramid SURE Challenge
Scale,
Cores Uncertainty
n
Resolution
tio
Sc
lu
ale
Environment
so
/C
Re
ov
y/
Well Logs
er
Information Knowledge &

l
in
ag
r ta
Experience Pyramid
e
Ce
Seismic Data
Experience
AI-DA Offers solutions to

SURE Challenge Knowledge
Data & Information
Figure 1.3 SURE Challenge: Having to deal with the wide ranges of Scale, Uncertainty,
Resolution and Environment of different data types when integrating them, (from
Aminzadeh [3]).
with different levels of uncertainties. For example, the direct measure-

ments of rock properties from the core data may involve little uncertainty.
The petrophysical information from well log data may be associated with
somewhat more uncertainty. The seismic data used to ascertain reservoir
properties, for their indirect nature of measurements involve much more
uncertainty. Thus, Uncertainty level and its variations with respect to dif-
ferent data types is poses another challenge in data integration.
In addition, having different data types with vastly different underlying
Resolution, also poses a challenge for data fusion. The resolving power of
different data types is drastically different. As shown in Figure 1.3, some
data types have very high resolving power. For example, while well log
data can resolve a reservoir unit of under an inch, seismic data, generally
speaking, may not be able to resolve a reservoir under 30 feet. Finally, the
effectiveness and usefulness of different data types are impacted by the geo-
logical conditions and reservoir “Environment”. This can be associated with
different reservoir types (carbonate, clastic, unconventional, heavy oil,) or
different reservoir conditions (High Pressure/High Temperature, or reser-
voir depth (shallow water, Deep, or Ultra Deep Water.)
Millimeters
Cores
Well
Logs SR2020’s World
Increasing Vertical Resolution
Bore Hole
Seismic
Surface
Seismic
10’s of
Meters
At the Well Increasing Areal Coverage Across the Asset
Figure 1.4 Areal coverage of well data is complemented by the larger areal sampling
of the geophysical methods. VSP vertical seismic profile and Crosswell seismic fill a
resolution “gap” between sonic log measurements and vertical seismic profiles. Courtesy
of SR2020 (now Optasens).
We refer to these four key challenges: Scale, Uncertainty, Resolution

and Environment as: the SURE Challenges. Top left side of Figure 1.3
illustrates three key data types: core, well log and seismic data. We will
refer to it as a data pyramid. The base of the pyramid is the seismic with
very large coverage but with limited resolution and lesser level of cer-
tainty. The top of the pyramid is the core data with very little coverage
(only at a particular well location involving a fraction of the well) but with
high level of certainty and resolution. Effective integration of all the data
types, in spite of the SURE challenge is what reservoir characterization is
all about. As we will show in the last chapter artificial intelligence and data
analytics can play a key role in offering solutions to the SURE challenge.
The bottom right-hand side in Figure 1.3 shows an upside-down pyra-
mid comprised of a different aspect of integration. That is, vast amount of
data needs to be combined with some technical knowledge and experience
to perform effective data mining and ultimately reservoir characterization.
As an aside, it must be pointed out that borehole geophysical data (e.g.
Vertical Seismic Profile and Cross-Well data) fills the gap between core
SEM
0.0001 The integration gap
light
microscopy
0.001
core
0.01
FMI
Vertical resolution, m
0.1
horizontal well log
well
1 log
cross- VSP
surface well
10 seismic
controlled-source
electomagnetics
100
grav-mag
1000
10 000 1000 100 10 1 0.1 0.01 0.001 0.0001
Low resolution Spatial resolution, m High resolution
Figure 1.5 Vertical and spatial resolution of various geophysical, well logs and laboratory
measurements. From www.agilegeoscience.com (left), and Optaense (right).
data and well log data on one side of the scale and 3D seismic data on the
other side.
In general. The resolution of different data types for reservoir charac-
terization and description varies considerably. Figure 1.5 illustrate such a
large variability for core to log to borehole geophysics and seismic, grav-
ity magnetics data and control source electromagnetics, among others.
This further demonstrates the importance finding a solution to the “SURE
Challenge” for reservoir characterization and other E&P problems. Also,
see Ma et al. [8] addressing integration of seismic and geologic data for
modeling petrophysical properties.
1.4 Reservoir Characterization in the Exploration,

Development and Production Phases
Reservoir characterization has different focus in different phase of the
life of a field. In what follows we briefly highlight the main objective of
Reservoir Characterization in Exploration, Development, Production
(primary recovery) and Production Enhancement (secondary and tertiary
recovery) phase. For the very reason, the notion of reservoir characteriza-
tion often times means different thing to geologists, geophysicists and res-
ervoir engineers. This is primarily due to the fact that their primary focus
is different phases of life of the field.
1.4.1 Exploration Stage/Development Stage

The pre-development (exploration) phase requires delineation of the res-
ervoir limits and assessment of its economic feasibility. Development stage
requires somewhat more accurate assessment of the reservoir extent, better
appraisal of the economic viability of the reservoir and placement of new
wells for further delineation of the reservoir. More detailed depth imaging
using advanced geophysical methods, borehole geophysics applications as
well as fault seal analysis, and better understanding of reservoir compart-
mentalization are some of the objectives of reservoir characterization at this
stage. At this stage it is necessary to determine the reservoir drive mech-
anism and the size and strength of the aquifer. At the development phase,
reservoir continuity rather than reservoir delineation becomes the focus.
The focus in this phase is to identify the structural extension or truncation
beyond well control in order to minimize the drilling of dry holes. While
the pre-development and development phase require considerable atten-
tion to reservoir characterization. Geophysicists construct an initial model
by correlating lithology, porosity, net pay thickness and other properties at

the well location using seismic attributes from surface 3D seismic, VSP and
synthetic seismograms. The initial model is then updated by computing
seismic attributes between the wells in order to predict reservoir properties
in 3D. Seismic attributes like acoustic impedance, variations in amplitudes,
frequencies, interval velocities and instantaneous phase are applied in the
computation.
1.4.2 Primary Production Stage

As the primary production of the reservoir begins, the goal is to position
wells at optimal locations that would maximize hydrocarbon recovery.
During secondary recovery and then enhanced recovery process, the engi-
neer’s objective is to maximize the volume of hydrocarbon contacted by
injected fluids. This is to achieve maximum volumetric sweep efficiency for
fluid production. To minimize cost and risk, engineers attempt to predict
reservoir performance—for both planning and evaluation of hydrocarbon
recovery projects. Reservoir description in terms of reservoir architecture,
flow paths, and fluid-flow parameters are the key to reservoir engineering.
Accurate prediction of reservoir production performance is predicated pri-
marily on how well the reservoir heterogeneities are understood and have
been modeled and applied for fluid-flow simulation. This stage requires
integration of reservoir characterization models with reservoir simulation,
history matching for production optimization. Reservoir management pro-
cess conducts reservoir related studies and applies the results from fluid
flow modeling in defining, updating and optimizing a development plan
for producing the reservoir and forecast the production profile. This phase
also involves optimization and management of reservoir performance eval-
uation, surveillance of fluid flow and changes in the reservoir that result in
changes in the original distribution of physical properties. The optimization
criteria can change during the life cycle of a producing reservoir. Managing
the reservoir depletion is a dynamic process and the reservoir engineers
constantly react and adapt to the changes as they evolve. Dynamic charac-
terization is a representation of the fluid flow in a static reservoir model and
needs to be validated with reservoir performance data.
1.4.3 Secondary/Tertiary Production Stage

When the reservoir ages, intervention to increase production through
reservoir stimulation and enhanced oil (or gas) recovery becomes neces-
sary. Here, the main objective shifts to plan the production and injection,
develop an optimum secondary/tertiary recovery mechanism and plan for

infill drilling well locations to maximize recovery rate and deliver oil and
gas at the planned rate and to extend the economic producing life of the
reservoir as much as possible. The finite resources that are available are
applied to plan and optimize the economic recovery of oil and gas from the
reservoir. Petroleum engineers are responsible for planning and executing
the development and of petroleum reserves. They seek to maximize petro-
leum recovery from the reservoirs. Effective use of reservoir characteriza-
tion and reservoir model updating play an important role in this process.
This is accomplished by incorporating physical property measurements of
the reservoir rock and fluid properties that describes the reservoir architec-
ture and its initial fluid distribution to make an optimum production and
reservoir stimulation plan. 4D seismic and/or micro-seismic monitoring
plays a role in assessing the effectiveness of well different well stimulation
approaches (for example EOR, artificial lift or hydraulic fracturing), giving
rise to the importance of “dynamic reservoir characterization” at this stage.
Engineers need to monitor the reservoir state of pressure, temperature
and fluid distribution during the producing life of a reservoir. This infor-
mation could 1) identify situations within the reservoir which may poten-
tially impact oil and gas recovery, and 2) locate problems that can cause
undesirable leakage or entry into wellbore. If these situations are not cor-
rected in a timely manner, irreversible damage might occur to the reservoir
affecting the ultimate oil and gas recovery. Inter-well monitoring of pro-
duction and injection processes using geophysical techniques also allow
improvement of field development plans and optimize reservoir manage-
ment. Field scale monitoring of reservoir drainage patterns would improve
the recovery factor.
1.5 Dynamic Reservoir Characterization (DRC)

Both in the primary and post-primary production phases, we need to have
an updated characterization of the reservoir. We will refer to this as DRC.
DRC can play a key role in production optimization and monitoring of
effectiveness of the EOR operation or hydraulic fracturing. It can help
surveillance of fluid flow which is an essential part of reservoir manage-
ment process. Likewise, changes in the reservoir pressure distribution is
also helpful to make important reservoir management decisions. Among
many data sets that are used to monitor fluid, pressure and other reservoir
properties is 4D seismic and Microearthquake data. For some examples,
see Kosco [7], Maity [9] and Maleki [10].
1.5.1 4D Seismic for DRC

Changes in the reservoir fluids or pressure distribution can be imaged by
4D. This is based on the impact of fluid saturation and reservoir pressure on
the changes in compressional and shear wave velocities of seismic waves. A
simplistic idea is to directly subtract two 3D seismic volumes of acquired
in say successive years. This is often referred to as 4D seismic, with the
fourth dimension being the time lapse between the two 3D seismic sur-
veys. An example of time lapse seismic is in Figure 1.6. As the reservoir
is produced, the distribution of the fluid properties changes with time but
the reservoir rock frame properties remain constant during the produc-
ing life of the reservoir. Therefore, by repeating the seismic measurements
top reservoir
OWC
OWC
(a)
top reservoir
(b)
Figure 1.6 Time-lapse seismic response changes caused by different positions of oil-
water contact (OWC) in Gullfaks field Tarbert reservoir. (a) Oil-water contact level
before production and (b) after production for 10 years. Zones with changes in seismic
impedance are circled. http://accessscience.com.
i.e., acquiring time lapse seismic data and computing the differences in the
seismic attributes between the measurements the changes in the distribu-
tion of reservoir fluid properties between wells could often be detected.
The changes in time lapse seismic data are due to the acoustic imped-
ance variation (in this case, mostly the compressional velocity change),
caused by the reservoir production or other changes (such as water or CO2
injection in the EOR process). Acoustic impedance is the product of veloc-
ity (V) and density (ρ), for time lapse seismic amplitudes are influenced
by the incompressibility (Krock) of the reservoir rock and the production-
generated changes in the incompressibility of the pore fluids (∆Kfluids).
The reservoir rocks must therefore be sufficiently compressible so that
there is a prominent and measurable contribution from the pore fluids.
Soft compressible rocks like unconsolidated sands (younger in geologic
times) are ideal for time lapse seismic while rigid or incompressible reser-
voir rocks such as carbonates do not lend themselves for effective applica-
tion of this technology.
The acquisition and processing parameters for different vintage 3D seis-
mic should either be the same or necessary calibration should be applied
to make them consistent. That is the seismic response should be identical
when no changes in the geologic formation due to injection of production
has taken place. Some of these difficulties may be mitigated by using per-
manent sensors in wells and recording time lapse data.
1.5.2 Microseismic Data for DRC

The permanent sensors also can record micro seismic or microearthquake
(MEQ) data in passive mode without any seismic source. They detect the
seismic events that are induced by hydrocarbon production due to change
in the reservoir stress with pressure changes. Many examples of MEQ data
applications in DRC and/or monitoring hydraulic fracturing or other well
stimulation processes have been reported.
The results of seismic monitoring (4D or MEQ or combination of the
two) along with the well log and production data can be used to constrain
the flow simulation model in the reservoir history matching and fluid flow
prediction process, as well as characterizing the fracture regime. As an
example, see Maity and Aminzadeh [8]. Figure 1.7 shows how integration
of conventional seismic, well log data and MEQ data have been able to cre-
ate a 3-D fracture distribution, using a neural network approach through
integrating “Fracture Zone Identifier” or FZI attributes.
C B'
A D'
400
C'
B A'
800
Depth (m)
1200
1600
2000
100
50 D 80
60
100 40
20
150
Figure 1.7 Use of conventional seismic, well log data and MEQ data to create a 3-D
fracture distribution volume, using a neural network by integrating “Fracture Zone
Identifier” or FZI attributes, Maity and Aminzadeh [9].
1.6 More on Reservoir Characterization and

Reservoir Modeling for Reservoir Simulation
No discussion on reservoir characterization is complete without under-
standing rock properties and the corresponding rock physics. Furthermore,
reservoir modeling could be considered as the last step for reservoir char-
acterization during different stages of the life of the reservoir. Indeed,
ideally, any reservoir simulation and could use reservoir models based
on static and dynamic reservoir characterization to improve the process.
4D seismic data can help with the reservoir model updating process, thus
enabling creation of a dynamic reservoir model.
Figure 1.8 shows how a static reservoir model with the associated struc-
tural earth model and corresponding geologic and reservoir properties
such as facies, porosities, and vshale among other parameters can be a
starting point (top left). Through reservoir simulation a flow model can
be created at different time points, with the respective information about
pressure, saturation and temperature. Output of the reservoir simulation
model and the earth model could be used to generate (invert for) the reser-
voir rock physics properties such as density, compressional and shear wave
Earth Model Flow Simulation

T1
T2
TN
Φ, vshale, facies...
Pp, So, Sw, Sg, Ss, T

4D Seismic
T1
T2 Rock Physics
Vp, Vs, vs, Pp, Phi Vp, Vs, vs, Sw, Phi
3.5
TN
0.2
3
Vp, Vs (kft/sec)
2.0 2.1
Vp, Vs (kft/sec)
2.5 2 1.5
Avg. pressure
2
1.0 1.2
1
4000 5000 6000 7000 8000 0 0.2 0.4 0.6 0.8 1

Pore pressure (psi) Water Saturation
Vp, Vs, ρ
Figure 1.8 The entire process of reservoir model updating through 4D seismic modeling
and reservoir simulation, (from Meadows, [11]).
velocities at different time periods. Such information can then be used to

generate the synthetic 4D seismic data. Comparison of the synthetic and
field 4D seismic data will then lead to an updated Earth model. The pro-
cess continuous until we establish a good match between real and simu-
lated seismic and reservoir flow models leading to an acceptable dynamic
reservoir characterization results from the accurately derived rock physics
properties as well as other reservoir properties such as porosity, permeabil-
ity, oil and gas saturation, and pressure among other properties.
In what follows we describe rock physics and reservoir modeling briefly.
1.6.1 Rock Physics

Rock physics investigates reservoir rocks properties that affect transmis-
sion of seismic waves through the rocks. These physical properties are
rigidity, compressibility, and porosity. This provides a connection between
elastic properties measured at the surface of the earth, within the borehole
environment or in the laboratory with the intrinsic properties of rocks,
such as mineralogy, porosity, pore shapes, pore fluids, pore pressures, per-
meability, viscosity, stresses and overall architecture such as laminations
and fractures.
Description of rock and fluid properties between the well control points
requires understanding of the linkage of bulk and seismic properties to
each other and their changes with geologic age, burial depth, and location.
This connection allows us to understand and model the petrophysical and
geometrical properties which give rise to the seismic signal. Rock physics
requires a knowledge and understanding of geophysics, petrophysics, geo-
mechanics and the causes of distribution of fluids in the subsurface reser-
voir between wells. From seismic fluid monitoring we can obtain valuable
information about reservoir fluid movements and geologic reservoir het-
erogeneities. The results can also resolve seal integrity issues and guide the
optimum placement of wells in complex reservoirs.
Rock physics uses sonic, density and dipole sonic logs to establish a rela-
tionship between the geophysical data and the petrophysical properties. In
‘80s and ‘90s many oil companies had their own rock physics laboratories.
Because of the longer-range objectives and the need to assemble large data-
bases, today such laboratories are found primarily within five or six univer-
sities and a few service companies. The focus of rock physics analysis started
with estimating porosity and permeability of sandstones and carbonates.
Today, much of the research is focused on unconventional reservoirs and
on estimating rock strength or “fracability” and the presence of total organic
carbon. For some detailed discussion on the value of rock physics analysis in
various aspects of reservoir characterization and reservoir property estima-
tion see Dvorkin and Nur [5] and Castagna et al. [4].
Integration of 3D seismic interpretation with well measurements provides
a powerful tool for characterization a reservoir for the 3D distribution of rock
properties and the geometric framework of the reservoir. While the cores,
wireline logs and outcrops provide the vertical resolution it is only geophys-
ical data like 3D seismic data that can provide detailed spatial information
between the wells for the geological model. Since 3D seismic is a measurement
made at the surface of the earth, the subsurface interpretation using seismic
data can be done only after proper calibration with available well information.
Seismic reflection data provide the gross acoustic properties within a volume
of rock and do not have the vertical resolution of wireline logs.
1.6.2 Reservoir Modeling

Quantification of rock properties and the fluids in three dimensions is the
process of reservoir modeling. The goal of reservoir modeling and fluid
simulation is increased hydrocarbon fluid production with an increased
rate of return. The 3D quantification is performed in a geo-cellular model
that consists or reservoir geometry, lithology, porosity, permeability and
initial fluid saturation. Integration of information from seismic data,

cores, wireline logs and outcrops provide the quantification of the static
reservoir model of the reservoir. A geological reservoir characterization is
performed using a cellular facies model. Rock properties are assigned to
model cells according to the defined facies. The geological models describe
the flow layers that account for fluid and displacement phenomenon in the
reservoir. It models the inter-well connectivity and continuity of flow units
in the rock facies present within the reservoir architecture.
Reservoir fluid simulation is the quantification of fluid flow over time in
the 3D reservoir model. The numerical model simulation and forecasts of
reservoir performance is based on the geo-cellular static model. Reservoir
simulation is performed to infer fluid flow behavior from a mathematical
model. The forecast of reservoir performance is improved with increased
accuracy in the geological model. Major decisions regarding the develop-
ment and production plans for the reservoirs e.g., location and spacing of
production and injector wells, depletion strategy, maximum production
rates are based on the reservoir simulation. As hydrocarbons remaining
in place become more difficult to recover, fluid movement in the reservoir
needs to be more closely monitored. The location of remaining hydrocar-
bons must be known to plan injection schemes. Also, the manner in which
injected fluids move and make contact with the target oil must be known in
order to evaluate and, if necessary, correct the recovery project.
Static reservoir model provides a representation of the structure, thickness,
lithology, porosity, initial fluids in the reservoir. As discussed in Section 1.5
on DRC, a dynamic reservoir model is a representation of the changes in fluid
flow in the reservoir model that needs to be validated with reservoir perfor-
mance data-pressure changes, production and injection rates. Rock proper-
ties defined in the reservoir rocks from 3D seismic interpretation include:
Lithology, Porosity, Net pay thickness (or porosity volume), Fluid type and
the respective fluid saturation, as well as the reservoir pressure. The heteroge-
neity within a petroleum reservoir has a profound influence on its production
performance. Structural deformations, fractures, lithological variations, and
diagenetic alternations all contribute to the creation or destruction of con-
duits and barriers to fluid flow through the reservoir matrix. Rock physics
is a key component of analyzing the reservoir properties. It is important to
monitor changes in the fluid flow or its composition during the producing life
of the field. Figure 1.9 illustrates different components of reservoir modeling.
These integrated reservoir models are critical for forecasting, monitor-
ing, and optimizing reservoir performance over the life cycle of the reser-
voir, from exploration, development, primary production and secondary/
tertiary production. They will enable reservoir engineers to more accurately
Well data Facies model Porosity model
Conceptual facies model Seismic attribute Permeability model
Figure 1.9 Reservoir modeling process workflow. The process takes control of the
data within its modeling framework and integrates the various types of data attributes.
Courtesy: Roxar-Emerson.
−600
−620
−640
−660
−680
−700
−720
0 10 20 30 40 50
Reservior Simulation Migrated map
Geophysical Monitoring
Reservior Modeling
Seismic Modeling
Static Geological Model
Figure 1.10 Integrated reservoir modeling, fluid simulation update and reiteration
by incorporating geophysical monitoring data. http://www.co2care.org/Sections.
aspx?section=538.5.
perform flow simulation studies, identify permeability flow-paths and bar-

riers, map bypassed oil, and monitor pressure and saturation fronts in the
reservoir. All of these are essential for effective reservoir management.
Figure 1.10 shows how the original (static) geological or reservoir model
based on the integration of geophysical data could be used to for reservoir
simulation which in turn it can be used for reservoir monitoring and res-
ervoir model updating.
In Part 7 of this volume, we will discuss Artificial Intelligence (AI) and
Data Analytic (DA) can help address some of the remaining complexities
associated with reservoir characterization results. For example, Nikravesh
and Aminzadeh [12] reported on the past, present and future of AI in res-
ervoir characterization. Twenty years later Aminzadeh [3], discussed how
human and machine intelligence can be combined to improve character-
ization results. It is firmly believed that AI- and DA offer hope solve the
issues related to the SURE Challenge discussed earlier.
1.7 Conclusion
Reservoir Characterization Is an important step in the entire life cycle of
the reservoir. Reservoir Characterization is aimed at assessing reservoir
properties and its condition, using the available data from different sources
such as core samples, log data, seismic surveys (3D and 4D) and produc-
tion data. This is done in different stages of the E&P process from high
grading reservoirs in exploration to their delineation, for their develop-
ment, as well as their description for optimum production to assessing
their evolution in their stimulation for enhance oil/gas recovery to extend
their economic life. An integrated approach for reservoir characteriza-
tion bridges the traditional disciplinary divides, leading to better han-
dling of uncertainties and improvement of the reservoir model for field
development. Among the main difficulties in reservoir characterization is
what I call “SURE” Challenge. The display here demonstrates the com-
plications involved in integrating different data types with different Scale,
Uncertainty, Resolution and Environment.
References
1. Aminzadeh, F., 2005, Meta-Attributes: A new concept detecting geologic fea-
tures and predicting reservoir properties, Second International Congress on
Geosciences Merida, Mexico September 2005
2. Aminzadeh, F. and Dasgupta, S., 2013 Geophysics for Petroleum Engineers,
Elsevier.
3. Aminzadeh, F., 2021, Reservoir Characterization: Combining Machine

Intelligence with Human Intelligence, E&P Plus, April 2021, Vol. 96 Issue 4,
E&P Plus, Hart Energy.
4. Castagna, J., Han, D., Batzle, M.L., 1995, Issues in rock physics and implica-
tions for DHI interpretation, The Leading Edge, August 1995.
5. Dvorkin, J., & Nur,A., 1993, Dynamic poroelasticity: A unified model with the
squirt and the Biot mechanisms, Geophysics 58, 524-533.
6. Fornel, A. and Estublier, A. 2013. To A Dynamic Update of The Sleipner
CO2 Storage Geological Model Using 4D Seismic Data. Energy Procedia. 37.
4902-4909. 10.1016/j.egypro.2013.06.401.
7. Kosco, K. & Schiøtt, C.R. & Vejbaek, Ole & Herwanger, Jorg & Wold,
Rune & Koutsabeloulis, N., 2010, Integrating time-lapse seismic, Reservoir
Simulation and Geomechanics. 231. 61-66.
8. Ma, Y. Z., Phillips, D. Gomez, E., 2020 Synergistic Integration of Seismic and
Geologic Data for Modeling Petrophysical Properties, The Leading Edge,
March 2020.
9. Maity, D., Aminzadeh, F., 2015. Novel Fracture Zone Identifier Attribute
Using Geophysical and Well Log Data for Unconventional Reservoirs,
Interpretation Journal, Vol.3, No. 3, P.T155-T167.
10. Maleki, M., 2018, Integration of 3D and 4D seismic impedance into the
simulation model to improve reservoir characterization. PhD Dissertation,
University of Compinas.
11. Meadows, M., 2012, Time-lapse seismic data for reservoir monitoring and
characterization Course notes on Advanced Oil Field Operations with
Remote Visualization, Guest Lecturer for F. Aminzadeh’s course, USC PTE
587.
12. Nikravesh, N. and Aminzadeh, F., 2001, “Past, present and future intelli-
gent reservoir characterization trends,” Journal of Petroleum Science and
Engineering, vol. 31, no. 2, pp. 67–79, 2001.
Part 2
GENERAL RESERVOIR
CHARACTERIZATION
AND ANOMALY DETECTION
2
A Comparison Between Estimated
Shear Wave Velocity and Elastic
Modulus by Empirical Equations and
that of Laboratory Measurements
at Reservoir Pressure Condition
Haleh Azizia1*, Hamid Reza Siahkoohi2, Brian Evans3,
Nasser Keshavarz Farajkhah4 and Ezatollah KazemZadeh4
1
Department of Geophysics, Science and Research Branch, Islamic Azad University,
Tehran, Iran
2
Institute of Geophysics, University of Tehran, Tehran, Iran
3
Department of Petroleum Engineering, Curtin University, Perth, Australia
4
Research Institute of Petroleum Industry, Tehran, Iran
Abstract
The objective of this study is to assess the accuracy of the empiricalequations in
estimating the shear wave velocity and elastic modulus of rock sample at reservoir
pressure condition. The evaluated relations are Gassman, Greenberg and Castagna
which have been in use by researchers for a long time and have shown acceptable
results. The plug sample investigated in this study is taken from Berea sandstone
reservoir southwest of Australia, which is a known reference sandstone for this
type of study. This plug in the laboratory was flooded with supercritical carbon
dioxide fluid, saturated and pressured under axial, radial and pore pressure com-
parable to oil reservoirs pressures, after which elastic wave velocity and elasticity
modules were determined. Then, using empirical relationships such as Gassman,
Greenberg and Castagna and measured values of P-wave velocity, shear wave
velocity, and elasticity coefficients were estimated. Comparison of theoretical val-
ues versus experimental values shows they compare very well, only in some cases
a difference between estimated and experimental values for the coefficients of elas-
ticity has been observed. We believe that the difference is due to the assumptions
*Corresponding author: [email protected]
25
that were made in those theories. The shear wave modulus didn’t remain constant
during fluid saturation. Also, the measured bulk modulus and the calculated val-
ues based on Gassman formula did not compare very well. This difference was
observed to be larger at higher pressures.
Keywords: Wave velocity estimate, elastic waves, the coefficients of elasticity,

rock physics, reservoir conditions, Gassman relations, Greenberg – Castagna
theory, injecting carbon dioxide, sandstone
2.1 Introduction
Study of propagation of shear and compressional waves give useful infor-
mation and constitutional characteristics of hydrocarbon reservoirs, such
as lithology and pore fluid type. This information is very important for res-
ervoir development and recovery, and especially for future decision mak-
ing. On the other hand, the behavior of reservoir rocks geomechanics, play
an important role in the design and implementation of drilling, production
planning and sustainability of oil and gas wells.
Having physical geology information such as density, porosity, com-
pressional and shear wave velocities are required to successfully perform
the above-mentioned projects. This is usually the case that the informa-
tion about shear wave velocity is not readily available compared with other
data. Therefore, theoretical or experimental approaches are necessary to
estimate this velocity.
In geomechanical evaluation of hydrocarbon reservoirs, several meth-
ods can be used to estimate shear wave velocity and elastic constants.
Conventional methods for estimating the shear wave velocity can be
Laboratory
measurements Core analysis
Direct measurement Dipole shear sonic imager
Methods to
Empirical relationships Greenberg - Castagne equations -
estimate shear
gassmn equations and...
wave velocity
Regression method Correlation between the data and

parameters
Intelligent method Method of artificial neural networks and fuzzy logic
Figure 2.1 Common methods for estimating the shear wave velocity.
Comparison Between Shear Wave Velocity and Elastic Modulus 27
divided into five categories which are [1–3]: laboratory measurements,

direct measurements, empirical relationships, the regression method and
the intelligent method (Figure 2.1).
Each of these methods has its own unique advantages that make it suit-
able to estimate the shear wave velocity. The dependency of wave propaga-
tion velocity to lithology, porosity, temperature, pressure and type of pore
fluid, has introduced a variety of investigation methods in geophysics and
rock physics.
Laboratory measurement of shear wave velocity in a core is known as
a standard procedure and the obtained velocities are comparable with
that of the other methods. There are other field methods to estimate
shear wave velocity such as Dipole Shear Sonic Imager and sonic logs.
Although these methods are common, they must be performed in a
large number of wells to obtain the velocity distribution in the entire
field, and core extraction or running of sonic tools in a large field is
very expensive [4]. Another common method to estimate the shear wave
velocity is based on theoretical evaluation and modeling. In the past few
decades, several empirical formulas have been introduced for estimat-
ing the shear wave velocity in rocks with different lithology, based on
physical parameters of rock, especially the P-wave velocity and poros-
ity. Many scientists, including Pickett [5], Milholand [6], Domenico [7],
Thomsen [8], Han [9], Krief [10], Castagna [11] and Greenberg [12]
have done very useful research in this area and various relationships
have been developed and presented. These relations are valid for the
saturated rocks with brine.
Gassman formulas have been introduced to extend these experimental
formulas to other fluids contents. In 1986, Han offered empirical regres-
sion formulas for elastic waves in laboratory condition which would esti-
mate the speed based on porosity and clay content. In 1989, Eberhart
added the pressure parameter to Han’s equation for shale sand rocks.
Years before, Tosaya and Nur [13] and Castagna et al. [11] presented
empirical formulas for shale sand rocks based on velocity, porosity and
clay parameters.
Other methods are artificial intelligence techniques such as neural
networks and fuzzy logic [14–16]. Although their estimates are associ-
ated with less error, these methods also present a specific model for each
different field, and the results from one field cannot be applied to other
fields.
There are two methods to calculate the elastic coefficients. The first
one is the static method (destructive) and the other one is the dynamic
(non-destructive) method. In the static method, elastic parameters of rock
are calculated from laboratory measurements such as the uniaxial com-

pressive strength of cores. But in a dynamic method, assuming that rocks
are elastic, the coefficients are determined from propagating shear and
compressional waves in rock and measuring waves travel time, without
damaging the rock [17].
Currently, there is no unique empirical formula or comprehensive the-
ory that could be utilized to determine the elastic wave velocity and elastic
coefficients at different environmental conditions, type of fluid and rock. In
this paper, experimental data (elastic wave propagation velocity) has been
gathered using a sample saturated with water and supercritical carbon diox-
ide at different reservoir pressure. Then while the common fluid (brine)
was replaced by critical CO2, the elastic wave velocity values were calculated
using Gassman formulas and Greenberg - Castagna empirical equations
and the results were compared with direct laboratory observations.
2.2 Methodology
2.1.2 Estimating the Shear Wave Velocity
A major part of the seismic signal analysis in regards to rock physics mod-
els relates shear wave velocity to mineralogy and porosity. Rock physics
analysis based on logs and cores and the relation of these to the geological
model, leads to the establishment of a relationship between velocity and
porosity. Formulation of the relation between rock velocity and rock prop-
erties like porosity was initiated by Gassman [18] and revised later on by
Mavko and Mukerji [19] and Mavko et al. [20]. Other studies on this sub-
ject include Wyllie et al. [21], Raymer et al. [22], Castagna et al. [11], Han
[23], Raiga-Clemenceau and colleagues [24], Eberhart [25], and the critical
porosity model of Wang and Nur [26].
Greenberg - Castagna model is utilized in this study to estimate the
shear wave velocity of a rock sample. Greenberg and Castagna (1993) pre-
sented an empirical formula for multi-mineral rocks saturated in brine:
 −1 −1 
L Ni
  L
 Ni
 
1  
VS =  
2  ∑ ∑ Xi aijV j P  + 
 
∑ ∑ Xi 

aijV − j P 



 , (2.1)

 i =1 j =0 i =1 J =0 
L
∑X =1 i
i =1
where L is the number of lithology in the formation, Xi is the percentage of

the volume of lithology, aij is regression coefficient, Ni is the degree of poly-
nomial regression for the targeted lithology, Vp and Vs are compression and
shear wave velocities (Km/s).
This formula estimates shear wave velocity using compressional wave
velocity in pure unit minerals, saturated in water. Regression coefficients
of the formula for four different lithologies were presented by Greenberg
and Castagna [12].
To estimate the shear wave velocity of a brine saturated rock using
Greenberg - Castagna, formula, one needs to find a way to replace the
existing fluid with brine as a common fluid. This work is done by utilizing
Gassman relations. In fact, by brine replacement, a similar condition is
assumed for the whole environment. Then the compressional wave veloc-
ity is obtained for brine saturated situation using the following formulas.
Finally, the shear wave velocity is obtained from the estimated compres-
sional wave velocity.
Replacing different types of pore fluids with brine, and keeping the rest
of the physical properties of the rock (e.g. porosity) intact, the compres-
sional wave modulus of the rock will also be changed [27]. Compressional
wave modulus is expressed as a linear combination of bulk modulus and
shear modulus:

M = K + 43 µ( )

(2.2)
The usual process is initiated by replacing the primary fluid with a fluid
with similar sets of velocities and rock densities, compared with the pri-
mary fluid. These velocities are usually obtained from logs, but sometimes
they may also be the results of theoretical models. In this study, the velocity
of the wave that has passed through the primary fluid (in our case, super-
critical dioxide is injected into the water) is obtained through laboratory
measurements. But the removal of the existing fluid effects and replacing
it by the common fluid (brine) has been achieved through the following
steps (Dvorkin, 2003):
In the first stage the effective bulk modulus of pore fluid composition,
(K fluid ) is calculated using:
1 Sgas Soil Sbr

= + + (2.3)
K fluid K gas K oil K br

where, Sgas, Soil, Sbr, indicate gas, oil and brine saturation and Kgas, Koil, Kbr,
correspond to the apparent modulus of gas, oil, and brine. In the next step,
bulk modulus of rock (Klog), is calculated by equation (2.4):
K log = ρb (Vp2 − 4Vs2 / 3) (2.4)

where, ρb is the bulk density and Vp is the compressional wave velocity of

the rock. At the next step, bulk modulus of dry rock (Kdry) is calculated
using the equation (2.5) and the mineral bulk modulus:
1 − (1 − φ)K log / K min eral − φK log / K fluid

K dry = K min eral (2.5)
1 + φ − φK min eral / K fluid − K log / K min eral

where, ϕ is porosity and Kmineral is the apparent modulus in the mineral

phase (Thomsen, 1986). The bulk modulus of rock saturated with brine
(Kcommon) is determined by:
φK dry − (1 + φ)K cf K dry / K mineral + K cf

K common = K mineral (2.6)
(1 − φ)K cf + φK mineral − K cf K dry / K mineral

where, Kcf is the bulk modulus. The compressional wave modulus of

the rock saturated with brine (Mcommon) is calculated using the following
formula:
M common = K common + ρb 4Vs2 / 3 (2.7)

The compressional wave velocity after removal of the primary fluid and
replacing it with brine is obtained by:
Vp′ = M common ρ (2.8)

b

In this case, when the shear wave data is not available, compressional
wave modulus (Mlog) is calculated from charts (logs) using the following
relation:
M log = ρbVp2 (2.9)

The compressional wave modulus of the dry rock (Mdry) is also calcu-
lated using compressional wave modulus of the rock’s minerals:
1 − (1 − φ)M log / M min eral − φM log / K fluid

M dry = Mmineral (2.10)
1 + φ − φMmineral / K fluid − M log / K min eral

Mmineral = K mineral + 4µmineral / 3 (2.11)
where, ϕ is porosity, μmineral is shear modulus and Kmineral is the apparent

modulus in the mineral phase. The changes in the elastic modules of dif-
ferent minerals as a whole have been estimated [28]. Finally, the compres-
sional wave modulus of the brine saturated rock (Mcommon) is calculated as
follows:
φM dry − (1 + φ)K cf M dry / Mmineral + K cf

M common = Mmineral
(1 − φ)K cf + φMmineral − K cf M dry / K mineral

(2.12)
Greenberg - Castagna formula is defined for rocks completely saturated

with brine. In this paper, the apparent modulus (Kcf) of 2/25 is assumed for
brine saturated cases.
It is worth mentioning that the fluid changes have no effect on shear
wave modulus, which is the same before and after complete saturation with
brine.
2.2.2 Estimating Geomechanical Parameters

Here to determine the bulk modulus, shear modulus, and Young’s modulus
of rocks, we assumed that they are elastic, homogeneous and isotropic and
used the following formula:

( ( ) )
K = ρ Vp2 − 4 3 Vs2

(2.13)
G = ρ × Vs2 (2.14)
3Vp2 − 4Vs2
E = ρ × Vs2 × (2.15)
Vp2 − Vs2

2.3 Laboratory Set Up and Measurements

The data used in this study were collected by a core flooding system at
the laboratory of petroleum engineering department, Curtin University
of Technology, Australia. The tested sandstone sample has the volume of
2.350 (gr/cm3), 3.79 cm in diameter, 7.98 cm in length and the porosity of
36/43 deg. After drying and preliminary preparation of this sample, it was
placed in a polymer sleeve to be protected from the fluid axial pressure
applied. The sample was placed inside the core holder, and piezoelectric
(transducers) were placed on the caps to send and receive the acoustic
wave. Figure 2.2 shows schematically, how the sample and exposure trans-
ducer were placed.
As seen in Figure 2.2, the enclosed sample is under Pr pressure (radial
pressure) and axial pressure, Pax, is applied. In these conditions, the pore
pressure is shown as Pp. In Figure 2.2, OD represents the outside diameter
of the cap, ID represents the inside diameter of the cap, r is the core radius
and Paxial is effective axial pressure that is calculated as in equation (2.16):
A1 = (π (OD )2 − π ( ID )2 ) 
σ = Paxial = Pax ∗ A1 / A2 (2.16)
A2 = π r 2 

Pr
OD PP Pax
Pax
Pr
Figure 2.2 The placement of the test device is shown in schemati.

Different parts of this chamber which is in direct contact with the fluid,
are neutral to chemical reaction. At the ends of this chamber, two caps are
placed were the transducers have been installed. As seen in Figure 2.3, the
receiver and transmitter are in direct contact with the sample.
The laboratory system is equipped with a back pressure regulator (BPR),
to make it possible to simulate the behavior of real reservoir pressure. For
axial and confining pressures, a hand pump has been used, and to achieve
pore pressure, a hydraulic pump has been utilized. This test has been per-
formed at room temperature. Signals are stored in a computer equipped
with particular software which has the capability to integrate into the oscil-
loscope. Shown in Figure 2.4a is the complete flooding system device with
three different fluid injection capsule, each with different kinds of fluids
and the oven chamber that maintains the desired test temperature (46 °C
in this experiment). Figure 2.4b indicates a picture of the core holder and
connected transducers.
It should be noted that the injected fluid in this experiment is CO2
supercritical, which has been injected under 1300 psi pressure and a
Transduce
Cap
Distance between transducer & cap
Figure 2.3 Schematic, placement of sample with transducer and the top cap.
(a) (b)
Figure 2.4 (a) The core flooding system, (b) Image of the holder connected to the core
and transducers.
temperature of 46 °C at a rate of 1 (ml/min) into the sample saturated with

distilled water.
2.3.1 Laboratory Data Collection

Travel times of compressional and shear waves through the super criti-
cally CO2 saturated core sample were measured at six different effective
pressures (within the reservoir pressure range). Later on, corresponding
compressional and shear wave velocities were calculated (Figure 2.5).
4000
3500
3000
Velocity (m/s)
2500
2000
1500
1000
1000 1500 2000 2500 3000 3500 4000 4500 5000
Effective pressure (psi)
Vp-experimental (m/s) Vs-experimental (m/s)
Figure 2.5 Compressional and shear wave velocity vs different effective pressure for super
critically saturated core sample.
4100
4000
3900
Velocity (m/s)
3800
3700
3600
3500
3400
3300
1500 2000 2500 3000 3500 4000 4500 5000
Vp-experimental (m/s) Vp-estimate (m/s)
Figure 2.6 P-wave velocity (experimental and estimated) at different effective pressures.
2.4 Results and Discussion

To see how closely estimated wave velocity correlates with the experimen-
tal velocity values; we replaced the primary CO2 saturated water with brine
(using apparent module of 2.25) using Gassmann equation. Then using the
Greenberg-Castagna formula we estimated the compressional wave veloc-
ity (Figure 2.6).
In Figure 2.6, the compressional wave velocity obtained from the labora-
tory experiment and the Gassmann equation is depicted versus effective pres-
sure. It can be noticed that at all effective pressures the estimated velocities
have higher values than experimental ones. The difference gradually enlarges
as the effective pressure is increased. The rate of velocity increment is faster
for estimated velocity than the experimental counterpart. This is because in
Gassmann theory environment (the core) is considered a homogeneous and
isotropic elastic environment. This theory is often true for Monomineralic
rocks such as pure quartz sandstone or clean limestone. But since the sample
used in this study developed naturally, possible impurities exist such as clay
or fine cracks; so wave scattering is not considered in the Gassmann equa-
tions. It is also effective on bulk modulus and often in this hypothesis, it is
assumed to be average. Another issue to the assumptions of this model is the
fluid replacement instead of several fluids, regardless of their distribution in
the environment and is often considered to be the average of the density of
the fluids. This is effective on the accuracy of the model especially while the
saturation is not uniform (Patchy saturation). Also with the pressure increas-
ing, joints and cracks closed more and reduced permeability. Therefore, fluid
flow and the opportunity to achieve balance are much reduced and the oper-
ating result is that Gassmann theory pushes farther than expected.
Shear wave velocity of the rock sample was determined at the different
pressure from the estimated compressional wave velocity, using Greenberg-
Castagna formulas.
Figure 2.7 shows the experimental and estimated shear wave velocity
values at different effective pressures. As it can be seen in contrary to the
compressional wave, the estimated shear wave velocity is very much in
accordance with the corresponding experimental values. Particularly this
compatibility is more obvious at low effective pressures. The difference
between experimental and estimated velocities and the rate of changes
increase as the applied pressure increases. This is because the shear waves
are only sensitive to changes solid in the environment. With the increase in
effective pressure and reduced porosity and micro cracks, stone resistance
increases against deformation (while volume or angle); as a result, it will
2120
2080
Velocity (m/s)
2040
2000
1960
1920
1500 2000 2500 3000 3500 4000 4500 5000
Vs-experimental (km/s) Vs-estimate (km/s)
Figure 2.7 S-wave velocity (experimental and estimated) at different effective pressures.
4
Y = 3.47 X – 8200
R² = 0.96
3.9
Vp-estimated (km/s)
3.8
3.7
3.6
3.4 3.42 3.44 3.46 3.48 3.5
Vp-experimental (km/s)
Figure 2.8 Cross plot of estimated P-wave velocities vs. laboratory measurements.
2.08
Y = 0.71 X + 0.57
2.06
R² = 0.96
Vs-estimate (km/s)
2.04
2.02
2
1.98
1.96
1.94
1.92 1.96 2 2.04 2.08 2.12
Vs-experimental (km/s)
Figure 2.9 Cross plot of the estimated S-wave velocities vs. laboratory measurements.
face increasing shear wave velocity. Since in Greenberg-Castagna equation

compressional and shear wave communication waves can be described
with porosity and modules, effective stress plays an important role in the
trend of estimation. But because of unconsidered pore shape, form, size,
and density of fractures, anisotropy, texture and especially effective pres-
sure on the environment in this relationship, differences although small
can be seen in the results.
Figure 2.8 shows cross plot of the compressional wave velocity measured
in the laboratory condition vs. the estimated values along with a best-fitted
line. The correlation of the values is 0.95. Figure 2.9 also shows the cross
plot of the experimental and estimated shear wave velocities along with a
best fitted line. The correlation of the values is 0.96.
VP ( Estimate ) = 3.47 VP ( Experimental ) − 8200 (km/s)

R 2 = 0.96
VS ( Estimate ) = 0.71 VS ( Experimental ) + 0.57 (km/s)
R 2 = 0.96

Figures 2.10 and 2.11 show the cross plots of the experimental and
estimated compressional and shear wave velocities. Mathematical rela-
tionship between two sets of wave velocities were obtained. As it shows
90% correlation observed in both modes between compressional and
shear wave velocity. The important note is that the slope of the curve is
2.12
2.1 Y = 1.40 X – 2.78
R² = 0.94
2.08
Vs-experimental (km/s)
2.06
2.04
2.02
2
1.98
1.96
1.94
3.38 3.4 3.42 3.44 3.46 3.48 3.5 3.52
Vp-experimental (km/s)
Figure 2.10 Plot of experimental shear wave velocity against compressional wave velocity.
2.08
Y = 0.29 X + 0.94
2.06
R² = 0.94
2.04
Vs-estimate (km/s)
2.02
1.98
1.96
1.94
3.55 3.6 3.65 3.7 3.75 3.8 3.85 3.9 3.95 4 4.05
Vp-estimate (km/s)
Figure 2.11 Plot of estimated shear wave velocity against compressional wave velocity.
approximately 5 times in the experimental measurements more than the

estimated measurement and this is due to larger estimated compressional
wave velocity.
Since Greenberg-Castagna model assumptions is ideal for a specially
designed environment, there is no great match between this model and
laboratory values. But as was mentioned earlier, correspondence between
the velocity, the pressure and shear are very close together. It seems envi-
ronment effective pressure is the essential factor that change is not con-
sidered in the model. This effect is highlighted on shear wave velocity and
shear wave module.
VS (Greenberg −Castagna ) = 0.86 VP (Greenberg −Castagna ) + 1.17

VS( Experimental ) = 1.40 VP ( Experimental ) − 2.78 (km/s)
R 2 = 0.94
VS( Estimate ) = 0.29 VP ( Estimate ) + 9.4 (km/s)
R 2 = 0.94

As it can be seen in Figure 2.12, the behavior of Vs-Experimental/
estimated are almost independent of variability of effective pressure. While
VP-Experimental/estimated increases with the rising effective pressure.
The difference between rates of VP’s is because of some assumptions of
Gassmann-Greenberg-Castagna equations, which they guess that media
is ideal.
4000
3500
Velocity (m/s)
3000
2500
2000
1500
1500 2000 2500 3000 3500 4000 4500 5000
Vs-experimental (m/s) Vp-experimental (m/s)
Vs-estimate (m/s) Vp-estimate (m/s)
Figure 2.12 Rate of variability of experimental/estimated velocities with increasing effective

pressure.
25
Y = –3.53 X + 72.42
23
R² = 0.11
K-estimate (MPa)
21
19
17
15
14.4 14.5 14.6 14.7 14.8 14.9 15
K-experimental (MPa )
Figure 2.13 Plot of Laboratory vs. estimated Bulk modulus (K) of rock sample.
At this stage, we compared the elastic coefficients of rock samples

obtained from experimental velocity values, and those estimated from
Greenberg-Castagna model. According to Figure 2.13, the correlation
between experimental and estimated values for volume or bulk modulus is
very low and is about 0.1. This is because the compressibility of the pores
media depends on the minerals type and texture of the rock. Also, it is a
function of the amount and shape of media porosity. When the pores are
filled with fluid, elastic modulus is affected by many parameters of fluid
such as the compressibility, the kind of fluid, distribution, viscosity and
also fluid incompressibility.
The fluid used in this study has a patchy saturation and has created
anisotropic, heterogeneous environments. Besides that, different effec-
tive pressures that applied on a frame stone, impressed the results of
Gassmann - Greenberg – Castagna equations.
K ( Estimate ) = −3.53K ( Experimental ) + 72.42 (MPa)

R 2 = 0.11

One of the assumptions of Gassmann equations is that the shear modu-

lus in the dry and wet state is constant. It seems that this assumption is not
applicable for most environments. The fact is that the changes to the tex-
ture of rock due to the reaction between rock and the fluid cause a change
in the shear modulus of saturated rocks. Changes in the shear modulus is
the main cause for the difference between the experimental velocity and
the calculated velocity utilizing the Gassmann equations. Anyway, these
differences in shear modulus cause a decrease in the use of Gassmann the-
ory to estimate the velocity. These differences can be observed in Figure
2.14 for both laboratory and estimated values. Although the correlation
between experimental and estimated values are very high and close to
0.96, by replacing the common fluid and use of the estimated shear wave
velocity, shear modulus values change a little as well. The reason behind
this phenomenon is the defect in the Gassmann hypothesis that the shear
modulus for rock is equal in both dry and saturated conditions. The reac-
tion between the fluid and the texture of the rock is consequently causing
a change in the shear modulus of saturated rocks.
10
9.8 Y = 0.70 X + 2.66

R² = 0.96
9.6
G-estimate (MPa)
9.4
9.2
8.8
8.6
8.6 8.8 9 9.2 9.4 9.6 9.8 10 10.2
G-experimental (MPa)
Figure 2.14 Plot of laboratory vs. estimated shear modulus.

26
Y = 1.08 X – 0.99
25.5
R² = 0.97
25
E-estimate (MPa)
24.5
24
23.5
23
22.5
22
21.5 22 22.5 23 23.5 24 24.5 25
E-experimental (MPa)
Figure 2.15 Plot of laboratory vs. estimated Young’s modulus.
G( Estimate ) = 0.70G( Experimental ) + 2.66 (MPa)

R 2 = 0.96

But based on the values in Figure 2.15, the experimental and the esti-
mated values of Young’s modulus are very close to one another and very
comparable (%97).
E( Estimate ) = 1.08 E( Experimental ) − 0.99 (MPa)

R 2 = 0.97

2.5 Conclusions
In this paper we evaluated the closeness of the estimated elastic wave veloc-
ities and elastic modulus of a rock sample via Gassmann - Greenberg -
Castagna formula to those obtained from laboratory measurements at
reservoir pressure.
• Estimating the physical parameters of rock samples from

Gassmann - Greenberg - Castagna equations at high pres-
sures is farther from the real situation. At high pressure,
velocity values are more differentiated from each other and
the estimated values are observed to be very far from the

experimental (laboratory) values. Although it appears that
at higher pressures cracks or fractures get closed, in sub-
stituting the fluid, especially brine for supercritical carbon
dioxide and water, elasticity values have very high varia-
tions, and this very strongly influences the estimated com-
pression wave velocities. And finally, this effect can be seen
on the shear wave derived from compression wave. But since
the shear wave velocity in the fluid and gas behave similarly,
the effect was observed less on this wave.
• In the Gassmann assumptions it has been stated that pore
fluid will not modify the elastic properties of the rocks. Thus,
this theory predicts that shear modulus is not affected by the
fluid saturation and will remain constant. In this study, the
shear modulus didn’t remain constant during fluid satura-
tion and during the transformation from gas-water satura-
tion state to water saturation state; changes were detected
to be around 2 percent. Because of this change in the shear
modulus, it makes one hesitate to utilize the Gassmann the-
ory to calculate the velocity. With the calculated and the
experimental values of the shear modulus being variable, it
can be concluded that Gassmann theory is not accurate for
estimating the compressional wave velocity in a saturated
state.
• The research found that the bulk modulus values obtained
from the laboratory experiments show a very weak com-
patibility with the values calculated using Gassmann for-
mulas. One of the probable reasons for this could be an
influence of the transformation of the rock skeleton by
brine (salt water). The overall conclusion from these stud-
ies is that, for samples saturated in brine at low pressures,
good agreement exists between bulk modulus values cal-
culated from Gassmann formula and experimental values
obtained in the laboratory. While at high pressures when
the rock is very strongly influenced by pressure, this agree-
ment is nonexistent. This change was about 9% at low pres-
sures and at high pressures, these differences reached up
to 23%. Therefore, one should contemplate before utilizing
Gassmann equation.
2.6 Acknowledgment
Authors extended their appreciation to Petroleum Engineering
Department, Curtin University of Technology, Australia that provided
the authors the opportunity to utilize the laboratory core flooding system.
Also, the authors thank and appreciate very much Prof. Vamegh Rasouli,
Dr. Amin Nabipour, Dr. Mohammad Sarmadi and Dr. Mohsen Ghasemi
for their help in conducting these experiments. Finally, the authors are
very grateful and extend their deepest appreciation to the respected faculty
of Petroleum Engineering and Geophysics departments, Curtin University
of Technology that cooperated in the design, manufacturing and installa-
tion of the transducers.
References
1. A. J. Engel and G. R. Bashford, A new method for shear wave speed estima-
tion in shear wave elastography. IEEE T. Ultrason. Ferr. 62(12), 2016–2114
(2015).
2. J. K. Jang, K. Kondo, T. Namita, M. Yamakava, and T. Shiina, Comparison
of techniques for estimating shear-wave velocity in arterial wall using
shear-wave elastography - FEM and phantom study, in IEEE International
Ultrasonic Symposium, Taipei, Taiwan, p. 10, (2015).
3. J. S. L. Heureux and M. Long, Correlations between shear wave velocity and
geotechnical parameters in Norwegian clays, in 17th Nordic Geotechnical
Meeting Challenges in Nordic Geotechnic, Reykjavik, Iceland, 299–308 (2016).
4. B. Widarsono and P. M. Wong, Estimation of rock dynamic elastic property
profiles through a combination of soft computing, acoustic velocity model-
ing, and laboratory dynamic test on core samples, SPE Asia Pacific Oil and
Gas Conference, 68712 (2001).
5. G. R. Pickett, Acoustic character logs and their application in formation eval-
uation. J. Petrol. Technol. 15, 650–667 (1963).
6. P. Milholand, M. H. Manghnani, S. O. Schlanger, and G. H. Sutton,
Geoacoustic modeling of deep-sea carbonate sediments. J. Acoust. Soc. Am.
68, 1351–1360 (1980).
7. S. N. Domenico, Rock lithology and porosity determination from shear and
compressional velocity. Geophysics 49, 1188–1195 (1984).
8. L. Thomsen, Weak elastic anisotropy. Geophys 51, 1654–1966 (1986).
9. D. Han, Empirical relationships among seismic velocity, effective pressure,
porosity and clay content in sandstone. Geophysics 54, 82–89 (1989).
10. M. Krief, J. Garat, J. Stellingwerf, and J. Venter, A petrophysical interpreta-
tion using the velocities of P and S waves (full wave from sonic). Log Anal,
31, 35–369 (1990).
11. J. P. Castagna, M. L. Batzle, and R. L. Eastwood, Relationship between com-

pressional and shear wave velocities in silicate rocks. Geophysics 50, 571–581
(1985).
12. M. L. Greenberg and J. P. Castagna, Shear wave velocity estimation in porous
rocks: Theoretical formulation, prelimining verifcation and applications,
Geophysics. Prospecting 40, 195–209 (1992).
13. C. Tosaya and A. Nur, Effects of diagenesis and clays on compression veloci-
ties in rocks. Geophysics, Res. Lett. 9, 5–8 (1982).
14. M. R. Rezaee, A. Kadkhodaie, and A. Barabadi, Prediction of shear wave
velocity from petrophysical data utilizing intelligent systems: An example
from a sandstone reservoir of Carnarvon Basin, Australia. J. Petrol. Sci. Eng.
55, 201–212 (2007).
15. M. A. Dezfoolian, Body wave velocities estimation from wireline log data
utilizing an artificial neural network for a carbonate reservoir, South Iran.
J. Petrol. Sci. Technol. 31(1), 32–43 (2013).
16. Sh. Malekia, A. Moradzadeha, R. Ghavami Riabia, R. Gholamib, and F.
Sadeghzadehc, Prediction of shear wave velocity using empirical correla-
tions and artificial intelligence methods. RIAG Journal of Astronomy and
Geophysics 3(1), 70–81 (2014).
17. F. G. Bell, Engineering Properties of Soil and Rocks, 4th ed., Back well science
Ltd, United Kingdom (2000).
18. F. Gassmann, On the elasticity of porous media: Viertelijahrsschrift der
Naturforschenden Gesellschaft in Zurich. 96, 1–23 (1951).
19. G.Mavko and T. Mukerji, Pore space compressibility and Gassmannn’s rela-
tion. Geophysics 60, 1743–1749 (1995).
20. G. Mavko, T. Mukerji, and J. Dvorkin, Rock Phsycis Handbook: Tools for
Seismic Analysis in Porous Media, Cambridge University Press, Cambridge
(1998).
21. M. R. J. Wyllie, A. R. Gregory, and L. W. Gardner, Elastic wave velocities in
heterogeneous and porous media. Geophysics 21, 41–70 (1956).
22. L. L. Raymer, E. R. Hunt, and J. S. Gardner, An improved sonic transit
time-to-porosity transform, Paper P, in 21st Annual Logging Symposium
Transactions, Lafayette, Lousiana (1980).
23. D. Han, A. Nur, and D. Morgan, Effects of porosity and clay content on wave
velocities in sandstones. Geophysics 51, 2093–2107 (1986).
24. J. Raiga-clemenceau, J. P. Martin, and S. Nicoletis, The concept of acoustic
formation factor for more accurate porosity determination from sonic tran-
sit time data. The Log Analyst, Jan-Feb (1988).
25. D. Eberhart, Ph., Investigation of crustal structure and active tectonic
processes in the coast ranges, central California, PhD. thesis in Stanford
University (1989).
26. Z. Wang and A. Nur, Seismic and Acoustic Velocities in Reservoir Rocks,
Volume 2: SEG Geophysics Reprint Series 10, Society of Exploration
Geophysicists (1992).
27. J. P. Dvorkin, and S. Alkhater, Pore fluid and porosity mapping from seismic,
First Break, 22, 53–57 (2004).
28. I. Takahashi, Rock physics as a quantitative tool for seismic reservoir charac-
terization, INPEX Corporation, p. 175 (2005).
29. P. Avseth, T. Mukerji, and G. Mavko, Quantitative Seismic Interpretation,
Cambridge University Press, Cambridge (2005).
3
Anomaly Detection within
Homogenous Geologic Area
Simon Katz1*, Fred Aminzadeh2, George Chilingar2 and Leonid Khilyuk1
Russian Academy of Natural Sciences, US Branch, Los Angeles, California

1
Petroleum Engineering Program, School of Engineering, University of Southern

2
California
Abstract
This paper describes the development and investigation of a methodology and
multi-step algorithms for the detection of geologic anomalies. The algorithms
presented in this paper belong to the group of machine learning one-class clas-
sification techniques. The developed methodology includes the following steps:
(a) compiling a single-class training set that includes data recorded in a known
homogeneous area, (b) construction of universal anomaly detection classifiers
(AD classifiers) and high-resolution classifiers specifically designed for detection
of an anomaly of a certain type, (c) construction of adaptive AD classifiers, and
(d) testing a hypothesis about an anomaly type using data from the part of the
anomaly detected by the AD classifier and from the training set.
Three basis AD classifiers were suggested and tested for identification of anom-
alies: (a) distance-from-the-center of the training set, (b) sparsity of neighbors
from the training set in the vicinity of a tested record, and (c) divergence from
the center of the training set. Distance-from-the-center of the training set and
sparsity of neighbors are universal and may be used for detection of anomalies
with unknown properties and for optimization of adaptive classifiers. Divergence
is a specialized classifier designed for detection of anomalies with known proper-
ties. New adaptive classifiers presented in the paper may be used for detection of
anomalies of unknown type with efficiency superior to universal classifiers. The
performed tests with the developed algorithms demonstrate the high efficiency of
the developed one-class classifiers in the detection of gas-filled sands as anoma-
lies even when moderate size training sets were utilized. A hypothesis test about
47
anomaly type had statistically significant results in the wide range of test set sizes
when an adaptive classifier was used for anomaly detection.
Keywords: Anomaly detection, one-class classification, posterior ROC-curve

analysis, optimization, seismic velocities, rock density, brine, gas, shale
3.1 Introduction
An anomaly detection problem arises in geological research and reser-
voir characterization when, for example, there is a need to locate gas-filled
sands in the vicinity of the area with identified brine-filled sands or shales,
Katz et al., [8]. Other examples of anomaly detection problems include
finding the location of an overpressure zone using data obtained from an
area with normal gas pressure Dvorkin et al., [6], Gurevich et al., [7] and
finding the location of fractured carbonates filled with gas, Chilingar et al.,
[5], Maity and Aminzadeh [12].
Detection of anomalous behavior in seismic and/or well log data, which
can be attributed to the presence of hydrocarbon, has been the subject
of many research activities. Seismic attributes introduced by Taner [16]
have been widely used to detect seismic anomalies. Subsequently, many
methods have been used to combine such attributes, either through clus-
tering, Aminzadeh and Chaterjee [10], or neural networks Aminzadeh
et al., [13]. Then, the meta attributes concept was introduced which com-
bines machine and human intelligence for a targeted anomaly detection
Aminzadeh [11−15]. For a comprehensive overview of seismic attribute
methodologies see Chopra and Murfurt [17].
The methodology for anomaly detection (AD) presented in this paper
is different from multiclass classification where a training set is formed as
a union of several subsets, and each subset contains records from a single
class. Records from all classes are mixed in the training set. It is also differ-
ent from the classical methodology for outlier detection. Outlier detection
methodology is aimed at analysis of a dataset that may contain both reg-
ular and outlier records which are not identified prior to outlier analysis,
so there is no training set (Barnett [2], Barnett and Lewis [3], Agrawal and
Raghavan [1], Breunig et al. [4]).
In this paper writers tested new algorithms that belong to the group
of one-class classification machine learning techniques. One-class classi-
fication is a comparatively new area of machine learning research (see for
example Tax and Duin [18], Muñoz-Marí [19], Marti et al. [24]). Machine
Anomaly Detection within Homogenous Geologic Area 49
learning one-class classification methodology presented in this paper has

several advantages compared to outlier detection. This methodology is
model-free and all anomaly detection decisions are based on the use of a
value of expected false discovery rate. The main advantage of this meth-
odology is its ability to work efficiently with training sets of moderate size.
Three basis AD classifiers are tested in this paper for identification of
geologic anomalies: (a) distance in multi-dimensional space from-the-
center of the training set, (b) sparsity of neighbors from the training set
in the vicinity of a tested record, and (c) Bregman type divergence from
the center of the training set. Distance and sparsity classifiers are universal
techniques. To apply them for anomaly detection one does not need prior
information about properties of a possible anomaly. Divergence is a highly
specialized AD classifier that takes into account differences in properties
of petrophysical parameters in a regular geologic area and those within a
potential anomaly. Universal classifiers were used in this paper for con-
struction of adaptive aggregated classifiers which structure is automatically
modified according to the estimated properties of the studied anomaly of
unknown type.
The writers used data on compression (Vp), shear (Vs) waves velocities,
and rock density ρ for various geologic conditions presented by Ramos
and Castagna [9]. Their dataset contains a total of 75 measurements of
Vp, Vs, and density ρ for gas-filled sands, brine-sands, and shales. There
are a total of 75 triplets (Vp, Vs, ρ) in their dataset. Part of brine-filled
sands and shales data was used in this paper for construction of multiple
bootstrap-generated training sets, whereas test sets included a combina-
tion of records from gas-filled and brine-filled sands, and shales.
3.2 Anomaly Detection Methodology

The geologic anomaly detection methodology developed and tested in this
paper relies on the use of a single-class training set that includes a fixed
number of multidimensional records Yk=(yk,1, yk,2 , …, yk,M). We constructed
anomaly detection classifiers (AD classifiers) in such a way that they tend
to have two main properties: (a) smaller values for the records within the
training set and the records outside the training set that are of the same
type as the training set records, (b) larger values for anomalous records
outside the training set. AD classifiers are constructed as functions of one
or several petrophysical parameters recorded at the location that is to be
checked for anomaly status:
AD(Yk ) = AD( y k ,1 , y k ,2 ,, y k , M ) (3.1)

The anomaly identification rule is constructed as follows:

Record Yk is classified as regular if
AD(Yk ) ≤ anomaly detection cutoff (3.2)
If AD(Yk) > anomaly detection cuttoff, then record Υ is classified as

anomalous. The anomaly cutoff is defined by the expected false discovery
rate (expFD):
N ( AD(Yk ) > anomaly cutoff ;Yk ⊂ TrainSet )

exp FD = (3.3)
K
N(AD(Yk) > anomaly cuttoff; Yk ⊂ TrainSet) is the number of records in

the training set with values of anomaly detection classifier exceeding cut-
off, K is total number of records in the training set.
For construction of anomaly detection classifiers we selected parame-
ters based on results of dissimilarity analysis (Katz et al., [8]). These param-
eters are Vp/Vs and Poisson’s Ratio.
3.3 Basic Anomaly Detection Classifiers

Three basic classifiers are introduced, analyzed and tested in this paper:
1. Distance from the center of the training set:
dist (Y , Center ) =
1
M ∑( y m − ctr ,m )2 (3.4)
m =1
where ym and ctr,m are coordinates of the tested record and

of the center of the training set respectively. The center of
the training set is defined as the mean over train set records.
Coordinates of the training set center are of the form:
ctr ,m =∑1
K k =1
y k ,m where yk,m is the m-th coordinate of the
k-th record in the training set, K is total number of records
in the training set.
2. Nearest neighbors sparsity:
spars(Y ) =
1
L ∑dist(Y ,neighbor ) l (3.5)
l =1
where dist(Y, neighborl) is the distance between tested record

Y and its l-th nearest neighbor from the training set. The
farther away in a parameter space tested records are from
the records in the training set, the larger are both the spar-
sity and the distance from the center of the training set.
These two classifiers are universal. Their performance is not
affected by the properties of records in the training set.
3. Divergence is defined as follows:
div(Y , Center ) =
1
M ∑a m
∗
(ctr ,m − ym ) (3.6)
m =1
The divergence defined by the Eq. 3.6 is of the “Bregman divergence”

type (Bregman [20]). It is similar to distance, but does not satisfy either
the triangle inequality or the symmetry conditions. Applications of
Bregman divergence to the solution of machine learning problems are
presented, for example in the Banerjee et al., [21]. Bregman type diver-
gence of Eq. 3.6 is a new highly specialized AD classifier with coefficients
am dependent on the anomaly type. It needs prior information about the
type of potential anomaly. This classifier may be efficient, for example, if
all coordinates of the anomaly records tend to be smaller than the respec-
tive coordinates of the records in the training set. This is to be the case
for such parameters as Vp/Vs and Poisson’s ratio, if the training set is a
compilation of the records obtained in brine-filled sands or shales, and the
anomaly of interest is gas-filled sands. In this case reasonable values for
coefficients in Eq. 3.6 are am=1.
We also construct and test adaptive aggregated anomaly classifiers

designed to identify anomalies with unknown properties. They are built as
a linear combination of measured parameters:
aggr (Y ) = ∑y m ∗ sm (3.7)
m =1
Weights sm in Eq. 3.7 should be adjusted according the properties of a

specific anomaly. In this paper, the writers showed a technique for the opti-
mization of these coefficients for detection of an anomaly with unknown
properties.
3.4 Prior and Posterior Characteristics of Anomaly

Detection Performance
To characterize anomaly detection quality, we introduced and distinguished
two types of quality characteristics: (a) Prior quality characteristics and (b)
Posterior (actual) quality characteristics. The only prior classification qual-
ity characteristic is an expected false discovery rate (expFD). Value of the
expected false discovery rate is assigned prior to performing data analysis.
It is used for calculation of anomaly detection cutoff (AD cutoff) on the
data in a training set. Posterior characteristics are calculated on the test set
with identified regular and anomaly records. They include true and false
discovery rates as functions of the AD cutoff. True and false discovery rates
form a posterior ROC curve, which is used for evaluation of area under
the ROC curve and comparative analysis of efficiency of several anomaly
detection classifiers.
The writers used bootstrap for statistical analysis of anomaly detection
results and did comparative analysis of properties of posterior efficiency
characteristics. At each bootstrap run, sampling with replacement was
done and a randomly formed pair of training and test set was constructed.
The training set was selected from a pool of regular records. Each test set
contained both regular and anomaly records. Multiple pairs of training
and test sets produced by random sampling were utilized for calculation of
quality characteristics of AD classifiers. They included mean and median
values, and width of the quantile region for analyzed AD characteristics.
They also included analysis of parameters characterizing relations between
expected false discovery rate and posterior AD characteristics. The ROC
curve analysis was done using multiple posterior ROC curves.
Figure 3.1 illustrates the methodology for one-class anomaly detection

technique that starts with an assignment of expected FDR and calculation
of anomaly classification cutoff. It also demonstrates a high level of perfor-
mance of the divergence classifier.
First 20 values in the Figure 3.1 marked by circles (Index≤20) are the
values of the divergence classifier on the records from the training set.
The points marked by triangles and crosses (Index>20) show values of the
divergence classifier on the records from the test set. The horizontal dashed
line shows anomaly detection cutoff. Records with the values above the
cutoff are classified as anomalous. One can observe that distribution of
the values of the divergence classifier on the regular records in the test set
is very similar to that in the training set. On the other hand, the values of
the divergence for anomaly records are systematically higher compared to
divergence for regular records. The anomaly detection cutoff corresponds
to the expected false discovery rate of 15%. Expected FDR is calculated as
percent of records in the training set that exceeds AD cutoff. Posterior true
and false discovery rates are, respectively, the percentages of regular and
anomaly records in the test set exceeding AD cutoff. In this particular case,
the posterior false discovery rate is smaller compared to the expected FDR
and equals 6.6%. True discovery rate is high and equals 84%. High true
discovery rate is due to the large proportion of anomaly records charac-
terized by positive divergence values. Low posterior FDR is due to the fact
1.5 Regular records, training set

Regular records, test set
Anomaly records, test set
1.0
Divergence
0.5
0.0
–0.5
0 20 40 60
Index
Figure 3.1 Divergence values for records in training and test set. The horizontal dashed
line is the classification cutoff for the expected false discovery rate of 15%. The test set
contains 30 regular records and 25 anomaly records from gas-filled sands. Each record
(Vp/Vs and Posson’s ratio).
that divergence values on a large proportion of regular records in the test

set are smaller than the classification cutoff.
Figure 3.2 illustrates an effect of the size of the training set on the possible
difference between posterior and expected false discovery rates. Black dots
on both plots show expected versus posterior FDR for randomly sampled
training and test sets. Continuous lines represent regression of the mean
expected vs. the mean posterior FDR. Both regression lines are averaged
over 1000 randomly sampled sets. On average, expected and posterior false
discovery rates are about the same for both sizes of the training set. On the
other hand, random differences between expected and posterior FDR are
much larger for the training sets with a smaller number of records.
Table 3.1 shows mean values of posterior false and true discovery rates of
three AD methods, applied for detection of the gas-filled sand anomaly. To
get mean values shown in the Table 3.1, 1000 randomly sampled training
and test sets were created and each pair of the training-test sets was used
for calculation of posterior true and false discovery rates. An expected false
discovery rate was assigned the value of 20%. According to the Table 3.1,
the mean value of the posterior false discovery rate is about the same and
very close to the expected false discovery rate for all three AD methods.
On the other hand, posterior true discovery rate is different using different
methods and is the largest for the divergence classifier.
Train set size = 25 Train set size = 5

1.0 1.0
0.8 0.8
Expected FDR
Expected FDR
0.6 0.6
0.4 0.4
0.2 0.2
0.0 0.0
0.0 0.4 0.8 0.0 0.4 0.8

Posterior FDR Posterior FDR
Figure 3.2 Expected versus posterior false discovery rate for two sizes of the training set.
Classifier: distance from the center of the training set. Coordinates of a record: Poisson’s
Ratio and Vp/Vs.
Table 3.1 Mean values of false and true discovery rates for three classifiers in
detection of gas-filled sands anomaly. Size of the training set: 20 records.
Mean of
Expected false Mean of posterior posterior true
AD discovery rate false discovery discovery rate
method (%) rate (%) (%)
Sparsity 20 20.44 60.85
Distance 20 20.44 71.32
Divergence 20 20.49 85.71
Since the divergence classifier was specifically designed for detection of

gas-filled sand anomalies, it outperforms two universal classifiers. Its mean
the posterior true discovery rate is as high as 85%.
3.5 ROC Curve Analysis

The ROC curve analysis presented in this section is based on the analysis of
results of anomaly detection with 1000 bootstrap sampled training and test
sets of different size. Two input petrophysical parameters, Poisson’s Ratio
and Vp/Vs, were used for calculation of AD classifiers for each of the pair
training - test sets. For every training set, the expected value of the false
discovery rate was assigned, and distributions of posterior true and false
discovery rates were calculated on anomaly detection results.
In the case of multiple bootstrap sampled training and test sets, ROC
curve analysis has to deal with multiple ROC curves. In that case, quantiles
of the true and false discovery rates, and quanties of area under ROC curve
may be used for a compact description of a set of multiple ROC curves.
The writers distinguish two types of ROC curve analysis: (a) posterior
ROC curve analysis, which is identical to traditional ROC curve analysis
and (b) expected - posterior ROC curve analysis. Posterior ROC curve is
a plot of posterior true discovery rate versus posterior false discovery rate,
both calculated on the test set. Expected-posterior ROC curve is a plot of
posterior true discovery rate (posterior TDR) versus expected false discov-
ery rate (expected FDR), where posterior TDR is calculated on the test data
and expected FDR assigned using training set data.
Figure 3.3 shows histograms of posterior AUC values for three c lassifiers -
divergence, sparsity, and distance. Sets of values of AD classifiers were cre-
ated with bootstrap random sampling of the training and test sets. Each
Divergence AUC
200
150
100
50
0
0.5 0.6 0.7 0.8 0.9
Sparcity AUC
40
Count
20
0
0.5 0.6 0.7 0.8 0.9
Distance AUC
60
40
20
0
0.5 0.6 0.7 0.8 0.9
AUC
Figure 3.3 Histograms of area under posterior ROC Curve (AUC) for three anomaly
detection classifiers. Size of each training set is 5 records.
training set contained five records. The total number of randomly gener-
ated pairs of training and test sets was 1000. The intersections of contin-
uous vertical black lines with x-axis at Figure 3.3 shows the median value
of the AUC. Dashed lines indicate lower and upper quantiles of the AUC
distribution calculated for quantile probabilities Plow=0.1 and Pupper=0.9.
According to the Figure 3.3, the divergence classifier is characterized by
the narrowest distribution of AUC values and the largest AUC median.
Distribution of this classifier is symmetric. Distributions of distance and
sparsity classifiers are asymmetric with long tails in the direction of smaller
values. As a result, lower quantiles for the distribution of distance and spar-
sity AUC are shifted towards smaller AUC values.
Two of the three AD classifiers with AUC shown in Figure 3.3 do not
rely on the use of information about properties of petrophysical parameters
within potential anomalies. This is their advantage since they may be used
for detection of any type of anomaly with unknown properties. Although
they underperform compared to the divergence classifiers that rely on the
use of known anomaly properties, they still can produce median AUC val-
ues as high as 0.75.
Figure 3.4 shows posterior AUC histograms for sparsity classifiers cal-
culated for two sizes of the training sets. Continuous vertical lines are
median values, intersection of doted lines with x-axis are the levels Plow=0.1
Size of training set: 25
40
Frequency
20
0
0.5 0.6 0.7 0.8 0.9
60 Size of training set: 5

Frequency
40
20
0
0.5 0.6 0.7 0.8 0.9
AUC
Figure 3.4 AUC histograms and quantile regions calculated from 1000 pairs of training-
test sets. Size of the training sets: 5 and 25 records. Sparsity classifier.
0.90
0.85
Median of posterior AUC
0.80
0.75
0.70
0.65
Divergence Distance
Sparsity
0.60
5 10 15 20 25
Size of the training set
Figure 3.5 Median of posterior AUC for three AD classifiers as a function of the size of
the training set.
and Pupper=0.9. According to Figure 3.4, the AUC quantile region in the case
of 25 records training sets is 40% narrower than the quantile region in the
case of the training sets of 5 records.
The median of posterior AUC values and the width of the AUC quantile
region for three AD classifiers are shown in Figures 3.5 and 3.6. According
0.2
Quantile width of posterior
AUC distribution
0.1
0.0
Divergence Distance
–0.1
Sparsity
5 10 15 20 25
Size of the training set
Figure 3.6 Quantile width of AUC distribution calculated on anomaly detection results
from 1000 randomly generated pairs of training-test sets.
to Figure 3.5, the median AUC of the divergence is higher than median
AUC of other classifiers with its values stable around 0.9. Importantly, it
is as high as 0.9 even for small training sets containing only five records.
The sparcity classifier has the lowest AUC median which may be as small
as 0.70.
According to Figure 3.6, divergence has the narrowest AUC quantile
region. AUC quantile regions of two other AD classifiers are significantly
wider than that of divergence. The width of AUC quantile regions for these
classifiers decreases with the increasing size of the training sets. It is still
about three times as wide as the AUC quantile width of divergence for the
size of training set 20, 25 records.
3.6 Optimization of Aggregated AD Classifier

Using Part of the Anomaly Identified by
Universal Classifiers
It was shown in the previous section that a divergence AD classifier is very
efficient in detecting an anomaly with known properties. It was noted also
that universal classifiers have lower efficiency compared to divergence.
The goal of this section is to develop a methodology for construction of

an adaptive AD classifier working as a universal classifier for detection of
anomalies of unknown type that is still almost as efficient as the special-
ized divergence classifier. The methodology for its synthesis is a two-step
procedure:
1. Detection of a part of the anomaly using universal classifiers,

such as the distance or the sparsity.
2. Optimization of aggregated classifier on the detected part of
anomaly.
The structure of the aggregated classifier is defined by a set of coeffi-

cients sm (Eq. 3.7). These coefficients are chosen so that they maximize the
ratio:
median(aggr (classifier Anomaly Re cords ))

ar (s1 , s2 ,s M ) = (3.8)
median(aggr (trainSet Re cords ))

where sm; 1 ≤ m ≤ M are weights and aggr() is defined by Eq. 3.7, classifier
Anomaly Records are records identified by a universal classifier as anomaly,
trainSetRecords are records from the training set.
To find coefficients sm that maximize efficiency criterion (3.8) we used a
multidimensional grid search. In the grid search, the coefficients sm in Eq.
3.7 take on a discrete set of values in the following region:
−1 ≤ sm ≤ 1 (3.9)
The efficiency criterion (3.8) is calculated for each combination of coef-

ficients sm at individual grid nodes. An adaptive aggregated AD classifier
maximizes efficiency criterion on the grid. As soon as the aggregated clas-
sifier is synthesized it may be used for anomaly detection using the test set.
We illustrate construction of an adaptive aggregated classifier using the
sparsity classifier at the first optimization step. The classifier to be synthe-
sized is of the following form:
aggr (s1 , s2 ) = s1 ∗ PoissonRation + s2 ∗ Vp /Vs (3.10)

Therefore, search is done on the two-dimensional grid.

Figure 3.7 shows values of the sparsity classifier used as the first step of
optimization of the aggregated classifier. Twenty regular records that form
1.5 Regular records, training set

Regular records, test set
Anomaly records, test set
1.0
SPARC
0.5
0.0
–0.5
0 20 40 60
Index
Figure 3.7 Sparsity values on the records of the training and test sets. Horizontal dashed
line - anomaly detection cutoff producing an expected false discovery rate of 20%.
500 Expected false discovery rate: 20%
200
0
0.55 0.60 0.65 0.70 0.75 0.80 0.85
True discovery rate
Expected
250
false discovery rate: 5%
100
0
0.55 0.60 0.65 0.70 0.75 0.80 0.85
True discovery rate
Figure 3.8 Anomaly detection. Histograms of posterior true discovery rate (TDR) for two
values of expected false discovery rate. AD method: aggregated. 20 regular records in each
training set.
the training set are randomly selected out of a set of 50 records. The test
set contains 30 regular and 25 anomaly records. Anomaly records are from
gas-filled sands. Regular ones are from brine-filled sand or shale. The
assigned value of the expected false discovery rate was 20%. Records clas-
sified as a potential anomaly include 13 actual anomaly records and 2 reg-
ular records. Thus the posterior true discovery rate is very moderate - 52%.
Table 3.2 Mean and three quantiles of distribution of AUC values.

Quantiles Width of
quantile
Mean P=0.05 Median P=0.5 P=0.95 region
Divergence 0.897 0.89 0.895 0.909 0.019
Aggregated 0.866 0.862 0.87 0.901 0.039
Distance 0.795 0.633 0.818 0.885 0.252
Sparsity 0.765 0.576 0.786 0.868 0.292
The results shown in Figure 3.8 characterize the performance of an adap-

tive aggregated classifier that was applied to 1000 bootstrap randomly formed
training - test datasets. Optimization of the aggregated classifier was done
on the part of the test records identified by the sparsity as anomaly. Figure
3.8 shows histograms of values of the posterior true discovery rate for
expected false discovery rates of 20% and 5%. According to Figure 3.8, a
higher expected false discovery rate of 20% leads to a higher posterior true
discovery rate in the range of 83%- 87%. In the case of lower expected FDR
of 5%, posterior discovery rate is within 70% -85%.
The results of comparative analysis of the performance of several clas-
sifiers in detection of gas-filled sand anomaly are shown in the Table 3.2.
This table shows mean, median, and two quantiles of posterior AUC values
for four classifiers. According to the Table 3.2, the highest quality results
are produced with the divergence classifier. Performance of the adaptive
aggregated classifier is slightly lower but close to performance of diver-
gence. The distance and the sparsity classifiers are characterized by smaller
median AUC values and wider widths of the quantile region.
3.7 Bootstrap Based Tests of Anomaly Type

Hypothesis
The hypothesis test discussed in this section is the first step in a more defi-
nite and automatic determination of anomaly type. It relies on the use of
the anomaly indicator defined by Eq. 3.11:
median (ϕ j (detectedAnomalyPart ))
J
anomalyIndicator = ∑ ηj ∗
median (ϕ j (trainSet ))
j =1

(3.11)
detectedAnomalyPart - is the dataset formed by records identified as

anomaly by an AD classifier, φj - anomaly characterization functions on
average taking larger values on regular records and smaller values on
records within hypothesised anomaly, ηj - positive coefficients, J is the total
number of anomaly characterization functions.
Further we use Poisson’s ratio as a single anomaly characterization func-
tion and assign a single η1 value 1.
In that case, anomalyIndicator takes the form:
median ( Poisson’ Ratio(detectedAnomaly ))

anomalyIndicator =
median ( Poisson’ Ratio(trainingSet ))

(3.12)
According to Ostander [22], high-porosity, brine-saturated sandstones

tend to have high Poisson’s ratios of 0.30 to 0.40, while gas-saturated sand-
stones may have a Poisson’s ratio as small as 0.1. Data presented by Jain
et al. [23] show that Poisson’s ratio for gas-saturated sands is in the range of
0.2-0.25, so that the largest possible value of the anomaly indicator (3.12)
in the case of a gas-filled sand anomaly is around 0.8. So the following
assumption might be valid for gas-saturated sands with porosity in a wide
range of values: If AnomalyIndicator is larger than 0.8, then the anomaly
type is not gas-saturated sands. Then the hypothesis H0 for anomaly type
with alternative H1 may take the form:
H0 : AnomalyIndicator ≥ significance cutoff

(3.13)
H1: AnomalyIndicator < significance cutoff

where significance cutoff = 0.8 and significance level α = 0.05.

Hypothesis H0 is rejected if the probability to exceed the significance
cutoff is smaller than a predefined significance level. Rejection of hypoth-
esis H0 does not contradict the assumption that the detected anomaly is
gas-filled sands. If hypothesis H0 is not rejected, then there is a contradic-
tion between the results of the test and the assumption that the detected
anomaly is gas-filled sands.
The anomalyIndicator in Eq. 3.12 is a random variable which value
is affected by many random factors. These factors include the noise that
distorts records in the training and test sets, the variability of the detec-
tedAnomalyPart, and the non-zero posterior false discovery rate. To test the
anomaly type hypothesis and to estimate the p-value we used a bootstrap
that generates multiple pairs of training-test sets and, respectively, multi-

ple bootstrap samples of anomalyIndicator. In the experiment presented in
this section, the number of bootstrap-generated samples of anomalyIndi-
cator calculated for four AD classifiers was 1000. These samples were used
80
40
0
0.4 0.6 0.8 1.0 1.2
(a)
80
40
0
0.4 0.6 0.8 1.0 1.2
(b)
50
20
0
0.4 0.6 0.8 1.0 1.2
(c)
Figure 3.9 Histograms of anomalyIndicator for three types of classifiers. (a)-adaptive

aggregated classifier, (b)-distance, (c)-sparsity. Intersection of vertical dashed lines with
x-axis - value of significance level
Table 3.3 Percent of values of anomalyIndicator exceeding significance cutoff α

(p-value).
Percent of values of anomalyIndicator
Number of Expected false exceeding significance level (p-value)
records in discovery
the train set rate (%) Sparsity Distance Aggregated
25 20 0.2 3.7 0.01
20 20 2.1 16.2 0.04
15 20 14.5 36.4 0.1
12 20 28.7 47.4 0.8
10 20 42.5 47.5 1.7
7 20 69.2 56.5 6.7
5 20 77.3 60.4 9.9
to calculate distributions of the anomaly indicators and to estimate the

probability of exceeding the significance cutoff by the anomalyIndicator.
Figure 3.9 shows three histograms of bootstrap generated values of
anomaly indicators for three types of AD classifiers. Vertical dashed lines
are drawn at the value of significance cutoff a equal to 0.8. Histograms were
calculated for the 20 records training sets. According to Figure 3.9, distribu-
tion of the anomaly indicator is the narrowest and is shifted in the direction
of small values when an adaptive aggregated classifier is used. Distribution
obtained with the distance classifier is characterised by the largest percent of
values of an indicator exceeding the significance level a. Generally, the per-
cent of anomalyIndicator samples exceeding the significance cutoff is higher
for universal classifiers compared to the aggregated classifier.
More detailed information on results of anomaly type hypothesis test is
given in the Table 3.3. It shows p-values for the hypothesis tests for three
types of AD classifiers and for multiple sizes of the training set. One can
observe that p-values are smaller than significance level for all three clas-
sifiers if the training set contains 25 records. With decrease of the size of
the training set p-value increases for all three AD classifiers. It becomes
larger than 40% for universal classifiers when the size of the training set
is 10 records or lower. P-value does not exceed 10% for all sizes of the
training sets in the case of aggregated classifier. It remains smaller than
the significance level for a sample size of 15 records or larger. For p-values
smaller than the significance level, hypothesis H0 is rejected so that statis-
tically significant results of the test are in accordance with suggestions that
the detected anomaly is gas-filled sands. Therefore, statistically significant
results for tests of anomaly type hypothesis were obtained for sizes of the
training set larger than 10 records in the case of the adaptive aggregated
classifier, for the size of a training set of 25 in the case of the distance clas-
sifier, and for sizes of 20 and 25 records in the case of the sparsity classifier.
3.8 Conclusion
New algorithms and a methodology for machine learning one-class anom-
aly detection are presented. Detection of gas-filled sandstones, detection of
abnormal pressure zones, and detection of gas-filled fractured carbonates
are examples of geologic anomaly detection problems.
Three groups of AD classifiers were developed and tested for anomaly
detection:
a. Universal classifiers applicable to detection of an anomaly

of arbitrary type. Included among them is the distance from
the center of the training set and the sparsity of neighbors

from the training set.
b. Specialized methods designed for detection of an anomaly
of a specifictype. An example of this type of method is the
Bregman type divergenceclassifier designed for the detec-
tion of gas-filled sand anomalies.
c. New types of adaptable classifiers. An example of a classifier
of this type isthe aggregated classifier.
d. All three groups of methods were tested for detection of the
gas-filled sandanomaly with regular records presented by
the records from the brine-filled sands and shales.
Statistical analysis presented in this paper illustrates that the specialized

divergence classifier outperforms universal AD classifiers in detecting gas-
filled sand anomalies. Its posterior discovery rate was as high as 84% with
a false discovery rate smaller than 20%. Adaptable classifiers that do not
need information about properties of anomalies have a posterior r discov-
ery rate around 80% with false discovery close to around 20%. Universal
methods have a lower posterior true discovery rate. The writers envision
the main role of universal AD methods in the detection part of the anom-
aly as a first step in the adaptation of aggregated classifiers.
A combination of bootstrap and ROC curve analysis was used for anal-
ysis of the efficiency of the developed algorithms. In the framework of
this approach multiple pairs of training and test sets were generated using
bootstrap resampling of the analyzed data set. Then multiple ROC curves
were generated. The set of generated ROC curves was subjected to statisti-
cal analysis with estimation of median AUC and lower and upper quantiles
for the AUC values.
Several tests of the hypothesis related to the identification of anomaly
type were completed. An instrument for testing the anomaly type hypoth-
esis is the parameter named anomalyIdentifier which multiple values were
generated via bootstrap. Poisson’s ratio was used as the anomaly charac-
terization function. When the adaptable classifier was used for anomaly
detection, test results were highly significant for sizes of the training set
exceeding 10 records.
References
1. R. Agrawal, and P. Raghavan, A Linear Method for Deviation Detection in
Large Databases Arning A. KDD-96, 164–169 (1996).
2. V. Barnett, The study of outliers: Purpose distance and model. Applied

Statistics 27(3), 242–250 (1978).
3. V. Barnett, and T. Lewis, Outliers in Statistical Data, p. 582, John Wiley, New
York, NY, (1994).
4. M. Breunig, H.-P. Kriegel, R. T. Ng, and J. Sander, LOF: Identifying density-
based local outliers, Proc. ACM SIGMOD Int. Conf. Dallas, 12 (2000).
5. G. Chilingar, S. Mazzulo, and H. Rieke, Carbonate Reservoir Characterization:
A Geologic-engineering Analysis, p. 639, part 1. Elsevier (1992).
6. J. Dvorkin, G. Mavko, and A. Nur, Overpressure detection from compres-
sional and shear-wave data. Geophysical Research Letters 26(22), 3417–3420
(1999).
7. A. Gurevich, G. Chilingar, and F. Aminzadeh, Origin of the formation fluid
pressuire distribution and ways to improving pressure prediction methods.
J. Pet. Sci. Eng. 12, 67–77 (1994).
8. S. Katz, G. Chilingar, F. Aminzadeh, and L. Khilyuk, Dissimilarity analysis
of petro-physical parameters as gas-sand predictors. Journal of Sustainable
Energy Engineering 2, 101–115 (2014).
9. A. Ramos, and J. Castagna, Useful approximations for converted-wave AVO.
Geophysics 66(6), 1721–1734 (2001).
10. F. Aminzadeh, and S. Chatterjee, Applications of cluster analysis in explora-
tion seismology. Geoexploration 23, 147–159 (1984).
11. F. Aminzadeh, Meta attributes: A new concept for reservoir characterization
and seismic anomaly detection, GCAGS 53th Annual Convention (2003).
12. D. Maity, and F. Aminzadeh, Novel fracture zone identifier attribute using
geophysical and well log data for unconventional reservoirs. Interpretation
Journal 3(3), 155–167 (2015).
13. F. Aminzadeh, J. Barhen, C. W. Glover, and N. B. Toomanian, Reservoir
parameter estimation using a hybrid neural network. Computers and
Geosciences 26, 869–875 (2000).
14. F. Aminzadeh, Applications of AI and soft computing for challenging prob-
lems in the oil industry. J. Petroleum Science and Engineering 47, 5–14 (2005).
15. F. Aminzadeh, A new concept for seismic anomaly detection, Offshore
Technology Conf., SPE (2005).
16. M. T. Taner, F. Koehler, and R. E. Sheriff, Complex seismic trace analysis.
Geophysics 44, 1041–1063 (1979).
17. S. Chopra, and K. J. Marfurt, Seismic attributes for prospect identification
and reservoir characterization. SEG Geophysical Developments 11, 464
(2007).
18. D. Tax, and R. Duin, Uniform object generation for optimizing one-class
classifiers. J. Machine Learning Research 2, 155–173 (2001).
19. J. Muñoz-Mari, F. Bovolo, L. Gomez-Chova, L. Bruzzone, and G. Camp-
Valls, Semisupervised one-class support vector machines for classification of
remote sensing data. Geoscience and Remote Sensing, IEEE Transactions on
48(8), 3188–3197 (2010).
20. L. Bregman, The relaxation method of finding the common points of convex
sets and its application to the solution of problems in convex programming.
Computational Mathematics and Mathematical Physics 7(3), 200–217 (1967).
21. A. Banerjee, S. Merugu, I. Dhillon, and J. Ghosh, Clustering with Bregman
Divergences. J. of Machine Learning Research 6, 1705–174 (2005).
22. W. Ostander, Plane-wave refection coefficients for gas sands at nonnormal
angles of incidence. Geophysics 49(10), 1637–1648 (1984).
23. P. Jain, C. Jambhekar, and P. Pandey, Identification of gas using Vp/Vs vis-
a-vis Poisson’s ratio. 9-th Biennial Int. Conf. and Exposition on Petroleum
Geophysics, Haiderabad (2012).
24. L. Marti, N. Sanchez-Pi, L. Molina, and A. Garcia, Anomaly detection
based on sensor data in petroleum industry applications. Sensors 15(2),
2774–2797 (2015).
4
Characterization of Carbonate
Source-Derived Hydrocarbons Using
Advanced Geochemical Technologies
Hossein Alimi *
University of Southern California USC, Global Conventional and Unconventional

Geochemistry (GCUG)
Abstract
Various crude oil classifications have been proposed by geochemists and petro-
leum refiners. Geologists and geochemists are more interested in identifying and
characterizing the crude oils, to relate them to source rocks and to measure their
grade of evolution using advanced geochemical technologies. Carbonate source
rocks studied are organically very rich, containing oil-prone kerogen capable of
generating significant quantities of hydrocarbons.
Tertiary and Cretaceous oils studied are of carbonate origin and classified as
aromatic-intermediate class. Most of the Tertiary oils studied are severely biode-
graded while those from Cretaceous reservoirs are non-degraded.
Almost all the oils investigated show the following biomarker parameters, char-
acteristic of oils derived from carbonate source rocks deposited in a marine envi-
ronment under reducing conditions:
• Predominance of phytane over pristane, (pristane/phytane ratio

below 1),
• Predominance of C29-norhopane over C30-hopane, (C29/C30
hopanes above 1),
• High C35-homohopane index,
• Low sterane/αβ-hopanes ratio, and
• High dibenzothiophene/phenanthrene ratio (mostly above 1).
69
Keywords: Pristane, phytane, biodegradation, C29-norhopane, C30-hopane,

carbonate-sourced oil, C35-homohopane index, biomarkers, dibenzothiophene/
phanathrene
4.1 Introduction
For obtaining a fundamental knowledge of the genetic history of crude oils
two types of geochemical methods should be used.
1. Those for determination of the genetic origin of the oil and

to unravel its mixtures, those methods that can see through
the effects of biodegradation are the most useful.
2. Those used to determine the biodegradation history and the
charging of a biodegraded reservoir. Methods that directly
analyze the effects of biodegradation are most useful.
All geochemical studies using molecular geochemical approaches

should include the classical methods of biomarker and paraffin analy-
sis. Application of these analyses, SARA, GC-FID, GCMS saturates, and
GCMS aromatics, form the foundation of advanced geochemical technol-
ogies (AGTs) necessary to solve the critical questions about the provenance
of the oil and history of the reservoirs.
Any geochemical oil or source rock correlation study should begin with
a liberal application of those classical molecular geochemical analyses
that have become routine. The basic SARA, GC-FID, GCMS saturates and
GCMS aromatics provide an initial assessment of extent of biodegradation,
possible source relationships and thermal maturities. However, the bio-
degradative alteration of heavy oil introduces complexities that are hidden
from the view of these classical methods.
Carbonate rocks (limestone and dolomite) contain major oil and gas
deposits throughout the world:
• 33 % of the North American Fields

• 50% of N. American Giant Fields
• ~ 40% of giant fields world wide
Carbonate rocks have been discounted as important source rock

because of their lower organic-carbon content and catalytic activity in
comparison to shales [1]. However, carbonate source rocks contain mostly
sapropelic (oil-prone, Type II)) organic matter [2], which yields a higher
Characterization of Carbonate Source-Derived Hydrocarbons 71
percentage of oil earlier than more humic organic matter of shales. The
unique Carbonate-Evaporite depositional and diagenetic environments,
without infux of major terrigenous and humic substances, produce
algal-sapropelic organic facies that yield sulfur-rich petroleum at lower
temperatures than humic-type shales. The worldwide presence of small to
very large, apparently immature, to marginally mature, non-biodegraded,
heavy oil/bitumen/asphalts deposits, rich in resins and asphaltenes, are oil
sourced from carbonate source rocks.
This study will provide an overview of the geochemical characteristics
of carbonate-derived oils and source rocks collected from different wells
and fields.
As with shales, the source potential of carbonate rocks depends pri-
marily upon the organic facies rather than the mineral matrix. Where
the depositional and early diagenetic environment is highly oxygen-
ated, the total-organic-carbon (TOC) content is low. The remaining
kerogen is highly oxygenated, with a negligible generative capacity for
hydrocarbons.
The early anoxic diagenetic depositional environment can result in
the deposition of organic-rich, fine-grained carbonate sediments that are
excellent potential source rocks [3]. Although they constitute a small per-
centage of all carbonate rocks, organic-rich, fine-grained carbonate rocks
are widespread in both time and space and are the probable source of
30-40% or more of the petroleum reserves of the world.
Gas-prone organic facies are rare in carbonate rocks, because they are
usually dominated by humic organic matter deposited in a dominantly
clay matrix. However, gas-prone organic facies may occur in carbonate
rocks as results of turbidite deposition [4] or by a mixture of kerogen
types II and III.
Oils derived from carbonate rocks are often richer in cyclic hydrocar-
bons and sulfur containing compounds than oils derived from shales,
owing to the lack of terrestrial-plant waxes in their source organic matter.
4.2 Samples and Analyses Performed

A total of 5 Neogene, Paleogene of Cenozoic (No. 1–5) and 9 Cretaceous
(No. 1–9) oils, as well as 6 Cretaceous calcareous core samples from wells A
and B have been selected for this study. The oils investigated were selected
from different formations and fields. Most of the Tertiary oils investigated
are very heavy and severely biodegraded. The bitumens studied are also of
Cretaceous age.
Total organic carbon content (TOC) and hydrocarbon generating

potential of the carbonate samples were analyzed by Leco carbon ana-
lyzer and Rock-Eval pyrolysis, respectively. Solvent extraction was used
to remove bitumen from the source rocks, while liquid chromatography
(LC), FID- gas chromatography (FID-GC), and gas chromatograph-mass
spectrometry (GC-MS) analyses were performed on both the oils and
selected bitumen. Selected results from these analyses are discussed in the
following chapter.
4.3 Results and Discussions

All 6 source rocks investigated are organically very rich (Table 4.1), showing
TOC content in the 3.35% to 11.31% range. Results of Rock-Eval pyroly-
sis, such as hydrogen index (HI) and Tmax values indicates that all carbo-
naceous source rocks analyzed are highly enriched in sapropelic (Type II)
kerogen (Figure 4.1) which based on Tmax-values appear to be thermally
within the oil window (Figure 4.2). The pyrolysis results also indicate that
the kerogen present in these source rocks show relatively high pyrolysis
S2-peak values ranging from 7.74 to 15.13 mg HC/g rock, indicating that
the kerogen present has a large remaining potential capable to generate
hydrocarbons at higher maturity stages [5, 6].
The LC-results presented in Figure 4.3 indicate that the majority of the
oils investigated are rich in aromatic hydrocarbons, with their polar com-
pounds (resins and asphaltenes) present in relatively high proportions,
characteristic of the aromatic-intermediate class. This class of oils includes
most crude oils from Jurassic and Cretaceous sources of the Middle East
[1].
Typical gas chromatograms (GC-fingerprints) of carbonate and shale-
derived oils are presented in Figures 4.4 and 4.5, respectively. Carbonate-
derived oil shows a pristane/phytane ratio below (i.e. a predominance of
phytane over pristane), indicative of a reducing environment, while that of
oil from shale facies shows a pristane/phytane above 1 [7–9].
As mentioned above almost all Tertiary oils studied (Figure 4.6 as an
example) are relatively heavy and severely biodegraded. This can be seen
from the absence of all normal alkanes, while Cretaceous oils analyzed are
relatively light and non-biodegraded (Figure 4.7 as an example), with all
normal and branced (iso-) alkanes being present.
Saturated and aromatic fractions of the oils and source rock bitumens
were further analyzed by GC-MS to investigate their aliphatic (saturated)
Table 4.1 Results of TOC and Rock-Evaol Pyrolysis for the rock samples analyzed.
Well Depth S1/
name (ft) % TOC* RE-SI* RE-S2* RE-S3* Tmax* HI* OI* TOC*100 PI
Well A 14144 11.31 3.76 15.13 1.09 454 134 10 33 0.20
Well A 14171 9.34 3.44 9.98 1.43 451 107 15 37 0.26
Well A 14193 6.25 2.92 7.9 1.37 454 126 22 47 0.27
WellB 13709 3.35 9.19 3.57 1.00 448 107 30 274 0.72
WellB 13718 9.82 15.67 12.9 1.00 454 131 10 160 0.55
WellB 13726 5.88 5.11 7.74 0.64 455 132 11 87 0.40
*TOC = Total organic Carbon (wt%); SI = mg HCs/g rock; S2 = mg HCs/g rock; S3 = C02/g rockmg; Tmax = deg C; HI = Hydrogen index (mg
HCs/g TOC); OI = oxygen index (mg/C02/g TOC).
1000
Well A Well B
900 l
Type I
800
Hydrogen index (HI, mg HC/g TOC)
700
II
600 Type II
500
400
300
200
III
Type III
100
IV
Type IV
0
0 10 20 30 40 50 60 70 80 90 100
Oxygen index (OI, mg CO2/g TOC)
Figure 4.1 Type of Kerogen present in source rocks.
and aromatic biomarkers. Table 4.2 shows biomarker parameters of car-

bonate facies oils compared to those of shale-derived hydrocarbons.
The following are the significant observations that could be obtained
from Table 4.2 and GC-MS biomarker results:
• Carbonate oils show Pristane/phytane ratios below one,

whereas pristane/phytane ratios of shale facies oils are
around or above one.
• Carbonate oils are more enriched in αβ-hopanes than ster-
anes, showing lower steranes/αβ-hopanes ratios, whereas in
shale-derived oils steranes are more abundant than hopanes,
showing high steranes/αβ- hopanes ratios [2].
1.0
Immature Oil window Dry gas window
Condensate -wet gas zone

0.9
Intensive
0.8 generation,
expulsion
Well A Well B
0.7
Production index (PI)
0.6
Stained or
0.5 contaminated
0.4
0.3
0.2 High level conversion
0.1
Low level conversion Overmature
0.0
400 425 450 475 500
Maturity (based on Tmax, °C)
Figure 4.2 It shows kerogen conversion and maturity (Tmax).
Increasing maturity Increasing biodegradation
AHC Ter. oil 1

Ter. oil 2
10 90 Ter. oil 3
20 80 Ter. oil 4
Ter. oil 5
30 70 Crt. oil 1
40 60 Crt. oil 2
50 50 Crt. oil 3
Crt. oil 4
60 40
Crt. oil 5
70 30 Crt. oil 6
80 20 Crt. oil 7
Crt. oil 8
90 10
Crt. oil 9
SHC 10 20 30 40 50 60 70 80 90 Polars Crt. SRExt
Figure 4.3 Bulk compositions of oils and bitumen studied.

IC4 NC4 IC5 IC4 NC4IC5
22DMB NC5 22DMB NC5
23DMB
CP 3MP 2MP CP23DMB 3MP 2MP
NC6 NC6
22DMP 24DMP
223TMB 22DMP
24DMP
223TMB MCP
BZ33DMP 23DMP 33DMP
BZ CH
11DMCP
3EP 2MH 3MH
1C3DMCP
1T3DMCP 11DMC P
1C3DMC
1T3DMC P 2MH
23DMP
3MH
1T2DMCPNC7 3EP1T2DMC
P P NC7
113TMCP MCH ISTD
ECP 113TMC
124TMCP
TOL 123TMCP 124TMC
ECPPP
123TMC P TOL
NC8 NC8
EB MXYL IP9 IP9 EB
PXYL PXYLMXYL
OXYL OXYL
NC9 NC9
PB IP10 PB IP10
Whole oil GC trace
Whole oil GC trace

NC10 NC10
IP11 IP11
NC11 NC11
NC12 NC12
IP13 IP13
IP14 IP14
NC13 NC13
IP15 IP15
NC14 NC14
IP16 IP16
NC15 NC15
NC16 NC16
IP18 IP18
IP19 NC17 IP19 NC17
PHEN PHEN
IP20 NC18 IP20 NC18
NC19 NC19
NC20 NC20
C25HBI NC21 C25HBI NC21

Phytane
NC22 NC22
Figure 4.4 GC-Fingerprint of a carbonate-derived oil.
Figure 4.5 GC-Fingerprint of a shale facies derived oil.

NC23 NC23
Phytane
NC24 NC24
NC25 NC25
NC26 NC26
NC27 NC27
NC28 NC28
NC29 NC29
NC30 NC30
NC31 NC31
Pr/Ph = 0.49
NC32 NC32
NC33 NC33
NC34 NC34
NC35 NC35
Pr/Ph = 1.44
NC36 NC36
NC37 NC37
NC38 NC38
NC39 NC39
NC40 NC40
NC41 NC41
G1090570.D
G2091217.D
Whole oiI GC trace G2091242.D
Ph
Pr
Figure 4.6 GC- Fingerprint of Tertiary Oil No. 3 which is severely biodegraded.
Whole oiI GC trace G20291210.D
Figure 4.7 GC- Fingerprint of Cretaceous Oil No. 3 which is not biodegraded.
Table 4.2 Typical biomarker characteristics of shale and carbonate-derived

hydrocarbons.
Biomarker parameter Shales Carbonates

Pristane/Phytane Often ≥ 1 Often ≤ 1
Steranes/αβ-hopanes High Low
Diasteranes/steranes High Low
C24 tetracyclic/C26 tricyclic terpanes Low-medium Medium-high
C29/C30 hopanes Often low Can be high (>1)
C35 Homophone Index Often low Often high (>0.1)
Hexacyclic hopanes & benzohopanes Low High
Dibenezothiophene/phenanthrene Low (<3) High (>1; Often
>3)
αββ steranes Low-medium Medium-high
• Carbonate-derived oils (Figure 4.8) often display a pre-

dominance of C29-hopane (norhopane) over C30-hopane
(hopane), with C29/C30 hopane ratios above one.
• In contrast, shale-derived oils show triterpene patterns
(Figure 4.9 as an example) in which C29-hopane is predomi-
nant over C30-hopane, with a C29/C30 hopane ratio being low
(mostly below one).
• The Dibenzothiophene/Phananthrene ratio [10] is relatively
high (mostly above 1) in carbonate oils (Figure 4.10), while
those in shale-derived oils are low (mostly below 1).
m/z 191:Tri-, tetra- and pentacyclic Terpanes

C29-Hopane
Tertiary oil # 1
m/z 191:Tri-, tatra- and pentacyclic Terpanes

Cretaceous oil # 1
Figure 4.8 Triterpane fingerprints of carbonate-derived oil.
C30-Hopane
C29-Hopane
Figure 4.9 Triterpane fingerprint of shale-derived oil.

Assesment of inferred source rock lithofacies for the oils studied

5.0
4.5 Ter. oil1

Zone 1A Maine carbonate
Ter. oil2
4.0 Zone 1A Zone 1B Maine carbonate, marine
Ter. oil3
Dibenzothiophene / phenanthrene
marllacustrine sulfate-rich
Ter. oil4
3.5 Zone 2 Lacustrine sulfate-poor
mixed shaje-carbonate Ter. oil5
3.0 Crt. oil1
Zone 3 Marine shale and other
lacustrine Crt. oil2
2.5 Crt. oil3
Zone 4 Fluvial/deltaic
Crt. oil4
2.0
Zone 1B Crt. oil5
1.5 Crt. oil6
Crt. oil7
1.0 Crt. oil8
Crt. oil9
0.5 Zone 2 Zone 3 Zone 4
Crt. SRExt
0.0
0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0 4.5 5.0
Pristane / phytane
Figure 4.10 Cross plot of dibenzothiophene/phenanthrene versus pristane/phaytane.
4.4 Summary and Conclusions

Carbonate rocks contain major oil and gas deposits throughout the world.
They commonly have been discounted as important source rock due to
lower organic-carbon content, however, carbonate source rocks contain
mostly sapropelic (oil-prone, Type II)) organic matter capable of generat-
ing significant quantities of oil.
The objective of the study was to provide an overview of the geochem-
ical characteristics of carbonate-derived oils collected from different wells
and fields. A total of 6 Cretaceous carbonate core samples and 14 oils
from Tertiary (5) and Cretaceous (9) reservoirs were investigated using
advanced geochemical technologies.
Almost all Tertiary oils in this study are severely biodegraded whereas,
the Cretaceous oils are not.
All the oils studied show biomarker parameters that are characteristic of
oils derived from carbonate source rocks, such as:
• Pristane/phytane ratios below 1

• High C29/C30-hopane ratios
• Low steranes/hopanes ratios
• High C35-homohopane index values
• High dibenzothiophene/phenanthrene ratio
References
1. B.P. Tissot and D.H. Welte, Petroleum Formation and Occurrence, Springer-
Verlag, New York (1984).
2. K.E. Peters, C.C. Walters, and J.M. Moldowan, The Biomarker Guide Volume
1 and 2 Biomarkers and Isotopes in Petroleum Exploration and Earth History,
pp. 492, Cambridge University Press, United Kingdom (2005).
3. GJ. Demaison and G.T. Moore, Anoxic environments and oil source bed gen-
esis. AAPG Bulletin 64, 1179-1209 (1980).
4. S.E. Calvert, Oceanographic controls on the accumulation of organic matter
in marine sediments, in Marine Petroleum Source Rocks, J. Brooks and A.J.
Fleet (Eds.), pp. 137-151, Blackwell, London (1987).
5. K.E. Peters, Guidelines for evaluating petroleum source rocks using pro-
gramed pyrolysis. AAPG Bulletin 70, 318-329 (1986).
6. D.K. Baskin, Comparison between H/C and Rock-Eval hydrogen index as
indicator of organic matter quality, in The Monterey Formation: From Rocks
to Molecules, CM. Isaacs and J.Rullkoetter (Eds.), pp. 230-240, Colombia
University Press, New York (2001).
7. D.H. Welte and D. Waples, Ueber die Bevorzugung gradzahliger n-Alkane in
Sedimentgesteinen. Naturwissenschaften 60, 516-517 (1973).
8. R. Alexander, R. Kagi, and G.W. Woodhouse, Geochemical correlation of
Windalia oil and extracts of Winning Group (Cretaceous) potential source
rocks, Barrow Subbasin, Western Australia. AAPG Bulletin 65, 235-250 (1981).
9. H. Ten Haven, J.W. de Leeuw, J. Rullkoetter, and J.S. Sinninghe damste,
Restricted utility of the pristane/phytane ratio as paleoenvironmental indi-
cator. Nature 330, 641-643 (1987).
10. W.B. Hughes, A. G. Holba, and L.I.P. Dzou, The ratios of Dibenzothiophene to
Phenanthrene and Pristane to Phytane as indicators of depositional environ-
ment and lithology of petroleum source rocks. Geochimica et Cosmochimica
Acta 59, 3581-3598 (1995).
5
Strategies in High-Data-Rate
MWD Mud Pulse Telemetry
Yinao Su1, Limin Sheng1, Lin Li1, Hailong Bian1, Rong Shi1, Xiaoying Zhuang2
and Wilson Chin2*
1
China National Petroleum Corporation (CNPC), Beijing, China
2
Stratamagnetic Software, LLC, Houston, Texas, United States
Abstract
Low data rates, typically 1-3 bits/sec or less, are the norm in “mud pulse”
Measurement-While-Drilling (MWD) systems. For example, “sirens” offering
high carrier frequencies produce low-amplitude signals subject to severe atten-
uation; and strong signals from “positive pulsers” are slow since large forces are
required to overcome mud inertia. Moreover, reflections, high pump noise, ero-
sion and rapid power consumption affect all pulsers. This paper describes an inte-
grated systems approach to high-data-rate telemetry. Source strength is enhanced
using downhole constructive interference, and surface signal processing elim-
inates downward reflections and pump noise through “directional filters,” both
drawing on wave-based methods. Hydraulic properties associated with torque,
erosion, aerodynamic stability and turbine performance, and subtleties found in
large-scale acoustic wave interactions, are studied using dynamic similarity. In
particular, short and long wind tunnels are introduced with significantly reduced
test times, cost and required labor. Experimental facilities, prototypes and signal
processing methods are presented in detail.
Keywords: Echo cancellation, mud pump noise, mud siren, MWD, noise
removal
81
5.1 Summary
Measurement-While-Drilling (MWD) systems presently employing mud
pulse telemetry transmit no faster than 1-3 bits/sec from deep wells con-
taining highly attenuative mud. The reasons – “positive pulsers” create
strong signals but large axial flow forces impede fast reciprocation, while
“mud sirens” provide high data rates but are severely lacking in signal
strength. China National Petroleum Corporation research in MWD telem-
etry focuses on improved formation evaluation and drilling safety in deep
exploration wells. A high-data-rate system providing 10 bits/sec and opera-
ble up to 30,000 ft is described, which creates strong source signals by using
downhole constructive wave interference in two novel ways. First, teleme-
try schemes, frequencies and pulser locations in the MWD drill collar are
selected for positive wave phasing, and second, sirens-in-series are used to
create additive signals without incurring power and erosion penalties. Also,
the positions normally occupied by pulsers and turbines are reversed. A
systems design approach is undertaken, e.g., strong source signals are aug-
mented with new multiple-transducer surface signal processing methods to
remove mudpump noise and signal reflections at both pump and desurger,
and mud, bottomhole assembly and drill pipe properties, to the extent pos-
sible in practice, are controlled to reduce attenuation. Special scaling meth-
ods based on dynamic similarity are developed to extrapolate wind tunnel
results to real muds flowing at any downhole speed. We also describe the
results of detailed acoustic modeling in realistic drilling telemetry chan-
nels, and introduce by way of photographs, CNPC’s “short wind tunnel” for
signal strength, torque, erosion and jamming testing, “very long wind tun-
nel” (over 1,000 feet) for telemetry evaluation, new siren concept prototype
hardware and also typical acoustic test results. Movies demonstrating new
test capabilities are available upon request.
5.1.1 High Data Rates and Energy Sustainability

Predictions that the world would “run out of oil” have worried consumers
and governments for decades; however, these concerns have always proven
unfounded. Although the planet’s supply of fossil-based energy is limited,
this resource is likely to be plentiful in the long term if prudent manage-
ment practices are in place. Sustainability can be enhanced by deepening
our understanding of geological processes and by augmenting this advan-
tage with detailed real-time data obtained during drilling and explora-
tion, e.g., see Boyd, Darling, Tabanou, Davis, Lyon, Flaum, Klein, Sneider,
Sibbit, and Singer [1].
Strategies in High-Data-Rate MWD Mud Pulse Telemetry 83
This is easier said than done. For example, in the early days of well log-
ging, formation evaluation dictated that gas or oil-filled rocks have higher
resistivities than water-filled rocks. Interpretation was simple. But through
the years, “low-resistivity pay” has become a worldwide phenomenon; with
lower oil prices driving the re-exploration of mature fields, new methods of
interpreting resistivity readings have proliferated.
Answers depend on many factors. It is known, for instance, that prop-
erties isotropic at small scales may be anisotropic macroscopically. Many
logging tools lack the resolution needed to resolve values for individual
thin beds of sand and shale, often taking averages that conceal useful reser-
voir information. Ideally, instruments would sample properties at multiple
scales and return their findings to on-site geologists. Similar arguments
apply to formation tester, acoustic, nuclear and NMR logs – more data,
especially if available during drilling, is always helpful.
The bottleneck in this quest for increasing information – again, aimed at
enhancing sustainability by reducing the likelihood of bypassing produc-
tive zones – is a direct consequence of low data transfer rates supported
by present MWD technology. In a digital age where “megabytes per sec-
ond” define the norm, data rates during drilling, typically just 1-3 bits/
sec or less, are limited by high mud attenuation, poor signal strength at
the source, and highly reverberant and distortive environments, all leading
to decreases in signal-to-noise ratio. Moreover, problems associated with
high levels of sand erosion and excessive power demands add to already
difficult engineering challenges.
Little effort has been undertaken in the past thirty years to approach
MWD mud pulse telemetry as a science based on acoustics and fluid-
dynamical principles. The present authors have approached this research
systematically over the past decade and results are reported in their recent
book, i.e., Chin, Su, Sheng, Li, Bian, and Shi [2]. This paper presents an over-
view of the new methodologies plus a major update on “direction-based”
(as opposed to frequency-oriented) signal processing techniques for sur-
face enhancement.
5.1.2 Introduction
The petroleum industry has long acknowledged the need for high-data-
rate MWD mud pulse telemetry in oil and gas exploration. This need is
driven by several demand factors: high density logging data collected by
more and more sensors, drilling safety for modern managed pressure drill-
ing and real-time decision-making, and management of economic risk by
enabling more accurate formation evaluation information.
Yet, despite three decades of industry experience, data rates are no bet-
ter than they were at the inception of mud pulse technology. To be sure,
major strides in reliability and other incremental improvements have been
made. But siren data rates are still 3 bits/sec in shallower wells and positive
pulser rates still perform at a dismal 1 bit/sec or less. Recent claims for data
rates exceeding tens of bits/sec are usually offered without detailed basis or
description, e.g., the types of mud used and the corresponding hole depths
are rarely quoted.
From a business perspective, there is little incentive for existing oil ser-
vice companies to improve the technology. They monopolize the logging
industry, maintain millions of dollars in tool inventory, and understandably
prefer the status quo. Then again, high data rates are not easily achieved.
Subtleties abound. Quadrupling a 3 bits/sec signal under a 12 Hz carrier
wave, as we will find, involves much more than running a 48 Hz carrier
with all else unchanged. Moreover, there exist valid theoretical consider-
ations (via Joukowski’s classic formula) that limit the ultimate signal pos-
sible from sirens. Innovative mechanical designs for positive pulsers have
been proposed by others and tested. Some offer extremely strong signals,
although they are not agile enough for high data rates. Unfortunately, the
lack of complementary telemetry schemes and surface signal processing
methods renders them hostage to strong reverberations and signal distor-
tions at desurgers.
One might surmise that good “back of the envelope” planning, from
a systems engineering perspective underscoring the importance of both
downhole and surface components, is all that is needed, at least in a first
pass. Acoustic modeling in itself, while not trivial, is a well-developed
science in many engineering applications. For example, highly refined
theoretical and numerical models are available for industrial ultrasonics,
telephonic voice filtering, medical imaging, underwater sonar for subma-
rine detection, sonic boom analysis for aircraft signature minimization,
and so on; several of them deal with complicated three-dimensional, short-
wave interactions in anisotropic media.
By contrast, MWD mud pulse telemetry can be completely described by
a single partial differential equation, in particular, the classical wave equa-
tion for long wave acoustics. This is the same equation used, in elementary
calculus and physics, to model simple organ pipe resonances and has been
the subject of research reaching back to the 1700s. Why few MWD design-
ers use wave equation models analytically or experimentally, by means of
wind tunnel analogies implied by the identical forms of the underlying
equations, is easily answered: there are no physical analogies that have
motivated scientists to even consider models that bear any resemblance
to high-data-rate MWD operation. For instance, while it has been possi-

ble to model Darcy flows in reservoirs using temperature analogies on flat
plates or electrical properties in resistor networks, such approaches have
not been possible for the problem at hand.
5.1.3 MWD Telemetry Basics

By far, the most comprehensive article available on MWD design from a
broad systems level perspective is due to Patton,Gravley, Godbey, Sexton,
Hawk, Slover, and Harrell [3]. The authors’ pioneering work at Mobil Oil in
the 1970s evolved into Schlumberger’s present commercial mud siren tech-
nology and has seen consistent improvement over the years. This decades-
old work differs from the present paper in several respects: (1) We have
introduced much broader use of the wave equation, both in surface signal
processing and in downhole source strength enhancement. (2) We have
developed the use of short and long wind tunnels in MWD testing, sig-
nificantly reducing costs, test times and labor, while at the same time facil-
itating real-time learning. (3) And we have addressed key issues related to
torque, erosion, jamming, aerodynamic stability, and other factors by using
predictive fluid-dynamical approaches that remove much of the uncer-
tainty that always accompanies engineering design. In each of these areas,
the present authors are not aware of any publicly available information.
Why is mud pulse telemetry so difficult to model? In all industry pub-
lications, signal propagation is studied as a piston-driven “high blockage”
system where the efficiency is large for positive pulsers and smaller for
sirens. The source is located at the very end of the telemetry channel (near
the drillbit) because the source-to-bit distance (tens of feet) is considered
negligible when compared to a typical wavelength (hundreds of feet).
For low frequencies, this assumption is justified. However, the mathe-
matical models developed cannot be used for high-data-rate evaluation,
even for the crudest estimates. In practice, as drilling mud passes through
them, a rapidly oscillating positive pulser or rotating siren will create pres-
sure disturbances that are antisymmetric with respect to source position.
For instance, as the valve closes, high pressures are created at the upstream
side, while low pressures having identical magnitudes are found on the
downstream side. The opposite occurs when the pulser valve opens.
The literature describes only the upgoing signal. However, the equally
strong downgoing signal present at the now shorter wavelengths will
“reflect at the drill-bit” (we will expand on this later) – with or without a
sign change – and travel through the pulser to add to upgoing waves that
are created later in time. Thus, the effect is a “ghost signal” or “shadow” that
haunts the intended upgoing signal. But unlike a shadow that simply fol-
lows its owner, the use of “phase-shift-keying” (PSK) introduces a random
element that complicates signal processing: depending on phase, the upgo-
ing and downgoing signals can constructively or destructively interfere.
Modeling of such interactions is not difficult in principle since the linearity
of the governing equation permits simple superposition methods.
However, it is now important to model the source itself: it must create
antisymmetric pressure signals and, at the same time, allow up and down-
going waves to transparently pass through it and interfere. It is also nec-
essary to emphasize that wave refraction and reflection methods for very
high frequencies (associated with very short wavelengths) are inapplicable.
The solution, it turns out, lies in the use of mathematical forcing functions,
an application well developed in earthquake engineering and nuclear test
detection where seismic waves created by local anomalies travel in multiple
directions around the globe only to return and interfere with newer waves.
Wave propagation subtleties are also found at the surface at the stand-
pipe. We have noted that (at least) two sets of signals can be created down-
hole for a single position-modulated valve action (multiple signals and
MWD drill collar reverberations are actually found when area mismatches
with the drill pipe are large). These travel to the surface past the stand-
pipe transducers. They reflect at the mud-pump and at the desurgers. For
high frequency, low amplitude signals (e.g., those due to existing sirens),
desurgers serve their intended purpose and the internal bladders “do not
have enough time” to distort signals. On the other hand, for low frequency,
high amplitude signals (e.g., positive pulsers), the effects can be disastrous:
a simple square wave can stretch, change in shape and literally become
unrecognizable.
Thus, robust signal processing methods are important. However, most
of the schemes in the patent literature amount to no more than common
sense recipes that are actually dangerous if implemented. These often sug-
gest “subtracting this, delaying that, adding the two” to create a type of
stacked waveform that hopefully improves signal-to-noise ratio. The dan-
ger lies not in the philosophy but in the lack of scientific rigor: true filter-
ing schemes must be designed around the wave equation and its reflection
properties, but few MWD schemes ever are.
Moreover, existing practices demonstrate a lack of understanding with
respect to basic wave reflection properties. For example, the mud pump is
generally viewed with fear and respect because it is a source of significant
noise. It turns out that, with properly designed multiple-transducer signal
processing methods, piston induced pressure oscillations can be almost
completely removed even if the exact form of their signatures is not known.
In addition, theory indicates that a MWD signal will double near a piston
interface, which leads to a doubling of the signal-to-noise ratio. This, in
fact, has been verified experimentally in the field, a result that has prompted
improved strategies for surface transducer placement.
5.1.4 New Telemetry Approach

A nagging question confronts all designers of high-data-rate mud pulse
systems: if sirens are to be the signal generator of choice (because lowered
torques enable faster direction reversals), how does one overcome their
inherently weaker signal producing properties? The Joukowski formula
(“p = ρUc”) provides an exact solution from one-dimensional acoustics
stating that the pressure induced by an end-mounted piston is equal to the
product of fluid density ρ, impact velocity U and sound speed c. It closely
describes the acoustic performance of the positive pulser. And because the
positive pulser brings the mud column to an almost complete stop – in a way
that mud sirens cannot and never will – the Joukowski formula provides the
upper limit for siren performance, at least as presently implemented.
This understanding prompts us to look for alternatives, both down-
hole and uphole. We first address downhole physics near the source. We
have observed that up and downgoing waves are created at the siren, and
that reflection of the latter at the drillbit and their subsequent interac-
tion with “originally upgoing” waves can lead to “random” constructive
or destructive wave interference that depends on the information being
logged. This is certainly the case with presently used phase-shift-keying
that position-modulates “at random” the siren rotor.
However, if the rotor is turned at a constant frequency, random wave
cancellations are removed. The uncertainty posed by reflections of phase-
shifted signals, whose properties depend on nozzle size, wavelength, annu-
lar geometry, logging data, and so on, is eliminated in the following sense:
a sinusoidal position modulation always creates a similar sinusoidal upgoing
pressure wave without “kinks” and possible sign changes. In fact, depend-
ing on the location of the source within the MWD drill collar, the geome-
try of the bottomhole assembly, the transmission frequency and the mud
sound speed, the basic wave amplitude can be optimized or de-optimized
and controlled with relative ease (e.g., “listening from the surface” and
adjusting frequencies as necessary).
Information in the form of digital 0’s and 1’s can therefore be transmitted
by changes in frequency, that is, through “frequency-shift-keying” (FSK).
But, unlike conventional FSK, we select our high frequencies by using
only values that optimize wave amplitude by constructive interference.
Neighboring low-amplitude waves need not be obtained by complete

valve slowdown, as in conventional PSK. If, say, 60 Hz yields a locally high
FSK amplitude, it is possible (and, in fact, we will show) that 55 Hz may
yield very low amplitudes, thus fulfilling the basic premise behind FSK.
The closeness in frequencies implies that mechanical inertia is not a lim-
iting factor in high data rate because complete stoppage is unnecessary, so
that power, torque and electronic control problems are minimal and not a
concern.
In order to make constructive interference work, the time delay between
the downgoing waves and their reflections, with the newer upgoing waves,
must be minimized. This is accomplished by placing the siren as close to
the drillbit as possible, with the downhole turbine now positioned at the
top of the MWD drill collar. This orientation is disdained by conventional
designers because “the turbine may block the signal.” However, this con-
cern is unfounded and disproved by all field experiments. This is obvi-
ous in retrospect. The “see through area” for turbines is about 50% of the
cross-section. If signals can pass through siren rotor-stator combinations
with lower percentages, as they have time and again, they will have little
difficulty with turbines.
5.2 New Technology Elements

The above discussion introduces the physical ideas that guided our
research. An early prototype single-siren tool designed for downhole test-
ing is shown assembled and disassembled in Figures 5.1a and 5.1b. To fur-
ther refine our approach and understanding of the scientific issues, math
models and test facilities were developed to fine-tune engineering details
and to obtain “numbers” for actual design hardware and software. We now
summarize the technology.
Figure 5.1a Prototype single-siren tool (assembled).

Figure 5.1b Prototype single-siren tool (disassembled).
5.2.1 Downhole Source and Signal Optimization

As a focal point for discussion, consider the hypothetical MWD drill collar
shown in Figure 5.2a. Here, physical dimensions are fixed while siren fre-
quency and position are flexible. Up and downgoing signals (with antisym-
metric pressures about the source) will propagate away from the pulser,
reflect at the pipe-collar intersection, not to mention the interactions that
involve complicated wave transfer through the drillbit and in the borehole
annulus.
A six-segment acoustic waveguide math model was formulated and
solved, with the following flow elements: drillpipe (satisfying radiation
conditions), MWD drill collar, mud motor or other logging sub, bit box,
annulus about the drill collar, and finally, annulus about the drillpipe (also
satisfying radiation conditions). The “mud motor” in Figure 5.2a could
well represent a resistivity-at-bit sub. At locations with internal impedance
changes, continuity of pressure and mass was invoked. The siren source
was modeled as a point dipole using a displacement formulation so that
created pressures are antisymmetric. Numerical methods introduce arti-
ficial viscosities with unrealistic attenuation and also strong phase errors
to traveling waves. Thus, the coupled complex wave equations for all six
sections were solved analytically, that is, exactly in closed form, to provide
uncompromised results.
Calculated results were interesting. Let “p” denote the actual pressure
traveling up the drillpipe (after all wave interferences are accounted for)
and “Dp” represent the siren signal strength due to mud flow, i.e., the mea-
sured differential pressure across the siren. Our objective is p/Dp >> 1 (Dp
is separately optimized in hardware and wind tunnel or mud analysis).
Figure 5.2b displays p/Dp on the vertical axis, a function of transmission
Drillpipe Mud
5.0" OD, 2.8" ID
MWD Collar 22 ft
8.0" OD, 4.5" ID Pulser
Mud Motor
8.0" OD, 3.5" ID 25 ft
Rotor, 2.5" dia.
Bit-Box/Nozzles
3.5" ID 2 ft
12.25"
Figure 5.2a Example MWD collar used for siren frequency and source placement
optimization analysis.
0.960
0.879
0.798
0.717
0.636 15 20
0.555 10 2
5 4
0.475 0
0.394 6
0.313 8
1.0
0.232 10
0.151 0.8
0.070 12
0.6 1.0
0.4 0.8
0.2 0.6
2
0.4
4
0.2
6
8 20
Frequency (Hz) 15
10
10
12 5
0 Source Position (Ft)
Figure 5.2b Drillpipe p/Dp to 12 Hz.

1.000
0.910
0.820
0.730 Frequency (Hz)
0.640 0 10 20 30 40
0.560 0 5 10 15 20 Source 50
0.460 0 20
0 Position (Ft) 15
0.320 F (ms) 10 10 20
0.280
0.190 (Hz) 20 20 10 15
0.100 30 30 5 10
0.010 40
40 0 5
50 50 1.0
1.0 1.0 0
1.0
0.5
0.5 0.5
0.5
0.0
0.0 0.0 0 10 20 0.0
0 5 10 15 20 30 40 50
Source Position (Ft)
Figure 5.2c Drillpipe p/Dp to 50 Hz.
2.00
1.82 OPTIMUM
1.64 Frequency (Hz) 20
00
80 1
1.46 15
1.28
1.10 60 10
0.91 80 1
00 40 5
0.73 Source Position 0 60 2.0 20
0.55 0 20 4 0 0
0.30 (Ft) 2 0 1.5
0.19 15 1.0 2.0 2.0
0.00 10
5 0.5
0 1.5 1.5
0.0
1.0 1.0
2.0
20 0.5
1.5 5 0.5
1 0.0
1.0 10 0.0
0
10
0.5
20
5 80
0.0 Source Position (Ft) 60
0 20 40 60 80 1000 0
40
Frequency (Hz)
20
0 0
Figure 5.2d Drillpipe p/Dp to 100 Hz.
frequency (left axis) and source position from the bottom (right axis). For
low frequencies less than 2 Hz, red zones indicate that optimal wave ampli-
tudes are always found whatever the source location. But at the 12 Hz used
in present siren designs, source positioning is crucial: the wrong location
can mean poor signal generation and, as can be seen, even “good locations”
can be bad. These calculations are repeated for upper limits of 50 Hz and
100 Hz in Figures 5.2c and 5.2d. In these diagrams, red means optimal
frequency-position pairs for hardware design and strong signal strength
entering the drillpipe.
That present drilling telemetry channels support much higher data rates
than siren operations now suggest, e.g., carrier waves exceeding 50 Hz, is
confirmed by independent research at Presco [4]. In our designs, we select
the frequencies and siren positions, or for sirens-in-tandem, in such a

way that high amplitudes are achieved naturally without power or erosion
penalties.
5.2.2 Surface Signal Processing and Noise Removal

Downhole signal optimization, of course, has limits. To complement efforts
at the source, surface signal processing and noise removal algorithms must
be developed that are robust. Our approach is based on rigorous mathe-
matics from first principles. The classic wave equation states that all “solu-
tions (measured at some point ‘P’ along the standpipe) are superpositions
of upgoing ‘f ’ and downgoing ‘g’ waves” – the objective is an equation or
expression for f alone. Two methods were devised. In the first, a differential
equation for f is constructed. It is then finite-differenced in space and time
as if a numerical solution were sought. However, it is not. The Dz and Dt
in the discretized result are re-interpreted as sensor spacing (in a multiple
transducer array) and time delay (for sampling), whose pressure parame-
ters are easily stored in surface data acquisition systems. The solution for
the derivative of the signal was given in the patent of Chin [5]. At the time,
it was erroneously believed that telemetered data could be retrieved from
spatial derivatives, but this proved difficult. In recent work, the method
was corrected by adding a robust integrator that handles abrupt waveform
changes. The successful recovery of “red” results to match “black” inputs,
using the seemingly unrelated green and blue (single transducer) inputs,
1
Pressure (psi)
0
100 200 300 400 500
–1 Time
–2
–3
Figure 5.3a Three step pulse recovery in noisy environment (pressure, vertical; time,
horizontal).
Pressure (psi) 1
0
100 200 300 400 500
–1 Time
–2
–3
Figure 5.3b Three step pulse recovery (very noisy environment).
is shown in Figures 5.3a and 5.3b. Mudpump generated noise can be almost
completely removed. Experimental validations are given later.
A second method for f, leading to a time-delay difference equation, was
also designed and does not bear the (small Dz and Dt) limitations of deriv-
ative models. Updated results for both models in which the effects of trans-
ducer separation and sampling time are evaluated under typical surface
standpipe constraints are presented in detail later.
5.2.3 Pressure, Torque and Erosion Computer Modeling

The mud siren, conceptualized in Figure 5.4a, is installed in its own MWD
drill collar and consists of two parts, a stationary stator and a rotor that
rotates relative to the stator. The rotor periodically blocks the oncoming
mud flow as the siren valve opens and closes. Bi-directional pressure pulses
are created during rotation. At the minimum, the cross-sectional flow
area is half-blocked by the open siren; at worst, the drill collar is almost
completely blocked, leaving a narrow gap (necessary for water hammer
pressure signal creation) between stator and rotor faces for fluid passage.
This implies high erosion by the sand-laden mud and careful aerodynamic
tailoring is needed. Because there are at least a dozen geometric design
parameters, testing is expensive and time-consuming. Thus, the compu-
tational method in Chin [6], which solves the three-dimensional Laplace
equation for the velocity potential in detail, is used to search for optimal
designs. Computed results, displayed for various degrees of valve closure,
are shown in Figures 5.4b and 5.4c. Other results include “resistive torque
vs angle of closure” important to the design of fast-action rotors. Results
are validated and refined by “short wind tunnel” analyses described later.
Mud Signal Mud Signal
Untapered Rotor
Upstream Stator
Straight Stator Tapered Rotor
Figure 5.4a Early 1980s “stable closed’ siren (left) and improved 1990s “stable-opened”
downstream rotor design.
While apparently simple in design, the rotor has unanticipated flow

effects. The early upstream rotor design produces numerous operational
hazards, the least of them being stoppage of data transmission. When rock
debris or sudden jarring occurs, the rotor is known to stop at a closed azi-
muthal position that completely blocks mud flow. This results in severe
tool erosion, extremely high pressures that affect well control, as well as
surface safety issues associated with high pressure buildup at the mud-
pump. Early solutions addressed the symptoms and not the cause, e.g.,
mechanical springs that unload the locked rotor, strong permanent mag-
nets that bias special steel assemblies to open positions (thus compromis-
ing direction and inclination measurements), and so on. It can be shown
that “stable closed” tendencies are a natural aerodynamic consequence
of upstream rotor configurations – the rotors tend to close even in clean
water. Numerous unsuccessful tests addressing this problem were per-
formed in the 1970s: Operational failures associated with jamming valves
were catastrophic.
Chin and Trevino [7] solved the problem by placing the rotor down-
stream as indicated in Figure 5.4a. The rotor, now “stable open,” is aug-
mented with special tapered sides (in fact, recent “self-spinning” sirens
drawing energy from the oncoming flow are described in [2]). Torques
required to turn, stop or speed up the rotor are much lower than those
Mud Siren Streamline Plot Mud Siren Streamline Plot
Figure 5.4b Streamline traces for erosion analysis.
associated with upstream rotors. From a telemetry standpoint, this means

faster position modulation requiring less torque and power, or much
higher data rate. Downhole signal enhancement and surface noise removal
are equally important, as noted earlier. In our research, all are addressed
and fine-tuned to work in concert to provide a fully optimized system.
Velocity Field Velocity Field
Figure 5.4c Velocities for erosion and pressure analysis.
5.2.4 Wind Tunnel Analysis: Studying New Approaches

While computer models are useful screening tools, they alone are not
enough. Gridding effects mask the finest flow details that can be uncovered
only through actual testing. The use of wind tunnels in modeling down-
hole mud flow was first proposed and used by Chin during his tenure with
Schlumberger. Technical details and justification were later disclosed in
Gavignet, Bradbury and Quetier [8] who used the method to study flows
beneath drill bits nozzles. This counter-intuitive (but correct) approach to
modeling drilling muds provides a strategically important alternative to
traditional testing and reduces the time and cost of developing new MWD
systems. Several oil service company wind tunnel systems have since been
designed and built by Chin.
The CNPC MWD wind tunnel test facility in Beijing consists of two
components, a “short flow loop” where principal flow properties and tool
characteristics are measured, and a “long loop” (driven by the flow in the
short wind tunnel) designed for telemetry concept testing, signal process-
ing and noise removal algorithm evaluation. Field testing procedures and
software algorithms for tool properties and surface processing are first
developed and tested in wind tunnel applications and then moved effort-
lessly to the field for evaluation in real mud flows. This provides higher
efficiency than with “mud loop only” approaches.
Our “short wind tunnel,” housed at an off-campus site, is shown in
Figure 5.5a. This laboratory location was selected because loud, low-
frequency signals are not conducive to office work flow. The created signals
are as loud as motorcycle noise (typically exceeding 100 db) and require
hearing protection for long duration tests. More remarkable is the fact
that internal pipe pressures are “1,000” times louder than the waves that
escape – another 1,000 arises from the ratio of mud to air density. The com-
bined ratio implies that careful and precise acoustic signal measurement
is required to accurately extrapolate those to mud conditions. Similarly,
torques acting on sirens are 1,000 times lower. In fact, air-to-mud torque
scaling is simply proportional to the dynamic head “ρU2” ratio, where U
is the oncoming speed. Thus, wind tunnel tests can be run at lower speeds
with inexpensive blowers, provided a quadratic correction is applied for
downhole flow extrapolation. The turbine, similarly designed and tested, is
not discussed here. Wind tunnel testing allows effects like high pitch angle,
wide annular clearances, strong turbulence and unsteady transients to be
modeled accurately – effects common to MWD tools that would challenge
the best computational fluid dynamics models. Details appear in [2].
In Figure 5.5a, a powerful water-cooled blower (blue) with its own power
supply pumps more or less constant flow rate air regardless of siren rotor
blockage. A sensitive flow meter records average flow rate. Flow straighten-
ers ensure uniform flow into the siren and to remove downstream swirl for
accurate differential pressure measurement. The siren test section deserves
special comment. The motion of the rotor is governed by its own electrical
controller and is able to affect position-modulated motions as required for
telemetry testing.
Blower with Observation

Muffler and Windows to
Water Coolant Outdoor
Flowloop for
Multiple
Independent
Transducer and
Power Supply
Wave
Flow Meter Interference
Flow Testing
Straighteners Data acquisition
For Flow Rate,
Siren Test Torque, Multiple
Section and and Differential
Electric Pressure
Controller Transducers
Torque Meter Siren p
Drive Motor Transducer
Total Length, Connection to

About 60-70 Outside Long
Feet Wind tunnel
Figure 5.5a Short “hydraulic” wind tunnel system.
Siren motion, again, is driven electrically as opposed to hydraulically;

azimuthal position, torque and Dp signal strength, i.e., the differential
pressure between upstream and downstream sides of the siren, are mea-
sured and recorded simultaneously. This data is important to the design
of control and feedback loops for actual modulation software. At the bot-
tom of Figure 5.5a, a black PVC tube is seen turning to the right into the
wall. This emerges outside of the test laboratory, as shown in Figure 5.5b,
into a long flow section more than 1,000 feet in length. Because the waves
are acoustically “long,” they reflect minimally at bends, even ninety-degree
bends. The long wind tunnel wraps itself about a central facility several
times before exhausting into open air. This end boundary condition is not,
of course, correct in practice; we therefore minimized its effect by reducing
signal amplitude so that end reflections are not likely to compromise data
quality. A special closed-end wind tunnel providing through-flow was sub-
sequently designed and tested to allow strong reflections “uphole” – this
required modifications only to the short driver section.
Also shown in Figure 5.5b are “a single transducer” close to the siren test
shed (middle left) and a three-level “multiple transducer array” (bottom right).
The former monitors the signal that actually leaves the MWD drill collar, as
it depends on constructive or destructive wave interference, while the latter
provides data for echo cancellation and noise removal algorithm evaluation.
Figure 5.5b Very long “acoustic” wind tunnel.
For the simplest schemes, only two transducers are required; three allow
redundancies important in the event of data loss or corruption. Additional
(recorded) noise associated with real rigsite effects is introduced into the
wind tunnel using low frequency speakers or woofers.
Numerous siren concepts and shapes were evaluated. Several of the
sirens shown in this paper are not practical but were purposely designed
to be impractical; a broad range of data was accumulated to enhance our
fundamental understanding of rotating flows as they affect signal, torque
and erosion.
In our work, we re-evaluated conventional four-lobe siren designs and
developed methods that incrementally improve signal strength and reduce
torque. Results reinforced the notion that the technology has reached its
performance limits. Radically different methods for signal enhancement
and minimization of resistive torque were needed.
As noted earlier, constructive wave interference provides “free” signal
amplitude without erosion or power penalties. This is cleverly implemented
Figure 5.5c A pair of ganged or tandem mud sirens.
Figure 5.5d Some sirens tested in wind tunnel.
in two ways. First, FSK with alternating high-low amplitudes is used. High
amplitudes are achieved by determining optimal frequencies from three-
dimensional color plots such as those in Figures 5.2b, c, d. Design param-
eters include sound speed, source position and frequency, MWD collar
design, and whenever possible, drill-pipe inner diameter and mud density.
This information is used in the waveguide model of Figure 5.2a and also in
a model for non-Newtonian attenuation applicable over the length of the
drillpipe. Low amplitudes need not be achieved by bringing the rotor to a
complete stop. If a high-amplitude is associated with 60 Hz, then a useful
low-amplitude candidate can be found at 55 Hz, as suggested by Figures
5.2b, c, d. Thus, FSK can be efficiently achieved while minimizing the
effects of mechanical inertia. Rotor torque reduction, while an objective in
Figure 5.5e Evaluation of hub convergence effects on signal strength and torque.
wind tunnel analysis, is useful but need not be the main design driver in
our approach since the rotor is never brought to a complete stop.
To make constructive wave interference work, the siren is located as
close to the most significant bottom reflector, normally the drillbit, as pos-
sible (intervening waveguides, e.g., mud motors, resistivity-at-bit subs, and
so on, do support wave transmission). Thus, the siren is placed beneath
the turbine in the MWD collar, in contrast to existing designs. Tests con-
firm that long waves pass effortlessly through turbines without reflection.
Detailed waveguide analyses suggest that signal gains of 1.5-2.0 are doable.
PSK methods, again, are undesirable because they cause wave cancella-
tions and ghost signals that hinder signal processing.
Additional signal enhancement is possible using constructive interference
of a different nature, specifically multiple sirens arranged in series or in tan-
dem. If the distance between sirens is small and siren apertures are properly
phased, signals will be additive. This idea was first proposed by the last author
Figure 5.5f Flow straighteners (PVC tubing) for upstream and downstream use.
(refer to Section 5.2) and a design from that publication is reproduced in

Figure 5.5c. This design, incidentally, is not CNPC’s preferred embodiment.
Two sirens, for instance, mean twice the signal. If the amplification
afforded in the previous paragraphs provides a modest signal gain of 1.5,
or 50%, the net would be a three-fold signal increase, more than enough
to overcome attenuation at the higher frequencies used. Performance is
determined by the single transducer at the middle left of Figure 5.5b, which
measures the signal leaving the MWD collar. The extent to which construc-
tive wave interference works is found by comparison with the differential
pressure Dp taken across the siren (e.g., see Figure 5.5h below). This Dp
depends on siren geometry, flow rate and rotation only: it is independent
of reflections since waves pass through without interaction.
Note that Figures 5.2b, c, d suggest that frequencies in the 50-60 Hz
range are not unrealistic, a conclusion independently reached at Presco
[4]. This use of higher frequencies is also supported by test results from
actual flow loop tests with real muds. We stress here that attenuation mea-
surements are subtle since the effects of acoustic nodes and antinodes
(which depend on frequency and flow loop boundary conditions) must
be properly accounted for. Almost all existing papers on signal attenuation
fail to recognize this problem, let alone correct for it.
Figure 5.5g Flow meter.
Figure 5.5h Siren test section with differential transducers.
Our systems approach to high-data-rate design requires an equal focus

on surface systems. As implied earlier, signal strength enhancement must
be accompanied by using the most sensitive piezoelectric transducers and
robust multiple-transducer echo cancellation methods. Figure 5.5n shows
a transducer array located far from the test shed to evaluate test algorithms.
Noise can be introduced by playing back actual field recordings. We have
Figure 5.5i Real-time data acquisition and control system.
Figure 5.5j Torque, position and rpm counter.
Figure 5.5k Short wind tunnel, “bird’s eye” view.

Figure 5.5l Test shed window overlooking long wind tunnel.
found, to our amusement, that the large firecrackers used at Chinese wed-
dings, e.g., see Figure 5.5o, provide a useful source of low-frequency, plane-
wave noise when all else is unavailable.
Conventional siren designs are built with four lobes cut along radial
lines. Rotating sirens with additional lobes would surely increase fre-
quency or data rate, but large lobe numbers are associated with much lower
Dp signals. For this reason, they are not used in designs to the authors’
knowledge. Because constructive interference now enhances our arsenal
of tools against attenuation, we have been able to reassess the use of higher
lobe numbers. Downhole and uphole telemetry concepts are easily tested
in our wind tunnels.
Wind tunnel usage enables a scale of knowledge accumulation, together
with cost, time and labor efficiencies, not previously possible. Numerous
parameters can be evaluated, first by computational models and then by
testing in air. Design parameters include lobe number, stator and rotor
Figure 5.5m Piezoelectric transducer closest to siren for constructive interference and
harmonic generation study.
thicknesses, stator-rotor gap, rotor clearance with the collar housing, rotor
taper angles, and so on.
Tests are not limited to signal strength. Torque is important, as is the
ability to pass lost circulation material (LCM) – this is assessed by intro-
ducing debris at the upstream end of the short wind tunnel and observing
Figure 5.5n Distant multiple transducer array setup.

the resulting movement. A closed-loop attachment (not shown) prevents

debris from flying across the laboratory floor. Erosion tendencies are
determined by noting the convergence effects of threads glued to solid sur-
faces – rapid streamline convergence implies high erosion, e.g., see Figures
5.4b, c.
Two new parameters were included in our test matrix. The bottom left
of Figure 5.5d shows a “curved siren” with swept-back blades. Research
was performed to determine the degree of harmonic generation associate
with constant speed rotations. Since the sound generation process is non-
linear, a rotation rate of ω will not only produce pressure signals with ω,
but those with frequencies 2ω, 3ω, 4ω and so on. Higher harmonics are
associated with acoustic inefficiencies we wish to eliminate, not to men-
tion surface signal processing problems. Figure 5.5e also shows conical
flow devices that guide inlet flow into the siren. Their effects on torque
and signal were studied. In Figures 5.5d-5.5o, we provide photographs of
actual sirens tested and devices used. Many are self-explanatory and are
given without explanation. Improvements to both short and long wind
tunnels have since been made in light of our experiences over the past
two years.
Figure 5.5o Fireworks for low frequency noise generation, when all else is unavailable.
5.2.5 Example Test Results

Here we highlight interesting test results. The first pertains to signal strength
as a function of rotation rate with flow speed fixed. Early Schlumberger
papers claim that Dp’s obtained at high frequencies are independent of fre-
quency, i.e., the siren functions as an orifice. We believed otherwise. As
the rotor turns, it brings oncoming mud to a halt, whatever the frequency.
However, the water hammer signal must weaken as rotation rate increases
because less time is available for fluid stoppage and rebound. The expected
monotonic decrease of Dp with increasing frequency is seen, for instance,
in Figure 5.6a, where we typically test up to 60 Hz as suggested by Figures
2b, c, d. The low Dp’s associated with existing “siren alone” approaches rein-
forced our efforts to seek more innovative signal enhancement methods.
In Figure 5.6b, pressure data from the near transducer in Figure 5.5m
appears at the left, while data from two far transducers in Figure 5.5n are
shown at the center and right. At the left, the pure sinusoid shows that
high-order harmonics have been completely eliminated by the siren design.
The two right figures, which contain additive noise, are almost identical.
Multiple transducer signal processing in Figure 5.6c shows how the red
signal is successfully extracted from the blue and green single-transducer
results to match the black upgoing waveform.
Complementary results obtained at 45 Hz are shown in Figures 5.6d, e.
The left diagram of Figure 5.6d is clearly not sinusoidal and provides
A B C D E
1
2 12.69 208 ∆P
3 13.03 303 0.04
4 18.41 398
0.03
5 27.68 583
6 36.88 798 0.02
7 50.01 993
0.01
8 55.32 1200
HZ
9 71.4 1406 0
10 83.01 1580 10 15 20 30 40 50 60 70 80 90
11 89.5 1767
12 12.47 197 ∆P
13 16.82 294 0.02
14 25.05 390 0.015
15 27.67 609
16 36.88 803 0.01
17 49.88 1004
0.005
18 55.32 1216 HZ
19 68.95 1391 0
20 78.55 1606 10 15 20 30 40 50 60 70 80 90
21 92.22 1811
Figure 5.6a Siren Dp vs ω at with flow rate fixed.

Figure 5.6b Low-frequency (10 Hz) long wind tunnel data.
Delta-P Signal Deconvolution
2
Series
1
Pressure (psi)
1
2
0 3
100 200 300 400 500 4
−1 Time
−2
−3
Figure 5.6c Low-frequency (10 Hz) signal recovery.
evidence of nonlinear harmonics. Their magnitudes are measured from

frequency domain analysis and efforts are made to determine their phys-
ical origin. Again, we have successfully extracted the MWD signal from a
noisy environment. Our work showed that transducer spacings of 10% of
a wavelength sufficed for signal extraction. Presently, not all noise sources
are included in our modeling efforts. Vibration and other sound mecha-
nisms will be included in future work.
Our experiences with constructive wave interference “at the drillbit”
are also worth noting. In Chin [9], where the use of downhole construc-
tive interference for signal enhancement was first suggested, the published
analytical model mistakenly assumed the bit as a solid reflector. In fact,
it is known that MWD signals are detectable in the annulus, where their
absence is used as an indicator of gas influx. The six-segment waveguide
2
Series
1
Pressure (psi)
1
2
0 3
100 200 300 400 500 4
−1 Time
−2
−3
Figure 5.6d High-frequency (45 Hz) long wind tunnel data.
2
Series
1
Pressure (psi)
1
2
0 3
100 200 300 400 500 4
−1 Time
−2
−3
Figure 5.6e High-frequency (45 Hz) signal recovery.
model now used to study typical MWD collars, e.g., see Figure 5.2a, is
more general and does not assume any particular reflection mechanism on
an a priori basis. Detailed calculations show that, more often than not, the
drillbit (because of its nozzles) actually acts as an open reflector – attesting
to the dangers of “common sense” and visual inspection. This model cre-
ates plots similar to Figures 5.2b, c, d. The wave characteristics of siren and
positive pulsers from present MWD vendors are consistent with those in
Figures 5.2b.
5.3 Directional Wave Filtering

In high-data-rate MWD mud pulse telemetry, surface reflections at the
mud pump and desurger can “confuse” standpipe transducers as wave
cancellations and reinforcements contaminate intended upgoing signals.
MWD pressure waves reflect without sign and shape change at pump pis-
tons; also, reflections at desurgers can lead to synchronization loss since
elastic membranes may unrecognizably distort signals (e.g., square waves
can reflect exponential-like form). In addition, mud pump noise traveling
opposite to the signal, typically many times the upcoming signal strength,
introduces further reductions to signal-to-noise ratio. Thus, the use of
robust “directional filtering” with multiple transducers, ideally employ-
ing minimal waveform information with the fewest practical constraints,
is motivated. Here applications for two schemes from [2] are evaluated
under severe operational conditions using synthetic signal and noise data.
In practice, the algorithms would be used with frequency, wavelet, white
noise and other filters.
5.3.1 Background Remarks

In MWD mud pulse telemetry, two practical constraints exist: (1) sound
speeds vary between 3,000 and 5,000 ft/sec, where the latter is achieved
with water or brine, and (2) transducer positions are restricted to a
thirty-feet standpipe on the rig floor where at most two pressure taps may
be drilled due to safety concerns. In high speed telemetry, base frequen-
cies used can be as high as 10-50 Hz and even up to 100 Hz; transducer
separations should represent small percentages of a wavelength but, at the
same time, cannot be “too close” since such measurements will be error-
prone. Similar considerations apply to time discretization. Microsecond
sampling, for instance, is out of the question but unnecessary; however,
coarser durations of 1 ms up to 10 ms are permissible for the frequencies
under consideration.
Consider, for example, a sound speed of 3,000 ft/sec and a 0.003 sec sam-
pling. A transducer separation that permits this would be 9 ft. If the time
duration is increased to 0.006 sec, then the separation increases to 18 ft.
If the sound speed is 5,000 ft/sec, the respective distances are 15 and 30
ft. All of these distances are doable on existing rig installations. Thus, we
ask, “Is it possible to design a good MWD echo cancellation and mud
pump noise removal algorithm that works under the restrictive environ-
ment described and performs quickly with a minimum of computation?”
Methods 4-3 and 4-4 for “directional filtering,” described in detail in [2]
are based on analytical wave equation properties. The former uses a differ-
ence equation delay formulation whereas the latter is hosted by a differential
equation for the upgoing MWD signal. The algorithms are completely dif-
ferent, but both fully eliminate propagating waves that travel in directions
opposite to the upgoing MWD signal. These can contain signal reflections
at positive displacement and/or centrifugal mud pumps, distorted reflec-
tions at the desurger, together with large-amplitude mud pump noise with
or without frequencies near those in the upgoing signal.
Information on downgoing waveform shape or amplitude is not neces-
sary. Application of these models is subject to requirements imposed by the
Nyquist-Shannon Theorem. In addition, the methods can be augmented
with damping to lessen earlier time effects and to improve computational
stability; processing times are almost instantaneous for the computations
with Intel i5 class chipsets. In practice, such “directional filters” are used
together with frequency, wavelet, white noise and other filters for robust
signal processor design. In this paper, the two methods are evaluated in
detail using synthetic datasets to examine their ability to remove all noise
traveling opposite to the upcoming signal. Both algorithms appear to work
well with large transducer separations and coarse time sampling – the
use of the two schemes together with multiple pressure transducer data
would offer redundancy that is likely to further minimize common syn-
chronization loss at the surface. The methods are also computationally effi-
cient, requiring few “multiplies and divides” compared to a typical FFT
calculation.
5.3.2 Theory
The ideas behind Method 4-3 are developed from wave equation properties.
The complete pressure disturbance is taken as the sum of two waves travel-
ing in opposite directions. Let “f ” denote the incident wave traveling from
downhole. Then, “g” will denote reflections of any type at the mudpump
(positive displacement pistons and centrifugal pumps are both allowed),
together with reflections at the desurger with any type of shape distortion
permitted, plus the mudpump noise itself. In general, we can write
p(x, t) = f(t – x/c) + g(t + x/c) (5.1)
where c is the measured speed of sound at the surface. Two transducers

are assumed to be placed along the standpipe. Note that the impedance
mismatch between standpipe and rubber rotary hose does introduce
some noise; however, since a portion of this effect propagates downward,
it is part of the “g” which will be filtered out in its entirety. The rotary
hose does not introduce any problem with our approach. Now let xa and
xb denote any two transducer locations on the stand-pipe. At location “b”
we have
p(xb, t) = f(t – xb/c) + g(t + xb/c) (5.2)
If we define τ = (xb – xa)/c > 0, it follows that
p(xb, t − τ) = f{t + (xa – 2xb)/c} + g(t + xa/c) (5.3)
But at location “a” we have
p(xa, t) = f(t – xa/c) + g(t + xa/c) (5.4)
Subtraction yields
p(xa, t) − p(xb, t − τ) = f(t – xa/c) − f{t + (xa – 2xb)/c} (5.5)
Without loss of generality, we set xa = 0 and take xb as the positive trans-

ducer separation distance
f(t) – f(t – 2xb/c) = p(0, t) – p(xb, t – xb/c) (5.6)
or
f(t) – f(t – 2τ) = p(0, t) – p(xb, t – τ) (5.7)
The right-side involves subtraction of two measured transducer pres-

sure values, with one value delayed by the transducer time delay τ, while
the left side involves a subtraction of two unknown (to-be-determined
upgoing) pressures, one with twice the time delay or 2τ. The deconvolution
solves for f(t) given the pressure values on the right and is solved by our
2XDCR*.FOR code.
Method 4-3 is extremely powerful because it eliminates any and all
functions g(t + x/c), that is, all waves traveling in a direction opposite to the
upgoing wave. Thus, g(t + x/c) may apply to mudpump noise, reflections of
the upgoing signal at the mudpump, and reflections of the upgoing signal
at a desurger, regardless of distortion or phase delay, reflections from the
rotary hose connections, and so on. The functional form of the downgoing
waves need not be known and can be arbitrary. This is not to say that all
downward moving noise sources are removed. For example, fluid turbu-
lence noise traveling downward with the drilling fluid is not acoustic noise;
it will not be removed and may degrade performance. Additional noise
sources and filters would be used together with Method 4-3. The order
in which filters are applied will affect the outcome of any signal process-
ing, and it is this uncertainty that provides the greatest challenge in sig-
nal processor design. The model in Equation 5.7 can be solved exactly in
closed analytical form and is implemented in our 2XDCR*FOR software
series. We note that this method is similar to the two-transducer delay-line
approach of Foster and Patton [10] who first solved a model analogous to
Equation 5.7 for MWD applications using approximate frequency domain
analysis. Other ad hoc derivations have since appeared, which are solved
by undisclosed methods, but to the authors’ knowledge, none have proven
successful and there have been no other publicly available validations of
delay approaches except that presented here.
Method 4-4 is similarly developed from wave equation properties but
follows a strategy different from the delay equation approach. Our deriva-
tion at first follows from Chin [5], which gave solutions for the derivative
of the signal only. The present method, which includes a robust integrator
to handle sharp pressure pulses, substantially changes the earlier work. In
Method 4-3, we used time delayed signals for which there was no restric-
tion on time delay size. For Method 4-4, we invoke time and space deriv-
atives; thus sampling times should at least be small compared to a period
and transducer separations should be small compared to a wavelength.
As in Method 4-3, the representation of pressure as an undamped upgo-
ing and downgoing wave is still very general (attenuation between close
transducer locations is minimal), and all waves in the downward direction
are removed with no information required about the mudpump, the desur-
ger or the rotary hose. Note that “c” is the mud sound speed at the surface
and should be measured separately. In our derivation, expressions for time
and space derivatives of p(x, t) are formed, from which the downgoing
wave “g” is explicitly eliminated, leaving the desired upgoing “f.” The steps
shown are straightforward and need not be explained.
p(x, t) = f(t – x/c) + g(t + x/c) (5.8)
pt = f ' + g' (5.11)
px = −c−1 f ' + c−1g' (5.12)
cpx = −f ' + g' (5.13)
pt − cpx = 2f ' (5.14)
f ' = ½ (pt − cpx) (5.15)
Equation 5.15 for “f,” which is completely independent of the down-

going wave “g,” however, applies to the time derivative of the upgoing
signal f(t – x/c) and not “f ” itself. Thus, if a square wave were traveling
uphole, the derivative of the signal would consist of two noisy spikes hav-
ing opposite signs. This function must be integrated in order to recover
the original square wave, and at the time the original patent was awarded,
a robust integration method was not available. The required integration is
not discussed in the patent, where it was simply noted that both original
and derivative signals in principle contain the same information. Here, a
special integration algorithm is given to augment the numerical represen-
tation in Equation 5.15.
At first glance, the two-transducer delay approach in Method 4-3 seems
to be more powerful because it does not require time integration, and since it
does not involve derivatives, there are no formal requirements for sampling
times to be small and transducer separations to be close. However, in any
practical high-data-rate application, the latter will be the case anyway, e.g.,
slow sampling rates will not capture detailed data. Thus, Method 4-4 is no
more restrictive than Method 4-3. However, the present method is powerful
in its own right because the presence of the ∂p/∂x derivative implies that
one can approximate it by more than two (transducer) values of pressure at
different positions using higher-order finite difference formulas – in oper-
ational terms, one can employ multiple transducers and transducer arrays
to achieve higher accuracy. Similarly, the presence of ∂p/∂t means that one
can utilize more than two time levels of pressure in processing to achieve
high accuracy. The required processing in space and time is inferred from
the use of finite difference formulas in approximating the derivatives shown
and numerous such computational molecules are available in the numeri-
cal analysis literature. The illustrative calculations used below assume two
transducers and pressures stored at two time levels. Higher order process-
ing does not add substantially to computational or storage requirements.
5.3.3 Calculations
Method 4-3 uses a difference equation time delay equation with damping
while Method 4-4 employs a differential equation approach. Both methods
(1) require at least two (piezoelectric) transducers mounted on the surface
standpipe, (2) operate with minimal information on waveform and noise
properties, and (3) robustly function in the less-than-ideal environment
prevalent on typical drilling rigs. The surface speed of sound is required
and is easily obtained by clocking the transit time of a sharp pulse traveling
between the two transducers. The results below are reported together with
software references cited in [2]. For all the calculations given, we duplicate
relevant Fortran source code describing the assumed signals and noise in
order to clearly document the test methodology employed.
Method 4-3, Difference Equation (Software reference, 2XDCR07D.FOR).

C CASE A. FSK MODULATION WITH SINUSOIDS
IF(T.LT.0.20) THEN
AMP = 0.25
FRQ = 20.
SIGNAL = AMP*SIN(2.*PI*FRQ*T)
ENDIF
IF(T.GE.0.20.AND.T.LT.0.35) THEN
AMP = 0.5
FRQ = 40.
ENDIF
IF(T.GE.0.35.AND.T.LT.0.50) THEN
AMP = 0.25 FRQ = 20.
ENDIF
IF(T.GE.0.50) THEN AMP = 0.5
FRQ = 40.
ENDIF
C CASE B. NARROW PULSE WIDTH

A = 10.0
R = 100.0
IF(T.LT.0.1) THEN
SIGNAL =A*(TANH(R*(T-0.100))-TANH(R*(T-0.101)))/2.
ENDIF
IF(T.GE.0.1) THEN
SIGNAL = 0.
ENDIF
C FUNCTION XNOISE(T)
C
Mud pump noise function may also include reflected MWD
signal, but
C
it is not necessary to add the wave reflection
to the total noise
C to demonstrate directional filtering.
C FRQPMP = Hertz freq of pump noise, propagates
downward.
PI = 3.14159
FRQPMP = 5.
AMP = 1.5
XNOISE = AMP*SIN(2.*PI*FRQPMP*T)
RETURN
END
Signal Deconvolution
2
Series
1
Pressure (psi)
1
2
0 3
0.1 0.2 0.3 0.4 4
−1 Time (sec)
−2
−3
Figure 5.7a Method 4-3, A-002 (8 feet).

2
Series
1
Pressure (psi)
1
2
0 3
0.1 0.2 0.3 0.4 0.5 0.6 4
−1 Time (sec)
−2
−3
Figure 5.7b Method 4-3, A-003 (12 feet).
2
Series
1
Pressure (psi)
1
2
0 3
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 4
−1 Time (sec)
−2
−3
Figure 5.7c Method 4-3, A-004 (16 feet).
In our Series A calculations, sinusoidal FSK (frequency-shift-keyed)

signals are assumed; the number, e.g., “002,” refers to millisecond
sampling, in this case, representing a coarse 2 ms. The sampling time
can be (crudely) taken as the travel time between two transducers.
2
Series
1
Pressure (psi)
1
2
0 3
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 4
−1 Time (sec)
−2
−3
Figure 5.7d Method 4-3, A-005 (20 feet).
2
Series
1
Pressure (psi)
1
2
0 3
0.05 0.10 0.15 0.20 4
−1 Time (sec)
−2
−3
Figure 5.7e Method 4-3, B-002 (8 feet).
The parenthesized figure number provides a “physical feeling” for trans-

ducer separation. For illustrative purposes, assuming a 4,000 ft/sec mud,
we find 4,000 ft/sec × 0.002 sec or 8 ft. The noise function is given above.
In both Methods 4-3 and 4-4, the derivations do not assume sinusoids
2
Series
1
Pressure (psi)
1
2
0 3
0.05 0.10 0.15 0.20 4
−1 Time (sec)
−2
−3
Figure 5.7f Method 4-3, B-004 (16 feet).
2
Series
1
Pressure (psi)
1
2
0 3
0.1 0.2 0.3 0.4 4
−1 Time (sec)
−2
−3
Figure 5.7g Method 4-3, A-002 (8 feet).
and no requirements for sinusoidal waveforms exist – all telemetry

schemes are supported. Black denotes the original upgoing MWD signal,
red the downgoing pressure wave, green the superposition of the two
at a transducer location (the MWD signals are almost unrecognizable),
2
Series
1
Pressure (psi)
1
2
0 3
0.1 0.2 0.3 0.4 0.5 0.6 4
−1 Time (sec)
−2
−3
Figure 5.7h Method 4-3, A-003 (12 feet).
2
Series
1
Pressure (psi)
1
2
0 3
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 4
−1 Time (sec)
−2
−3
Figure 5.7i Method 4-3, A-004 (16 feet).
while blue shows a successfully extracted signal (using only green data)
which is almost identical to the black. Note that high-data-rate telemetry
requires closer transducer separations along with finer time samples than
in more conventional applications. As shown in our source code, the FSK
2
Series
1
Pressure (psi)
1
2
0 3
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 4
−1 Time (sec)
−2
−3
Figure 5.7j Method 4-3, A-005 (20 feet).
2
Series
1
Pressure (psi)
1
2
0 3
0.05 0.10 0.15 0.20 0.25 0.30 4
−1 Time (sec)
−2
−3
Figure 5.7k Method 4-3, B-003 (12 feet).

2
Series
1
Pressure (psi)
1
2
0 3
0.1 0.2 0.3 0.4 0.5 0.6 4
−1 Time (sec)
−2
−3
Figure 5.7l Method 4-3, B-006 (24 feet).
2
Series
1
Pressure (psi)
1
2
0 3
0.1 0.2 0.3 0.4 4
−1 Time (sec)
−2
−3
Figure 5.7m Method 4-3, A-002 (8 feet).

2
Series
1 1
Pressure (psi)
2
3
0
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 4
−1 Time (sec)
−2
−3
Figure 5.7n Method 4-3, A-004 (16 feet).
2
Series
1
Pressure (psi)
1
2
0 3
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 1.1 1.2 4
−1 Time (sec)
−2
−3
Figure 5.7o Method 4-3, A-006 (24 feet).

2
Series
Pressure (psi)
1 1
2
0 3
0.1 0.2 0.3 0.4 0.5 0.6 4
−1 Time (sec)
−2
−3
Figure 5.7p Method 4-3, B-006 (24 feet).
2
Series
1 1
Pressure (psi)
2
0 3
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 4
−1 Time (sec)
−2
−3
Figure 5.7q Method 4-3, B-008 (32 feet).

2
Series
1
Pressure (psi)
1
2
0 3
100 200 300 400 500 4
−1 Time (sec)
−2
−3
Figure 5.8a Method 4-4, Run C-1.
2
Series
1
Pressure (psi)
1
2
0 3
100 200 300 400 500 4
−1 Time (sec)
−2
−3
Figure 5.8b Method 4-4, Run C-2.

2
Series
1
Pressure (psi)
1
2
0 3
100 200 300 400 500 4
−1 Time (sec)
−2
−3
Figure 5.8c Method 4-4, Run C-3.
2
Series
1
Pressure (psi)
1
2
0 3
100 200 300 400 500 4
−1 Time (sec)
−2
−3
Figure 5.8d Method 4-4, Run C-4.

3
2
Series
Pressure (psi)
1 1
2
0 3
100 200 300 400 500 4
−1 Time (sec)
−2
−3
Figure 5.8e Method 4-4, Run C-5.
frequencies used are 20 and 40 Hz. The Series B results assume single
non-sinusoidal pulses.
Method 4-3, Difference Equation (Software reference, 2XDCR07E.FOR).

In the next set of difference equation calculations for Series A, the previ-
ous 20-40-20-40 frequency sequence is replaced by 60-40-60-40, with all
else unchanged. However, the single pulse used earlier in Series B is now
replaced with the more complicated four-pulse sequence shown below.
A = 10.0
R = 100.0
SIGNAL = A*(TANH(R*(T-0.100))-TANH(R*(T-0.101)))/2.
1 +1.5*A*(TANH(R*(T-0.200))-TANH(R*(T-0.201)))/2.
2 +2.0*A*(TANH(R*(T-0.300))-TANH(R*(T-0.301)))/2.
3 +3.0*A*(TANH(R*(T-0.400))-TANH(R*(T-0.401)))/2.
Method 4-3, Difference Equation (Software reference, 2XDCR07F.FOR).

Finally, we consider a noise function that is not sinusoidal in our calcula-
tions. In fact, the downward propagating sine wave defined previously in
FUNCTION XNOISE(T) is replaced by a single isosceles triangle, noting
that multiple triangles are easily constructed, as follows:
FUNCTION XNOISE(T)
C TYPE 3: NOISE MODEL IS A TRIANGLE
A = 7.
B = 0.2
XNOISE = TRIANGLE(T, A, B)
RETURN
END
FUNCTION TRIANGLE(T, A, B)
IF( T.LE.0.0) TRIANGLE = 0.

IF(T.GT.0.0.AND. T.LE.B) TRIANGLE = A*T
IF(T.GT.B .AND. T.LE.2*B) TRIANGLE = A*(2.*B-T)
IF(T.GT.2*B) TRIANGLE = 0.
RETURN
END
Method 4-4, Differential Equation (Software reference, SAS14D.FOR

Option 3 identical to SIGPROC-1.FOR).
Next, consider the differential equation method in [2]. Inputs appear in
bold red font. In the first four simulations, a three-pulse “canned” signal is
used for convenience. For the fifth, a seven-pulse calculation is performed
– the software allows any number of rectangular pulses with arbitrary
widths, separations and amplitudes, although general waveforms are per-
mitted that require only minor source code modification. In our figures,
black represents the original amplitude and frequency modulated upgoing
MWD signal, blue and green denote pressures obtained at the two trans-
ducer locations (the large noise function is omitted for brevity), while red,
displaying the signal extracted from blue and green data, reproduces the
black curve in many instances. Additional conventional processing would
improve comparisons. Because significant noise is assumed, the blue and
green signals do not differ substantially, at least visually – thus, it is remark-
able that usable red lines are achievable at all. Algorithm details are avail-
able in [2] and only the input screen menus are shown for runs denoted
“C.” In the first four calculations, sound speeds of 5,000 and 3,000 ft/sec are
considered for transducer separations of 30 and 10 feet. Again, in the fifth,
a high-data-rate seven-pulse simulation with pulses of varying amplitudes
is shown – the plot is stretched horizontally to show pulse shape details.
Run C-1
Internal MWD upgoing (psi) signal available as
P(x, t) = + 5.000 {H(x- 150.000-ct) - H(x- 400.000-ct)}
+ 10.000 {H(x- 600.000-ct) - H(x- 1000.000-ct)
+ 15.000 {H(x- 1400.000-ct) - H(x- 1700.000-ct)}
Units: ft, sec, f/s, psi ...
Assume canned MWD signal? Y/N: Y
Downward propagating noise (psi) assumed as
N(x, t) = Amplitude * cos {2╥ f (t + x/c)} ...
o Enter noise freq “f” (hz): 15
o Type noise amplitude (psi): 30
o Enter sound speed c (ft/s): 5000
o Mean transducer x-val (ft): 1750

o Transducer separation (ft): 30
Run C-2
P(x, t) = + 5.000 {H(x- 150.000-ct) - H(x- 400.000-ct)}
+ 10.000 {H(x- 600.000-ct) - H(x- 1000.000-ct)}
+ 15.000 {H(x- 1400.000-ct) - H(x- 1700.000-ct)}
Run C-3
P(x, t) = + 5.000 {H(x- 150.000-ct) - H(x- 400.000-ct)}
+ 10.000 {H(x- 600.000-ct) - H(x- 1000.000-ct)}
+ 15.000 {H(x- 1400.000-ct) - H(x- 1700.000-ct)}
Run C-4
P(x, t) = + 5.000 {H(x- 150.000-ct) - H(x- 400.000-ct)}
+ 10.000 {H(x- 600.000-ct) - H(x- 1000.000-ct)}
+ 15.000 {H(x- 1400.000-ct) - H(x- 1700.000-ct)}

Run C-5
P(x, t) = + 5.000 {H(x- 150.000-ct) - H(x- 400.000-ct)}
+ 10.000 {H(x- 600.000-ct) - H(x- 1000.000-ct)}
+ 15.000 {H(x- 1400.000-ct) - H(x- 1700.000-ct)}
Assume canned MWD signal? Y/N: n
Number of rectangular pulses: 7
o Left “x” coord, Pulse 1: 50

o Right “x” coord, Pulse 1: 100
o Amplitude (psi), Pulse 1: 5






Input waveform summary,

P(x, t) = + 5.000 {H(x- 50.000-ct) - H(x- 100.000-ct)}
+ 10.000 {H(x- 150.000-ct) - H(x- 200.000-ct)}
+ 15.000 {H(x- 250.000-ct) - H(x- 300.000-ct)}
+ 20.000 {H(x- 350.000-ct) - H(x- 400.000-ct)}
+ 5.000 {H(x- 450.000-ct) - H(x- 500.000-ct)}
+ 10.000 {H(x- 550.000-ct) - H(x- 600.000-ct)}
+ 15.000 {H(x- 650.000-ct) - H(x- 1700.000-ct)}
Closing remarks. In high-data-rate MWD mud pulse telemetry utiliz-

ing high transmission frequencies, the effects of signal cancellations and
distortions at the mudpump and desurger imply severe detection problems
and synchronization loss. These are worsened by strong pump transients
and weaker (siren) pulser signals relative to conventional positive pulse
poppet devices. In this paper, we have addressed propagating noise effects
and shown that it is possible to design robust “directional filters” that oper-
ate with short transducer spacings and large time sampling durations in
typical noisy drilling rig environments. The latter support fast calculations
using computers with modest hardware resources. In practice, the filters
will be used together with conventional frequency, wavelet, white noise,
adaptive and other filters. Our signal processing efforts are complemented
by low-torque and high-amplitude siren hardware, where signal strengths
are augmented by using constructive wave interference methods. The
authors believe that a fully integrated system addressing both surface and
downhole concerns stands the best chance in confronting the challenges
offered by noisy and attenuative environments.
5.4 Conclusions
We have summarized our strategy for high-data-rate mud pulse telemetry
and means for developing the technology. Our target objective of 10 bits/
sec at 30,000 feet appears to be doable. The signal amplification approach
used, together with new surface signal processing techniques, plus the use
of specially designed tools that are integrated with mud and drillpipe prop-
erties, provide a systems oriented process that optimizes data transmission.
Needless to say, we have acquired much in our testing program, and we are
continually learning from our mistakes and developing new methods to
improve the technology.
Acknowledgments
The project “Downhole High-Data-Rate Continuous Wave Mud Pulse
Telemetry” was sponsored by China National Petroleum Corporation’s
Technology Management Division under Contract 2008C-2101-1. The
authors thank the management of CNPC for permission to publish
this research and for its support and interest throughout the effort. We
also thank the University of Petroleum (East China), whose College of
Electromechanical Engineering provided test facilities for research exper-
iments, and Professor Jun Fang, Professor Xueshi Gao and Dr. Peng Jia,
who diligently contributed their expertise during numerous wind tunnel
test sessions.
References
1. A. Boyd, H. Darling, J. Tabanou, B. Davis, B. Lyon, C. Flaum, J. Klein, R.
Sneider, A. Sibbit, and J. Singer, The lowdown on low-resistivity pay. Schlumb.
Oilfield Rev. Autumn, 4–18, (1995).
2. W.C. Chin, Y. Su, L. Sheng, L. Li, H. Bian, and R. Shi, MWD Signal Analysis,
Optimization and Design, John Wiley & Sons, New Jersey (2014).
3. B.J. Patton, W. Gravley, J.K. Godbey, J.H. Sexton, D.E. Hawk, V.R. Slover,
and J.W. Harrell, Development and successful testing of a continuous-wave,
logging-while-drilling telemetry system. J. Petrol. Technol. 1215–1221 (1977).
4. Presco, Inc., Measurement while drilling, http://www.prescoinc.com/science/
drill-ing.htm (2011).
5. W.C. Chin, Multiple transducer MWD surface signal processing, US Patent
5969638, assigned to Halliburton (October 19, 1999).
6. W.C. Chin, MWD siren pulser fluid mechanics. Petrophysics 45. 363–379,
(2004).
7. W.C. Chin and J. Trevino, Pressure pulse generator, US Patent 4785300,
assigned to Schlumberger (November 15, 1988).
8. A.A. Gavignet, L.J. Bradbury, and F.P. Quetier, Flow distribution in a roller
jet bit determined from hot-wire anemometry measurements. SPE Dril. Eng.
19–26, (1987).
9. W.C. Chin, Measurement-while-drilling system and method, US Patent
5583827, assigned to Halliburton (December 10, 1996).
10. M.R. Foster and B.J. Patton, Apparatus for improving signal-to-noise ratio
in logging-while-drilling system, US Patent 3742443 assigned to Mobil Oil
Corporation (June 26, 1973).
6
Detection of Geologic Anomalies with
Monte Carlo Clustering Assemblies
Simon Katz1*, Fred Aminzadeh2, George Chilingar2, Leonid Khilyuk2
and Matin Lockpour1
Russian Academy of Natural Sciences, US Branch, CA

1
2
Petroleum Engineering Program, School of Engineering, University of
Southern California, Los Angeles, CA
Abstract
Authors present new clustering-based algorithms for detection of geologic anom-
alies and results of their testing on the data containing anomalies of two types:
(a) high permeability anomaly with regular records containing smaller perme-
ability values, and (b) gas-filled sand anomaly with regular records containing
data from brine-filled sands. Results of algorithms testing, presented in the paper,
demonstrate high stability of anomaly detection with false discovery rate below
20% and with the true discovery rate exceeding 73%.
Keywords: Anomaly detection, clustering assembly seismic velocities, rock

density, brine, gas, true discovery, false discovery
6.1 Introduction
The goal of this paper is to present new clustering assembly algorithms for
detection of geologic anomalies of various types. These algorithms may be
utilized as one of the steps in location of oil and gas reservoirs, overpres-
sure zones, and other geologic anomalies. List of publications on detection
of geologic anomalies includes finding the location of fractured carbonates
filled with gas [4], finding the location of an overpressure zone [5], and
one-class methodology for anomaly detection within the homogeneous
135
geologic area [7]. Key element of the proposed methodology is repeated

clustering of analyzed data. Clustering technique has been used for geo-
logic applications [2] and for detection of single outliers or clusters of small
size [8, 10, 11]. In geological applications, size of anomaly area may not
be necessarily small. Besides, records in the training set may form several
clusters of different size.
This makes the problem of detection of geologic anomaly more com-
plex, compared to the problem of outlier detection. The authors overcome
this problem via construction of multiple randomized train and test sets
and clustering them. As a consequence, obtained multiple cluster sets form
clustering assembly. The clustering assembly is used for calculation of
irregularity index of an individual clusters and anomaly indexes for each
cluster set, each cluster, and anomaly index of individual records in the test
set. A decision for anomaly identification is done via analysis of anomaly
indexes and selection of the threshold for anomaly identification.
Evaluation of efficiency of algorithms was done using data with
anomalous records of two types. First one is the dataset with anomalous
records containing high permeability values, exceeding 1000 mD, and
regular records with smaller permeability values. Permeability dataset
was published by Aase et al. in [1] and posted as an open source at the
pubs.usgs.gov website. It contains 99 records. with eleven of them with per-
meability exceeding 1000 mD. Parameters used for clustering this dataset
are porosity and grain size. Another anomaly type is anomaly with records
collected from gas-saturated sands and regular records collected from
brine-filled sands. In this case, number of anomalous and regular records
both equal 25, and each record contains three parameters – VP, Vs, and
rock density ρ. These data were published by Ramos. and Castagna, in S9].
Authors used NbClust R clustering package [3]. This package includes
9 clustering methods and a large number of criteria for selection of the
optimal number of clusters. Euclidean distance in detection of anomalies
of both types was used. In the case of high permebility anomaly, decision
about optimal number of clusters was defined by majority of all criteria, and
by using clustering criterion “silhouette” in the case of gas-sand anomaly.
6.2 Analysis of Inhomogeneity of the Training

and Test Sets and Instability of Clustering
Ideally, all regular records in the union of the train and test sets form a
single cluster, so that records outside this cluster are anomalous. In fact,
Detection of Geologic Anomalies with Monte Carlo 137
this usually does not happen, and both the train and the test sets may
be broken into several clusters. Another issue, that complicates anom-
aly detection clustering methodology, is instability of clustering process,
where minor change in the clustered data may lead to significant change
in the number of clusters and in their size. This is illustrated by Figure
6.1 that shows histograms of the number of clusters and their sizes in
repeated clustering of the dataset with records collected from brine-filled
sands areas. Clustering was performed on the same dataset with repet-
edly randomized reordering of the clustered records and removing five
records from the dataset. One can observe, that the dataset ‘brine’ is inho-
mogeneous containing clusters of different size, which number varies in
the wide range. Figure 6.1 also shows, that there are clusters as small as
one or two records. Therefore, using small cluster size as indicator of
geologic anomaly is problematic. To overcome the problem of clustering
instability and questional relations between cluster size and its potentional
anomaly, authors developed anomaly detection algorithms, that rely on the
use of clustering assemblies containing large number of individual cluster
sets. Each cluster set is generated by clustering the randomized test set, that
contains records, some of which labeled as regular and others - unlabeled
records. Clustering assembly is utilized to formulate criteria for identifica-
tion of anomalous cluster sets, clusters, and anomalous records.
Due to instability of individual clusters, anomaly detection, that rely
on the use of individual cluster set, produce unreliable results. This is
illustrated by Figure 6.2, that shows three clusters and ellipses, drawn
around clusters with confidence level 0.95 (R function dataEllipse).
Clustering was done on randomized test set, that includes anomalous
Number of Clusters Size of Clusters

150
120
100
Frequency
Frequency
80
50
40
0
2 3 4 5 6 2 4 6 8 10 12 14
Figure 6.1 Histograms of the number of clusters and the number of records in individual
clusters (cluster size) in the dataset containing records from brine-filled sands.
0.40
Anomaly
Regular
0.30
Grain.size
Regular
0.20
0.10
5 10 15 20 25
Porosity
Figure 6.2 Clustering of a randomized test set that includes regular and anomalous
records.
and regular permeability data. Two of the clusters are regular and one is
anomalous. One can observe, that cluster ellipses overlap each other and
poorly separated.
6.3 Formation of Multiple Randomized Test Sets

and Construction of the Clustering Assemblies
Algorithms presented in this paper rely on the use of the initial train set,
that contains only regular records, labeled as “regular” and initial test set,
that contains unlabeled records. Some of the unlabeled records may be
regular others be anomalous. To obtain reliable identification of anom-
alous and regular records in the initial test set, the authors constructed
multiple randomized test sets, built as the union of randomly formed
subsets of the initial train and test sets. Randomization is done by Monte
Carlo resampling of records of the train and test set. Clustering of m ultiple
randomly-formed test sets results in the formation of an assembly of clus-
ter sets sCl(j), 1 ≤ j ≤ J, that contains J cluster sets. Due to randomization,
each of the records in the initial test set will appear multiple times in dif-
ferent cluster sets, so that multiple values of index j may be assigned to
the same record. If the number of Monte Carlo runs is large enough, the
number of appearances of individual records in the clustering assembly
also will be large. This is illustrated by Figure 6.3. It shows values of the
number of appearances of individual records in the clustering assembly
build via repeated clustering of the randomized test set. Each randomized
210
BRINE GAS
Number of presence
in different clusters
190
170
150
0 10 20 30 40 50
Records
Figure 6.3 Number of appearances of individual records in different clusters of the

clustering assembly.
test set includes 15 randomly selected records from gas set and the same
number of records from the brine set. Number of Monte Carlo runs is 300.
One can observe, that the number of appearances for any individual
record exceeds 160. So large number of appearances of individual records
in different clusters of the clustering assembly opens the way for building
stable and reliable procedures for identification of anomalous records.
6.4 Irregularity Index of Individual Clusters in the

Cluster Set
Further in this paper, the authors used the following notations: m is the
index of the record in the union of the initial train and test sets, j is the index
of the cluster set formed via clustering randomized test set, r is the index
of the cluster in the cluster set with index j. Thus, each cluster is indexed
by pair of indexes (r, j ). Records in he initial trainset are labeled as regular,
whereas records in the initial test set are unlabeled.
Key element of methodology of identification of anomalous and regular
records, is the assumption, that due to the difference between properties of
records of these two types, they will form different clusters. Based on this
assumption, the authors introduce following parameters, that characterize
possibility (not probability) that a given individual cluster and cluster set
contain anomalous records:
a. Irregularity index, Irreg(j, r), of a cluster with index r in the

cluster set with index j:
n.unlabeled( j, r ) (6.1)
Irreg ( j, r ) =
n.cluster( j, r )

where n.unlabeled(j,r) and n.cluster(j,r) are the numbers of unlabeled

records from initial test set and the total number of records in the cluster.
According to Eq. 6.1, this parameter satisfies the following conditions:
0 ≤ Irreg (j, r) ≤ 1 (6.2)
If irregularity index is close to 1, the majority of records in this cluster

are unlabeled and a number of regular records is small. Consequently, this
cluster may be identified as anomalous. On the other hand, if this index
is close to zero, majority of its records are labeled as regular, and cluster
should be identified as regular.
b. Index of anomaly presence in a cluster set sCl(j) is defined
by Eq. 6.3:
AnClset ( j ) = max ( Irreg ( j , r ))− min( Irreg ( j , r )) (6.3)

r r
According to Eq. 6.3, index of anomaly presence in the cluster set satis-
fies the following constraints:
0 ≤ AnClset (j) ≤ 1 (6.4)
If irregularity index is the same for all clusters in the cluster set, then:
AnClset (j) = 0 (6.5)
On the other hand, if there is at least one cluster, that contains only
labeled records from the train set and at least one cluster containing only
unlabeled records, then:
AnClset (j) = 1 (6.6)
Values of anomaly index close to 1 are strong indication that there is

anomalous cluster in the cluster set.
c. Anomaly index of an individual cluster is defined by Eq. 6.7:
AnCl(j, r) = Irreg(j, r) * AnClset(j) (6.7)
According to Eq. 6.7, anomaly index of a cluster is close to 1, if both its

irregularity index and anomaly index of the cluster set are close to 1.
6.5 Anomaly Indexes of Individual Records

and Clustering Assemblies
Specific feature of clustering assemblies, is that each record in the initial
train and test sets will appear in a number of different clusters of the clus-
tering assembly, and each cluster will be characterized by different values
of irregularity index. This opens the way for construction of stable anomaly
index for individual records in the studied dataset. In the following sec-
tions the authors test mean-aggregated anomaly index defined by Eq. 6.8:
J (m )
rAnom(m) =
1
J (m) ∑ Irreg (m, j,r ) (6.8)
j =1

where m is the index of the record, r is the index of the cluster containing
record with the index m at Monte Carlo run with index j, J(m) is the total
number of appearances of record with index m in the clustering assembly.
Similarly, anomaly index of the whole clustering assembly is defined as:
anS =
1
J ∑anClset( j) (6.9)
j =1

Record with the index m will be assigned label “anomaly” or ”regular”

according to the following rule:
 “anomaly ”; rAnom(m) > tr

label(m) =  (6.10)
“regular ”; rAnom(m) ≤ tr

where tr is the threshold, defined using prior false discovery rate (section
6.6).
6.6 Prior and Posterior True and False Discovery

Rates for Anomalous and Regular Records
Concept of prior false discovery rate for individual records was introduced
in [7]. Bellow, we expand this concept to prior estimates of false discovery
rates of anomalous records, clusters, and anomalous cluster sets.
False discovery rates of anomalous records, clusters, and cluster sets in
clustering assemblies are defined, respectively, as the fractions of individ-
ual records, clusters, and cluster sets, identified as anomalous, when all
records in the clustering assembly are regular. Analysis of prior false dis-
covery rates (FD) is done via repeated formation of a random test sets,
built as the random subsets of the initial train set, and formation of the
clustering assembly that contains only regular records. Identification of
anomalous clusters, anomalous cluster sets, or anomalous records in this
clustering assembly is the act of false discovery. Prior false discovery rate
is used for selection thresholds in anomaly detection procedures to guar-
antee low posterior (actual) false discovery rate. Posterior false (FD) and
true (TD) discovery rates are calculated using clustering assembly, built
via repeated clustering of randomized test sets, that contain regular and
anomalous records. Posterior (true) FD and TD are defined, respectively,
as the fractions of regular and anomalous records identified as anomalous
in this clustering assembly.
6.7 Estimates of Prior False Discovery Rates for

Anomalous Cluster Sets, Clusters, and Individual
Records. Permeability Dataset
In this section the authors present results of analyses of prior FD obtained
using train set, that includes records with the regular permeability values
smaller than 1000 mD. Figure 6.4 shows histograms of the three prior
anomaly parameters, calculated using clustering assembly built via Monte
Carlo resampling of this set. According to this figure, all three anomaly
indexes vary in the range of 0.0 - 0.5 with maxima of histograms below 0.3.
Low range of these parameters indicate absence of anomalous records in
this clustering assembly.
Table 6.1 shows estimated prior false discovery rate of the cluster sets
and clusters calculated as functions of threshold.
Irregularity index
8
6
4
2
0
Anomaly index of individual clusters
Percent of total
8
6
4
2
0
Claster sets anomaly index
8
6
4
2
0
0.0 0.2 0.4 0.6 0.8 1.0
Values of anomaly parameters
Figure 6.4 Histograms of three prior anomaly parameters.
Table 6.1 Prior rates of false discovery of

anomalous clusters, and cluster sets.
Prior false discovery rate
Threshold Cluster sets Clusters
0.05 0.864 0.079
0.1 0.717 0.033
0.15 0.626 0.012
0.2 0.426 0.005
0.25 0.273 0.002
0.3 0.162 0.001
0.35 0.098 0.001
As shown in the Table 6.1, the prior false discovery rates for anomalous
cluster sets and individual clusters in the cluster set drop below 0.01 at the
threshold equal to 0.35.
6.8 Posterior Analysis of Efficiency of Anomaly

Identification. High Permeability Anomaly
Assembly of individual cluster sets analyzed in this section and utilized for
reliable anomaly identification includes 300 of individual cluster sets. Each
individual cluster set is generated via clustering of a randomized test set con-
structed as the union of initial test set and randomly-generated subset of the
training set. Initial test set contains 11 anomalous records and 10 regular
ones. Figure 6.5 shows histograms of the distribution of irregularity index
and anomaly indexes of cluster sets and individual clusters. Importantly, all
300 cluster set anomaly indexes are large and exceed 0.6. This is strong indi-
cation of the presence of anomaly in every randomized test set. Histogram
of the anomaly index of individual clusters shows presence of two clus-
ter groups – clusters with high anomaly index exceeding value of 0.5 that
are potentially anomalous, and those with anomaly index as small as 0.1.
Histogram of the irregularity index shows similar pattern with the presence
of a group of irregularity indexes with values exceeding 0.8.
Cluster set anomaly index

12
10
8
6
4
2
0
Anomaly index of individual clusters
Percent of total
12
10
8
6
4
2
0
Irregularity index
12
10
8
6
4
2
0
0.0 0.2 0.4 0.6 0.8 1.0
Histograms of three anomaly indexes
Figure 6.5 Histograms of three posterior parameters characterizing presence or absence

of anomalous records in the randomized test sets.
Table 6.2 Parameters of clusters, labeled as anomalous with individual cluster

anomaly index exceeding 0.5.
Mean values
Number
Cluster of Irregularity Anomaly index Anomaly index
size clusters index of clusters of cluster sets
1 9 1.000 0.814 0.814
3 1 1.000 0.833 0.833
7 215 1.000 0.813 0.813
8 368 0.900 0.646 0.714
9 395 0.908 0.659 0.723
10 1 1.000 0.923 0.923
Table 6.2 shows parameters of clusters with posterior cluster anomaly

index exceeding 0.5. According to this table, there is a wide range of sizes
of clusters of this type, whereas three anomaly indexes are all in the narrow
ranges. This is another indication, that the size of the cluster alone is not
reliable indicator of anomaly of records in this cluster.
According to the Table 6.2, majority of clusters with large anomaly inde-
ces are not small and contain at least 7 records.
Efficiency and reliability of identification of anomalous records are
illustrated in Figure 6.6, which shows plots of anomaly index of all
records (regular and anomalous) in the permeability dataset. Horizontal
dashed line is drawn at the threshold level equal to 0.4. There is only
one falsely identified regular record with anomaly index as high as 0.8.
Anomaly index of other regular records is smaller 0.4. Among eleven
anomalous records there are eight records with anomaly index exceed-
ing 0.4. These records correctly identified as anomalous. There are also
three anomalous records with anomaly index smaller than 0.4. These
records are falsely identified as regular. True discovery rate in this case
is around 73%.
Anomaly index
Anomalous permeability
1.2
of records
0.6
0.0
Records
Anomaly Index
Regular permeability
0.8
of records
0.4
0.0
Records
Figure 6.6 Anomaly indexes of individual records. High permeability anomaly.
6.9 Identification of Records in the Gas Sand Dataset

as Anomalous, using Brine Sand Dataset as Data
with Regular Records
The authors present in this section results of analysis of accuracy of detec-
tion of gas sand anomaly, with initial test and train sets formed as gas and
brine datasets. Randomized test sets were repeatedly built as the unions of
the randomly formed subsets of 15 records from initial train and test sets.
The main complicating factor of detecting gas-sand anomaly is inhomoge-
neous structure of gas and brine data-sets, so that both train and test set
contain several clusters.
Figures 6.7 and 6.8 show distributions of prior and posterior values of
irregularity index and anomaly indexes of individual clusters and cluster
sets. To calculate the prior indexes, randomized train and test sets were
formed as not intersecting subsets of the brine dataset. Posterior indexes
were calculated using brine and gas datasets as initial train and test sets.
One can observe significant differences in distributions of prior and pos-
terior indexes shown at these figures. According to Figure 6.7, absolute
majority of all three prior indexes do not exceed threshold of 0.6. On the
other hand, large number of values of posterior indexes, shown in Figure
6.8, is larger than this threshold. This indicates the possibility of detection
of anomalous records, that form the initial test set.
Figure 6.9 illustrates efficiency of detection of gas-sand anomaly using
clustering assemblies. It shows values of mean-aggregated anomaly index
of individual records (Eq. 6.8) in brine-filled and gas-filled sand datasets.
Cluster set Anomaly of Irregularity

anomaly clusters
400
800
400
600
300
300
Frequency
Frequency
Frequency
400
200
200
100
200
100
0
0
0.0 0.4 0.8 0.0 0.4 0.8 0.0 0.4 0.8
cl.set.anom cl.anom irr.prior
Figure 6.7 Histograms of three prior anomaly indexes.
250
Cluster set Cluster Irregularit y
250
anomaly anomaly
300
200
250
200
200
150
Frequency
Frequency
Frequency
150
150
100
100
100
50
50
50
0
0.0 0.4 0.8 0.0 0.4 0.8 0.0 0.4 0.8
Figure 6.8 Histograms of three posterior anomaly indexes.
According to this figure, absolute majority of values of anomaly index of

the records in the brine dataset is smaller than 0.31, whereas majority of
the records in the gas dataset are characterized by values of this index
exceeding this value. Accuracy and stability of detection of gas sand anom-
aly is illustrated by Table 6.3. It shows thresholds, true, and false discovery
rates of anomalous records calculated using three independently gener-
ated clustering assemblies. Thresholds were calculated as quantile values
of anomaly indexes for records in the brine dataset with the same quantile
probabilities in all three clustering assemblies.
Brine
0.6
Anomaly index
0.4
of records
0.2
Gas
0.0
0 10 20 30 40 50
Records
Figure 6.9 Values of anomaly index of individual records in gas-sand and brine-sand
datasets.
Table 6.3 Thresholds, true, and false discovery rates of anomalous records
using three independently generated clustering assemblies. 300 Monte
Carlo runs.
Quantile
probabilities Thresholds False discoveries True discoveries
1 2 3 1 2 3 1 2 3
0.6 0.32 0.32 0.34 0.32 0.36 0.36 0.8 0.84 0.96
0.7 0.35 0.34 0.36 0.24 0.2 0.32 0.8 0.8 0.8
0.8 0.38 0.36 0.38 0.2 0.2 0.2 0.8 0.8 0.8
0.9 0.38 0.38 0.39 0.04 0.04 0.04 0.8 0.8 0.8
0.95 0.38 0.38 0.39 0.04 0.04 0.04 0.8 0.8 0.8
According to Table 6.3, values of true and false discovery rates obtained
using three independently generated clustering assemblies show minor
differences. For all three clustering assemblies, true discovery rates not
smaller 0.8, with false discovery rates as low as 0.04.
6.10 Notations
Gas set - dataset that contains records from gas-filled sands, brine set
- dataset that contains records from brine-filled sands.
TD and FD - true and false discovery rates.
j is the index of the cluster set formed via clustering randomized test
set, r is the index of the cluster in the cluster set with index j. Thus,
each cluster is indexed by pair of indexes (r, j).
Irreg(j, r) - irregularity index of the cluster with index r within cluster
set with the index j.
n.unlabeled(j, r) and n.cluster(j, r) are the numbers of unlabeled and
the total number of records in the cluster.
AnClset(j) - index of anomaly presence in a cluster set sCl(j).
AnC(j, r) - anomaly index of individual cluster.
rAnom(m) anomaly index of individual record.
J(m) is the total number of appearances of the record with index m in
the clustering assembly.
TD(tr), FD(tr) - posterior true and false discovery rates, fr-threshold
in anomaly detection rules.
6.11 Conclusions
• Authors present new algorithms for detection of geologic anom-
alies and results of their testing. Key element of the developed
algorithms is construction of multiple randomized test sets, and
construction of multiple cluster sets, that form clustering assem-
bly. The clustering assembly is used for calculation of irregularity
index of individual records and anomaly indeces for each cluster
set, each cluster, and individual records in the test and train sets.
• The algorithms were tested on the data with anomalies of two
types: (a) high permeability anomaly with regular records con-
taining smaller permeability values, and (b) gas-filled sand anom-
aly with regular records containing data from brine-filled sands.
Results of algorithms testing, presented in the paper, demonstrate
high stability of anomaly detection with true discovery rate higher
than 73% with false discovery rates equal or even lower than 0.2
for both anomaly types.
References
1. Aase, N., Bjorkum, P., and Nadeau, P., 1996, The effect of grain-coating
microquartz on preservation of reservoir porosity. Bull. Am. Assoc. Petrol.
Geologists 80, (1): 1654-1673.
2. Aminzadeh, F. and Chatterjee, S., 1984, Applications of cluster analysis in
Exploration Seismology, Geoexploration, 23: 147-159.
3. Charrad M, Ghazzali N., Boiteau V., Niknafs A., 2014. NbClust: An R
Package for determining the relevant number of clusters in a data set. Journal
of Statistical Software, 61 (6): 1-36.
4. Chilingar G., Mazzulo S., Rieke, H., 1992. Carbonate Reservoir
Characterization: A Geologic-engineering Analysis. Elsevier, 639 pp.
5. Dvorkin J, Mavko G., Nur A., 1999, Overpressure detection from com-
pressional and shear-wave data. Geophysical Research Letters, 26, (22):
3417–3420.
6. Gurevich A., Chilingar G, and Aminzadeh F., 1994. Origin of the formation
fluid pressure distribution and ways to improving pressure prediction meth-
ods. J. Pet. Sci. Eng., 1: 67-77.
7. Katz S., Aminzadeh F., Chilingar G., Khilyuk L., 2015. Anomaly detection
within homogeneous geologic area. J. of Sustainable Energy Engineering, 3,
(2): 169-186.
8. Lian Duan, Lida Xu, Ying Liu., 2008. Cluster-based outlier detection. Annals
of Operations Research, 168, (1): 151–168
9. Ramos, A. and Castagna, P., 2001. Useful approximations for converted-wave
AVO. Geophysics, 66: 1721–1734.
10. Zimek, A., Campello, R., Sander, J.. 2014. Ensembles for unsupervised outlier
detection. ACM SIGKDD Explorations Newsletter. 15: 11–22.
11. Zengyou He, Xiaofei Xu, Shengchun Deng. 2003. Discovering cluster-based
local outliers. Pattern Recognition Letters, 24(9-10):1641-1650.
7
Dissimilarity Analysis of Petrophysical
Parameters as Gas-Sand Predictors
Simon Katz1*, George Chilingar2, Fred Aminzadeh2 and Leonid Khilyuk1
Russian Academy of Natural Sciences, US Branch,

1
Los Angeles, California

2
Petroleum Engineering Program, School of Engineering, University of Southern
California
Abstract
Multiple petrophysical parameters, whose values may be used as indicators of gas
presence in sands and sandstones, are analyzed here. Bulk and shear moduli, Lame
parameter λ, Vp/Vs, and Poisson’s ratio are among parameters subjected to com-
parative analysis as potential gas indicators. Rock density ρ, product ρ*λ and ratio
of Lame Parameter λ to shear modulus were also tested as potential gas indicators.
Additionally, two one-dimensional aggregated parameters constructed as combi-
nations of several parameters and multidimensional parameters were tested using
machine learning techniques and ROC curve analysis.
Ranking parameters and evaluation of their efficiency was based on two meth-
ods: (1) Dissimilarity-based method independent of classification technique and
(2) ROC curve analysis specific to individual classification methodology. Both
types of methods work in terms of true and false discovery rates. The parameter
built as a combination of five other parameters (Lame parameter λ, Vp to Vs ratio,
velocities of compressional and shear waves, and rock density) get high rankings
using both methods. This parameter demonstrates a higher true discovery rate
and a limited false discovery rate in identification of gas-sands versus brine-sand
or shale. The parameters Poisson’s ratio, Vp/Vs, bulk over shear modulus, Lame
parameter λ over shear modulus also rank highly.
Cross validation was used to control over classification with undue increase
of estimate of true discovery rate. Classification with most efficient parameters of
gas-sand vs. brine-sand or shale leads to the true discovery rate exceeding 80%
and a false discovery rate not exceeding 20%.
151
Keywords: Classification, probability of true discovery, dissimilarity, ROC-curve

analysis, seismic velocities, rock density, brine, gas, shale
7.1 Introduction
Evaluation of petrophysical parameters as classifiers of lithology and fluid
content is presented in multiple publications. Wide attention was given to
Vp/Vs ratio as a possible lithology classifier. Garotta (1987), replace with
Ramos and Castagna (2001), Margrave et al. (1998) and Stewart demon-
strated that changes in lithology can lead to a change in the Vp/Vs. Picket
(1963) determined Vp/Vs values for sandstones, dolomites and limestones.
Goodway (1997, 2001) suggested that the parameters derived from Lame
parameters (λ * ρ, μ* p and λ/μ) may be used for identification of gas-sands.
Ostrander (1984) studied the effect of changing Poisson’s Ratio on reflec-
tion coefficients of seismic waves.
The goal of this paper is development and testing of methodology for
lithologic and fluid content classification based on the use of multiple pet-
rophysical parameters. The authors performed comparative analysis of
efficiency of multiple petrophysical parameters as potential indicators of
changing lithology and gas saturation. Advancement in this direction was
achieved with construction of the dissimilarity measures and analysis of
aggregated parameters built as combinations of several individual parame-
ters. In addition, evaluation of the false and true discovery rates with ROC
curve analysis improves the reliability of estimation of efficiency of gas
indicators for both individual and aggregated parameters.
7.2 Petrophysical Parameters for Gas-Sand

Identification
The starting set of petrophysical parameters obtained from seismic data are
compressional and shear waves velocities Vp, Vs, and rock density ρ.
The writers tested several additional parameters for identification of gas
sands. All the studied parameters are derived from Vp and Vs equations:
4
K+ µ
Vp = 3 = λ + 2µ ; V = µ (7.1)
s

ρ ρ ρ
Dissimilarity Analysis of Petrophysical Parameters 153
The following parameters were the subject of further analysis presented

in this paper:
µ = ρ ∗ Vs2 (7.2)
 4  (7.3)
K = ρ ∗ Vp2 − Vs2
 3 

λ = ρ ∗ (Vp2 − 2 ∗ Vs2 ) (7.4)

Vp2 4
K/µ= − (7.5)
Vs2 3
ρ ∗ (Vp2 − 2 ∗ Vs 2 ) Vp2
λ/µ= = 2 −2 (7.6)
ρ ∗ Vs2 Vs
Poisson’s ratio [8]
δ=
(Vp / Vs ) − 2
2
(7.7)
2 ∗ (Vp / Vs ) − 1
2

λ*ρ (7.8)
μ*ρ (7.9)
The bulk modulus K
K = λ + 2 * μ / 3 (7.10)
In addition two new aggregated parameters:
Group_B = Vp + 1.5 * (Vp/Vs – Vs) (7.11)

Group_A = Vp − Vs * 1.5 + ρ + VpVs + λ/6 (7.12)
were also tested.

The authors used data on Vp, Vs, and rock density presented by Ramos
and Castagna (2001). Their dataset contains total of 75 measurements of
Vp, Vs, and density for gas-filled sands, brine-sands, and shales. At each
condition there are 25 measurements of the triplet (Vp, Vs, ρ). Inasmush
as the data are collected from several geologic areas, the results of this data
set analysis may have general significance.
7.3 Lithologic and Fluid Content Dissimilarities

of Values of Petrophysical Parameters
The writers define the dissimilarity between two datasets of a petrophysical
parameter in two different geologic conditions as a normalized difference
of median or mean values of this parameter:
median ( param|condition j ) − median ( param|lconditionk )

diss = 2 ∗
std ( param|condition j ) + std ( param|conditionk )

(7.13)
median ( param|condition j ) − median ( param|lconditionk )

diss1 = 2 ∗
std ( param|condition j ) + std ( param|conditionk )

(7.14)
where median(param|conditionj) mean(param|conditionj), std(param|condi-

tionj), median(param|conditionk) mean(param|conditionk) and std(param|con-
ditionk) are median, mean, and standard deviations of the set of values of
parameter calculated for geologic conditions conditionj and conditionk,
param - the studied parameter. Conditions take one of the four values - (1)
brine-filled sand, (2) gas-filled sand, (3) shale, and (4) brine-filled sand or
shale.
The authors further used the dissimilarity constructed as a normal-
ized difference of the medians (Eq. 7.13). Advantage of the normalized
median differences is that they are more stable with respect to rare outliers.
Generally, if the dissimilarity of the parameter between two conditions is
smaller than 1, then this parameter is inefficient as a classifier.
7.4 Parameter Ranking and Efficiency

of Identification of Gas-Sands
We ranked parameters and evaluated their efficiency in identification of
gas-sands using two methods. In the first method, we used the dissimilar-
ity between datasets and ranked parameters according to the values of esti-
mated dissimilarities. Advantage of this method is that it is independent of
classification technique. In the second method, we calculated ROC curves
that give probabilities of true and false discovery for each of the parame-
ters. ROC curves are different for different classification techniques. Hence
results of the parameter ranking done with the second method might be
different for different classification methods.
The parameters in Table 7.1 are ranked according to the dissimilar-
ities between parameter values in gas-sands vs. brine-sands or shales.
Group_A parameter has the highest rank for all combinations of gas-sand
vs. other conditions. Five parameters have dissimilarities larger than 1.0
for gas-sand versus other conditions. These distinguished parameters are
Group_A, ρ * λ, Poisson’s.ratio, λ, and Group_B. Vp/Vs has dissimilarity
larger than 1.0 for combination of gas-sand versus brine sand and is close
to 1.0 for gas-sand versus shale. All parameters have dissimilarities sig-
nificantly lower than 1.0 for the combination of brine-sand versus shale.
Importantly, two aggregated parameters - Group_A and Group_B outper-
form individual parameters that were used for aggregation.
Table 7.1 shows values of dissimilarities for 4 different combinations of
gas-filled sand, shale, and brine-filled sand.
Trends in dissimilarities for all 14 parameters are illustrated by Figure 7.1.
According to the Figure 7.1, there is no significant difference in dissimi-
larities for three combinations: gas-filled sand versus any of the three other
conditions. Dissimilarities for brine-sand versus shale are all low. Four
parameters (Vp, Vs, μ, and ρ* μ) show the lowest levels of dissimilarities for
all four combinations of conditions.
Effect of high performing parameter Group_A on dissimilarity between
gas-filled sand and brine-sand or shale is illustrated by two histograms
shown in Figure 7.2.
Histogram for the values of Group_A parameter calculated for gas sand
is in the range of relatively smaller values compared to the histogram of
this parameter for brine-filled sand or shale. Thick vertical line at Group_
A=5.6 in Figure 7.3 is the line of separation of two histograms.
The sort of mutual separation of two histograms shown in the Figure
7.2 leads to the following classification rule: when Group_A parameter is
Table 7.1 Dissimilarities among values of 14 parameters in 4 combinations of gas-filled sand, shale, and brine-filled sand,
Group_A and Group_B are parameters defined by Equations 7.11 and 7.12.
Gas-filled sand vs. shale Gas-filled sand vs. Gas-filled sand Shale vs. brine-
Parameters Rank or brine-filled sand brine-filled sand vs. shale filled sand
Group A 1 2.058 1.903 2.549 0.628
ρ* λ 2 1.526 1.515 1.644 -0.593
Poisson’s ratio 3 1.507 1.435 1.609 -0.509

λ 4 1.464 1.432 1.401 -0.346
Group B 5 1.291 1.428 1.319 0.338
Vp/Vs 6 0.918 1.064 0.98 -0.305
ρ 7 0.912 0.964 0.883 -0.295
(Continued)
Table 7.1 Dissimilarities among values of 14 parameters in 4 combinations of gas-filled sand, shale, and brine-filled sand,
Group_A and Group_B are parameters defined by Equations 7.11 and 7.12. (Continued)
Gas-filled sand vs. shale Gas-filled sand vs. Gas-filled sand Shale vs. brine-
Parameters Rank or brine-filled sand brine-filled sand vs. shale filled sand
K 8 0.745 0.876 0.691 0.182
λ/μ 9 0.642 0.876 -0.64 -0.177
K/μ 10 0.642 0.734 0.584 0.167
Vp 11 0.036 0.039 -0.198 0.072
ρ*μ 13 -0.048 -0.043 -0.245 0.072
μ 14 -0.172 -0.127 -0.392 0.101
Vs 15 -0.393 0.324 0.584 0.123
2.5
Gas vs Shale & Brine
Gas vs Brine
Gas vs Shale
2.0
Shale vs. Brine
1.5
Dissimilarity
1.0
0.5
–0.5 0.0
2 4 6 8 10 12 14
Ranks of the Parameters
Figure 7.1 Dissimilarities for 14 parameters in four combinations of conditions.

Horizontal axis: ranks of the parameters shown in Table 7.1.
Group_A Histogam. Gas-Sand.

3.0
Frequency
1.5
0.0
3 4 5 6 7 8 9
Group_A
Group_A Histogam. Brine-sand or Shale.

Frequency
8
4
0
3 4 5 6 7 8 9
Group_A
Figure 7.2 Histograms of the values of the Group_A parameter calculated for gas-sand
and brine-sand or shale.
smaller than 5.6, it indicates gas-sand; otherwise it is a brine-filled sand or

shale. Rates of true and false discovery are estimated as fractions of param-
eter values that are smaller than 5.6 in the histograms for gas- and brine-
sand. Their values calculated according to this rule are:
Group_A Histogam. Shale.
4
Frequency
2
0
3 4 5 6 7 8 9
Group_A
Group_A Histogam. Brine-sand.

Frequency
4
2
0
3 4 5 6 7 8 9
Group_A
Figure 7.3 Histograms of the Group_A parameter calculated for brine-sand and shale.
 rate of true discovery = 0.84

 (7.15)
rate of false discovery = 0.16

Relations 15 indicate high efficiency of the Group_A parameter in iden-

tification of gas-sands with high rate of true discovery and low false dis-
covery rate.
Figure 7.3 illustrates significant overlap in the histograms of the
Group_A parameter in classification of brine-filled sand vs. shale. In this
case, the dissimilarity between two groups is small.
Overlap of histograms is very significant and rate of false discovery
would be large for moderate values of true discovery rate. Thus the brine-
sand vs. shale is not reliably classifiable with this parameter.
7.5 ROC Curve Analysis with Cross Validation

Detailed analysis of rates of true and false discovery is done with analysis
of ROC curves. Area under ROC curve analysis was initially developed as
part of a signal processing technique and later found multiple applications
in many different areas of engineering, scientific research and geology
(Murphy and Monteiro, 2012) and geophysics (Everett, 2013). ROC curve
analysis shows a true discovery rate as a function of the rate of false discov-
ery for a given classification rule. Area under ROC curve (AUC) charac-
terizes performance of a classification procedure. The larger the AUC the
better is the performance of classification algorithm. Another advantage of
ROC curve analysis is its ability to adjust classification rule to guarantee an
appropriate level of false discovery rate.
We used the leave one out cross validation technique to avoid producing
over optimistic results. This technique is performed in four steps:
a. Remove a record from the input data set and optimize pre-
dicting model onthe remaining records.
b. Make prediction for removed record with optimized pre-
dictive model.
c. Return the record back to the data set.
d. Repeat steps a to c with another record.
e. Go through all records in the data set repeating steps a-d.
This cross validation methodology allows to avoid overfitting by pre-

dictive model since dataset for its construction does not include data from
the test set.
ROC curve construction starts with calculation of some classification
measure that defines tested record possibility to belong to one of two data
groups. Example of such a measure is the value of a posterior probabil-
ity to belong to the gas-sand class. The classification measure is compared
with a threshold and the number of record from the class 1 that exceed the
threshold defines the true discovery rate. The false discovery rate is defined
by the number of records from the class 2 that exceeds the same threshold.
Both true and false discovery rates are calculated for a range of threshold
values. Ranking both parameters and classification methods is done with
calculation of area under ROC curve (AUC). We used ROCR R software
package (Sing et al., 2005) for calculation of this measure to rank both
machine learning techniques and parameters used as gas-sand predictors.
Figure 7.4 shows examples of ROC curves produced by Linear Discriminate
Analysis for four parameters. The ROC curves were calculated using leave
one out cross validation.
All four ROC curves exhibit similar trends. All four parameters show
high true discovery rates larger than 0.8 and smaller than 0.2 for false dis-
covery rate.
1.0
0.8
True Discovery Rate
0.6
0.4
Group_A
Poisson’s ratio
0.2
Group_B
Vp/Vs
0.0 0.2 0.4 0.6 0.8

False Discovery Rate
Figure 7.4 ROC curves for four high performing parameters. Method: LDA.
7.6 Ranking Parameters According to AUC Values

The main measure of classifier efficiency in the framework of ROC curve
analysis is the area under ROC curve (AUC). The bigger the AUC values,
the more efficient the classifier is. Table 7.2 shows area under ROC curve
(AUC) for fourteen parameters. AUC values were calculated for ROC
curves of the following classification algorithms - Random Forest (RF),
General Linear Model (GLM), Linear Discriminant Analysis (LDA),
and K-Nearest Neighbor (KNN). Parameters are ranked according to
the mean AUC value. In spite of very different algorithms structure,
AUC values are similar for any parameter across all four methods. Five
parameters listed in the Table 7.2 have AUC values exceeding 0.8 for all
four methods. Another five parametes (ρ, μ, Vs, Vp, and ρ * μ) perform
poorly and have AUC values lower than 0.7 for all four classification
methods.
Figure 7.5 shows plot of mean AUC values averaged across four classifi-
cation techniques.According to Figure 7.5, there are seven parameters with
mean AUC exceeding 0.8 and five parameters with mea AUC smaller than
0.6. The high performing parameters are Group_A, Group_B, Poisson’s
ratio, Vp/Vs, K/μ, λ/μ. Importantly, four top parameters identified using
dissimilarity criterion are among five top parameters ranked by mean AUC
values.
Table 7.2 AUC values for four classification methods and 14 predictor
parameters.
Parameter Mean
rank Parameter RF GLM LDA KNN AUC
1 Group A 0.92 0.89 0.89 0.91 0.90
2 Group B 0.81 0.89 0.89 0.84 0.86
3 Poisson’s ratio 0.78 0.88 0.88 0.84 0.85
4 Vp/Vs 0.78 0.88 0.87 0.84 0.84
5 λ 0.86 0.83 0.84 0.84 0.84
6 λ/μ 0.78 0.88 0.85 0.85 0.84
7 K/μ 0.78 0.88 0.85 0.85 0.84
8 ρ* λ 0.83 0.83 0.83 0.79 0.82
9 K 0.67 0.70 0.70 0.74 0.70
10 ρ 0.46 0.67 0.67 0.56 0.59
11 Vs 0.58 0.50 0.50 0.49 0.52
12 Vp 0.55 0.49 0.49 0.53 0.52
13 μ 0.36 0.41 0.41 0.34 0.38
14 ρ*μ 0.24 0.28 0.28 0.28 0.27
Mean Area Under Rock Curve.

Average over KNN, GLM, RF, LDA.
0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
Mean AUC
High Performing Parameters

Low Performing Parameters
2 4 6 8 10 12 14
Parameter’s Rank
Figure 7.5 Mean area under ROC curve for 14 parameters. Area under Rock Curve
averaged over four machine learning techniques. Horizontal axis shows parameters’ ranks
given in the Table 7.2.
7.7 Classification with Multidimensional Parameters

as Gas Predictors
The goal of this section is to extend one dimensional aggregation of param-
eters to the multi-dimensional case. Group_A and Group_B are examples
of 1D aggregated parameters. Their advantage is in high sensitivity to the
changing type of sand saturation and in simplicity of interpretation. Their
disadvantage is in their fixed structure. Multidimensional parameters are
more flexible. Disadvantage of multidimensional aggregated parameters
is in a more complex interpretation. Machine learning classification tech-
niques are necessary to use full potential of the parameters of this type.
2D parameters may be used to build 2D cross sections to visualize sep-
aration and/or overlap of classified data. Examples of these cross sections
are shown in the Figures 7.6 and 7.7.
According to the Figures 7.6 and 7.7, Group_A - Poisson’s ratio cross sec-
tion shows better separation of two groups of data. Therefore, 2D aggregated
parameter (Group_A, Poisson’s ratio) is more effective in identification of
gas-sands. Both cross sections indicate that there is a group of anomalous
gas-sand values of parameters deep inside brine-sand or shale area.
Efficiency of the 2D aggregated parameter with components (Group_A,
Poisson ratio) is illustrated by Figure 7.8. It shows true and false discovery
rates obtained with KNN and LDA classifiers and 2D input predictor.
Intersection of two vertical and horizontal lines on plots of ROC curves
for KNN and LDA classifications shows critical values of false discovery
rate not exceeding 20% and true discovery rate not smaller 80%. High true
0.4
0.3
Passion’s Ratio
0.2
0.1
0.0
Gas-Sand
Brine-Sand or Shale
–0.1
3 4 5 6 7 8 9
Group_A
Figure 7.6 Group_A vs. Poisson’s ratio cross section of gas sand vs. brine sand or shale.
20
15
Lambda
10
5
Gas-Sand
0
Brine-Sand or Shale
1.5 2.0 2.5 3.0 3.5 4.0 4.5

Vp/Vs
Figure 7.7 Vp/Vs versus λ cross section of gas-sand vs. brine-sand or shale.
0.8
True Dsaicovery Rate
0.6
0.4
0.2
LDA
KNN
0.0 0.1 0.2 0.3 0.4 0.5

False Discovery Rate
Figure 7.8 ROC Curves for KNN and LDA classification techniques for 2D predictor
(Group_A, Poisson’s ratio).
discovery rate with limited rate of false discovery is the result of high sepa-
ration of the values of this predictor shown in the Figure 7.6.
7.8 Conclusions
1. Fourteen petrophysical parameters were analyzed as poten-
tial predictors of gas presence in sand formations. Vp to
Vs ratio, bulk and shear moduli, Lame parameter λ, and

Poisson’s ratio were among parameters subjected to compar-
ative analysis as gas-sand identification tools. We also tested
ratio of bulk to shear modulus and ratio of Lame parameter
λ to shear modulus. Ranking parameters and evaluation of
their efficiency was based on methods of two types: dissimi-
larity based method independent of classification technique
and ROC curve analysis specific to individual classification
methodology. Both groups of methods work in terms of true
and false discovery rates.
2. To get unbiased estimate of parameters efficiency and to
control undue increase of true discovery rate we used ‘one
out cross validation’. Even with cross validation, classifica-
tion of gas sand vs. brine sand or shale lead to true discov-
ery rate exceeding 80% and false discovery rate smaller than
20% with several high efficiency parameters.
3. We constructed and tested three types of aggregated param-
eters that were built as combinations of other parameters.
a. One dimensional parameters with clear and straight for-
ward interpretation. Their application as classification
tool does not need advanced classification techniques.
b. Multidimensional parameters with less clear interpreta-
tion. Parameters of this type are to be used as predictors
in machine learning classification methods.
c. Two dimensional parameters in 2D cross sections were
used to visualize separation and overlap of data sets
obtained in different geologic conditions.
4. Parameter Group_A built as a combination of five parame-
ters (Vp, Vs, ρ, Vp/Vs, and Lame parameter λ) got the high-
est rank in classifi cation testing.
5. Parameters Poisson’s ratio, Group_B, λ, Vp/Vs, k/μ, and λ/μ
also have high ranking. Classification algorithms working
with these parameters produced high true discovery rate
with limited rate of false discovery.
6. Parameters Vp, ρ * μ, μ, Vs have low ranking either due to
strong variability or small differences in median values for
gas- and brine-sand or shales.
7. Results of classification of brine-filled sand versus shale were
poor with limited true discovery rate and high rate of false
discovery for all tested parameters and four classification

algorithms.
Definitions and Notations

Vp, Vs - compressional and shear velocities
K - bulk modulus
μ - shear modulus,
λ - Lame parameter,
ρ - rock density.
δ - Poisson’s ratio.
RF - Random Forest
GLM - General Linear Models
LDA - Linear Discriminant Analysis
KNN - Nearest Neighbor
References
1. Aminzadeh F., Dasgupta S., 2013. Geophysics for Petroleum Engineering.
Elsevier, 282 pp.
2. Buryakovsky, L., Chilingar G., Rieke H., and Shin S., 2012. Fundamentals of
the Petrophysics of Oil and Gas Reservoirs. Wiley. 314 pp.
3. Garotta, R.J. and Granger, P.Y., 1987, Comparison of responses of compres-
sional and converted waves on a gas sand. 57th Ann. Internat. Mtg., Soc.
Expl. Geophys., Expanded Abstracts; 627-630.
4. Goodway, W., Chen, T., and Downton, J., 1997. Improved AVO fluid detec-
tion and lithology discrimination using Lamé petrophysical parameters;
“Lambda-Rho”, “Mu-Rho”, & “Lambda/Mu fluid stack”, from P and S inver-
sions. from P and S inversions: 67th Ann. Internat. Mtg., Soc. Expl. Geophys.,
Expanded Abstracts; 183-186.
5. Goodway W. 2001. AVO and Lame’ constants for rock parameterization and
fluid detection. Recorder, 26; 39-60.
6. Dumitrescu C., 2009. Case study of a heavy oil reservoir interpretation using
Vp/Vs ratio and other seismic parameters. SEG Annual Meeting, October
25-30, Houston, Texas.
7. Dvorkin, J. 2008. Can gas sand have a large Poisson’s ratio?. Geophysics
Volume 73 (Issue 2).
8. Everett, M., 2013. Near-Surface Applied Geophysics., Cambridge University
Press. 400pp.
9. Gardner, G.H.F., Garner, L.W., and Greogory, A.R., 1974, Formation velocity
and density – The diagnostic basics for stratigraphic traps: Geophysics, 39;
770-780.
10. Margrave, G. F.,Lawton, D. C. and Stewart, R. R., 1998,. Interpreting channel
sands with 3C-3D seismic data: The leading edge, 17 (04); 509-513.
11. Murphy R.J. and Monteiro S.T., 2012. Evaluating classification techniques
for mapping vertical geology using field-based hyperspectral sensors;
Geoscience and Remote Sensing, IEEE Transactions. 50, (3066–3080)
(Issue: 8).
12. Ostrander W., 1984. Plane-wave reflection coefficients for gas sands at non
normal angles of incidence. Geophysics, 49; 10 (1637–1648).
13. Pickett, G.R., 1963. Acoustic character logs and their applications in forma-
tion evaluation. Journal of Petrol. Technol., 15, 659-667.
14. Ramos A., and Castagna J., 2001. Useful approximations for converted-wave
AVO. Geophysics, 66 (6). 1721–1734.
15. Sing T, et al., ROCR: visualizing classifier performance in R. Bioinformatics.
2005, Oxford Journals V. 21(20), 3940-3941.
8
Use of Type Curve for Analyzing
Non-Newtonian Fluid Flow Tests
Distorted by Wellbore Storage Effects
Fahd Siddiqui* and Mohamed Y. Soliman
University of Houston, Gulf Freeway, Houston TX
Abstract
Objective: This study reconsiders the problem of wellbore storage and skin dis-
tortion for the transient flow of non-Newtonian power law fluids through porous
media injection tests. A better method of pressure analysis data from such tests is
derived which corrects some mathematical inconsistencies present in the litera-
ture. Methodology: Because the problem being considered results in non-linear
boundary conditions, Laplace transform and other solution devices cannot be
used to obtain analytical solutions. Hence the finite element method was chosen
to solve the problem numerically. This study uses the finite element method to
generate type curves for the partial differential equation resulting from the flow of
non-Newtonian power law fluids through porous media and the associated nonlin-
ear boundary condition that accounts for the wellbore storage distortion and skin
effects. Results and conclusions: The mathematical reasons for the inaccuracies in
the previously presented solutions are described. A corrected step by step analysis
procedure for obtaining the formation properties, skin factor and the associated
wellbore storage constant is described using type curves. The application is aug-
mented by one real field test data, and three simulated test data examples.
Keywords: Non-Newtonian fluids, wellbore storage, fluid flow through porous

media
169
8.1 Introduction
Pressure data analysis is routinely performed on injection and production
tests to determine the reservoir properties such as permeability, and well-
bore properties such as skin and wellbore storage. This problem becomes
complicated if injected fluid is a non-Newtonian power law liquid. Correct
knowledge and evaluation of these properties is essential to develop the
petroleum reservoirs in the most efficient manner and to sustain an opti-
mum recovery of energy resources.
The flow of the non-Newtonian fluids has been successfully described
by previous researchers (Odeh & Yang, 1979; Ikoku & Ramey, 1979;
Siddiqui, Soliman, & House, 2014). Other authors (Ikoku & Ramey, 1980;
Vongvuthipornchai & Raghavan, 1987; Olarewaju, 1992; Igbokoyi & Tiab,
2007) have also described the wellbore storage and skin distortion.
The problem of wellbore storage and skin distortion is being reconsid-
ered here because the subject needs an effective mathematical treatment
to generate correct solutions. Previously presented solutions have mathe-
matical inconsistencies that must be addressed. The mathematical reasons
for those inaccuracies in the previously presented solutions are described
below.
1
∂p  ∂p  n
CD wD −  D  =1 (8.1)
∂t D  ∂rD  rD =1

1
 ∂p  n
pwD = pD − s  D  (8.2)
 ∂rD  rD =1

Ikoku (Ikoku, 1978) applied Laplace transform to Equations 8.1 and

8.2, however it was applied incorrectly because this boundary condition
is non-linear (as evident by the power on the spatial derivative of pres-
sure), and seemingly the Laplace transform was applied to the linear part
while ignoring the non-linearity. Appendix A shows the mathematical
steps leading to this oversight. Later Ikoku and Ramey (Ikoku & Ramey,
1980) presented a wellbore storage simulator based on Duhamel’s princi-
ple. As noted by Vongvuthipornchai & Raghavan (Vongvuthipornchai &
Raghavan, 1987), the use of Duhamel’s principle is analogous to lineariz-
ing the partial differential equation along-with its boundary conditions.
Analyzing Non-Newtonian Fluid Tests 171
Therefore, such application of Duhamel’s principle results in the deviation

presented in Figures 8.1 and 8.2.
Vongvuthipornchai and Raghavan (Vongvuthipornchai & Raghavan,
1987) developed numerical solutions using the finite difference method.
However, the definition of the dimensionless wellbore storage coeffi-
cient (SD in the reference equation 8.6) seems to be missing an ‘n’ in the
denominator. If the authors intentionally defined the wellbore storage
1E + 2
pwD and tD.dpwD/dtD
1E + 1
1E + 0
pwD (WBS simulator) pwD' (WBS simulator)

pwD (equation 2) pwD' (equation 2)
pwD (this study) pwD' (this study)
1E – 1
1E + 1 1E + 2 1E + 3 1E + 4 1E + 5 1E + 6 1E + 7
tD
Figure 8.1 Comparison of solutions for n = 0.3.
1E + 2
pwD and tD.dpwD/dtD
1E + 1
1E + 0
pwD (WBS simulator) pwD' (WBS simulator)

pwD (equation 2) pwD' (equation 2)
pwD (this study) pwD' (this study)
1E – 1
1E + 1 1E + 2 1E + 3 1E + 4 1E + 5 1E + 6 1E + 7
tD
Figure 8.2 Comparison of solutions for n = 0.8.

coefficient in that manner, then they had to consider placing the ‘n’ in
the denominator of the boundary condition (equation 8.16 in the refer-
ence) to ensure algebraic compatibility. Admittedly the authors assumed
a steady state viscosity profile in the boundary condition which yields
a good agreement only for n>0.8. For other values of n, the deviation
is ‘large’ (Vongvuthipornchai & Raghavan, 1987) in the transition zone
from storage dominated to radial flow. This is mainly because for values
of n closer to unity, the non-linearity is diminished. They showed that by
assuming in Equation 8.3, their solution is identical to the Laplace space
solution in Equation 8.4.
1
−1
µ ∗ ∂ pD n
= (8.3)
µa ∂rD
 2   2 
K 1− n  z  + s zK 2  z
1 
3− n 3 − n
 
3− n 3 − n

pwD =
z  2    2   2 
zK 2  z  + C D z  K 1− n  z  + s zK 2  z
3−n  
 3− n 3 − n  
3− n 3 − n

3− n

(8.4)
Recent researchers (Olarewaju, 1992; Igbokoyi & Tiab, 2007) have

used inner boundary conditions without the non-linearity (Equations
8.5 and 8.6). However, no reason is supplied for this assumption. The
resulting solution of this approximation is the same as presented in
Equation 8.4.
∂ pwD  ∂ pD 
CD − =1 (8.5)
∂t D  ∂rD  rD =1

 ∂p 
pwD = pD − s  D  (8.6)
 ∂rD  rD =1

Even though the inversion of Equation 8.4 retains certain features of the
solution (for example, unit slope line and long-time solutions are correct),
the transition between the two flow regimes would be incorrect (Figure 8.1
and 8.2). Since both the studies were aiming to generate type curves, the
transition zone would be of paramount importance to perform analysis of

tests that do not have a complete data recorded.
A comparison of the solutions of: Ikoku and Ramey’s (Ikoku &
Ramey, Well bore Storage and Skin Effects During the Transient Flow
of Non-Newtonian Power-Law Fluids in Porous Media, 1980) well-
bore storage (WBS) simulator; numerical inversion of Equation 8.4
(Vongvuthipornchai & Raghavan, 1987; Olarewaju, 1992; Igbokoyi &
Tiab, 2007); and the finite element solutions presented in this study is
displayed in Figures 8.1 and 8.2 for two different cases. This comparison
shows that for small values of n the transition region of the solution does
not match the actual (this study) non-linear boundary condition solu-
tion. However, for larger values of n the solutions yield a good match
which is congruent to reduction in non-linearity because of n being
closer to unity. However, the wellbore storage simulator solution does
not match any of the other solutions.
8.2 Objective
The objective of this study is to use numerical methods to solve the prob-
lem of non-Newtonian fluid flow through porous media in a correct way
and devise a step by step analysis methodology using type curves for esti-
mating formation permeability, wellbore storage and skin factors from
injection pressure test data. The problem is defined in the following sec-
tions as a partial differential equation (Equation 8.11) in dimensionless
form along with the associated boundary conditions (Equations 8.1 and
8.2). The resulting solution is presented as type curves that can be used
by the step by step analysis methodology which gives reliable formation
properties because it removes the mathematical inconsistencies described
above.
8.3 Problem Analysis

The mathematical equations used in this study are defined by the following
dimensionless variables (unless explicitly stated).
pwf − pi
pwD = n 1−n (8.7)
 q  rw
 2π h  λeff

t
tD = n −1 (8.8)
 q  rw3−n
nϕ ct
 2π h  λeff

C
CD = (8.9)
2π nhφct rw2
k
λeff = (8.10)

µeff
Odeh and Yang (Odeh & Yang, 1979) and Ikoku and Ramey (Ikoku &
Ramey, Transient Flow of Non-Newtonian Power-Law Fluids in Porous
Media, 1979) independently developed similar partial differential equa-
tions and proposed similar analysis techniques for practical well test anal-
ysis. Equation 8.11 is the partial differential equation proposed by Ikoku
and Ramey:
∂ 2 pD n ∂ pD ∂p
2 + = rD1−n D (8.11)
∂rD rD ∂rD ∂t D
This study uses the equation above along with proper boundary condi-
tions to generate the correct solutions (in form of type curves) for analyz-
ing tests distorted by wellbore storage.
8.3.1 Model Assumptions
1. Homogeneous and isotropic porous medium

2. Negligible gravity effects
3. Radial flow
4. Isothermal conditions prevail throughout the injection
period within the porous medium
5. Fluid has small, constant compressibility
6. The non-Newtonian behavior of the fluid can be character-
ized by a power-law relationship
8.3.2 Solution Without the Wellbore Storage Distortion

Using the inner boundary condition in Equation 8.12 (that did not involve
well-bore storage and skin effects), the solution to Equation 8.11 was
obtained in Laplace space as Equation 8.13.
 ∂ pD 
 ∂r  = −1 (8.12)
D rD =1

 2 
K 1−n z

3− n 3 − n

pwD ( z ) = (8.13)
 2 
z 3/2 K 2 z
 3 − n 
3− n
Siddiqui et al. (Siddiqui, Soliman, & House, 2014) generated dimen-

sionless curves based on numerical inversion of Equation 8.13 and showed
that the real-time solution obtained by Ikoku and Ramey was only correct
for very long times and for n<0.6.
8.3.3 Wellbore Storage and Skin Effects

To consider the wellbore storage and skin, the inner boundary condition
must be modified. A correct inner boundary condition was formulated by
Ikoku (Ikoku, Transient Flow of Non-Newtonian Fluids in Porous Media,
1978) and is given by Equations 8.1 and 8.2.
Using the partial differential equation and the above described bound-
ary conditions, commercial finite element software, COMSOL, was used to
solve for the dimensionless wellbore pressure and its derivative.
8.3.4 Solution by Mathematical Inspection

Before solving the problem by finite element method, a guess for the solu-
tion can be made by inspecting the inner boundary condition. For the
wellbore storage dominated region, the gradient of pressure, across the
sand-face, would be zero. Hence the resulting solution (Equation 8.14)
would be the usual unit slope straight line.
tD
pwD = (8.14)
CD
After the wellbore storage effects have diminished completely, the gra-
dient of pressure across the sand face would become negative unity - hence
the inner boundary condition would reduce to Equations 8.12. Therefore,
the solution would reduce to Equation 8.13 after the wellbore storage
effects have diminished completely.
8.3.5 Solution Verification

The above two features of the solution were used to check the numerical
stability and accuracy of the solution. Moreover, the log-derivative of the
pressure from the numerical solution was also subjected to these checks. It
can be observed from the type curves (Appendix B) that in fact the solu-
tion follows the unit slope line described by Equation 8.14. Moreover, for
each n value, the dimensionless pressure curves converge to the inversion
of Equation 8.13 after wellbore effects vanished for the respective values of
skin. Finally, the problem for Newtonian fluid (n = 1) was also solved using
the finite element and compared against the numerical inversion of the
well-known solution (Equation 8.15) to confirm the numerical accuracy
of the solution.
1 K 0 ( z ) + s z K1 ( z )
pwD ( z ) = (8.15)
K1 ( z ) + CD z  K 0 ( z ) + s z K1 ( z ) 
3/2
z

This verifies the presented type curves against known existing solutions.
8.4 Use of Finite Element

Commercially available finite element software, COMSOL, was used to
generate a numerical solution of the non-linear problem. The problem
was described as an axis-symmetric one dimensional flow. The mesh was
selected to be logarithmically increasing away from the wellbore. This is
because much of the pressure drop is occurring near the wellbore hence the
mesh was closely spaced in that region. The outer boundary was modeled
by letting the interval be sufficiently large so that pressure transients never

“see” the boundary.
The solutions were generated for various values of n and are included
as Appendix B. These type curves can be used to analyze the pressure test
data without having to solve the equations using numerical methods. The
following section describes the step by step method to effectively use these
type curves.
8.5 Analysis Methodology

We present a simplified analysis methodology based on type curve match-
ing. The natural log derivative is used to reduce uncertainty in analysis
and to achieve unique results. Such a derivative is defined in Equation 8.16
and is used in the following analysis methodology because it preserves the
slope of the original function. Because an inverse problem is being solved,
multiple type curves can ‘fit’ the observed pressure data. Pressure derivative
is thus useful because the data must be matched on two curves of pressure
and pressure derivative simultaneously to reduce matching uncertainty.
dpwD dp
= t D wD = t D pD ′ (8.16)
d ln t D dt D
8.5.1 Finding the n Value

The first step is to determine n. Estimate the slope (m) by plotting log
(t.dpwf /dt) vs. log (t) (Figure 8.3) from the long-time data. If the derivative
is not smooth then it is suggested to plot log (pwf–pi) vs. log (t). Use this
slope (m) to estimate the value of n:
3m − 1
n= (8.17)
m −1
Based on this value, the nearest type curve can be selected (from
Appendix B: Type Curve Charts for Various Power Law Indices). If, how-
ever long time data are not available, the value of n obtained from labora-
tory measurements can be used with caution. If a satisfactory match is not
obtained with the laboratory value, a different value of n should be used
2.4
Log [t.dpwf/dt (psi)]
2.3
p' = 58.11 t0.1648
R2 = 1
2.2
2.1
2.5 2.75 3 3.25 3.5
Log [t (mins)]
Figure 8.3 Determining n for Example 4.
to get a better match on type curve. This is because the in-situ value of n
might be different from the one obtained under laboratory conditions.
8.5.2 Dimensionless Wellbore Storage

The dimensionless wellbore storage coefficient can be found in the conven-
tional method and requires the early time data. The USL (Unit Slope Line)
is identified and, from Equation 8.14, CD=tD/pωD; hence:
q  t 
CD = (8.18)
2π nhφct rw  pwf − pi  USL
2

If a unit slope line is not present, then C (hence CD) can be calculated
from well-bore properties, but it may not be representative of the test con-
ditions. In any case the CD=0 curve can also be used to analyze the data free
from wellbore storage distortion, without calculating CD.
8.5.3 Use of Type Curves

Based on CD and n, the appropriate type curve is selected and the data is
plotted on the same size log-log grid. After plotting, horizontal and vertical
shifting is done to match the plotted data on the type curve. We suggest
shifting along the pre-determined CD USL (if it exists) and matching the
derivative first. Once this is accomplished, the nearest skin curve can be
selected from the (pwf-pi) and pwD match. Interpolation or graphical esti-
mation of skin is also recommended. Selecting the nearest skin value is
recommended only if the curves are undistinguishable.
8.5.4 Match Point

Once a satisfactory fit is obtained by horizontal and vertical shifting, a
match point is chosen from any point on the graphing paper. By record-
ing (t, tD) and (pwf-pi, pwD) and using the definitions of dimensionless
variables:
n
 q  1−n  pwD 
λeff = r (8.19)
 2π h  w  pwf − pi 
MP
n −1
 q   tD 
λeff = nϕ ct rw3−n (8.20)
 2π h   t  MP

Either of the above equations can be used to determine the effective

mobility. It should be noted that these equations do not yield an estimate
independent from each other, and reproduce the same value of λeff .
After finding the effective mobility (λeff) and skin (s), effective mobil-
ity can be split into permeability and effective viscosity if the consistency
parameter (H) for the non-Newtonian fluid is known from laboratory
experiments. The expression for effective viscosity in porous media is
given by (Bird, Stewart, & Lightfoot, 1960; Savins, 1969):
n
H   3n + 1   1−n
µeff = 3 (150 kϕ ) 2 (8.21)
12   n  
By substituting into effective mobility λeff (Equation 8.10), and solving

for permeability:
2
n
 H   3n + 1   1−n  1+n
k =  λeff 3  (150kϕ ) 2  (8.22)
 12   n   
If a satisfactory match is not obtained, finer adjustments to the value

of n should be made, because simultaneous matching on pressure and
pressure derivative should be given preference. It should be noted that
the in-situ values of H and n might be different from the ones obtained
with laboratory experiments. Even though the value of n can be estimated
from the pressure data with the presented technique, a value H may still be
needed to calculate permeability. The value of H obtained from laboratory
experiments may be used to estimate the value of permeability. However,
it is suggested to fine tune the value of H to obtain permeability. This is
necessary in the interest of getting parameters that match better with the
field experience.
8.5.5 Uncertainty in Analysis

Like all pressure transient analysis problems, the quality of analy-
sis depends on the quality and accuracy of the input parameters: net
pay thickness, porosity, compressibility, power law index, consistency
parameter etc. The values of n and H should be measured as accurately
as possible through laboratory experiments. Even though the proposed
analysis technique can calculate the in-situ value of n (which reduces
uncertainty), the value of H still needs to be measured accurately to eval-
uate correct permeability of the medium. Pressure transient analysis is
an inverse problem and therefore there is always a chance of non-unique
interpretation. The additional uncertainty, arising from non-Newtonian
behavior, has been mitigated by including pressure derivatives in this
analysis technique.
8.6 Test Data Examples

The above discussed technique was applied to one real field data and
three simulated data. The real field data (Ikoku, Practical Application of
non-Newtonian Transient Flow Analysis, 1979) and one of the simulated
data (Vongvuthipornchai & Raghavan, 1987) were obtained from the liter-
ature. Two of the simulated data were generated using the equations in this
study solved using the finite element method.
Example 1: Real Field Data

Example 1 is a real field injection/fall off data obtained from the paper pre-
sented by Ikoku (Ikoku, Practical Application of non-Newtonian Transient
Flow Analysis, 1979). Table 8.1 shows the input parameters used for the
analysis.
This example did not have sufficient early time data to obtain CD
correctly. However, this is where the power of type curve matching
plays an important role. CD was obtained after carefully matching the
pressure derivative curve first on the radial regime and then finding
the closest CD curve to match the recorded data. Figure 8.4 shows the
1E + 2
pwD and tD.dpwD/dtD
1E + 1
pwD test data pwD type curve

pwD' test data pwD' type curve
USL
1E + 0
1E + 3 1E + 4 1E + 5 1E + 6 1E + 7
tD
Figure 8.4 Matching cart for Example 1.
Table 8.1 Input parameters

for Example 1.
Input parameters
ϕ 0.25
ct 1.31E-09 Pa–1
rw 7.62E-02 m
h 9.144 m
q 9.20E-04 m3/s
H 0.019 Pa.sn
Table 8.2 Comparison of analyses for Example 1.

Parameter This study (Ikoku, 1979) (Igbokoyi & Tiab, 2007)
n 0.45 0.40 0.52
CD 633 231
C (m3/Pa) 5.28E-06 1.26E-08
λeff(m(1+n)/ 1.04E-08 1.82E-08 1.20E-09
Pa.s)
k (md) 26.6 21.0 33.6
s 4.5 4.5 4.05
resulting match. Table 8.2 summarizes the results from three analyses.
Because of the differences in the transition zone (as described above)
the resulting CD values are different. However, since we used the correct
(non-linear) boundary condition, the CD value presented in this study
is correct.
8.6.1 Match Point

After type curve matching following match point was obtained:
 pwD  5
 pwf − pi  = 1.69 × 10 (8.23)
MP
Example 2: Simulated Data 1

Example 2 has been obtained from the published literature. Table 8.3 dis-
plays the input parameters for this example.
Table 8.4 shows the results of our analysis. It should be noted that
Vongvuthipornchai and Raghavan used a CD of 100; however, because of
the definition difference (described in introduction section) the obtained
values differ by a factor of n. The match obtained in Figure 8.5 is not sat-
isfactory for the transition region because this data was simulated using
a linearized boundary condition, however all other analyses match com-
pletely with the simulated data.
Table 8.3 Input parameters for

Example 2.
Input parameters
ϕ 0.1978
ct 1.45E-09 Pa–1
rw 6.09E-02 m
h 15.24 m
q 3.68E-04 m3/s
H 0.0314 Pa.sn
Table 8.4 Comparison of analyses for Example 2.

(Vongvuthipornchai
Parameter This study & Raghavan, 1987) (Igbokoyi & Tiab, 2007)
n 0.50 0.50 0.50
CD 208 100 212
C (m3/Pa) 2.85E-06 1.37E-06 1.08E-08
λeff (m(1+n)/ 5.10E-09 3.7E-09 1.30E-09
Pa.s)
k (md) 60.4 60.0 60.0
s 4.0 4.0 2
8.6.2 Match Point

 pwD  4
 pwf − pi  = 9.48 × 10
MP
1E + 2
pwD and tD.dpwD/dtD
1E + 1
pwD test data pwD type curve

pwD' test data pwD' type curve
USL
1E + 0
1E + 2 1E + 3 1E + 4 1E + 5 1E + 6 1E + 7
tD
Figure 8.5 Matching chart for Example 2.

Data for injection of polymer into the reservoir with known characteristics
(given in Table 8.5) was generated using the correct boundary conditions
from the finite element software. Only early time data was generated delib-
erately to demonstrate the analysis of such an instance. Since the long-time
data (from radial flow) are not available, the power law index n would need
to be obtained from routine laboratory measurements. Since this data is
simulated, it is assumed that the n was available with relative accuracy. If
Table 8.5 Input parameters for Example 3.

Input parameters
ϕ 0.13
ct 4.35E-09 Pa–1
rω 6.10E-02 m
h 33.2 m
q 4.97E-05 m3/s
n 0.8
H 0.2 Pa.sn
such an n value is not available, it would have to be assumed and would

cause much uncertainty in analysis.
8.6.3 Analysis Recommendations

Type curve matching can be performed by obtaining the CD value first
(using the assumed/measured n = 0.8). By selecting the appropriate chart,
simultaneous matching on pressure and pressure derivative should be car-
ried out on all available data from the transition zone. The same should be
attempted for nearby (±0.1) n values. Such exercise would result in a range
of formation properties and it would be up to the analyst to make the best
use of the scarce data and report properties that match better with the field
experience. It should be noted that a mere change in n of 0.1 can cause up
to 25% error in analysis. This error occurs not only for the early time data,
but also for the radial flow regime (calculation of permeability and skin).
This is expected because n appears as an index in the partial differential
(Equation 8.11) describing the fluid flow. Therefore, obtaining the correct
value should be stressed. The uncertainty can only be reduced partially by
employing pressure derivative. The uncertainty is inherent in the problem
itself and does not depend on analysis technique because of partial record-
ing of pressure data.
8.6.4 Match Point

 pwD  5
 pwf − pi  = 9.40 × 10
MP
Figure 8.6 and Table 8.6 show the match and results respectively.

This data was generated for the same way as in Example 3. Table 8.7 lists
the formation and fluid properties. The n value was determined from the
plot of log (t.dpwf/dt) vs. log (t) displayed in Figure 8.3.
1E + 2
pwD and tD.dpwD/dtD
1E + 1
1E + 0
1E – 1
1E + 0 1E + 1 1E + 2 1E + 3 1E + 4 1E + 5 1E + 6
tD
Table 8.6 Simulation input vs. analysis results for

Example 3.
Parameter Simulated Analysis (± 25%)
n 0.80 0.80±0.1
CD 100 1.15E+02
C (m3/Pa) 9.45E-06 1.09E-05
λeff(m(1+n)/Pa.s) 3.06E-12 3.06E-12
k (md) 23.0 23.0
s 10.0 10.0
8.6.5 Analysis Recommendations

It should be noted that determining n depends on long time data as well as the
smoothness of data. This example shows injection data for 45 hours. With the
availability of this long-time data n can be estimated from the slope. Afterwards
the usual steps of matching and obtaining a match point can be used.
8.6.6 Match Point

Table 8.7 Input parameters for Example 4.

Input Parameters
ϕ 0.15
ct 2.90E-09 Pa–1
rw 7.62E-02 m
h 22.9 m
q 1.84E-04 m3/s
H 0.1 Pa.sn
1E + 2
pwD and tD.dpwD/dtD
1E + 1
1E + 0
1E – 1
1E + 1 1E + 2 1E + 3 1E + 4 1E + 5 1E + 6 1E + 7
tD
Table 8.8 Simulation input vs. analysis results

for Example 4.
Parameter Simulated Analysis

n 0.60 0.605
CD 10 1.04E+01
C (m3/Pa) 3.75E-07 3.95E-07
λeff(m(1+n)/Pa.s) 2.99E-10 2.85E-10
k (md) 50.0 51.0
s 10.0 10.0
 pwD  5
 pwf − pi  = 3.48 × 10
MP
Figure 8.7 and Table 8.8 show the match and results respectively.
8.7 Conclusion
The problem was reconsidered for wellbore storage effects on the flow of
non-Newtonian power law fluids through porous media to correct some of
the inaccuracies present in the literature. A new analysis methodology was
developed, based on type curves, which gives reliable results and its appli-
cation was demonstrated for a few cases. We stress that this new method
be used for analyzing power-law fluid flow pressure tests to obtain accurate
results.
Nomenclature
ct Total compressibility, Pa-1
C Wellbore storage coefficient, m3/Pa
CD Dimensionless wellbore storage coefficient
h Formation thickness, m
H Consistency (power law parameter), Pa.sn
k Permeability, m2 [1 md=9.86923x10-15 m2]
Kv Modified Bessel function of the second kind of order v
n Flow behavior index (power law parameter)
pD Dimensionless pressure drop, non-Newtonian fluid
pi Initial Pressure, Pa
pwD Dimensionless pressure drop at wellbore, non-Newtonian fluid
pwD’ Dimensionless wellbore pressure log-derivative, non-Newtonian fluid
pwf Flowing wellbore pressure, Pa [1 psi=6894.76 Pa]
q Flow rate, m3/s
rD Dimensionless radial distance
rw Wellbore radius, m
s van Everdingen-Hurst skin factor
tD Dimensionless time, non-Newtonian
USL Unit Slope Line
z Laplace parameter
λeff Effective mobility, mn+1/Pa.s

µ* Characteristic viscosity, Pa.s
μa Apparent viscosity, Pa.s
μeff Effective viscosity, Pa.sn.m1-n
Ф Porosity
References
R.B. Bird, W.E. Stewart, and E.N. Lightfoot, Transport Phenomena, John Wiley &
Sons Inc, New York City (1960).
A.O. Igbokoyi and D. Tiab, New type curves for the analysis of pressure transient
data dominated by skin and Wellbore Storage—Non-Newtonian fluid. SPE
(2007, January).
C. Ikoku, Transient Flow of Non-Newtonian Fluids in Porous Media, Stanford
University (1978).
C. Ikoku, Practical application of non-Newtonian transient flow analysis. SPE
(1979).
C. Ikoku and H.J. Ramey, Transient flow of non-Newtonian power-law fluids in
porous media. SPE Journal (1979).
C. Ikoku and H.J. Ramey, Well bore storage and skin effects during the transient
flow of non-Newtonian power-law fluids in porous media. SPE Journal
(1980).
A.S. Odeh and H.T. Yang, Flow of Non-Newtonian power-law fluids through
porous media. SPE Journal (1979, June).
J.S. Olarewaju, A reservoir model of Non-Newtonian fluid flow. SPE (1992).
J.G. Savins, Non-Newtonian flow through porous media. Industrial and Engineering
Chemistry, 18–47 (1969, October).
F. Siddiqui, M.Y. Soliman, and W. House, A new methodology for analyzing Non-
Newtonian fluid flow tests. J. Petrol. Sci. Eng. 173–179 (2014, December).
S. Vongvuthipornchai and R. Raghavan, Well test analysis of data dominated
by storage and skin: Non-Newtonian power-law fluids. SPE Formation
Evaluation (1987, December).
Appendix A: Non-Linear Boundary Condition

and Laplace Transform
The Laplace transform of a function is given by:
∞

∫
p(z) = L ( p(t) ) = e ez p(t)dt
0
Ikoku (Ikoku, Transient Flow of Non-Newtonian Fluids in Porous

Media, 1978) applied the transform on Equation 8.1 as follows:
 1

∂ pwD  ∂ pD  n
L  CD −  = L(1)
 ∂tD  ∂tD  rD =1 
 

 1

 ∂ pwD    ∂ pD  n 
L  CD − L  = L(1)
 ∂tD    ∂rD  rD =1 
 

This step is correct because of the linearity of the transform; however,

the following step is an oversight on author’s part:
1
 ∂ pwD    ∂ pD   n
CD L  − L = L(1)
 ∂t D    ∂rD   r =1
D
1
 dpD  n 1
CD zpw ( z ) −  =
 drD  rD =1 z
This is because
1
 1

 ∂ pD n   ∂ pD   n
L  ≠ L
  ∂rD  rD =1    ∂rD   rD =1
 
Since
1
 1
 ∞ 1
n  ∞ ∂p  n  dp  n
1
 ∂ pD n  ∂ pD 
L 
∫
= e 
∫
− zt zt D D
dt ≠  e dt  = 
  ∂rD  rD =1   ∂rD  rD =1  ∂rD   drD  rD =1
  0
0

Appendix B: Type Curve Charts for Various Power

Law Indices
1E + 2
s = 30
pwD and tD.dpwD/dtD
s = 10
1E + 1
s=5
s=0
1E + 0
C = 10
D CD = 102 CD = 103 CD = 104
1E – 1
1E + 0 1E + 1 1E + 2 1E + 3 1E + 4 1E + 5 1E + 6 1E + 7 1E + 8 1E + 9
tD
Figure B1 Pressure and pressure derivative type c.
1E + 2
s = 30
pwD and tD.dpwD/dtD
s = 10
1E + 1
s=5
s=0
1E + 0
C = 10
D D C = 102 C = 103 C = 104
D D
1E – 1
1E + 0 1E + 1 1E + 2 1E + 3 1E + 4 1E + 5 1E + 6 1E + 7 1E + 8 1E + 9
tD
Figure B2 Pressure and pressure derivative type curve for n=0.4.
1E + 2
s = 30
pwD and tD.dpwD/dtD
s = 10
1E + 1
s=5
s=0
1E + 0
C = 10
D DC = 102 C = 103 C = 104
D D
1E – 1
1E + 0 1E + 1 1E + 2 1E + 3 1E + 4 1E + 5 1E + 6 1E + 7 1E + 8 1E + 9
tD
1E + 2
s = 30
pwD and tD.dpwD/dtD
s = 10
1E + 1
s=5
s=0
1E + 0
C = 10
D DC = 102 C = 103 C = 104
D D
1E – 1
1E + 0 1E + 1 1E + 2 1E + 3 1E + 4 1E + 5 1E + 6 1E + 7 1E + 8 1E + 9
tD
1E + 2
s = 30
s = 10
pwD and tD.dpwD/dtD
1E + 1
s=5
s=0
1E + 0
CD = 10 CD = 102 CD = 103 CD = 104

1E – 1
1E + 0 1E + 1 1E + 2 1E + 3 1E + 4 1E + 5 1E + 6 1E + 7 1E + 8 1E + 9
tD
s = 30
s = 10
1E + 1
pwD and tD.dpwD/dtD
s=5
s=0
1E + 0
CD = 10 CD = 102 CD = 103 CD = 104

1E – 1
1E + 0 1E + 1 1E + 2 1E + 3 1E + 4 1E + 5 1E + 6 1E + 7 1E + 8 1E + 9
tD
s = 30
s = 10
1E + 1
pwD and tD.dpwD/dtD
s=5
s=0
1E + 0
CD = 10 CD = 102 CD = 103 CD = 104

1E – 1
1E + 0 1E + 1 1E + 2 1E + 3 1E + 4 1E + 5 1E + 6 1E + 7 1E + 8 1E + 9
tD
s = 30
s = 10
1E + 1
pwD and tD.dpwD/dtD
s=5
s=0
1E + 0
CD = 10D C = 102 C = 103 C = 104

D D
1E – 1
1E + 0 1E + 1 1E + 2 1E + 3 1E + 4 1E + 5 1E + 6 1E + 7 1E + 8 1E + 9
tD
Part 3
RESERVOIR PERMEABILITY
DETECTION
9
Permeability Prediction Using
Machine Learning, Exponential,
Multiplicative, and Hybrid Models
Simon Katz1*, Fred Aminzadeh2, George Chilingar2 and M. Lackpour2
1
2
Petroleum Engineering Program, School of Engineering,
University of Southern California
Abstract
The authors present a unified methodology for permeability prediction with non-
linear multiplicative, exponential, and hybrid multiplicative-exponential nonlin-
ear models. Due to logarithmic transform of these models they may be used for
prediction with both linear regression and various machine learning methods. It
was demonstrated that enhancement of prediction accuracy is achieved with new
two-level adjustable committee machines. The new prediction methodology was
tested on data from sandstone and carbonate reservoirs with similar pattern of
improvement of prediction accuracy due to using nonlinear prediction models
and two-level committee machines.
Keywords: Permeability prediction, machine learning, exponential,

multiplicative, hybrid nonlinear models, exhaustive models search, neural
networks, two-level committee machines
9.1 Introduction
Modeling of relations between permeability and other petrophysical
parameters, such as porosity or grain size, leads usually to construction of
nonlinear permeability models. Results presented in multiple publications
197
demonstrate, that nonlinear models accurately reproduce nonlinear rela-

tions between permeability and other parameters included in the model.
On the other hand, permeability prediction is usually done with simple
additive models, so that nonlinear relationships between permeability
and predictors may be taken into account by nonlinear machine learning
method [1−5]. Herein we present review of the new methodology of per-
meability forecast with nonlinear models that may be used both with linear
multiple regression and nonlinear machine learning methods. The main
advantages of this methodology are in its efficiency of selection of a set of
predictors and improved prediction accuracy due to nonlinear nature of
the prediction model.
We tested this methodology on three datasets: dataset with data col-
lected in sandstone reservoir and two datasets with data collected in car-
bonate reservoirs. The sandstone dataset [6] is presented as an open source
at the pubs.usgs.gov website. It includes data collected in the Norwegian
sector of the North Sea. Total number of records with no missing val-
ues is 99. Quantitative parameters in this dataset are porosity, grain size
and burial depth. Ordered categorical variables qualitatively characterize
content of the following mineralogical components: microquartz, macro-
quartz, clay, and carbonates. Datasets from carbonate rocks were published
by Chilingarian et al. [7]. These are datasets with 16 records each. They
include data on permeability, effective porosity, irreducible fluid satura-
tion, and specific surface area.
Authors use extended Monte Carlo cross validation developed by Katz
et al. [8] for analysis of accuracy of permeability prediction. We character-
ize prediction accuracy by prediction bias, defined as mean absolute dif-
ference between output of the committee machine used for prediction and
actual permeability. Authors of this paper used R package “leaps” [9] for
exhaustive search of high accuracy linear regression models and R func-
tion lm() for linear regression.
9.2 Additive, Multiplicative, Exponential, and Hybrid

Permeability Models
This section gives a brief review of four prediction permeability models.
The simplest one is linear, additive model of the form:
k = a0 + a1x1 + … + aRxR (9.1)

Permeability Prediction 199
where k is permeability xr; 1 ≤ r ≤ R are recorded parameters or some func-

tions of recorded parameters. Models of this type may be equally used for
prediction with linear regression and nonlinear machine learning meth-
ods. On the other hand, mostly nonlinear models have been used so far for
analysis of relations between permeability and other parameters. We ana-
lyze in this section three types of nonlinear permeability models and their
transformation into additive model of the form of Eq. 9.1. These models
are: multiplicative, exponential, and hybrid models. Hybrid model is the
most universal and flexible. Exponential and multiplicative models may be
understood as special cases of the hybrid model.
Hybrid multiplicative-exponential permeability model is defined by
Eq. 9.2.
k = λ * M(x1,..,xn) * E(z1,..,zn) (9.2)
where k is permeability, λ is a constant, xr and zr are predictor parameters,

M(x1,...,xn) and E(z1,...,zm) are nonlinear functions of predictor parameters
of the form:
M(x1,..,xn) = (x1)q1 * … * (xn)qn (9.3)
E(z1,..,zm) = exp(α1 * z1 + … + αm * zm) (9.4)
Exponential and multiple models are of the form:
k = a * E(z1,..,zm) (9.5)
k = b * M(x1,..,xn) (9.6)
The goal of this paper is to design and test permeability forecast meth-
odology characterized by two seemingly controversial features: (a) techni-
cally, forecast relies on the models with linear structure, and (b) results of
forecast reflect complex nonlinear relationships between permeability and
predictor parameters.
According to Eq. (9.2), hybrid model is transformed into additive model
of the form (1) by log transform of both sides of Eq. 9.2:
H = log(λ) + log(M) + log(E); H = log(k) (9.7)
log(M) = q1 * s1 + … + qn * sN (9.8)
log(E) = α1 * z1 + … + αm * zM (9.9)
where sr are transformed variables of the form: sr = log(xr), 1 ≤ r ≤ n;

Therefore, variable H = log(k) is defined by the linear additive equation:
N M
H = δ0 + ∑δ ∗ s + ∑δ
r r p ∗zp (9.10)
r =1 p =1

where δ0 is intersect, δr and δp are coefficients in linear additive equation.

Log transforms of multiplicative and exponential models are special
cases of log transform of a hybrid model. They are defined by Eqs. (9.11)
and (9.12):
H = δT + q1 * s1 + … + qn * sn (9.11)
H = δE + α1 * z1 + … + αm * zm (9.12)
According to Eqs. 9.10, 9.11, and 9.12, all three nonlinear models are
transformed into linear additive models with dependent variable - loga-
rithm of permeability. Forecast methodology presented in this paper is
based on this type of logarithmic transforms. It includes three basic steps:
(a) logarithmic transform of nonlinear permeability model into linear one,
(b) forecast of values of log(permeability), and (c) transform of forecast
values of log(permeability) into forecasts of permeability via exponential
transform. Additional elements of our permeability forecast methodology
are construction and use of two-level committee machines, basis function
expansion, and exhaustive search for optimum subset of predictors.
9.3 Combination of Basis Function Expansion

and Exhaustive Search for Optimum Subset
of Predictors
Basis function expansion [5] was developed as a tool, that takes into account
nonlinear relationships between predicted parameter and predictors. It is
used to present predicted parameter as a function of base predictors and
multiple nonlinear components:
yn = aO + ∑a x (n) + ∑b q (n);
r r l l ql = θ l ( x l1 (n),, x lm (n))
r l
(9.13)
where xr(n) are base predictors, n is index of the record, xls, 1 ≥ ls ≤ lm, are
parameters from a subset of base parameters, θl are nonlinear functions
of recorded parameters, ql = θl(x11(n), ... , xlm(n)) are newly constructed
expanded base predictors. Basis function expansion allows efficient use
of linear regression, when relations between response and predictors are
nonlinear. Negative side of the use of basis function expansion may be in
a large number of constructed predictors which will result in overfitting
and instability of the forecast. To avoid potential overfitting and forecast
instability the authors combine basis function expansion with exhaustive
search for optimum subset of predictors [9] (R package ‘leaps’). Exhaustive
search presented in this package analyses possible subsets of predictor
variables and finds best subsets of predictors for approximation. The unre-
solved issue with this approach is that a subset of predictors, that guaran-
tees excellent approximation, may not produce equally accurate forecasts.
So, the authors use exhaustive search for selection candidate subsets of pre-
dictors, and then evaluate accuracy of the forecast with selected predictors
with enhanced Monte Carlo cross validation [8].
Equations 9.14 and 9.15 present models resulting from exhaustive search
methodology for additive, Eq. 9.14, multiplicative, Eq. 9.15, for three pre-
dictors in a subset.
k = δa + δ1ϕ * g2 + δ2ϕ * clay + δ3g2 (9.14)
log(k) = δm + a1log(ϕ*g) + a2log(clay) + a3log(ϕ*depth) (9.15)
where ϕ is porosity, g is grain size. One can observe, that large part of
selected predictors are nonlinear functions of recorded basis predictors.
9.4 Outliers in the Forecasts Produced with Four

Permeability Models
Individual forecasts with models of all four types may produce outliers
with values far exceeding possible permeability values in the studied area.
We review in this section the approach for identification of outliers, that
relies on the use of available geological information. This information is

defined by the value of the expected upper boundary of permeability val-
ues in the studied area, defined by the parameters coefOutlier and cutoff:
cutoff = max(k) * coefOutlier (9.16)
Here max(k) is the maximum permeability value in the training set.

Individual permeability forecast is identified as outlier, if its value is larger
than cutoff. There is some uncertainty in assigning value to parameter coe-
fOutlier. If expected maximum permeability in the studied area is about the
same as maximum permeability in the training set, then coefOutlier equals
1 or slightly higher, otherwise, it may be much larger. On the other hand,
if coefOutlier is unnecessary large, then actual outliers will be not identi-
fied. This is illustrated by Figure 9.1, that shows sorted values of individual
forecasts produced with hybrid model. Outliers are identified as individual
forecasts with values plotted above respective line. One can observe that
the number of identified outliers decreases with increase of coefOutlier.
Figure 9.1 also illustrates that forecasted permeability values and their
outliers are all positive. Individual permeability forecasts are produced
by exponential transform of log(permeability) forecasts. Results of expo-
nential transform are always positive. This is one of the advantages of the
forecasts with three nonlinear models compared to forecasts with standard
additive model, that may be equal to zero or even negative.
Decrease of percent of individual forecasts identified as outliers with
increase of coefOutlier is illustrated by Table 9.1.
According to the Table 9.1, forecast with additive model is the most
stable with no outliers among individual forecasts. Percent of identified
12000 coefOutlier=1
Individual predictions
coefOutlier=2
8000 coefOutlier=3
4000
0
0 500 1000 1500 2000 2500 3000
Index
Figure 9.1 Plot of sorted values of individual permeability forecasts generated with
hybrid model. Dashed horizontal lines are drawn at three values of parameter coefOutlier.
Table 9.1 Percent of identified outliers among individual forecasts for five values
of coefOutlier. Number of predictor variables in models of four type: 7.
Models
Cutoff
coefOutlier (mD) Additive Exponential Miltiplicative Hybrid
1 2620 0 1.26 4.01 2.04
2 5240 0 1.19 3.10 1.94
3 7860 0 0.03 1.39 1.94
4 10480 0 0.00 0.95 0.75
5 13100 0 0.00 0.07 0.71
outliers in forecasts with nonlinear models decreases with increase of

parameter cutoff.
9.5 Additive, Multiplicative, and Exponential

Committee Machines
Committee machines (CM) are manifestation of machine learning through
a set of individual forecasts, each operating independently, but their out-
puts are combined through a “gate keeper”. Some of the applications of
committee machines in oil and gas problems are reviewed in [10]. We con-
structed and used committee machines of two levels. First-level committee
machines are Monte Carlo committee machines [8] modified in this paper,
so that they may work with individual forecast produced with different
permeability models. Inputs to committee machine of the first level are
multiple individual forecasts produced with the same model and randomly
perturbed training sets. Inputs of the second level committee machines are
weighted outputs of the first level CM produced with several forecast mod-
els. Weights in committee machines of both levels are used to deal with
outliers in individual forecasts and outliers at the outputs of the first level
committee machines. Output of Monte Carlo committee machine intro-
duced in [8] is defined by the Eq. 9.17:
Y(n) =
1
M (n) ∑ F(n,m) (9.17)
m
where F(n, m) are individual forecasts produced with perturbed training

set with index m for the record with index n, M(n) is the number of indi-
vidual forecasts for the record with index n.
Inputs to additive, multiplicative, and exponential committee machines
of the first level may be written as:
Fadd(n,m) = Uadd(n,m) (9.18)
Fexp(n,m) = exp(Uexp(n,m)) (9.19)
Fmult(n,m) = exp(Umult(n,m)) (9.20)
Fh(n,m) = exp(Uh(n,m)) (9.21)
Here Uadd(n,m), Umult(n,m), Uexp(n,m) and Uh(n,m) are individual fore-

casts produced by linear regression with additive, and logarithmically
transformed exponential, multiplicative, and hybrid models.
Additive, multiplicative, exponential, and hybrid committee machines
are weighted average of individual forecasts defined by Eqs. 9.18, 9.19,
9.20, and 9.21.
As it was shown in section 9.4, there might be outliers among inputs to
the committee machine. We consider two options for dealing with outli-
ers: traditional approach of removing outliers and two methods for outlier
replacement. Results presented in the sections 9.6 and 9.7 were obtained
with removing of outliers. Efficiency of two algorithms of outlier replace-
ment is analyzed in section 9.8.
We tested two algorithms for outlier replacement: (a) with first level
committer machine and (b) with the second level CM.
Outlier replacing committee machine of the first level is defined by Eqs.
9.22, 9.23.
N N
Y (n) =
1
∑
W (n) m=1
F (n, m) ∗ w(n, m); W (n) = ∑w(n,m) + δ 2
m =1
(9.22)
where δ is a small number.
 1; F (n, m) ≤ cutoff
w(n, m) =  (9.23)
 mp / F (n, m); F (n, m) > cutoff
Parameter mp is the mean value of permeability over several nearest

neighbors of the record with index n. Results shown in Table 9.8 were
obtained with five nearest neighbors for calculation of this parameter.
Parameter δ in Eq. 9.22 was chosen equal to 0.01.
Another version of equation for weights w(n,m) designed for the use in
the second level committee machines is given by Eq. 9.24:
 1; F (n, m) ≤ cutoff
w(n, m) =  (9.24)
 0; F (n, m) > cutof

If for some records the number of outliers equals to the number of indi-
vidual forecasts, then, according to Eq. 9.22 and 9.24, first level committee
machine produces forecast value equal to zero, which is equivalent to not
producing permeability forecast. Zero forecasts are dealt with by the sec-
ond level committee machine. Its simplest form is given by Eq. 9.25:
P(n) = Y1(n) * ω(n) + Y2(n) * (1 − ω(n)) (9.25)
where n index of the record in the test set, Y1(n) and Y2(n) forecasts by
first level committee machines defined by Eqs. 9.22 and 9.24 and produced
with two different permeability models. It is also assumed, that Y1(n) is
produced either with multiplicative, exponential, or hybrid model, so that
zero values among set of values of Y1(n) are indicators of outliers. Weight
function ω(n) in Eq. 9.25 is defined as:
 1; Y (n) > 0
w(n) =  1 (9.26)
 0; Y1 (n) = 0
Level 2
- Designed to deal with zero values of 1st level

- Input: weighted outputs of 1st level CM
Level 1
- Monte Carlo committee machines

- Input: Multiple individual forecasts produced with the same model and randomly perturbed training sets.
Models can be additive, multiplicative, exponential or hybrid.
Figure 9.2 Diagram of the structure of the second level committee machine. Weights are
to deal with outliers. CM is manifestation of machine learning through a set of individual
forecasts, each operating independently, outputs are combined through a gate keeper.
It follows from Eqs. 9.25 and 9.26, that output of the second level com-
mittee machine is of the form:
 Y (n) if Y1 (n) > 0

P(n) =  1 (9.27)
 Y2 (n) if Y1 (n) > 0
Let S1 and S2 are two sets of indexes n at which forecasts Y1(n) and
Y2(n) have outliers, indicated by zero values. It follows from Eqs. 9.25 to
9.27, that the number of outliers at the output of second level committee
machine will be equal to the size of intersection of S1 and S2. Therefore, if
the size of S2 is smaller than the size of S1, then the number of outliers in
P(n) will be smaller than in the Y1(n).
Diagram shown at Figure 9.2 illustrates mechanism of forming second
level committee machine.
9.6 Permeability Forecast with First Level Committee

Machines. Sandstone Dataset
We characterize accuracy of permeability forecast by root mean square
error of individual forecasts (RMSE), mean absolute bias of forecast by
first and second level committee machine (MAB), and mean instability of
individual forecasts (MINST):
N M
RMSE =
1
N ∑ ∑(k(n) − F(n,m))
1
M
2
(9.28)
N =1 m =1
N
MAB =
1
N ∑ k(n) − Y (n) (9.29)
N =1
N M
MINST =
1
N ∑ ∑(Y (n) − F(n,m))
1
M
2
(9.30)
N =1 m =1
where k(n) is actual permeability, n -index of the record in the studied data
set, m is index of individual forecast F(n,m) for the records with index n.
According to Eq. 9.29, parameter MAB may also be interpreted as mean
absolute error of the forecast by first level committee machine.
Plots of mean absolute bias of permeability forecast done with first level
committee machines and four forecast models are shown in the Figure 9.3.
Bias was calculated for nonoutlier records with cutoff equal to max(k)*1.25.
According to Figure 9.3, bias of the forecast with additive models is higher
than bias of the forecasts with nonlinear models. On the other hand, out-
put of the first level committee machine produced with additive model
has the smallest number of outliers. According to Table 9.2, the number of
detected outliers tend to decrease with increase of the cutoff, while bias of
200 Additive
Mean absolute bias (mD)
“Exponential
Multiplicative
160
Hybrid
120
80
1 2 3 4 5 6 7
Number of predictions
Figure 9.3 Mean absolute bias of permeability forecast with additive multiplicative,
exponential, and hybrid models. Sandstone dataset.
Table 9.2 Number of outliers at the output of the first level committee
machine. Sandstone dataset.
Permeability models
coefOutlier Additive Multiplicative Exponential Hybrid
1.0 1 4 2 2
1.15 0 3 2 2
1.25 0 2 2 2
1.5 0 1 2 2
2.0 0 1 1 2
Table 9.3 Bias of the forecast with four predictors for four permeability models
and different outlier cutoffs. Number of predictors: 4.
Permeability models
coefOutlier Additive Multiplicative Exponential Hybrid
1.00 132.4 98.6 117.9 109.3
1.15 132.4 117.4 117.5 106.7
1.25 133.6 127.4 112.6 110.4
1.5 133.2 155.6 110.4 109.1
2.0 133.1 151.7 154.5 138.5
the forecast, as shown in the Table 9.3, tends to increase with increase of
cutoff.
Tables 9.4 and 9.5 present estimates of RMS error and bias generated
with hybrid model. RMS error is significantly larger than the absolute bias.
Both parameters show two trends – decrease of its value with increase of the
number of predictors and increase of its value with increase of coefOutlier.
Instability of individual forecasts and advantage of the first level com-
mittee machines is illustrated by Figure 9.4. According to this Figure 9.4,
difference between actual permeability and individual forecasts and their
range of values may be as large as 700 mD, while difference between actual
permeability and output of the committee machine is significantly smaller.
Wide range of values of individual forecasts indicates their instability.
Table 9.4 Absolute bias of the output of first level committee machine produced
with hybrid model.
Cutoff
Number of
predictors 1 1.25 1.5 2 100
2 134 134.8 134.7 133.9 424.4
3 113.2 141.8 145.1 143.5 239.6
4 106.8 108.4 149.8 151.9 232.1
5 104.1 103.1 103.3 158.4 265.5
6 96.2 96.8 97.5 159.4 250.5
7 106.4 104.4 104.4 157.8 248.4
Table 9.5 Root Mean Squared error (mD) for individual forecasts with hybrid
models.
coefOutlier
Number of
predictors 1 1.25 1.5 2 3 100
2 276.4 289.9 285.7 288.4 290.3 1972.4
3 259.2 251.4 299 401.1 433.7 995.9
4 258.8 251.2 279.6 367.7 525.5 928.5
5 227.5 240.9 211 242.9 634.8 1201.9
6 213.6 218 219.8 215.2 727.6 1167.2
7 235.7 250.4 240.6 245.4 489 1163.1
Table 9.6 illustrates pattern of changing of mean instability of individual

forecasts with changing number of predictors in the model and changing
parameter coefOutlier. Similarly to the patterns of change of RMSE and
mean absolute bias of the forecasts, men instability tends to increase with
increase of coefOutlier. Specific feature of parameter mean instability is its
increase with increase of the number of predictors.
Commitee machine output

3000 Permeability
its forecasts (mD) Individual forecasts
Permeability and
2500
2000
1500
1 2 3 4 5 6
Index
Figure 9.4 Individual forecasts, output of the first level committee machine, and actual
permeability in six records from sandstone data set. Number of predictors is 7, coefOutlier
is 1.25, hybrid permeability model.
Table 9.6 Mean instability of individual forecasts with hybrid model for the
records with permeability exceeding 700 mD. Sandstone dataset.
coefOutlier
Number of
predictors 1 1.25 1.5 2 3 100
1 34.4 39.8 47.2 62.9 63.1 60.2
3 75.6 98.3 84.2 122.7 108.5 172.2
5 91.8 93 102.1 94.3 131.9 192.2
7 103.1 133.1 115.4 144.1 173.1 173.6
9.7 Permeability Prediction with First Level

Committee Machines. Carbonate Reservoirs
Figures 9.5 and 9.6 illustrate strong difference in prediction accuracy pro-
duced with linear additive and two nonlinear models. Forecast with addi-
tive model has higher bias compared to bias of the forecast with two other
models.
High accuracy of the forecast with two nonlinear models and first level
committee machine is illustrated by Figure 9.7. Forecasts were done with
carbonate dataset Kuybyshev with three predictors. One can observe
excellent forecast results done with multiplicative model. The forecast with
the exponential model shows some differences with actual permeability,
while the forecast with additive model is significantly different from pre-
dicted permeability.
400 Forecast models

Additive
Multiplic ative
300
Forecasts bias (mD)
Exponential
200
100
0
1 2 3 4 5
Number of predictors
Figure 9.5 Bias of permeability forecast by the first level committee machine. Outlier
cutoff equal max(permeability)*1.25. Kuybushev dataset.
700 Forecast models

Additive
Multiplic ative
Forecasts bias (mD)
500 Exponential
300
100
0
1 2 3 4 5
Number of predictors
Figure 9.6 Bias of permeability forecast by the first level committee machine. Outlier
cutoff equals max(permeability)*1.25. Central Asia carbonate dataset.
2500
Permeability
Forecast with exponential model
Forecast with multiplicative model
2000
Forecast with additive model
Permeability (mD)
1500
1000
500
0
5 10 15
Index of the records
Figure 9.7 Permeability forecast with multiplicative, exponential, and additive models.
Kuybushev dataset.
9.8 Analysis of Accuracy of Outlier Replacement by

The First and Second Level Committee Machines.
Sandstone Dataset
We analyze in this section efficiency of outlier replacement by the first and
the second level committee machines. Outliers in the output of the first
level committee machine (Eq. 9.22) were suppressed using weights w(n,m)
defined by Eq. 9.23 with five nearest neighbors from the training set used
for weights calculation. Committee machine of the second level was con-
structed according to Eq. 9.25 with forecasts by additive, and multiplicative
permeability models and weights w(n,m) defined by Eq. 9.24. Forecast with
additive and multiplicative models have respectively zero and four outli-
ers. Therefore, output of this committee machine includes all not outlier
forecasts with multiplicative model and four forecasts produced with the
additive model, that replace outliers in the multiplicative model. Efficiency
of such replacement is illustrated by Table 9.7. It shows actual permeability
values, values of outliers, and their replacements by the first and second
level committee machines. One can observe significant improvement of
forecast accuracy due to replacement of outliers by both methods, with
forecast bias decreased by about five times with both outlier replacement
methods.
Table 9.7 Outliers of permeability forecasts with multiplicative model and their replacements at the output of the second level
committee machine.
Index of the records Actual Forecast Absolute bias after replacement Absolute bias due to replacement
with forecast permeability outliers of outlies by the first level of outliers by second level
outliers (mD) (mD) committee machine (mD) committee machine (mD)
4 1013 3358 1761 1337
76 1019 2821 998 1499
84 2144 3259 1636 1452
85 2620 9324 1583 1817
Mean Absolute Bias 2882 578 550
mean(absoluteBias ) 1.97 0.35 0.34

mean( permeability )
2500 Permeability
Regular forecast values
Permeability (mD)
Replacement of outliers
1500
500
0
0 20 40 60 80 100
Index
Figure 9.8 Permeability forecast with second level committee machine.
Figure 9.8 illustrates efficiency of outlier replacement by the second

level committee machine. It shows ordered permeability values, respective
nonoutlier forecasts, and outliers replacements. Outliers are beyond the
range of permeability in the Figure 9.6. Their replacements, marked by tri-
angles, are close to the actual permeability.
9.9 Conclusion
The authors introduced new methodology for permeability forecast with
nonlinear permeability models of three types: multiplicative, exponen-
tial, and hybrid multiplicative-exponential. It includes three basic steps:
Present predicted parameter as a

function of base predictors and Check accuracy of prediction by
multiple nonlinear components Exhaustive
search prediction bias
Limitation: possible overfitting and Randomly create a pair of test set-
instability of the forecast training set to generate multiple
Use equations 14 and 15
predictions of 1 sample
Find optimum set of base predictors
Basis function R package leaps Extended monte
expansion Limited: not necessarily best predictors carlo cross
Use equation 13 validation
2-Level adjustable
committee machine
Figure 9.9 Workflow for permeability prediction.

(a) logarithmic transform of nonlinear permeability model into linear one,

(b) forecast of values of log(permeability), and (c) transform of forecasted
values of log(permeability) into forecast of permeability via exponential
transform. Figure 9.9 illustrates the work-flow for the permeability predic-
tion process.
Two levels committee machines are designed for enhancement of accu-
racy of the forecasts and for dealing with forecasts outliers. The first level
CM are Monte Carlo committee machines, modified to work with three
nonlinear models and to deal with forecasts outlies. Inputs to the second
level committee machines are the outputs of committee machines of the
first level.
Developed permeability forecast methodology was tested on one sand-
stone and two carbonate datasets. The authors demonstrated, that accuracy
of permeability forecast with multiplicative, exponential, and hybrid mod-
els is higher compared to accuracy of the standard linear additive model,
with smallest bias achieved with exponential and hybrid models and with
number of predictors from 5 to 7.
Notations and Definitions
k- permeability, ϕ-porosity; g - grain size. n index of the record in the

training and test sets, xr predictor variable used in construction of permea-
bility model, H- logarithm of permeability, clay, microqurtz, macroquartz –
ordered factor variables with values 1 to 7. Increase of factor variable value
means increase in content of respective component. ql(n) - synthetic pre-
dictor variable, constructed as nonlinear function of one or several base
predictors.
F(n,m) - individual forecast for the record with index n produced with
the use of perturbed training set with index m., Y(n) - output of the first
level Monte Carlo committee machine, w(n,m) - weights in first level com-
mittee machine used to suppress outliers. P(n) - output of the second level
committee machine, w(n) - weights in the second level committee machine
used to suppress outliers.
Uadd(n,m), Umult(n,m), Uexp(n,m) and Uh(n,m) are respectively individual
forecasts produced with additive, and logarithmically transformed expo-
nential, multiplicative, and hybrid models. Fadd(n,m), Fmult(n,m), Fexp(n,m)
and Fh(n,m) - inputs to additive, multiplicative, and exponential committee
machines of the first level.
References
1. C. Fung, K. Wong, and H. Eren, Modular artificial neural network for predic-
tion of petrophysical properties from well log data. IEEE T. Instrum. Meas.
46(6), 1295–1299 (1997).
2. A. Bhatt and H. Helle, Committee neural networks for porosity and permea-
bility prediction from well logs. Geophys. Prospect. 50, 547–674 (2002).
3. M. Nikravesh and F. Aminzadeh, Past, present and future intelligent res-
ervoir characterization trends (editors’ view points). J. Petrol. Sci. Eng. 31,
67–79 (2001).
4. Z. Huang, J. Shimeld, M. Williamson, and J. Katsube, Permeability predic-
tion with artificial neural network modeling in the Venture gas Feld, offshore
eastern. Geophysics 61(2), 422–436 (1996).
5. T. Hastie, R. Tibshirani, and J. Friedman, The Elements of Statistical Learning:
Data Mining, Inference and Prediction, 2nd edition. Springer-Verlag, New
York, p. 745 (2009).
6. N. Aase, P. Bjorkum, and P. Nadeau, The effect of grain-coating microquartz
on preservation of reservoir porosity. Am. Assoc. Petr. Geol. B. 80(10), 1654–
1673 (1996).
7. G. Chilingarian, J. Chang, K. Bagrintseva, Empirical expression of permea-
bility in terms of porosity, specific surface area, and residual water saturation
of carbonate rocks. J. Petrol. Sci. Eng. 4(4), 317–322 (2006).
8. S. Katz, F. Aminzadeh, W. Long, G. Chilingar, and M. Lackpour, Rock per-
meability forecasts using machine learning and Monte Carlo committee
machines. J. Sustain. Energy Eng. 4(2), 182–200 (2016).
9. P. Cortez, Data Mining with Neural Networks and Support Vector Machines
Using the R/rminer Tool. Chapter in Advances in Data Mining. Applications
and Theoretical Aspects. Volume 6171 of the series Lecture Notes in
Computer Science, 572–583 (2010).
10. F. Aminzadeh and P. de Groot, Neural Networks and Soft Computing
Techniques, with Applications in the Oil Industry, EAGE Publishing, p. 164
(2006).
11. R. Gholami, A. Shahraki, and A. Paghaleh, Support vector regression for pre-
diction of gas reservoirs permeability. J. Min. Environ. 2(1), 41–52 (2011).
10
Geological and Geophysical Criteria
for Identifying Zones of High Gas
Permeability of Coals (Using the
Example of Kuzbass CBM Deposits)
A.G. Pogosyan *
Kuban State University, Geophysitist, St. Uralskaya, Krasnodar, Russia
Abstract
The article reviews the results of an analysis of changes in the petrophysical prop-
erties of coals and the dynamic conditions existing during their formation. The
combined effect of physical properties and dynamic parameters on the perme-
ability of coalbeds is evaluated; their relation to the distribution of flow rates in
producing wells is determined. The feasibility of downhole multiwave VSP obser-
vations in CBM fields is substantiated.
Keywords: Coal Bed Methane (CBM), Elastic Deformation Modules (EDM),

fractures, permeability, petro physical properties, Vertical Seismic Profiling
(VSP)
10.1 Introduction
Geological exploration is increasingly focused on the non-traditional
resources, such as shale gas, shale oil, and coalbed methane. Thus, reserves
and resources of these sites will be a priority as key exploration areas, both
on land and offshore.
In recent years, developing coalbed methane resources has been a pri-
ority in Russia. Among the coalbeds methane (CBM) basins of Russia,
217
Kuzbass is the largest coalbed methane basin in the world ready for large-
scale production of methane. The basin’s recoverable methane resources
are estimated at 13 trillion m3. These methane resources occur at depths
of 1800–2000 m [1]. The first Russian pilot production field for extracting
methane gas from coalbeds was launched in 2010 at the Taldinskaya area
of the Kuzbass.
Current practice of pilot development of coalbed methane at the
Taldinskaya area in Kuzbass shows that even at high concentrations of
methane resources (resources in the area are 95 billion m3, with an aerial
density of 3 billion m3/km2), methane production may be uneconomic
because of low gas recovery.
Whereas, if the gas contained in sandstones escapes to the surface due
to reservoir pressure, special channels (fractures) must be created for its
movement within coalbeds. This approach is required because coalbed
properties differ significantly from the corresponding properties of tradi-
tional sandstone and carbonate gas reservoirs.
In addition to artificially created channels, gas recovery values are also
affected by the original natural gas permeability of coal due to fracturing.
Natural gas permeability of coalbeds is the most important property nec-
essary for commercial production of methane. Without initial natural gas
permeability, and without the ability to increase the gas recovery from
coalbeds, there is no methane production potential—despite the huge
methane resources in coalbeds.
As a result, the criteria characterizing coalbed permeability should be
considered the most important of all criteria for assessing the prospects
for developing of coal and gas production. The development of geological
and geophysical criteria for identifying the zones of high gas permeability
of coalbeds and their changes within the area is the most urgent geological
task at all stages of CBM field development, making the application of var-
ious geophysical methods a priority.
The important issue is not only to define the geological and geophysical
gas permeability evaluation criteria, but also to develop an effective meth-
odological basis for studying the physical properties of thin coalbeds in
interwell space.
The physical and geological factors affecting the distribution of well flow
rates determined from an analysis of physical and mechanical properties
of coals and dynamic conditions of their occurrence are examined below
using the example of the Taldinskoye Field of the Kuznetsk Basin, com-
pared with the results of production tests of coal/methane wells.
Geological and Geophysical Criteria 219
10.2 Physical Properties and External Load

Conditions on a Coal Reservoir
Coal and gas production experts consider porous (capacitive and filtra-
tion) coal-bed media as a set of fractured and mesoporous systems of coal
matrix. The average size of pores in the matrix is estimated to be 0.5–2 nm
(5–20 A). However, in these pores, almost in the coal matrix itself, the great
bulk of methane is contained, forming a “solid coal and gas solution” in the
opinion of many experts [2, 3].
Coal has a dual pore system of macropores and micropores [4]. Unlike
conventional gas deposits, coalbeds have two kinds of permeability that
“coexist:” microporous permeability in the coal at the matrix level and
macroporous permeability at the level of fractures and macropores. The
macropores system in coal is also referred to as the cleat/fracture system
[5]. Microporous coal permeability is very low: about 10–9 μm2 (10–6 mD),
according to the data of Halliburton Co, and about 10–5–10–7 μm2 (10–2–
10–4 mD), according to Russian researchers [2].
Macroporous permeability is a measure of the degree of coalbed frac-
turing. According to Ayruni et al. [2], macroporous permeability of pro-
ductive coalbeds in the San Juan Basin, US, ranges from 5 to 30 mD, and
higher values are reached in some areas. This is suffi cient for cost effective
methane production using sorbed gas intensification technologies, namely
hydraulic fracturing. Natural coal fracturing and the predisposition of coal-
beds to form secondary fractures under the influence of natural endoge-
nous (carbonization) and exogenous (geodynamic, neotectonic) processes,
as well as man-induced factors (hydraulic fracturing), are determined by
their properties such as elastic deformation moduli (EDM), including the
volume compressibility factor K, Young’s modulus E and the Lamé con-
stants m [6, 7].
Correlations of elastic deformation moduli, densities, and velocities of
seismic waves make it possible to obtain elastic parameters from data of
the multiwave acoustic velocity log and the multiwave VSP by formulas
given below on the basis of the elasticity theory:
- Young’s modulus:
v s2 (3 ∗2p −4 ∗ v s2 )
E = ρ∗ (10.1)
Vp 2 − Vs 2

- shear modulus:
G = ρ * Vs2 (10.2)
- bulk modulus K (compressibility reciprocal)
4
K = ρ ∗ (Vp 2 − Vs 2 ) (10.3)
3
where ρ - density of the medium, g/cm3, Vp – velocity of longitudinal

waves, Vs – velocity of transverse waves S, dim Vp,Vs = LT-2 [Vp,Vs] = m/s.
Dimension and unit of elasticity moduli: dim E,G,K = L-1 MT-2 [E,G,K]
= 1 Pa.
However, the actual permeability of the coalbed may vary considerably
under the influence of the external load on the reservoir and the overbur-
den pressure.
Thus, when pressure drops and the coalbed is no longer in the equi-
librium state, the gas flow begins with the gradual transition of gas mol-
ecules from the absorbed to the adsorbed state, and then to the free
phase. Furthermore, the most intense methane flow, both laminar and
turbulent, occurs in submacropores (10–5 to 10–4 cm), macropores (10–4
to 10–2 cm), and pores and fractures of both endogenous and exoge-
nous genesis visible to the eye (larger than 10–2 cm). Thus, even a small
external load on the coalbed primarily affecting the sizes of these pores
leads to a significant reduction in volumetric flow rate and a drop in coal
permeability.
Ayruni and Malyshev [6] showed that the volume of the smallest pores
(12.5–75A) remained almost unchanged during the compression of coal at
20 MPa. At the same time, the volume of large pores forming the filtration
capacity was reduced 27 times.
Thus, the effect of external pressure on different size pores is not the
same; larger pores (slit-shaped pores, fractures of endogenous and exoge-
nous origin) are more susceptible. However, the sorption capacity of coal-
bed methane, determined mainly by micropore volume, remains fairly
constant with changes in external load on the reservoir; these pores are
resistant to external influence due to their configuration (spherical).
Thus, the physical properties and the coal overburden pressure are among
the most important factors controlling the permeability of coals, but are
poorly known. It is likely that the above factors will have a decisive influ-
ence on the coalbed gas recovery zones (all other conditions being equal).
Therefore, these factors must be examined by using available geophysical

data.
An experimental research site for developing technologies for CBM
section studies and developing a scientific and methodological basis for
field development has been established in the Taldinskaya area of the
Kuzbass Basin. Analysis of elastic deformation moduli (EDM) of the coal
and dynamic conditions of a set of pilot production wells drilled in the
Taldinskaya area was made, based on longitudinal and transverse wave
propagation velocities (full-wave sonic logging) to assess their relationship
with the distribution of flow rates.
The calculated EDM values of coalbeds show a significant spread not
only in the section of an individual well (Figure 10.1 a, b, c), but also over
the area. This indicates the heterogeneity of physical properties of coal-
beds. Each coalbed layer within the same well section is characterized by
different EDM values layers of same metamorphism level. Thus, each sec-
tion exhibits an alternation of coalbeds with different filtration capacity.
Studies by Zemtsova and Pogosyan [7] established that the physical
properties of coalbeds are mainly determined by the coefficient of volume
compressibility K (parameter K classifies coals by their moisture and ash
contents). The influence of moisture and ash content on the Young’s modu-
lus of elasticity E and Lamé constants and μ is minimal. Physical properties
are characterized by a significant spread in the section of wells penetrating
the section at a depth interval of 150–800 m.
When the EDM values are compared with well flow rates, it is found that
the same values of Young’s modulus E and shear modulus G are typical of
coalbeds penetrated by wells with both high (wells 2, 4, 7) and low flow
rates (wells 3, 4, 6, 8, 9, 10), whereas the volume compressibility factor K is
relatively low (0.1–0.2 GPa–1) for high rate wells and relatively high for low
rate wells (see Figure 10.1 c).
This suggests a more complex multivariate relationship of filtration
properties of coals with the physical state of the medium, taking into
account not only its intrinsic properties (physical properties, and coal
quality indicators, i.e., ash content, moisture content), but also the external
dynamic conditions discussed below.
Thus, the low flow rates in coalbeds with increased compressibility
(wells 3, 4, 6, 8, 9, 10) are possibly due to the influence of overburden pres-
sure of overlying high density rocks. Under this pressure, as liquid and
gas is displaced from the pore space during production, coalbed compac-
tion occurs. The more compressible the coals, the more they are subject
to “structural” compaction (in addition to shrinkage of the coal material
itself). This process occurs at the contact of low-density rocks (e.g., coal
16
E (GPa)
14
12
10
8
6
2
0
66 NN 65 64 63 62 61 60-59 58 57 56 55? 54 NN 52a 52 51 50 48 45 44 43
(a) Index of coalbed layer
G (GPa)
7
6
5
4
3
2
1
0
66 NN 65 64 63 62 61 60-59 58 57 56 55? 54 NN 52a 52 51 50 48 45 44 43
(b) Index of coalbed layer
K (GPa-1)
0,5
The compressibility factor of high rate coalbeds
0,4 has relatively low values (0.1-0.2 GPa-1)
0,3
0,2
0,1
0
66 NN 65 64 63 62 61 60-59 58 57 56 55? 54 NN 52a 52 51 50 48 45 44 43
(c) Index of coalbed layer
Coalbed parameter value High rate wells Low rate
layers exposed to hydraulic fracturing well 2 well 5 well 9
Low rate coalbeds well 4 well 6 well 10
High rate coalbeds well 7 well 8
Figure 10.1 Distribution of elastic deformation modules of coalbeds: (a) distribution

of Young’s modulus values (b) distribution of shear modulus values (c) distribution of
volume compressibility factor values.
or clay) with high-density sandstones and siltstones, and is controlled

by the geodynamic factor. Strata with high flow rates appear to be more
resistant to the effects of external loads (wells 2, 4, 7), due to their low
compressibility.
Based on the above, it is obvious that in order to evaluate the coalbed gas
recovery potential, we must take into account not only diagnostically sig-
nificant intrinsic physical properties of coal, but also the impact of external
conditions determined by the most recent geodynamic processes. Above
all, we must assess the distribution of the external load on the coalbed.
The distribution of external load on a coalbed is one of the key indicators

defining the balance of static and dynamic methane volumes in coal and
reflecting the gas dynamic properties of the coalbed. When determining
the mechanisms of pressure formation in a coalbed, we must start from the
concepts of pressure formation as a resultant influence of many physico-
chemical (sorption-desorption), hydraulic and geological (moisture con-
tent in the coalbed) and geodynamic (overburden pressure) factors. The
first two factors mainly affect the micropores forming the static volume of
methane (sorbed gas). The effect of the geodynamic factor is substantial
and controls the distribution of gas permeability.
The results of calculating rock pressure parameters (vertical pressure,
lateral thrust) characterizing the conditions for fracture development in
coalbeds are given below. A correlation between the data obtained and flow
rates was found (Figures 10.2, 10.3, 10.4).
The calculated values of geostatic (vertical) pressure and dynamic values
of the lateral thrust made it possible to establish the following. Wells 2 and
4, characterized by elevated flow rates and relatively low coalbed compress-
ibility, were drilled with abnormally low geostatic rock pressure and ele-
vated values of lateral thrust (Figures 10.2, 10.3). Coalbeds of wells No. 3, 5,
6, 8, 9, 10 were subject to a greater geostatic load of 0.6–0.8 MPa on average
at the respective depths than those of wells No. 2, 4, 7 (Figure 10.2), while
the difference in pressure increases with depth. Thus, if near-wellbore res-
ervoir pressure drops during methane extraction, a considerable flow rate
reduction will occur due to the closing of horizontal and slightly inclined
fractures under the action of the vertical pressure component.
Geostatic pressure (MPa)
21 40
44 43-42
19 ? 43
48-45 44
39 40
45
17 High rate wells
48 43
39 well 2
15 50 44
well 4
48 45 51
13 50
44
52 well 7
45 52 51
48
53 51
52 Low rate wells
11 50 54
52 52 52 well 5
9 56
50
52
5 well 6
54 5 54
56
well 8
7 57
60-59 58 well 9
Index of coalbed layer
5 well 10
Figure 10.2 Geostatic pressure in wells.

The calculated dynamic values of the lateral thrust (λ) of coalbeds vary from
0.1 to 0.8 of the vertical overburden pressure, with high rate wells differing
from the low rate wells by high lateral thrust values of 0.5–0.8 (Figure 10.3).
Such a significant variation in λ values in rocks of the same lithotype is a better
reflection of the internal macroscopic structural features of the structure of
coalbeds; in particular, the nature and orientation of exogenous fracturing.
Thus, a comparison of data on the geostatic and lateral thrust of over-
burden pressure with the flow rates (Figures 10.2, 10.3, 10.4) and the dis-
tribution of volume compressibility factor K (Figure 10.1 c) reveals the
correlation of the “external” rock and dynamic and “internal” physical
properties of coalbeds with gas permeability.
Therefore, the effectiveness of hydraulic fracturing depends on the com-
pleteness of geological information on dynamic rock conditions and the
structure of the borehole environment, combined with the physical and
0.9
Lateral thrust
High rate wells

0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
Low rate wells
0 Index of coalbed layer
66 NN 65 64 63 62 61 60-59 58 57 56 55? 54 NN 52a 52 51 50 48 45 44 43
Figure 10.3 Distribution of the lateral thrust of coalbeds.
7000
flow rate, thousand m3/day
M 1:50 000
well 6
S N
6000
well 4
5000
well 7
well 10
well 8
well 5
4000
well 9
3000
well 2
2000
1000
0
well 4 well 2 well 7 well 5 well 8 well 6 well 10 well 9
Figure 10.4 Gas flow rates of Taldinskaya area wells.

mechanical properties of coalbeds. The probability that a well with a high

flow rate is controlled by the geodynamic scenario affects the distribution
of geodynamic compression and tension fields.
High flow rates and the presence of local “gas zones” (all other condi-
tions being equal) should be expected in areas with the smallest increase
in external load on a coalbed, which is resistant to external impact (low
compressibility) and has a developed system of optimally oriented endo-
and exogenous fractures, unless the gas has already escaped to the surface
through fractures.
Production wells with low flow rates drilled without considering the
characteristics of the mining dynamic factor and physical and mechani-
cal properties of coals were less cost effective, even if from the traditional
point of view, they were favorably located in areas of high methane density
(2–3 billion m3/km2).
Evaluation of coalbeds as commercial methane extraction sites must be
based not only on inferred methane resources, but also on the ability of
their economic recovery, taking into account the natural permeability of
the coalbed.
10.3 Basis for Evaluating Physical

and Mechanical Coalbed Properties
in the Borehole Environment
At the stage of pilot well drilling, there is a need to obtain timely and reliable
information on the distribution of gas permeability of the target coalbed
in the borehole environment and determine the boundaries of increased
gas permeability areas. Even with a large volume of well data (e.g., logging
and core data) in the area, these problems cannot be solved with sufficient
accuracy and reliability due to the obvious methodological limitations of
logging data and variability of the physical properties of coal. Detecting the
high permeability zones in a CBM section and selecting the optimum loca-
tion for coalbed penetration for subsequent hydraulic fracturing is rather
complicated.
In the study [8] there was an example of a predicted area of high accu-
mulated coalbed methane based on surface seismic data (elastic modules,
frequency spectrum) in one of coabed methan area in north China. The
given approach is an example of a qualitative interpretation of seismic data
and not providing a quantitative estimate of physical properties of coals
and permeability as more important parameters.
In the case of Kuzbass basin, the use of traditional surface multiwave

seismic survey solutions to these problems is extremely complicated due to
the complex seismological conditions (complex structure of the upper part
of the section, rugged terrain, and fine lamination). The seismic reflection
surveys conducted in the Kuzbass did not resolve the question of tracing
thin coalbeds (3–7 m., in thickness) in order to study their structure and
physical properties [7−9].
However, at present, there is a considerable accumulated experience
of down-hole seismic surveying for studying structurally complex media
using traditional wave field attributes derived from seismic records [10].
The heterogeneity of the geological section and, above all, its hydrocarbon
saturation, are clearly shown in some cases by multiwave VSP (MW VSP)
data. Attribute analysis of coal events from the zero-offset VSP surveys,
walkaway VSP surveys and surface seismic is anticipated to provide useful
information regarding the physical properties of the coal and thus, its suit-
ability for CBM development [11].
An example of implementing multiwave VSP technology in a CBM section
and evaluating the physical properties is given below. State of the art equip-
ment and monitoring technology at internal points of the medium made it
possible to obtain high-quality multiwave seismic data (Figure 10.5 a) with a
frequency range of longitudinal waves up to 125 Hz. This allowed a detailed
tie-in of seismic reflections to thin coalbeds and made it possible to study the
structure of target formation together with the identification of small ampli-
tude faults (undetectable by conventional surface seismic surveys) (Figure 10.5
b) as a main factor that affects the local distribution of gas concentration [12].
It is particularly important to note that the use of a modern three-
component downhole receiver (VSI —Versatile Seismic Imager, multi-
probe downhole seismic device from Schlumberger) allowed high-quality
registration not only of the longitudinal (PP), but also the converted (PS)
and transverse (S) incident and reflected waves. The physical properties of
the coalbeds in the borehole are quantified using the data on the speed of
propagation of longitudinal and transverse waves.
According to the VSP data, minimum values of the bulk modulus K are
found far from the well, which is indicative of increased compressibility of
the medium (Figure 10.5 c). The observed compressibility distribution in
the area made it possible to predict the possible development of high com-
pressibility coalbeds in the borehole environment far from the wells in the
southwestern sector. Thus, multiwave modification of VSP technology is
one of the possible methodological principles that can serve as a database
for quantitative evaluation of the physical properties of a section in the
borehole environment.
(a) Well 2 Well 2

NW SP2 X M SP6 SE NW SP2 X M SP6 SE
-160 -140 -120 -100 -80 -60 -40 -20 0 0 20 40 60 80 100 120 140 160 180 200 -160 -140 -120 -100 -80 -60 -40 -20 0 0 20 40 60 80 100 120 140 160 180 200
50 50
100 Bed 68
Bed 67
PS 100 Bed 68
Bed 67 PP
150 150
Bed 66 Bed 66
200 200
Bed 62-61 Bed 62-61
250 250
Bed 60-59 300 Bed 60-59
300
350 Bed 56 350 Bed 56
400 Bed 54 400 Bed 54
450 450
Bed 52 a Bed 52 a
500 500
Bed 51 Bed 51
550 550
Bed 50 Bed 50
600 600
650 650
700 700
750 750
800 800
H M H M
Sandstones Interlayered fine- and coarse-grained siltstones
Interlayered sandstones and fine-grained siltstones Bed 60-59 Index of coalbed layer
Structural interpretation Distribution of bulk modulus β = 1/K

84
H
(b) PP
(c) 16
73
H
16 SP
51HPP
PP
SP SP
10
20 30 SP
0 20
62HPP 20
0 10
Well 14
-60 0 10
30
10
CMS_2
0
20 SP
0
0
10
SP 12 1
15
51HPP
84
H
0
SP SP
PP
62HPP
17 13
73
H SP
PP
73HPP
Reflected PP wave tracing lines Well 2 EDM parameters (GPa)
Increased coalbed
Small amplitude disjuctive faults VSP shotpoints compressibility area
Figure 10.5 Results of experimental studies using the multiwave VSP technique:
(a) structure of coalbeds in the borehole environment in the wave field of longitudinal
PP and transverse PS waves (b) identification of small amplitude faults (c) distribution
of bulk modulus (inverse of compressibility factor) in the borehole environment for the
target coalbed.
The opportunity of studying the wave field at interior points of the coal
deposit made it possible to eliminate the effect of surface interference
waves, thus improving the data quality while also widening the energy
spectrum of wave fields by adding new waves of different nature and other
seismic processes to the traditional composition of useful waves. These
processes reflect the nonlinear properties of deposits, particularly its emis-
sion activity due to internal micro-structural heterogeneity.
The appearance of methods based on the construction of fundamen-
tally new physical models of real media, taking into account the seismic
non-linearity of oil and gas reservoirs, is characteristic of the development
of modern geophysics [9−13]. The tasks associated with the study of geo-
dynamic objects include fractured gas-saturated waveguide zones, rhe-
ologically weakened zones and media with a complex stress-strain state.
Thus, in terms of its studied properties, the CBM reservoir of Kuzbass has
all of these characteristics.
The presence of microstructural irregularities in coalbeds, e.g., those
significantly larger than atomic size but still small on the elastic wavelength
scale, such as microfractures, may lead to abnormal shows of nonlinear
medium properties, significantly increasing the intensity of nonlinear
acoustic parameters and substantially changing the qualitative nature
of the nonlinearity (the appearance of a distinct frequency or amplitude
dependence) itself.
Furthermore, the linear acoustic characteristics of the medium may stay
almost constant. Thus, the “structural sensitivity” (at the microscopic level)
of the medium’s nonlinear properties is significantly higher than that of lin-
ear elasticity parameters [14]. Desorption processes form mechanical force
fields affecting the coalbed, and the cumulative effect of these forces, leads
to acoustic radiation (UV MS) of sources of increased gas dynamic activity.
Obviously, studies of weak nonlinear processes in the medium are pos-
sible only with high-quality geophysical data registration. Downhole seis-
mic technology successfully tested in a CBM section can provide a reliable
basis for research on these processes in a seismic wave field.
10.4 Conclusions
Geological and geophysical criteria influencing the formation of high gas
permeability zones in coals are established and their relationship with the
distribution of flow rates of pilot CBM wells is physically justified.
The author provides technical solutions for increasing the efficiency
of the studies of the physical properties of coalbeds and broadening the
understanding of the distribution of permeability of CBM reservoirs.
Acknowledgement
The writer is greatly indebted to academician George V. Chilingarian for
his invaluable help.
References
1. A.M. Karasevich, V.T. Khryukin, B.M. Zimakov, G.N. Matvienko, S.S.
Zolotih, V.G. Natura, and T.S. Popova, Kuzbass Basin—The Largest Resource
Base of Commercial Methane Production from Coal-beds, pp. 64, Mining
Science Academy, Moscow. (2001).
2. A.T. Ayruni, M.A. Iofias, and L.M. Zenkovich, The Scientific Basis for
Determining the Gas Permeability of Coal-beds under Varying Filtration
Parameters, pp. 52, Research Institute of Comprehensive Exploitation of
Mineral Resources, Academy Sciences USSR, Moscow. (1982).
3. Y.N. Malyshev, K.N. Trubetskoy, and A.T. Ayruni, Fundamentally Applied
Methods of Solving the Problem of Coal-bed Methane, pp. 519, Mining Science
Academy, Moscow. (2000).
4. Halliburton, Chapter 2, in Coal-bed Methane: Principles and Practices,
pp. 126, R. E. Rogers, K. Ramurthy, G. Rodvelt, M. Mullen (Eds.), Oktibbeha
Publishing Co., LLC, Starkville, MS 39759, ISBN 978-0-9794084-1-0 (2008).
5. L.V. Baltoiu, B.K. Warren, and T.A. Natras, State of the art in coalbed meth-
ane drilling fluids. Drilling Completion 23, 250–257 (2008).
6. G.N. Boganik and I.I. Gurvich, Seismic Surveying, pp. 754, AIS, Tver. (2006).
7. D. P. Zemtsova and A. G. Pogosyan, The possibility of using multiwave seis-
mic surveys to assess the elastic deformation modules of CBM sections,
Abstracts of the XIth Scientifi c Conference and Exhibition “Geomodel-2009”,
Gelendzhik, EAGE, Volume 145, 145 (2009).
8. Tang, H. Seismic prospecting technique for coal bed methane accumulating
area. Procedia Earth Planet. Sci. 3, 224–230 (2011).
9. O.L. Kuznetsov, I.A. Chirkin, Yu.A. Kuryanov, G.V. Rogotsky, and V.P.
Dyblenko, Experimental studies, in Seismoacoustics of Porous and Fractured
Geological Media: Volume 2, O.L. Kuznetsov, (Ed.), pp. 320, State Research
Center of VNIIGeosystems Research Institute, Moscow. (2004).
10. E.I. Galperin, Vertical Seismic Profiling, pp. 344, Nedra, Moscow. (1982).
11. S.E. Richardson, D.C. Lawton, and G.F. Margrave, Seismic applications in
coalbed methane exploration and development, 01 (2003).
12. M. Li, B. Jiang, S. Lin, F. Lan, and J. Wang, Structural controls on coalbed
methane reservoirs in Faer coal mine, Southwest China. J. Earth Sci. 24, 437–
448 (2013).
13. D.P. Zemtsova, A.A. Nikitin, and A.G. Pogosyan, Study of fractured zones of
a CBM section of Kuzbass basin based on seismic emission of geodynamic
noise, Bulletin of Higher Educational Institutions, Geology and Exploration,
Moscow, No. 3, 51–55 (2009).
14. V.Y. Zaytsev, N.V. Pronchatov-Rubtsov, and S.N. Gurbatov, “Non-classical”
Structure-induced Acoustic Nonlinearity: Experiments and Models, pp. 223,
Novgorod, Nizhny. (2007).
11
Rock Permeability Forecasts Using
Machine Learning and Monte
Carlo Committee Machines
Simon Katz1, Fred Aminzadeh2, Wennan Long2*, George Chilingar2
and Matin Lackpour1
Russian Academy of Natural Sciences, U.S. Branch,

1
2
Abstract
We developed new concepts of extended Monte Carlo cross validation and Monte
Carlo committee machines. We subsequently used those concepts to predict per-
meability by linear regression and machine learning methods such as Neural
Networks, Support Vector machines, and Regression Tree. Among the parameters
we calculated using extended Monte Carlo cross validation are: root-mean squared
error of individual forecasts, forecast bias, correlation between forecast and actual
permeability, and forecast instability as a measure of sensitivity to perturbations
of the training set. Output of Monte Carlo committee machines is constructed
as the average of machine learning outputs generated from multiple versions of
perturbed training sets. We observed that Monte Carlo committee machines pro-
duced high stability forecasts, while individual machine learning forecasts (e.g. a
single ANN) were characterized by lower stability. Higher accuracy forecasts were
achieved when we applied machine learning methods and linear regression using
permeability models that included quantitative and categorical predictors and
second-order interactions among the predictors.
Keywords: Permeability forecast, linear regression, machine learning, Monte

Carlo cross validation, committee machines
231
11.1 Introduction
Current research in modeling and forecasting rock permeability includes
a variety of methods and models. The structure of permeability models
is defined by several factors which include a set of petrophysical param-
eters utilized as predictors, heterogeneity of lithology in the studied area,
a model that defines relations between permeability and predictors, and
methods used for model construction. The means used for rock perme-
ability forecasts are linear, log-linear regression, and machine learning
methods such as neural networks and support vector machines. An
excellent analysis of correlations amongst permeability, porosity, confin-
ing pressure, cementation, and grain size was done by AlHomadhi [1].
Results presented in this paper indicate the importance of factors other
than porosity in reliable permeability forecast. Log-linear regression
model for permeability with porosity, specific surface area, and irreduc-
ible fluid saturation were developed and analyzed by Chilingarian et al.
[2] (See Addendum). The model was tested on data from several carbon-
ate reservoir rock areas in the former USSR. One specific feature of the
regression models utilized in that paper is the presence of second-order
interactions. The inclusion of irreducible fluid saturation in the set of
predictors led to a high correlation between actual and predicted perme-
ability values. Generally, the advantage of linear regression permeability
models is in their interpretability. Their weak point is their rigid struc-
ture. More flexible but more difficult to interpret are machine learning
techniques, such as neural networks [3–6] and support vector machines
[5, 7] used for permeability forecasts. Additional enhancement of the
efficiency of machine learning permeability forecasts might be done with
committee machines [4, 8–11]. Efficiency of forecast is estimated with
different versions of cross validation. Cross validation version ‘leave one
out’ was used by Gholami [7] to validate the permeability model built
with support vector machines. Monte Carlo cross validation was pro-
posed in Qing’s paper [12] and used for estimating the number of com-
ponents in the calibration model. General review of machine learning
methods, such as neural networks and soft computing for reservoir char-
acterization, was presented by Nikravesh and Aminzadeh [13].
Authors of this paper concentrated on development and testing of two
algorithms: (a) extended version of Monte Carlo cross validation and (b)
algorithm for Monte Carlo committee machines. Although we will not
delve into the details of the ANN committee machines, we will include a
brief description from Aminzadeh and de Groot [11] in Appendix 2.
Rock Permeability Forecasts Using ML 233
Classic methods of cross validation, such as leave one out or k-fold have
estimated single parameter - forecast error. The extended version of Monte
Carlo cross validation is aimed at detailed analysis of the forecast and
includes estimation of additional parameters such as forecast bias, fore-
cast instability, correlation between forecast and target parameters, and
comparative analysis of accuracy of individual and committee machine
forecasts. Developed methods and algorithms have wide applications for
forecasts of different types of reservoir characterization. They may be uti-
lized, for example, for analysis and forecast of reservoir porosity distribu-
tion and in the forecast of trends in hydrocarbon production [14].
In this paper, the analyses of forecasts were obtained using a dataset
[15] published as open source at the pubs.usgs.gov website. That data-
set includes data from Jurassic sandstones collected from 15 wells in the
Norwegian sector of the North Sea in the depth range of 3300–4250 m.
Each record in the dataset contains a permeability value and respective
values of quantitative and categorical parameters. The total number of
records with no missing values is 99. Quantitative parameters in the data-
set are porosity, grain size and burial depth. Categorical variables in this
dataset qualitatively characterize content of the following mineralogical
components: microquartz, macroquartz, clay, and carbonates. They have
the following discreet values: 0 - not observed, 1 - very minor, 2 - minor,
3 - medium, 4 - much, 5 - very much.
Authors of this paper use R rminer package [5] to run machine learning
methods and R function lm() for linear regression.
11.2 Monte Carlo Cross Validation and Monte Carlo

Committee Machines
In Monte Carlo cross validation the test set is formed as a set of records
that are randomly removed from the analyzed dataset. The remaining
records form a training set. A key element of Monte Carlo cross validation
is the formation of the multiple randomly formed pairs (training set-test
set) and generation of forecasts for each pair. If the number of randomly
formed pairs (training set-test set) is large enough, any record in the ana-
lyzed dataset will appear in multiple test sets. As a result, multiple forecasts
will be produced for each record in the analyzed dataset. One of the results
of Monte Carlo cross validation is the average, over multiple forecasts, for
an individual record in the analyzed dataset. The average over multiple
forecasts may be interpreted as the output of the committee machine.
Monte Carlo committee machines designed and tested in this paper

are analogs for bagging committee machines [16]. The authors designed
MC committee machines in two applications: (a) compact description
of Monte Carlo cross validation results, (b) high stability forecast not
sensitive to possible perturbations of the training set. MC committee
machines rely on the use of multiple randomly formed training sets con-
structed by removing a fraction of randomly selected records from the
studied dataset. A test set may be formed as the union of the records
randomly removed from the training set, so that its records will include
values of predicted parameters. Then output of the committee machines
will be part of the results produced by Monte Carlo cross validation. The
test set may also be a predefined dataset with records that do not include
values of a predicted parameter. In this case, the goal of Monte Carlo
committee machines will be to produce forecasts of the parameter to be
predicted in a new geologic area.
Output of the Monte Carlo committee machine is in the form below:
1 N
Y (k ) = Σ F (k , m) (11.1)
N m=1
where k is the index of the record in the analyzed dataset, m is the index of
a randomly formed training set that was utilized to produce a forecast F(k,
m). N is the total number of randomly generated training sets that are used
to produce individual forecasts for a record with index k.
Mean absolute bias (mabF) and mean absolute error (maErF) of an
individual forecast for the set of records S and Bias of individual forecasts
(bF(k)), absolute bias averaged over a set of records S, mabF(S), and averaged
absolute error of individual forecasts, mabErF(S), are defined as:
1
bF (k ) = P(k ) − Y(k ); mabF (S ) = ∑ bF (k ) ;
N (S ) k ⊂ S
1 (11.2)
mabErF (S ) = ∑ P(k ) − F (k , m)
N (S ) k ⊂ S
Root mean squared error of the forecast by the committee machine for the
group of records is shown as:
0.5
 1 N (S) 
RMSE(S ) ∑ ( P(k ) − Y (k ))2  (11.3)
 N (S ) k ⊂ S 
where N(S) is the number of records in the subset S.

To analyze bias of the forecast produced by the committee machine we
construct multiple committee machines and define mean output over mul-
tiple committee machine forecasts calculated for the same record:
1 R
YR (k ) = ∑ Y (k , r ) (11.4)
R r =1
where r is the index of the committee machine forecasts, R is the total of

their number.
Bias of the forecast by the committee machine for the individual value of
the forecasted parameter and mean absolute bias within a subset of ana-
lyzed records are defined by Eq. 11.4 and 11.5:
1
bCM (k ) = YR (k ) − P( K ); mabCM (S ) = Σ bCM (11.5)
N (S) k⊂ S
1 M
F (k ) = ∗ Σ Y (k ) − F (k , m) (11.6)
M − 1 m=1
Calculation of instability index for the Monte Carlo committee machine

forecast, instCM(k) involves multiple committee machine forecasts Yr(k)
(Eq. 11.1); 1 ≤ r ≤ R using multiple Monte Carlo cycles.
1 R
instICM (k ) = ∗ Σ YR (k ) − Y (k , r ) (11.7)
R r =1
where R is total number of Monte Carlo cycles used to produce R forecasts

for records with index k.
Values of the above parameters depend on the parameter named ‘per-
turbation index’ that should be defined prior to the start of a Monte Carlo
cycle.
n.record(Train) (11.8)
prInd = 1 −
n.records(dataSet )
where n.records (dataSet) and n.records (Train) are respectively the number
of records in the analyzed data set that includes all available records with
known value of predicted parameter and in the randomly formed training
sets, Train in Eq. 11.8 is a randomly formed train set. This parameter satis-
fies the following constraints: 1> prInd>= 0, and for stable, not perturbed
training sets prInd = 0.
Major application of the MC committee machines forecast with a fixed
test set with records that do not include predicted parameter, so that per-
turbations are done only to the train set. In that case the goal of the Monte
Carlo committee machine is to decrease instability of the forecast of pre-
dicted parameters.
11.3 Performance of Extended MC Cross Validation

and Construction MC Committee Machines
This section presents a sequence of steps that form a complex algorithm
of the extended MC cross validation and construction of MC committee
machines. Its goal is analysis of accuracy of the forecasts, and an increase
of forecast stability.
1. Assign two parameters: (a) number of MC runs in the whole

MC cycle and (b) perturbation index.
2. Build randomly formed pair of train-test sets with their sizes
assigned in accordance with value of perturbation index.
3. Perform MC run that produces predictions of the forecasted
parameter using a pair of train-test sets. Save results of the
predictions as the data-set that includes actual values of the
predicted parameter and respective predictions.
4. Perform full MC cycle by repeating multiple times steps
2 and 3. If a number of individual MC runs in the cycle is
large enough, multiple predictions will be produced for each
value of forecasted parameter in the analyzed dataset.
5. Calculate forecasts by the committee machines for the
record with index k as the average of individual forecasts for
that record. Analyze accuracy and instability of individual
forecasts, and accuracy of the forecast by the committee

machines in accordance with equations 1 to 7.
6. Run multiple MC cycles to calculate the instability of com-
mittee machine forecasts.
To use MC committee machines for forecasts with a predefined test set

with records that do not contain a predicted parameter, keep the test set
unperturbed and do perturbations only to the training set by randomly
removing from it a certain number of records.
11.4 Parameters of Distribution of the Number

of Individual Forecasts in Monte Carlo
Cross Validation
Further, for this paper, we did Monte Carlo cross validation of permeability
forecasts with 1000 Monte Carlo runs within each MC cycle. Therefore,
1000 pairs (train-test) are created for analysis of the forecast efficiency. The
total number of records in the studied data set is 99. At each Monte Carlo
run, 9 records are randomly assigned to the test; remaining records go to
the training set. So perturbation index for the results presented herein is
0.091. Since each pair of train-test sets are randomly constructed, the num-
ber of individual forecasts for each record in the studied dataset is a random
number. Figure 11.1 shows a histogram of a number of individual forecasts
in Monte Carlo cycle of 1000 runs. Horizontal line marks 74 forecasts.
Number of forecasts
120
Forecasts number
80
40
0
0 20 60 100
Record index
Figure 11.1 Histogram of a number of individual forecasts in Monte Carlo cycle of

1000 runs.
Table 11.1 Parameters of distribution of a number of individual

forecasts for individual records by MC cross validation in four MC
cycles. Each MC cycle includes 1000 MC runs.
Monte Carlo cycles
Parameters of distribution of the
number of individual forecasts 1 2 3 4
Minimum 81 75 80 78
Lower quantile P: 0.25 94 95 96 93
Median 101 101 101 101
Mean 101 100 101 101
Upper quantile P: 0.75 108 107 107.5 108
Maximum 125 128 123 127
One can observe that the number of individual forecasts is rather stable
and, in this example, does not go below 74.
Parameters of distribution of the random number of individual fore-
casts, calculated for four MC cross validation cycles, are shown in Table 11.1.
Each cycle had 1000 cross validation runs. As seen, the estimated parame-
ters of the distribution are similar across all four cycles. Minimum number
of forecasts within four cycles varies from 75 to 81. Lower quantiles vary
within a range of 93 to 96. Upper quantiles are in the range of 107–108.
Therefore, if a number of Monte Carlo runs in the MC cycle is not less
than 1000, one may expect that about 75 forecasts may be used to evaluate
errors, bias and instability of the output of the committee machine and of
individual forecasts.
11.5 Linear Regression Permeability Forecast

with Empirical Permeability Models
The main advantages of multiple linear regressions when used for permea-
bility forecasts are in fast data processing and in an explicit form of approx-
imation permeability model. Conversely, the selection of appropriate
permeability models and rigidity of the selected model may be problem-
atic in the application of linear regression. The authors partially overcame
these complications by considering several permeability models with dif-
ferent levels of complexity and different sets of permeability predictors.
Complex models that include interactions among many predictors may

over fit predicted variables. The tool for control of over fitting used in this
paper is Monte Carlo cross validation. To be able to perform compara-
tive analysis of the accuracy of linear regression and forecasts with several
machine learning methods, we work with the same models for both linear
regression and machine learning methods. These models are:
Model 1: Permeability ~ Porosity + Grain.Size (11.9)
Model 2: Permeability ~ Porosity + Grain.Size + Porosity*Grain.Size

(11.10)
odel 3: Permeability ~ Porosity + Grain.Size + Depth + Microquartz

M
+ Porosity*Grain.Size + Porosity ^ Depth + Grain.Size*Depth (11.11)
Model 4: Permeability ~ Porosity + Grain.Size + Depth + Clay

(11.12)
Model 5: Permeability ~ Porosity + Grain.Size + Depth +

Microquartz (11.13)

Microquartz + Clay (11.14)
Model 7: Permeability ~ Porosity + Grain.Size + Grain.Size*Depth

(11.15)

Porosity*Grain.Size + Porosity^Depth + Grain.Size*Depth (11.16)

Microquartz + Clay + Porosity*Grain.Size + Porosity^Depth + Grain.
Size*Depth + Porosity*(Clay + Microquartz) + Microquartz*(Grain.
size + Porosity) (11.17)
Model 1 includes only two quantitative predictors with no interactions.

Models of this type are often used in permeability forecasts with machine
learning methods. Model 2 is an extension of Model 1 with interaction
between two predictors in the models. Models 1, 2, and 7 include only

quantitative predictors. Other models, in addition to quantitative predic-
tors, include one or more categorical predictors that take discrete sets of
values. They also interacted among predictors.
Individual forecasts produced by linear regression and machine meth-
ods may be negative, although predicted parameter in Eqs. 11.9–11.17 is
not smaller than zero. To improve accuracy of the forecast, all forecasts
with negative values are assigned value 0.
Bias, instability, errors of individual linear regression forecasts and cor-
relations of the output of MC committee machines with permeability are
shown in Table 11.2. These parameters were estimated using MC cross val-
idation with 1000 Monte Carlo cross validation runs in each MC cycle.
Values of mean squared error of forecast with all 9 models are about 80%
larger compared to mean absolute errors. This might be due to the effect
of outliers. According to Table 11.2, bias, individual forecast errors and
instability are sizable for the simplistic Model 1. Values of these parame-
ters for Models 3, 8 and 9 are significantly smaller than values of the same
parameters obtained with Model 1. Thus, the effects of second-order inter-
action between porosity and Grain.Size and the effects of categorical pre-
dictors are significant.
Table 11.2 presents only partial information on the forecast accuracy
produced by extended Monte Carlo cross validation. Important charac-
teristics of the forecast accuracy are trends of accuracy characterization
parameters as functions of a predicted parameter. This is illustrated by
Figure 11.2, which shows bias of the individual permeability forecast as a
function of the forecasted permeability.
Forecasts relying on the use of Model 1 have a large negative bias at big-
ger values of permeability, which may be as sizable as -1300 mD. Therefore
Table 11.2 Errors, bias, instability, and correlations of forecasts by MC

committee machine with forecasted permeability.
Permeability models 1 2 3 4 5 6 7 8 9
Mean Squared Error 475 284 274 414 413 422 374 267 264
RMSE (S)
Mean Absolute Error 234 165 154 216 220 219 204 149 143
maErF(S)
Mean absolute bias 228 165 147 215 221 212 201 144 142
mabF(S)
Instability instIF(k) 37 21 28 37 38 39 33 27 31
1500
Model 1
Model 3
500
Bias
–500
–1500
1400 1800 2200 2600

Permeability
Figure 11.2 Bias of the forecast with permeability Models 1 and 3. Bias is calculated for
permeability values in the range of 1100-2650 mD. Horizontal scale – permeability.
Model1
1500
0
0 20 40 60 80 100
Model3
1500
0
0 20 40 60 80 100
Figure 11.3 Ordered values of permeability and their forecasts with permeability
Models 1 and 3. Continuous line - ordered permeability, small dots - individual
forecasts. Vertical axis - permeability in mD horizontal scale - index of the record.
Model 1 is not appropriate for forecast and identification of outsized per-

meability zones. Absolute values of the bias of the forecasts with Model 3
are smaller and not systematically negative. The difference in the proper-
ties of the forecasts produced with these two models is illustrated in Figure
11.3 which shows actual permeability values and their forecasts with two
models. One can observe different signs of the bias produced with Model
1 and significant deviations of forecast values from actual permeability. As
for forecasts with Model 3, they are much closer to forecasted permeability
especially at large permeability values.
According to Figure 11.3, bias of the forecast with Model 1 is large and
positive for permeability values smaller than 500 mD. It becomes neg-
ative for larger permeability values, so that permeability of 2600 mD is
predicted as 1500 mD. A forecast with Model 3 is much more accurate and
has smaller bias in a full permeability range.
11.6 Accuracy of the Forecasts with Machine

Learning Methods
This section presents results of analyses of individual forecasts with six
machine learning methods. These methods are: neural networks (NN),
support vector machines (SVM), regression tree (RT), principal compo-
nent regression (PCR), k-nearest neighbor (KNN), and regression with
partial least squares (PLSR). Tables 11.3 and 11.4 show forecast bias and fore-
cast instability for the same permeability models that were used for analysis
of accuracy of linear regression. Both tables show similar patterns as those
for linear regression forecast. The smallest bias was obtained for forecasts
with Models 3, 8, and 9. Shown in Table 11.3, regression tree committee
machines produced forecasts with the least bias when compared to other
methods. Instability of forecasts by this method is not small and exceeds
instability of linear regression forecasts. Instability of the individual
Table 11.3 Mean absolute bias of the forecasts by Monte Carlo committee
machines with six machine learning methods. Mean absolute bias is calculated
for forecasts with nine permeability models for each forecast method.
Methods
Models PLSR KNN SVM NN PCR RT
1 231.9 153.0 169.5 152.1 230.3 118.8
2 164.0 154.5 166.5 153.7 164.0 121.1
3 146.0 139.7 143.6 126.5 146.7 124.7
4 217.6 142.7 142.2 127.0 215.9 125.1
5 221.7 157.3 146.3 133.3 221.7 121.9
6 214.8 147.7 155.7 143.5 215.0 123.5
7 205.4 151.6 164.4 137.0 204.3 120.5
8 144.2 133.4 141.4 133.4 143.2 125.4
9 143.1 145.5 142.7 148.5 142.2 123.3
Table 11.4 Instability index of individual forecasts with six machine

learning methods.
Methods
1 30.6 47.8 89.6 127.7 31.6 95.7
2 14.8 45.3 84.9 131.6 14.9 107.8
3 21.4 45.1 82.2 173.8 22.2 105.8
4 31.7 48.4 88.8 163.1 33.0 102.2
5 36.6 51.6 81.9 142.0 32.3 95.7
6 34.0 49.9 93.4 168.2 30.7 94.1
7 25.6 48.1 81.9 141.9 28.0 93.4
8 21.7 40.4 83.0 156.6 23.2 98.2
9 28.6 49.6 84.8 209.1 30.2 94.6
SVM
1000
–2000
0 500 1500 2500
RT
1000
–2000
0 500 1500 2500
Figure 11.4 Forecast bias as a function of permeability values for two machine learning
methods with permeability model (Model 1).
forecasts is the largest for neural network. Hence, more individual fore-
casts may be necessary to build neural network or regression tree commit-
tee machine forecasts with low instability.
Smaller regression tree bias compared to the bias of support vector
machines is illustrated by Figure 11.4. As seen here, bias of forecast by the
committee machines relying on support vector machines is systematically
negative at permeability values larger than 1500 mD. Therefore, a support
Table 11.5 Correlation coefficients between outputs of Monte

Carlo committee machines and forecasted permeability.
Machine learning methods
1 0.65 0.86 0.86 0.88 0.67 0.910
2 0.88 0.86 0.86 0.89 0.87 0.911
3 0.89 0.85 0.90 0.91 0.89 0.909
4 0.72 0.85 0.90 0.89 0.72 0.909
5 0.73 0.83 0.90 0.90 0.72 0.911
6 0.73 0.85 0.90 0.89 0.73 0.910
7 0.79 0.84 0.86 0.89 0.78 0.909
8 0.89 0.87 0.90 0.90 0.89 0.913
9 0.89 0.84 0.90 0.90 0.89 0.913
vector machine is not suitable for forecasts of large permeability values.

.
Bias of forecasts with regression tree is smaller and not systematically

negative.
Correlation between outputs of MC committee machines and forecasted
permeability is shown in Table 11.5. The highest correlation coefficient at
the level of 0.91 is for forecasts with regression tree. Regression adapts
well to different permeability models so that even forecasts with simplistic
Model 1 are characterized by a correlation of 0.91.
11.7 Analysis of Instability of the Forecast

This section and the one following illustrate the importance of analysis of
instability in individual forecasts. They also elucidate an increase of sta-
bility of the forecast by the MC committee machines via an increase of a
number of individual forecasts that shape those committee machines.
Table 11.6 and Figure 11.5 present results of detailed analyses of accu-
racy and instability of NN forecasts for eight values of permeability.
Individual forecasts were produced with permeability Model 2. The first
row of the Table below shows actual values of predicted permeability, while
other rows present results of analyses of individual and committee machine
Table 11.6 Instability of individual forecasts for eight permeability records.

Parameters Output of MC cross validation
Index of the 1 2 3 4 5 6 7 8
record
Permeability 755 808 917 933 1013 1019 1353 1461
MC committee 486 712 493 436 1600 1886 1857 1078
machine
forecast
Individual 272 398 581 606 1137 924 567 521
Forecast
Error
Maximum of 875 4624 2479 5655 6870 6743 4911 2573
Individual
Forecasts
Minimum of 179 179 0 0 0 126 607 0
Individual
Forecast
6000
Permeability
CM forecast
Individual
Permeability (md)
2000 0 4000
1 2 3 4 5 6 7 8
Record index
Figure 11.5 Individual and committee machine permeability forecasts in comparison

with actual permeability.
forecasts. Committee machines output is built as an average of 75 individ-

ual neural network forecasts. According to Table 11.6, the range of individ-
ual forecasts is huge, and the difference between maximum and minimum
individual forecasts for the same permeability value may be as large as 6000
mD. The committee machine forecast is closer to actual permeability.
Graphic illustration of instability of individual forecasts is given at

Figure 11.5. Small dots denote the values of individual forecasts, which
may even equal zero for actual permeability values exceeding 900 mD.
11.8 Enhancement of Stability of the MC Committee

Machines Forecast Via Increase of the Number
of Individual Forecasts
The effect of the number of forecasts that form the output of the committee
machines on instability of the forecast is illustrated by Figure 11.6. It shows
instability index of individual forecasts by support vector machines, instIF
(k) in comparison to instability index, instCM (k) of the forecast by the
SVM MC committee machines. The committee machine output is built as
the average over several individual SVM forecasts. Values of instCM (k) are
calculated for the records with permeability within the range of permeabil-
ity greater than 500 mD. Parameter N at Figure 11.6 is the number of individ-
ual forecasts that form the SVM committee machine. The total number of
committee machines R is used for the calculation of average YR(k ) . Results
shown in Figure 11.6 were obtained for a small perturbation index equal to
0.1. According to this Figure, individual forecasts are sensitive to minor
perturbations of the training set, with instability index of individual fore-
casts as high as 600 mD. The instability index of the committee machines
declines with an increasing number of individual predictions in a commit-
tee machine. At 50 or more individual forecasts that form the committee
Individual
Committee, N = 5
200 400 600 800
Committee, N = 50
Committee, N = 100
STD
0
500 1500 2500

Permeability (md)
Figure 11.6 Instability indexes of individual permeability forecasts by support vector

machines and by MC committee machines. N is the number of individual forecasts in a
committee machine. Individual in Figure 11.5 is instability index of individual forecasts.
machine, an instability index of the committee machines forecast is about

6 times less than that of individual predictions.
11.9 Conclusions
New methods of extended Monte Carlo cross validation and Monte Carlo
committee machines are introduced in this paper and tested with rock
permeability forecasts. Extended Monte Carlo cross validation is designed
for detailed analysis of bias and instability of the forecasts, forecast error,
and correlation between forecast and forecasted parameter. Monte Carlo
committee machines are expected to enhance forecast stability, which is
improved with an increase of a number of individual forecasts that form
the committee machine. Forecast bias is a basic characteristic of forecast
accuracy. It cannot be diminished.
Comparative analysis of the accuracy of individual permeability fore-
casts and forecasts by Monte Carlo committee machines relying on linear
regression, and several machine learning methods, was performed using
extended Monte Carlo cross validation protocol. A list of analyzed machine
learning methods includes neural networks, k-nearest neighbor, support
vector machine, principal component regression, partial least squared,
and regression tree. Committee machines relying on neural network and
regression tree outperformed linear regression and are characterized by
smaller forecast bias. The highest accuracy, smallest bias and highest cor-
relation between forecast and forecasted permeability was obtained with
regression tree.
Nomenclature
• MC cross validation = Monte Carlo cross validation, MC
committee machine - Monte Carlo committee machine, MC
run - Monte Carlo run.
• F(k, m) = Individual forecast for record with index k with
model built using randomly formed train set with index m.
• mabF(S) = Mean absolute bias of individual forecast.
• bCM(k) = Mean absolute bias of the forecast by the commit-
tee machine for the value of the forecasted parameter in the
record with index k.
• instIF(k) = Instability index for individual forecasts.
• instCM(k) = Instability index for the MC committee machine

forecast.
• prlnd = Perturbation index defining degree of random per-
turbations of the train set.
• SVM = Support vector machine, NN - neural net, KNN -
k-nearest neighbor regression, PCR- principal component
regression, PLSR - partial least squares regression, LR -
linear regression, RT - regression tree.
• Notations for the models = Linear models for response vari-
able for forecasted parameter are of the form:
responce = a0 + a1 * predictorn + … +an * predictorn
• Neither linear regression nor machine learning forecast
functions in R accept models presented in this form. Instead,
they use compact form of linear model.
responce ~ predictor1 + … + predictorn
This form of linear model is utilized in this paper.
• MC run = Forecast for records in a single pair randomly
generated train and test sets.
• Individual forecast = Forecast for individual record in the
test set produced as part of Monte Carlo run.
• MC cycle = Full set of MC runs performed for analysis of accu-
racy of the forecast or construction of MC committee machine.
• Instability index = Forecast characterization parameter that
defines potential changes of the forecast due to perturba-
tions of the training set, such as a small increase of the size
of the training set.
Appendix 1- Description of Permeability Models from

Different Fields
Many carbonate reservoirs have very high irreducible water saturation. As
much as 50-60% of pore space is occupied by this immovable water, which,
for the most part, does not participate in the flow of fluids.
As the fluids move within the reservoir, they consider the water in the
dead-end pores and fractures as being part of the solid rock. Thus, the defi-
nition of effective porosity that is used in many countries is incorrect. In the
United States the definition of effective porosity is their “effective” porosity
minus the irreducible fluid saturation. Russia defines it in the same way,
but refers to it as open (inter-communicating) porosity. When we use the
word ‘effective’ properly we will achieve more accurate results when assess-
ing the correlation between porosity and permeability.
In addition, specific surface area (per unit of pore volume), which is
a measure of the degree of fracturing, must be considered when evaluat-
ing the relationship between porosity and permeability. Fractures do not
contribute much to porosity but they do substantially increase the per-
meability. A few near-perfect correlations were obtained by Chilingarian,
Bagrintseva and Chang [2] for several carbonate reservoirs by adding two
additional variables: irreducible fluid saturation (Swr) and specific surface
area (Ss):
1. Vuktylskiy Gas-Condensate Field, Russia
logk = 0.9532 – 2.7880*10−2Swr−5.5597*10−4SS + 1.3309*10−1φ +

1.7707*10−5SwrSS (R = 0.997)
2. Kuybyshev, Along-Volga Region, Russia
logk = 2.1085 – 5.0777*10−2Swr−4.3785*10−4SS + 7.9959*10−2φ +

7.6326*10−6SwrSS (R = 0.998)
3. Orenburg Field, Russia
l ogk = 3.4351 – 2.0442*10−1Swr−9.5086*10−6SwrSs + 8.0217*10−3Swrφ +

2.3892*10−5SSφ (R = 0.981)
Where SS = Specific Surface Area (per unit of pore volume), Swr = irreduc-
ible water saturation and ϕ = fractional porosity. The role of the insoluble
residue (IR) content is also being investigated by the writer in determining
the relationship between porosity and permeability.
Appendix 2- A Brief Overview of Modular Networks or

Committee Machines*
Modular Neural Networks (Committee Machines)*
Modular neural networks or dynamic committee machines are comprised
of a cluster of individual neural networks, referred to as local experts,
which are connected to each other. Usually they are controlled or governed
by a global expert manifested by a gating network that acts as a referee or
y1
Expert 1
w1
y2
Expert 2
w2
Integrator
x y
Input Final output
yN
Exert N
wN
Global expert (gate keeper)
Figure A2.1 Configuration of a Committee Machine.
Input Hidden Output Input Hidden Output Input Output

Hidden block
layer layer layer layer layer layer block block
(a) MLP (b) MLP/w short cuts (c) CCP
Figure A2.2 Multilayer Perceptron (a) without and (b) with short cut connections and
passive input neurons (c) Completely Connected Perceptron (CCP) with active input
neurons (from Fruhwirth and Steinlechner, 2005).
committee chair. The role of the global expert is to make rulings or judg-
ments pertaining to the local experts, assign a significance factor or weight
to their respective outputs, and to determine what role each should play in
the final outcome of the combined network. The number of outputs of the
gating network is the same as the number of individual networks that are
combined. Figures 11.1–11.6 is an example of a network.
Dynamic committee machines are capable of handling more compli-
cated problems. Examples of applications of committee machines in the oil
industry include those in the inversion of induction of log data Zhang and
Poulton [17] and prediction of shear wave logs from sonic and other suites
of logs Fruhwirth and Steinlechner [18].
Fruhwirth and Steinlechner [18] used a Completely Connected
Perceptron (CCP) with active input neurons, a common MLP as well as
an MLP with short cuts for log prediction application. They used the net-
work to predict shear wave logs from a suite of other logs in a well. The
training was done using an existing shear wave and other suite of logs from
another well. The training started using solely the CCP architecture for 10
network generations starting without any hidden units (Figures 11.1–11.6).
This operation is equivalent to a multi-linear regression in many attribute
analysis applications except that the nodes in CCP can account for any
non-linear relationship between the known (in this case the known log
suites) versus the unknown (the shear wave log).
The final CCP had 9 hidden units. In each of these generations 20 differ-
ent and randomly initialized networks in parallel were trained, which can
be considered another type of realization of a modular neural network. The
goal was to prevent getting trapped too much in local error minima.
*This Appendix is adopted from Aminzadeh and de Groot [11].
References
1. E.S. AlHomadhi, New correlations of permeability and porosity versus con-
fining pressure, cementation, and grain size and new quantitatively correla-
tion relates permeability to porosity. Arabian Journal of Geosciences 7(7),
2871–2879 (2014).
2. G. Chilingarian, J. Chang, and K. Bagrintseva, Empirical expression of per-
meability in terms of porosity, specific surface area, and residual water satu-
ration of carbonate rocks. J. Petrol. Sci. Eng. 4(4), 317–322 (2006).
3. C.C. Fung, K.W. Wong, and H. Eren, Determination of a generalised BPNN
using SOM data-splitting and early stopping validation approach, Proceedings
of Eighth Australian Conference on Neural Network (ACNN’97), pp. 129–133
(1997).
4. B. Alpann and H.B. Helle, Committee neural networks for porosity and per-
meability prediction from well logs. Geophysical Prospecting. 50(6), 547–
674 (2002).
5. P. Cortez and P. Perner, Data mining with Neural Networks and Support
Vector Machines using the R/rminer Tool. Chapter in Advances in Data
Mining Applications and Theoretical Aspects, Volume 6171 of the series
Lecture Notes in Computer Science, 572–583 (2010).
6. W. Long, D. Chai, and F. Aminzadeh, Pseudo Density Log Generation using

Artifcial Neural Network. Society of Petroleum Engineers (2016, May 23).
doi: 10.2118/180439-MS.
7. R. Gholami, A. Shahraki, and A.M. Paghaleh, Support vector regression for
prediction of gas reservoirs permeability. J. Min. Environ. 2(1), 41–52 (2011).
8. C-H. Chen and Z-S. Lin, A committee machine with empirical formulas
for permeability prediction. Computers & Geosciences 32(4), 485–496, ISSN
0098-3004 (May 2006).
9. R. Sadeghi, A. Kadkhodaie, B. Rafei, and S. Khodabakhsh, A committee
machine approach for predicting permeability from well log data: A case
study from a heterogeneous carbonate reservoir, Balal oil Field, Persian Gulf.
Geopersia 1(2), 1–10 (2010).
10. L. Heng and Y. Daoyong, Estimation of multiple petrophysical parameters
for hydrocarbon reservoirs with the ensemble-based technique. Advances in
Petroleum Exploration and Development 4(2), 1–17 (2012).
11. F. Aminzadeh and P. de Groot, Neural Networks and Soft Computing
Techniques with Application in the Oil Industry, EAGE Publications, The
Netherlands (2006).
12. Q-S. Xu and Y-Z. Liang, Monte Carlo cross validation. Chemometr. Intell.
Lab.. 56(1), 1–11 (2001).
13. M. Nikravesh and F. Aminzadeh, Past, present and future intelligent reser-
voir characterization trends. J. Petrol. Sci. Eng. 31, 67–79 (2001).
14. S. Katz, G. Chilingar, and L. Khiliyk, Forecast and uncertainty analysis of
production decline trends with bootstrap and Monte Carlo Modeling.
J. Sustain. Energy Eng. 2, 350–367 (2015).
15. N.E. Aase, P.A. Bjorkum, and P.H Nadeau, The effect of grain-coating micro-
quartz on preservation of reservoir porosity. AAPG Bull. 80(10), 1654–1673
(1996).
16. L. Breiman, Bagging predictors. Machine Learning 24(2), 123–140 (1996).
17. L. Zhang and M. Poulton, Neural Network Inversion of EM39 induction
log data, in Geophysical Applications of Artifcial Neural Networks and Fuzzy
Logic, W. Sandaham and M. Leggett (Eds.), pp. 231–247, Kluwer Academic
Publishers, Dordrecht (2003).
18. R.K. Fruhwirth and S.P. Steinlechner, A systematic approach to the optimal
design of feed forward neural networks applied to log-synthesis, EAGE 66th
Conference & Exhibition, Expanded Abstracts (2004).
19. W.N. Venables and B.D. Ripley, Modern Applied Statistics with S, Fourth edi-
tion, Springer, The Netherlands (2002).
Part 4
RESERVES EVALUATION/
DECISION MAKING
12
The Gulf of Mexico Petroleum
System – Foundation for
Science-Based Decision Making
Corinne Disenhof 1,2*, MacKenzie Mark-Moser1,3 and Kelly Rose1
1
United States Department of Energy, National Energy Technology Laboratory,
Albany, Oregon, USA
2
URS Corporation, Albany, Oregon, USA
3
ORISE, Albany, Oregon, USA
Abstract
As offshore hydrocarbon development in the Gulf of Mexico has moved into
deeper waters and more technically challenging subsurface environments, the
tools to evaluate and reduce risks and potential impacts of drilling continue to
evolve. Science-based decision-making, risk reduction, and identification of tech-
nology gaps are key to the responsible development of extreme offshore hydro-
carbon resources. This paper specifically focuses on providing a review of data
and information related to the subsurface petroleum system for the U.S. Gulf of
Mexico. This information is vital to understanding the current state of knowl-
edge about the subsurface geology and hydrocarbon system for this region, and
for quantifying and assessing knowledge gaps and uncertainty. This review paper
summarizes relevant peer-reviewed and open-source publications, as well as pub-
licly available databases, focusing on regions associated with deepwater (>500’
water depth) and ultra-deepwater (>5000’ water depth) settings.
Keywords: Deepwater, reservoirs, hydrocarbons, risk reduction, data sources,

deposition, structure
255
Introduction
The Gulf of Mexico (GOM) basin is a petroleum province of global and
domestic economic importance. The United States’ Bureau of Ocean
Energy Management (BOEM) estimates that in the federal offshore GOM,
undiscovered technically recoverable resources total 87.5 billion barrels
of oil equivalent [1], and hydrocarbon companies are moving into deeper
water and riskier plays in search of the profitable prospects needed to
supply global demand that is projected to grow to 27 million incremental
barrels per day by 2020 [2]. Uncertain subsurface conditions in explor-
atory wells, already one of the riskiest phases of hydrocarbon development
[3, 4], are exacerbated by extreme environments that include water depths
of up to 10,000 feet, total well depths nearly 30,000 feet below the mudline,
pressures of almost 30,000 psia (pounds per square inch, absolute), and
temperatures over 300 degrees Fahrenheit [5]. In addition, recent events,
including impacts from natural events (e.g. Hurricanes Katrina and Rita)
and anthropogenic events (e.g. Deepwater Horizon oil spill), have high-
lighted gaps in our ability to predict risks and effectively prevent delete-
rious outcomes throughout the lifecycle of hydrocarbon development in
extreme offshore settings [3].
To reduce uncertainty, identify knowledge and technology gaps, support
risk assessments, and improve flow rate estimates in the GOM region, a
sufficient understanding of the subsurface is required for a variety of stake-
holders. While proprietary data and information are available to specific
entities, they are typically associated with restrictions for use and access.
However, information from published literature and publicly available data
sources can be integrated to support good comprehension of the petroleum
system across the GOM, and when appropriate, offer insights into field- or
site-specific areas, as well. Information from published literature includes
site-specific reservoir descriptions and basin-wide syntheses, studies of
basin structure and salt, analyses of the petroleum system, and reservoir
fluids found within hundreds of references. Publicly available databases
containing information such as subsurface attributes are less abundant
and typically focus on field or well specific scale information. However,
BOEM does release a well-populated GOM-wide public database of res-
ervoir sands characteristics that contains field-specific information and
production statistics [6]. While information about the subsurface GOM
petroleum system ranges in quantity, accessibility, and resolution, there is a
foundation from existing studies that can be utilized to reduce uncertainty
and evaluate for spatial and temporal trends at the field to regional scale.
The Gulf of Mexico Petroleum System 257
These can be used by a range of stakeholders (e.g. industry, regulators, and

scientists) to support science-based decision-making, while also serving to
identify knowledge and technology gaps. In support of ongoing and future
science-based studies pertinent to the GOM, this paper reviews aspects
of the GOM petroleum system that are relevant basin-wide and presents
references and resources to access for more information. Many sources ref-
erenced also contain in-depth field- or site-specific analyses detailing the
reservoirs, structures, and petroleum systems in the GOM. (e.g. [7–9]).
Basin Development and Geologic Overview

The Geological Society of America’s 1991 volume The Gulf of Mexico Basin
[10] is a thorough synthesis of information on every aspect of the GOM
basin’s geology. In 2008, Galloway published a chapter in The Sedimentary
Basins of the United States and Canada [11] to provide an updated descrip-
tion of the current knowledge of the northern Gulf ’s stratigraphy and dep-
ositional history. These discuss in detail the origin of the GOM basin, its
structural evolution and stratigraphic framework, and its resources. The
structural framework for the GOM is well laid out in a widely-cited 1995
paper by Diegel et al. [12], and is augmented by more recent studies of
basement structure and salt tectonics [11–15]. Basin evolution, as it per-
tains to the petroleum system of the GOM, is summarized here.
During the Mesozoic Era, breakup of Pangea initiated the formation of
the GOM basin. Intermittent connections to the Pacific Ocean, concurrent
with rifting in the Jurassic Era, led to periodic inundation of the forming
basin by shallow seawater that laid down a thick layer of evaporites now
known as the Louann Salt. Continued rifting, followed by seafloor spread-
ing, separated the salt into northern and southern basins and permanently
connected the basin to the ocean in the east [11, 16, 17]. Late Mesozoic depo-
sition is marked by interbedded deposition of terrestrial clastic sediments,
marine shales, and carbonates associated with prograding continental mar-
gins and the development of carbonate platforms around the basin [18, 19].
River systems, including the proto-Mississippi River, deposited terrigenous
sediments to the Basin starting in the Late Jurassic to the present [11, 20].
During the late Cretaceous and into the Cenozoic Era, continental
uplift due to the Laramide Orogeny and later crustal movements in North
and Central America increased the sediment supply to the Gulf dramat-
ically [11]. Thick, sandy strata prograded onto the continental shelf and
spilled onto the continental slope and basin floor, resulting in interbedded
sands and shales within the basin itself [7, 11, 17, 20–22]. In some places,
Cenozoic basin deposits accumulated to more than 10,000 meters thick
[17]. Deposition in the Pliocene and Pleistocene is also characterized by
thick, interbedded terrigenous deposits, which themselves are more than
3000 meters thick in some areas [17, 20].
The result of this large and rapid deposition of sediment was destabili-
zation and gravitational detachment of the sediment fill on the underly-
ing shale and salt layers, in addition to mobilization of the autochthonous
Louann salt [17]. Large detachment fault systems parallel the current basin
shoreline in Texas and Louisiana, including Paleocene and Eocene fault
systems under the Texas coastal plain and Oligocene and Miocene detach-
ments on the offshore Texas-Louisiana shelf [12] (Figure 12.1). Periods of
extension caused coeval compression at the toe of the continental slope,
creating provinces of compressional salt-cored anticlines in the western
and central basin (the Perdido and Mississippi Fan fold belts) (Figure 12.1).
MFFB
Folds
Fault Zones
PDFB Salt
Bathymetry (ft)
Value
0
-13300
0 1530 60 90 120
Miles
Figure 12.1 Map of regional fault zones, folds, and salt occurrences in the northern
GOM. Regional fault zones are down-to-the-basin. PDFB: Perdido Fan Fold Belt, MFFB:
Mississippi Fan Fold Belt. Data supporting this map are publicly available for download
from the USGS Energy Data Finder [23]. Bathymetry data from [24].
Differential loading on the Louann salt displaced it towards the center

of the basin and up through the overlying sediment, creating a complex
system of salt welds and allochthonous diapirs, sheets, and canopies. The
Louisiana-Texas continental slope is characterized by many minibasins,
where ponded sediment and salt withdrawal have created advantageous
environments for hydrocarbon accumulation [1, 11, 12, 22].
Petroleum System
Classic elements of a petroleum system include the geologic and hydro-
logic components and processes necessary to generate and store hydrocar-
bons, including a hydrocarbon source, migration pathway, reservoir rock,
trap and seal, and appropriate timing [25, 26]. For the GOM, information
about the basin’s petroleum system and sub-systems can be found in many
published articles and volumes, including the Atlas of Northern Gulf of
Mexico Gas and Oil Reservoirs [8, 9] and the AAPG Bulletin theme issue
Gulf of Mexico Petroleum Systems [27]. Updated literature with descrip-
tions of recent plays are presented in publications such as a 2012 study of
traps and reservoirs in the central northern Gulf of Mexico by Weimer and
Bouroullec [7] and a review of the properties of the deep water Wilcox
formation by Oletu et al. [5]. Reports from BOEM update the estimated
hydrocarbon reserves annually [28, 29], and their 2012 Outer Continental
Shelf Assessment summarizes the current knowledge of the northern Gulf ’s
plays, including their geology and estimated hydrocarbon resources and
reserves [1, 30].
For publicly available information pertinent to reservoir properties and
hydrocarbon properties, one of the most extensive available resources is
BOEM’s Atlas of Gulf of Mexico Gas and Oil Sands Data [31], which is pub-
licly downloadable and updated annually. This expansive data set presents
data aggregated and averaged to the sand level from non-releasable, res-
ervoir-specific data sources. This includes production statistics, estimated
reserves, age, physical and reservoir properties of the sands, and properties
of the hydrocarbons produced.
Reservoir Geology
Deepwater and ultra-deepwater reservoirs are dominantly associated with
turbidite deposits, particularly with the channel and levee systems and
fan/lobe sheet sand mass-transport deposit components of these turbidite
systems [1, 7, 32, 33]. The GOM is a notably prolific basin in that hydrocar-
bon plays span the vertical and lateral extent of the system. Plays occur ver-
tically throughout the stratigraphic column, and laterally across the basin,
the multitude of established large and small fields illustrate the breadth
of the system as well [11, 34]. Miocene-age reservoirs, many of which
are associated with deepwater and ultra-deepwater plays, have been the
most productive reservoir system, producing more than 23 billion barrels
of oil equivalent as of 2009 [1]. In addition, notable deepwater and ultra-
deepwater hydrocarbon plays in the GOM are also found within Pliocene
and Pleistocene sediments. In the future, the Paleocene, Eocene, and
Oligocene sediments that comprise Lower Tertiary plays on the continen-
tal slope are expected to contain the most undiscovered, technically recov-
erable resources, estimated at nearly 18 billion barrels of oil equivalent [1].
The Lower Tertiary Wilcox trend has produced at least a dozen discoveries
in the last decade [32, 35] (Figure 12.2).
Reservoir quality and type varies largely as a function of depositional
and structural environment, but can also be a function of secondary alter-
ation and diagenetic processes. Turbidite sands associated with Miocene to
Miocene Trend
Lower Tertiary Trend
Bathymetry (ft)
value
0
-13300
015 30 60 90 120
Miles
Figure 12.2 The general location of identified deepwater hydrocarbon trends in the
GOM. Based on data from [31, 35]. Bathymetry from [24].
Pleistocene reservoirs generally have high porosities and good to excellent

reservoir quality [2, 6], but can be compartmentalized by faults or inter-
bedded low permeability strata. In comparison, older Lower Tertiary sands
are generally deeply buried with good lateral extent, but in comparison to
the Miocene reservoirs have much lower porosities and permeabilities [5].
Reservoir properties associated with GOM plays can vary widely across the
basin [2, 5, 33], thus analysis of field- to reservoir-specific attributes, when
available, is key to improving understanding of different scale differences
in reservoir attributes.
Hydrocarbons
Hydrocarbons are produced from almost 1,300 fields in the federal offshore
GOM, 209 of which are in the deepwater and ultra-deepwater. According
to BOEM, more than half of the hydrocarbon reserves remaining in the
GOM are in the deepwater and ultra-deepwater reservoirs. As of 2011,
21.91 billion barrels of oil and 192.4 trillion cubic feet of gas have been
produced [29]. The northern GOM is more prone to gas than oil, though
many fields produce both [29, 33, 36]. Oil and gas compositions, as well as
associated fluids and gases (e.g. CO2 or H2S) are important as indicators for
reservoir connectivity, age, source, and timing of fluid migration [37–41].
Average gas-oil ratios, oil and gas gravities, and production estimates are
available in public data from BOEM [6, 31].
Hydrocarbon sources in the GOM have been heavily influenced by the
timing of sediment deposition. Lower sediment influx to the basin in the
Mesozoic Era, followed by the rapid, voluminous deposition in the Cenozoic
Era, has left much of the modern continental slope over-pressured and with a
depressed temperature gradient, leaving Mesozoic rocks on the slope within
the necessary temperature regimes to generate hydrocarbons [42]. Analyses
of regional hydrocarbon sources can be found in [33, 36, 37, 43].
The primary sources throughout the GOM basin are the carbonates
and shales deposited in the Upper Jurassic, Upper Cretaceous, and early
Cenozoic Eras [33, 37, 42]. Hood et al.’s analysis of GOM hydrocarbon sys-
tems interprets the source rocks supplying the most of the deepwater GOM
reservoirs to be Upper Jurassic sediments of Tithonian age, and links sea-
floor seeps in the ultra-deepwater to Oxfordian sources [37]. Reservoirs
in the western offshore basin and the Texas-Louisiana continental shelf
are charged with hydrocarbons from Paleocene and Eocene source rocks.
Hood et al. [37] subdivide these systems further by original composition.
Data from both wells and seeps were used in Hood et al.’s [37] analysis
to map the locations and maturity of the sources and hydrocarbons. GOM
deepwater and ultra-deepwater reservoirs tend to be less mature than res-
ervoirs with similar sources closer to the basin rim [31, 37]. Detailed chem-
ical analyses of seeps and reservoir fluids can be used to provide insights
into hydrocarbon sources and migration pathways [37–41]. Research
and data on the presence and indications of seeps throughout the Gulf of
Mexico is also available from BOEM [41, 44].
Salt and Structure

Subsurface faults and fractures are important as migration pathways, traps,
and seals throughout the GOM, but the publicly available data related to
these structural elements, when available, are scattered and difficult to
assemble. Basin-wide studies of structural elements for the GOM include
Diegel et al.’s [12] description of the larger framework of structural trends
and features across the GOM, along with Hood et al.’s [37] analysis of
hydrocarbon systems and migration pathways.
Detailed subsurface structural information is often interpreted from
2D and 3D geophysical surveys that are commonly commercial or pro-
prietary. However, publicly available geophysical data sets for some areas
of the GOM are accessible from several sources, including the National
Geophysical Data Center [45] and BOEM [6]. In addition, interpreted seis-
mic lines are often published in field-specific studies as they are heavily
relied upon for field assessments such as those by Weimer et al. [46] for
fields in the Green Canyon and Ewing Bank region.
Salt is a distinctive element of the GOM petroleum system. Salt defor-
mation, injection, and withdrawal are responsible for many of today’s exist-
ing traps and seals. Sediment loading and salt withdrawal have produced
numerous miniba-sins, creating accommodation space for sediment accu-
mulation, and the evacuated salt pierces and overhangs younger sediments
as diapirs and canopies [7, 11]. The salt and sediment movement creates
faults, salt welds, and turtle structures, and compressional folding at the
base of the continental slope has formed series of salt-cored anticlines. All
of these structures are common hydrocarbon traps [1, 7], but also often
serve as fluid migration pathways. Recent studies by Hudec et al. [47] and
Hudec et al. [14] connect structural styles and trap types in the deep water
to salt provinces influenced by the original emplacement and movement of
the salt related to rifting in the Jurassic Era.
Hydrocarbon reservoir placement, size, and thickness in the GOM can

be strongly influenced by salt and structural complexity [7, 14]. Weimer and
Bouroullec [7] describe the control structure and salt have on the size and
location of reservoirs in the region around the Mississippi Fan and detail
trap styles and patterns throughout the area. Deepwater reservoirs in this
area produce primarily from Miocene or younger sediments, and traps are
typically combinations of structural and strati-graphic elements. Structural
traps are formed by anticlines or faults, including turtle structures, caused by
extension and structural inversion during salt evacuation. Thunder Horse,
the largest field in the GOM, produces from a turtle structure [7]. The salt-
bounded minibasins prevalent on the continental slope often have combi-
nation traps due to stratigraphic pinchout, confining or overhanging salt, or
turtle structures [36] and have produced discoveries such as the Mars, Ursa,
and Devils Tower fields [7]. On the ultra-deep water Louisiana-Texas slope,
the Paleocene and Eocene Wilcox reservoirs are often buried below the
Sigsbee salt canopy. Anticlinal, faulted structures dominate the traps here,
including the salt-cored folds of the Perdido and Mississippi Fan fold belts.
The Great White field produces from the Perdido Fold Belt, while Cascade
and Chinook fields produce from pillow folds near the toe of the Sigsbee salt
[5, 32]. Information about individual field structure and traps can be found
in many detailed publications, including [7–9, 27].
Conclusions
Knowledge of the GOM geology and petroleum system, particularly from
publicly available sources, is continuously evolving. When used appropri-
ately, these information sources offer industry, regulators, and scientists
opportunities to study the system, identify knowledge and technology gaps,
and reduce risks related to exploration and development of the hydrocar-
bon system to mutual benefit. Challenges to drilling and production are
only going to continue to grow as operations move farther offshore [2, 4]
and to increasing sub-seafloor total depths. Insights into the subsurface
from myriad publicly available resources, including publications and data
sets, are key to reducing uncertainty in models and therefore risks [3], and
helping to ensure responsible and enduring access to domestic hydrocarbon
resources. Direct detection and characterization of the subsurface is costly
and limited to locations where sediment cores, sidewall cores, and proper-
ties from wellbore devices offer in situ measurements. The vast majority of
the data and information available for use in interpreting and predicting
subsurface conditions are gathered from indirect detection techniques that

require interpretation, including seismic surveys and geophysical logs.
Technology has developed immensely in the last few decades, increasing
the usefulness of seismic analyses in deepwater and subsalt settings and
increasing the accuracy and resolution of formation evaluations and logs
[2, 43]. However, while geophysical data are far more widely available than
cores, they are still relatively scarce in comparison to the area of the basin
as a whole. This is true particularly in the deepwater and ultra-deepwater:
borehole data from BOEM indicates that only ten percent of the wells in
the GOM are in the deep and ultra-deepwater [6].
A range of stakeholders, including industry, regulators, and researchers,
can and should continue to utilize the ever-growing body of publicly avail-
able data to identify solutions to gaps in knowledge and technology related
to offshore exploration, production and operations in the GOM. While pro-
prietary data sets and resources are often coveted for the information they
can offer, the breadth and range of publicly available data and published lit-
erature about these systems offers significant opportunities to advance the
public state of knowledge about the subsurface GOM hydrocarbon system,
particularly in deep water and the deep subsurface. Published and publicly
available information forms a significant foundation for modeling and risk
assessment. As organizations make decisions based on science-based pre-
dictions, existing and available data should always be leveraged to reduce
uncertainty and risks for all stakeholders.
Acknowledgments and Disclaimer

The authors would like to acknowledge the support of BSEE and BOEM
personnel with the U.S. Department of Interior for conversations and dis-
cussions about the subsurface system in the Gulf of Mexico. We would also
like to acknowledge and thank Jennifer Bauer, Nicolas Huerta, and Roy
Long, U.S. Department of Energy, National Energy Technology Laboratory,
for their insights and discussions about this publication and the broader
NETL offshore risk assessment project as a whole.
This technical effort was conducted as part of the National Energy
Technology Laboratory’s ongoing offshore and ultra-deepwater research,
including support under the RES contract DE-FE0004000.
This project was funded by the Department of Energy, National Energy
Technology Laboratory, an agency of the United States Government,
through a support contract with URS Energy & Construction, Inc. Neither
the United States Government nor any agency thereof, nor any of their
employees, nor URS Energy & Construction, Inc., nor any of their employ-
ees, makes any warranty, expressed or implied, or assumes any legal liability
or responsibility for the accuracy, completeness, or usefulness of any infor-
mation, apparatus, product, or process disclosed, or represents that its use
would not infringe privately owned rights. Reference herein to any specific
commercial product, process, or service by trade name, trademark, manu-
facturer, or otherwise, does not necessarily constitute or imply its endorse-
ment, recommendation, or favoring by the United States Government or
any agency thereof. The views and opinions of authors expressed herein do
not necessarily state or reflect those of the United States Government or
any agency thereof.
References
1. R. Klazynski, E. Klocek, L. Nixon, A. Petty, and P. Post, Assessment of
Technically Recoverable Hydrocarbon Resources of the Gulf of Mexico
Outer Continental Shelf as of January 1, 2009. U.S. Bureau of Ocean Energy
Management Gulf of Mexico OCS Regional Office, Office of Resource
Evaluation, New Orleans, LA. (2012).
2. H. Elshahawi, Deepwater exploration and production in the Gulf of Mexico-
challenges and opportunities. Petrophysics 55, 81–87 (2012).
3. K. Rose, F. Aminzadeh, L. Sim., R. Ghanem, C. Disenhof, J. Bauer, M.
Mark-Moser, C. Thimmisetty, N. Jabbari, and A. Khodabakhshnejad, Risks
and impact assessment for deepwater and ultra-deepwater Gulf of Mexico
resources, Offshore Technical Conference, 25364-MS (2014).
4. D. Izon, E.P. Danenberger, and M. Mayes, Absence of fatalities in blowouts
encouraging in MMS study of OCS incidents 1992–2006. Well Control July/
Aug (2007).
5. J. Oletu, U. Prasad, A. Ghadimipour, S. Sakowski, K. Vassilellis, B. Graham,
N. Park, and L. Li, Gulf of Mexico wilcox play property trend review, Offshore
Technical Conference, 24186-MS (2013).
6. Bureau of Ocean Energy Management Data Center, www.data.boem.gov
(2014).
7. P. Weimer and R. Bouroullec, Petroleum Geology of the Mississippi Canyon,
Atwater Valley, Western Desoto Canyon, and Western Lloyd Areas, Northern
Deep Gulf of Mexico: Traps, reservoirs, and their timing. 32nd GCSSEPM
Research Conference Proceedings. 32, 110–132 (2012).
8. S.J. Seni, T.F. Hentz, W.R. Kaiser, and E.G.J. Wermund, (Eds.), Miocene and
Older Reservoirs: Atlas of Northern Gulf of Mexico Gas and Oil Reservoirs,
Volume 1, Bureau of Economic Geology, The University of Texas at Austin.
(1997).
9. T.F. Hentz, S.J. Seni, and E.G.J. Wermund, (Eds.), Pliocene and Pleistocene
Reservoirs: Atlas of Northern Gulf of Mexico Gas and Oil Reservoirs, Volume
2, Bureau of Economic Geology, The University of Texas at Austin. (1997).
10. A. Salvador (Ed.), The Gulf of Mexico Basin. The Geology of North America.
Geological Society of America, Boulder, Colorado. (1991).
11. W.E. Galloway, Depositional evolution of the Gulf of Mexico sedimentary
basin, in The Sedimentary Basins of the United States and Canada, A.D. Miall
(Ed.), pp. 505–549, Elsevier. (2008).
12. F.A. Diegel, J.F. Karlo, D.C. Schuster, R.C. Shoup, and P.R. Tauvers, Cenozoic
structural evolution and tectono-stratigraphic framework of the Northern
Gulf Coast continental margin, in Salt Tectonics: A Global Perspective, M.P.A.
Jackson, D.G. Roberts, and S. Snelson (Eds.), 109–151, AAPG Mem. 65
(1995).
13. S.H. Hall, The role of autochthonous salt inflation and deflation in the north-
ern Gulf of Mexico, Mar. Pet. Geol. 19, 649–682 (2002).
14. M.R. Hudec, M.P.A. Jackson, and F.J. Peel, Influence of deep Louann struc-
ture on the evolution of the Northern Gulf of Mexico. AAPG Bull. 97, 1711–
1735 (2013).
15. T.P. Dooley, M.P.A. Jackson, and M.R. Hudec, Coeval extension and short-
ening above and below salt canopies on an uplifted, continental margin:
Application to the Northern Gulf of Mexico. AAPG Bull. 97, 1737–1764
(2013).
16. A. Salvador, Triassic-Jurassic, in The Gulf of Mexico Basin: The Geology
of North America, A. Salvador (Ed.), pp. 131–180, Geological Society of
America, Boulder, Colorado. (1991).
17. A. Salvador, Origin and development of the Gulf of Mexico basin, in The Gulf
of Mexico Basin: The Geology of North America, A. Salvador (Ed.), Geological
Society of America, Boulder, Colorado. (1991).
18. E.J. McFarlan, and L.S. Menes, Lower cretaceous, in The Gulf of Mexico Basin:
The Geology of North America, A. Salvador, (Ed.), pp. 181–204, Geological
19. N.F. Sohl, R.E. Martínez, P. Salmerón-Ureña, and F. Soto-Jaramillo, Upper
Cretaceous, in The Gulf of Mexico Basin: The Geology of North America,
A. Salvador, (Ed.), pp. 205–244. Geological Society of America, Boulder,
Colorado. (1991).
20. J.M. Coleman, H.H. Roberts, and W.R. Bryant, Late quaternary sedimenta-
tion, in The Gulf of Mexico Basin: The Geology of North America, A. Salvador,
(Ed.), pp. 325–352, Geological Society of America, Boulder, Colorado.
(1991).
21. A. McDonnell, R.G. Loucks, and W.E. Galloway, Paleocene to eocene
deep-water slope canyons, Western Gulf of Mexico: Further insights for
the provenance of deep-water offshore Wilcox Group plays. AAPG Bull. 92,
1169–1189 (2008).
22. W.E. Galloway, D.G. Bebout, W.L. Fisher, J.B.J. Dunlap, R. Cabrera-Castro,
J.E. Lugo-Rivera, and T.M. Scott, Cenozoic, in The Gulf of Mexico Basin:
The Geology of North America, A. Salvador, (Ed.), pp. 245–324, Geological
23. United States Geological Survey, USGS Energy Data Finder, http://certmap-
per.cr.usgs. gov/geoportal/catalog/main/home.page (2014).
24. NOAA National Geophysical Data Center, GEODAS Grid Translator,
ETOPO11-minute Global Relief data, http://www.ngdc.noaa.gov/mgg/gdas/
gd_designagrid.html (2012).
25. L. Magoon, The petroleum system—A classification scheme for research,
exploration, and resource assessment. U.S. Geol. Surv. Bull. 1870, 2–15
(1988).
26. L.B. Magoon and Z.C. Valin, Overview of petroleum system case studies.
AAPG Mem. 60, 329–338 (1994).
27. N. Hurley and P. Weimer, Gulf of Mexico petroleum systems. AAPG Bull. 82,
865–1112 (1998).
28. Bureau of Ocean Energy Management, Reserves Inventory Program- Gulf
of Mexico (GOM) OCS Region, http://www.boem.gov/Reserves-Inventory-
Program-Gulf-of-Mexico-OCS-Region (2014).
29. E.G. Kazanis, D.M. Maclay, and N.K. Shepard, Estimated Oil and Gas
Reserves, Gulf of Mexico OCS Region, December 31, 2011. OCS Report
2014-656, U.S. Bureau of Ocean Energy Management Office of Resource
Evaluation, New Orleans, La. (2014).
30. Bureau of Ocean Energy Management, Resource Assessment- Gulf of
Mexico OCS Region, http://www.boem.gov/Oil-and-Gas-Energy-Program/
Resource-Evaluation/ Resource-Assessment/RA-Gulf.aspx (2014).
31. Bureau of Ocean Energy Management, Atlas of Gulf of Mexico Gas and Oil
Sands Data, http://data.boem.gov/homepg/data_center/gandg/gandg.asp
(2014).
32. D. Meyer, L. Zarra, and J. Yun, From BAHA to Jack, evolution of the lower
tertiary wilcox trend in the deepwater Gulf of Mexico. Sedimen. Rec. 5, 4–9
(2007).
33. R. Nehring, Oil and gas resources, in The Gulf of Mexico Basin: The Geology
of North America, A. Salvador, (Ed.), pp. 445–494, Geological Society of
America, Boulder, Colorado. (1991).
34. B.J. Bascle, L.D. Nixon, and K.M. Ross, Atlas of Gulf of Mexico Gas and
Oil Sands as of January 1, 1999, OCS Report MMS 2001-086, U.S. Minerals
Management Service, New Orleans, LA. f(2001).
35. Halliburton, Deepwater Gulf of Mexico, 2000–2008 Deepwater Discoveries,
http://www.halliburton.com/public/solutions/contents/Deep_Water/
related_docs/GOM_ DWMap.pdf (2008).
36. A.C.J. Huffman, P.D. Warwick, and W.I. Finch, Energy resources of the
Northern Gulf of Mexico basin, in Gulf of Mexico Origin, Waters, and Biota:
Volume 3, Geology, N.A. Buster and C.W. Holmes, (Eds.), pp. 229–245, Texas
A&M University Press, College Station, Texas. (2011).
37. K.C. Hood, L.M. Wenger, O.P. Gross, and S.C. Harrison, Hydrocarbon sys-
tems analysis of the Northern Gulf of Mexico: delineation of hydrocarbon
migration pathways using seeps and seismic imaging, in Surface Exploration
Case Histories: Applications of Geochemistry, Magnetics, and Remote Sensing,
D. Schumacher and L.A. LeSchack (Eds.), pp. 25–40, AAPG (2002).
38. I.R. MacDonald, J.F. Reilly, Jr., S.D. Best, R. Venkataramaiah, F. Sassen, N.L.
Guinasso, Jr., and J. Amos, Remote sensing inventory of active oil seeps
and chemosynthetic communities in the Northern Gulf of Mexico, in M66:
Hydrocarbon Migration and its Near-Surface Expression, D. Schumacher and
M.A. Abrams (Eds.), pp. 27–37, AAPG Memoir (1996).
39. O.C. Mullins, J.Y. Zuo, K. Wang, P.S. Hammond, R. De Santos, H. Dumont,
V.K. Mishra, L. Chen, A.E. Pomerantz, C.L. Dong, H. Elshahawi, and D.J.
Seifert, The dynamics of reservoir fluids and their substantial systematic
variations. Petrophysics 55, 96–112 (2014).
40. B.N. Orcutt, S.B. Joye, S. Kleindienst, K. Knittel, A. Ramette, A. Reitz, V.
Samarkin, T. Treude, and A. Boetius, Impact of natural oil and higher hydro-
carbons on microbial diversity, distribution, and activity in Gulf of Mexico
cold-seep sediments. Deep-Sea Research II 57, 2008–2021 (2010).
41. W. Shedd, R. Boswell, M. Frye, P. Godfriaux, and K. Kramer, Occurrence
and nature of “bottom simulating reflectors” in the Northern Gulf of Mexico.
Mar. Pet. Geol. 34, 31–40 (2012).
42. J.A. Nunn and R. Sassen, Framework of hydrocarbon generation and migra-
tion, Gulf of Mexico Continental Slope. AAPG Bull. 70, 1187–1187 (1986).
43. T. Matava, A petroleum systems study of the Northern Gulf of Mexico. Lead.
Edge. 25, 478–482 (2006).
44. Bureau of Ocean Energy Management, Seismic Water Bottom Anomalies
Map Gallery, http://www.boem.gov/Seismic-Water-Bottom-Anomalies-Map-
Gallery (2014).
45. NOAA National Geophysical Data Center, Marine Trackline Geophysical
Data, http:// www.ngdc.noaa.gov/mgg/geodas/trackline.html (2014).
46. P. Weimer, J.R. Crews, R.S. Crow, and P. Varnai, Atlas of petroleum fields and
discoveris, Northern Green Canyon, Ewing Bank, and Southern Ship Shoal
and South Timbalier Areas (Offshore Louisiana), Northern Gulf of Mexico.
AAPG Bull. 82, 878–917 (1998).
47. M.R. Hudec, I.O. Norton, M.P.A. Jackson, and F.J. Peel, Jurassic evolution of
the Gulf of Mexico salt basin. AAPG Bull. 97, 1683–1710 (2013).
13
Forecast and Uncertainty Analysis
of Production Decline Trends with
Bootstrap and Monte Carlo Modeling
Simon Katz1*, George Chilingar2 and Leonid Khilyuk1

1
2
Abstract
A new multi-step procedure was developed and tested for the analysis and forecast
of production decline curves. It includes multidimensional grid search over values
of nonlinear least squares errors, nonlinear least squares approximation, Monte
Carlo and block bootstrap simulation of production trends.
Several decline curve models were tested in grid search and iterative minimiza-
tion: SEPD, extended Hyperbolic, Duong and the Power-Exponential. Grid search
and iterative minimization worked equally well on all tested models.
Multidimensional grid search finds a starting point for nonlinear iterative pro-
cess. Then, iterative minimization finds parameters of an optimum approximating
model. The approximating model is used for the forecasting of production trends.
Block bootstrap and new Monte Carlo simulation methods produce third-level
aggregated forecasts of production rate. In addition, these two methods character-
ize the range of possible values of predicted production (uncertainty range).
Comparative analysis of Monte Carlo and block bootstrap simulation indicates
that these methods are characterized by different widths of the uncertainty regions
and by a certain mutual shift of predicted production curves. Thus, the joint use of
both techniques may result in a more reliable forecast of future production.
Keywords: Decline curves, uncertainty analysis, forecast, nonlinear least squares,

bootstrap, Monte Carlo
269
13.1 Introduction
Production decline curve analysis uses empirical decline curve models for
forecasting hydrocarbon production trends. Construction of the produc-
tion decline models is based on the use of production data in individual
wells or in a group of wells. Widely used methods of the decline curve anal-
ysis and forecast of production often rely on a graphical approach and may
include slope analysis of certain modified curves and estimates of decline
curve value at zero time. Another approach is minimization of approxi-
mation error via nonlinear regression. Darwis et al. [1] and Towler et al.
[2] used a robust linear regression technique for the prediction of decline
curve production. Nonlinear regression techniques seem more flexible and
usable with models of different type in a unified manner.
There are two important issues in the nonlinear regression: (a) selec-
tion of a starting point for stable and converging iterative minimization of
approximation error and (b) probabilistic analysis of uncertainty of final
results due to the effect of random fluctuations in input data. The first issue
is resolved if the starting point is close enough to the minimum of the
approximation error. To locate the starting point of this kind we developed
and utilized multidimensional grid search.
The probabilistic analysis of decline curves is discussed and developed
in a number of publications. A priori assumptions about the distribution of
a model’s parameters were introduced in Lin et al. [3]. These assumptions
were then used for analysis of production decline uncertainty. Jochen et al.
[4] applied the bootstrap method to develop stochastic reserves estima-
tion. Cheng et al. [5] applied a modified bootstrap method for analysis of
future production of an oilfield, using the Arps hyperbolic model.
Probabilistic analysis done in this paper relies on the unified approach
of Monte Carlo and block bootstrap methods that generates multiple pre-
dicted decline curves.
We developed and tested the multi-step forecasting and probabilistic
analysis on several decline curves models defined by Eqs. 13.1 to 13.4:
q0
Modified Hyperbolic model: u(t ) = (13.1)
(1 + g ∗ t )c

Modified Hyperbolic model of Eq. 13.1 is identical to the standard

hyperbolic model with parameters c = 1/b and g = b * D. We used the
modified version of Hyperbolic model to provide better convergence of
Forecast and Uncertainty Analysis of Production Trends 271
iterative minimization of nonlinear least squares and stabilize iterative

minimization of the approximation error.
SEPD model: u(t ) = a ∗ exp(−b ∗ (t c )) (13.2)
 a 
Duong model: u(t ) = q1 ∗ t −m ∗ exp  ∗ (t å −m − 1) (13.3)
 1− m 
New Power-Exponential model u(t )=q1 ∗ t −m ∗ exp(−b ∗ t ) (13.4)

Power-Exponential model is an extension of Exponential model. It

reduces to exponential model if m = 0 in Eq. 13.4.
The following definitions will be used in construction and approxima-
tion of decline curves:
Approximated model - defines the structure of the decline curve to be
approximated and forecasted.
Approximating model - used to construct an approximation of the
recorded or simulated decline curve.
Approximated and approximating models are not necessarily the same.
For example, a decline curve generated with the model SEPD may be
approximated by the decline curve generated by Hyperbolic model (SEPD
- approximated model, Hyperbolic - approximating model) and vice versa.
Effects of different approximating models on approximation error for a
given approximated model are analyzed in the Sections 13.4 and 13.5 of
this paper.
This paper is devoted to development and verification of new multi-
step algorithms for analysis and forecast of production decline curves. It
includes rigorous verification of new algorithms using synthetic data and
quantitative analysis of uncertainty of multistep forecast. Next step will be
using real field production data with analysis of uncertainty of forecast of
cumulative production, estimation time to economic limit, and analysis of
uncertainty of expected total production.
13.2 Simulated Decline Curves

Simulated decline curves were used in this paper for analysis of nonlinear
forecasting efficiency and analysis of uncertainty of the forecast.
A decline curve is presented as a sum of deterministic and random com-

ponents of the form:
u(t )=model(t )+υ (t ) ∗ g (t )+ς (t ) (13.5)

where: model(t) is the deterministic component defined by one of the Eq.

13.1 to 13.4, v (t) is the stationary correlated random component with zero
mean, correlation matrix cor (t) and constant variance.
g(t) is the deterministic function that defines trend in standard devia-
tion of fluctuations of the reservoir performance.
ς(t) is the random uncorrelated variable with standard deviation σς. It
imitates measurements errors.
Random component a(t) = v (t) * g(t) models fluctuations in reservoir
performance, Example of the SEPD decline curve calculated in accordance
with Eq. 13.2 and distorted by a combination of correlated and uncorrelated
random components (Eq. 13.5) is shown in the Figure 13.1. According to
the Figure 13.1, values of the decline curve are correlated in the lag range
of (+5, -5) months.
Autocorrelation
1.00
Autocorrelation
0.75
0.50
0.25
0.00
-5 0 5
Decline curve
Production (bbl/day)
80
60
40
20
20 40 60
Months
Autocorrelation
Decline curve
Figure 13.1 SEPD decline curve and autocorrelation of its random component.
13.3 Nonlinear Least Squares for Decline Curve

Approximation
The authors developed a two-step procedure for iterative minimization of
weighted residual of nonlinear least squares a (X) that defines approxima-
tion error:
J
a( x ) =
1
J ∑W ∗(Model(t , X ) − u(t ))
j j j
2
(13.6)
j

where: j is the time index, ωj are positive weighs, X is the vector of model
parameters, X = (x1,x2,..,xn). u(tj) are values of the decline curve at time
moments tj, model (tj, X) is the continuous nonlinear or linear function of
time and vector of parameters X. In the following sections of this paper,
model (tj, X) is defined by one of the Eqs. 13.1 to 13.4. Vector parameter
of the approximating model X is defined as one that minimizes weighted
residual of Eq. 13.6. The multi-step minimization procedure includes
multi-dimensional grid search followed by iterative minimization of the
approximation error (Eq. 13.6) and random generation of multiple decline
curves via Monte Carlo or bootstrap methods.
13.4 New Method of Grid Search for Approximation

and Forecast of Decline Curves
Both approximation and forecast of decline curves are based on minimiza-
tion of the nonlinear least squares (NLS) that defines approximation error.
Estimated parameters are those corresponding to the NLS minimum. The
most straightforward method for minimization of the nonlinear least squares
is the grid search. In a grid search, developed and reviewed in this paper,
nonlinear least squares (Eq. 13.6) is approximated by a set of grid values
a( X i,k,n ) = a( x1 + ∆x1 ∗ i , x 2 + ∆x 2 ∗ k , x 3 + ∆x 3 ∗ n) (13.7)
Estimated vector parameter Xi0,k0,n0 is defined as one satisfying the fol-

lowing equation:
a( X å 0 ,0 ,0 ) = å (a( x1 + ∆x1 ∗ i , x 2 + ∆x 2 ∗ k , x 3 + ∆x 3 ∗ n)) (13.8)

å, ,
The main advantage of the grid search is its flexibility. The grid search
works with any approximating model of any complexity. It always results in
a location of a global minimum on a grid. Although the grid minimum does
not coincide with the actual global minimum, it gets closer to the actual
global minimum when the distance among grid nodes decreases. Another
advantage of the grid search is that it does not need a starting point. The
range for model parameters and distances between grid nodes are the only
parameters that should be pre-defined. Unfortunately, a decrease in the
distance among the grid nodes leads to an increase of the number of nodes
and, respectively, the computer time necessary for performing the grid
search. Thus, a certain balance is necessary. The authors envision the role
of the grid search as an instrument for finding the location of a starting
point for iterative methods of minimization of nonlinear least squares.
Efficiency of the grid search for several approximating models is illus-
trated in Figure 13.2. It shows four curves approximating SEPD decline
curve distorted by random noise. Dots in the Figure 13.2 show this approx-
imated curve. The approximating curves were obtained via minimization
of the approximating error (Eq. 13.7). Resulting approximating curves
shown as continuous lines are constructed with SEPD, Power Exponential,
Hyperbolic, and Duong models. According to Figure 13.2, approxima-
tion with all four models results with approximating curves close to the
15
14
Production (bpl/day)
13
12
11
10
9
0 10 20 30 40 50 60 70
Months
Figure 13.2 Approximation of the SEPD curve via multidimensional grid search. Four
approximating models.
approximated one. Two of the approximating curves in Figure 13.2 prac-

tically coincide.
13.5 Iterative Minimization of Least Squares

with Multiple Approximating Models
This section illustrates the efficiency of approximating the decline produc-
tion curves with the iterative minimization Levenberg-Marquard algo-
rithm. Function nlsLM of R programming language uses this algorithm
for iterative minimization of weighted residual of nonlinear least squares
defined by Eq. 13.6. Function nlsLM may also use constraints on the model
parameters of the form:
xi,1 ≤ xi ≤ xi,2 to stabilize iterative minimization of the nonlinear least
squares.
Figure 13.3 shows an approximated Hyperbolic curve distorted by ran-
dom noise. It also shows four curves produced with Hyperbolic, SEPD,
PowerExp, and Duong models. The curves shown at this figure were cal-
culated with parameters used as starting points for minimization. One can
observe significant differences between the approximated noise distorted
Approximate curve
500
Sepd
Hyperbolic
Duong
Production (bb/day)
Power. Exp
300
100
0 10 20 30 40 50 60 70
Months
Figure 13.3 Approximated decline curve and four starting decline curves with
parameters different from parameters of approximated model. Model for approximated
curve - Hyperbolic. Four starting approximated curves are constructed using Hyperbolic,
Power. Exp, SEPD, and Duong models.
Approximated curve
200
Sedm
Hyperbolic
Duong
Power. Exp
Production (bb/day)
150
100
50
0 10 20 30 40 50 60 70
Months
Figure 13.4 Approximating curves produced by Levenberg-Marquardt minimization of

nonlinear least squares. Approximated model: Hyperbolic. Four approximating models.
model and each of the starting curves. Approximation results are shown in
the Figure 13.4. All four approximating models lead to excellent approx-
imation, and four approximating curves for four approximating models
are practically indistinguishable. This indicates potential flexibility in the
selection of approximating models for real data applications where actual
the model for a decline curve is uncertain. It also indicates potential prob-
lems since different models can produce equally good approximations and
at the same time may generate different forecasts.
The problem with the Levenberg-Marquardt type minimization is that
for certain starting parameters approximation fails. Thus, the selection
of starting parameters is important. A two-step procedure that includes
search on the multidimensional grid and iterative Levenberg-Marquardt
minimization is developed and reviewed in the following sections. It over-
comes the starting-point problem and avoids troublesome search for the
set of starting parameters.
13.6 Grid Search Followed by Iterative Minimization

with Levenberg-Marquardt Algorithm
The two-step approximation of the decline curves is the most efficient
approach in building approximating production decline curves. The first
10
Approximation error Grid search
8 Grid search +
Iterative minimization
6
4
2
Sepd Hyperb Duong Power. Exp

Approximation model
Figure 13.5 Approximation errors for grid search followed by iterative minimization.
step is the grid search. Results of the grid search are used as a starting point
for Levenberg-Marquardt iterative minimization.
Figure 13.5 illustrates a decrease of the approximation error by the
iterative Levenberg-Marquardt procedure that follows the grid search.
According to Figure 13.5, approximation errors are systematically lower
after the second step approximation compared to the grid search errors at
the first step.
Iterative minimization of least squares produces an approximating
curve with parameters Xapprox. These parameters are used to extrapolate the
approximation model and to construct a forecasted decline curve with val-
ues for time t outside the approximation segment:
decla (T ) = u0( X approx, t ) (13.9)

where: u0(t) is the decline curve produced by the model used in the grid
search and iterative approximation.
13.7 Two Methods for Aggregated Forecast

and Analysis of Forecast Uncertainty
Results of the decline curve approximation done by Levenberg-Marquard
minimization (nlsLM R function) are utilized in this and the following
sections for statistically justified forecasting of the decline curves and esti-
mation of the range of forecasted values. We analyze here two techniques
for joint forecasting and analysis of forecasted uncertainty:
a. Monte Carlo simulation of the forecasted decline curves,

b. Block bootstrap simulation of the forecasted decline curves.
Both approaches are used for generation of multiple randomly simu-

lated decline curves, calculation of aggregated forecasted curves and anal-
ysis of uncertainty of the forecasts.
Randomly simulated decline curves are formed as decline curves with
randomly generated parameters.
uk (t ) = u(t , X k ) (13.10)
where: Xk is the simulated vector of parameters of approximating model.

In the case of Monte Carlo simulation, random vector Xk is derived from
estimate of covariance matrix of the estimated parameters produced by
nlsLM R function. In the case of bootstrap based simulation, multiple
approximated data sets uk(t) are formed. Vector Xk is defined as vector of
parameters of the decline curve approximating uk(t).
In addition to decla(T) (Eq. 13.9) forecasted decline curve with param-
eters of the approximating model, the set of randomly simulated decline
curves (Eq. 13.10) is used to produce two types of forecasted production
curves:
declMeanR(t ) =
1
K ∑u(t , X )k (13.11)
k =1
and
declMedianR(t ) = median(u(t , X k )) (13.12)

u
where mean and median are calculated for a set of values u(t, Xk) at the
fixed time t.
The aggregated forecasts of Eq. 13.11 and 13.12 are different from each
other if distributions of randomly simulated values u(t, Xj) are not sym-
metric which is often the case for Monte Carlo simulated decline curves.
13.8 Uncertainty Quantile Ranges Obtained Using

Monte Carlo and Bootstrap Methods
The authors used estimates of quantiles of distribution of values of simu-
lated curves calculated for a discrete set of time moments t to give a com-
pact description of the possible range of predicted values of the production
trend. The quantiles were used to calculate the uncertainty quantile range
for daily or monthly production [production(t)], defined by the following
equation:
R( Plower , Pupper , t ): q( p = Plower , t ) ≤ production(t ) ≤ q( p = Pupper , t )

(13.13)
where: R(Plower,Pupper, t) is the uncertainty quantile range, q is the quantile,

Plower,Pupper are the two quantile probabilities, t is the time.
Uncertainty quantile range covers 100 * (Plower –Pupper) percent of the total
number of randomly generated decline curves. It may be interpreted as a
confidence interval at each time moment t for the unknown actual produc-
tion value. Accordingly, probability for simulated production value to get
outside of the uncertainty quantile range is:
åoutside R1 (13.14)
lower upper

Figure 13.6 shows histogram of production values generated by

Monte Carlo simulation for SEPD decline curve at Month=350. The
histogram shown at Figure 13.7, is obtained from 10,000 Monte Carlo
samples of predicted production values. According to the Figure 13.6,
the histogram is asymmetric with a long tail at large production values.
Two quantiles calculated for probabilities (Plower = 0.1 and Pupper = 0.9)
are marked by dashed vertical lines. Median is shown as a vertical con-
tinuous line. Locations of the quantiles with respect to the median are
asymmetric. The quantile pair shown in the Figure 13.6 defines a quan-
tile range R for production rate as functions of quantile probabilities p
during t=350Months.
R(0.1,0.9, t ): q( P = 0.1, t ) ≤ production(t ) ≤ q( P = 0.9, t ) (13.15)

Estimated probability for predicted production to be within the quantile

range defined by Eq. 13.14 and 13.15 is 0.8.
Quantile(P=0.1)
Quantile(P=0.9)
Median
1000
Count
500
8 10 12 14
Production(bbl/day)
Figure 13.6 Example of histogram of production values generated by Monte Carlo

simulation at Month=350.
Selection of the width of the quantile region is contradictory. The lower

the width of the quantile region, the smaller is uncertainty of predicted
results. On the other hand, probability of the actual production value to
locate outside of the quantile range is to be larger for narrower quantile
regions.
13.9 Monte Carlo Forecast and Analysis of Forecast

Uncertainty
In the case of Monte Carlo simulation of multiple decline curves, a ran-
dom vector Xk is derived from the estimate of covariance matrix of esti-
mated parameters produced by nlsLM R function. Monte Carlo analysis of
uncertainty of prediction of decline curve starts with generation of multi-
ple model parameters derived from model estimated parameters and esti-
mated covariance matrix.
Example of estimated parameters used for calculation of a forecasted
decline curve (Eq. 13.9) is given in the Table 13.1. This table also contains
indications of the reliability of the forecasted deterministic decline curve
of Eq. 13.9.
Three columns of the Table 13.1 give estimates of the parameters of the
SEPD curve, standard deviation of the estimates, and P-values for esti-
mated parameters. In this example the P-values are extremely small and
standard deviations are much smaller than the estimates of the respective
parameters. This indicates that the estimate of the parameters of approxi-
mating model is highly reliable.
Table 13.2 shows the example of estimated covariance matrix of the
coefficients of SEPD model.
Examples of several simulated vectors of parameters for SEPD model
derived from estimated covariance matrix are given in the Table 13.3
The first 10 rows in this table show 10 Monte Carlo simulated vector
coefficients produced using data illustrated by the Tables 13.1 and 13.2.
Mean_10,000 in the Table 13.3 is the average over 10,000 generated coef-
ficients. Deterministic is the actual vector of coefficients shown in the first
column of the Table 13.1. Table 13.3 illustrates that randomly generated
coefficients fluctuate around its deterministic value.
Table 13.1 Example of the estimated parameters of the SEPD-type decline curve.
Parameter Estimate Std. error t value P-value

A 360.78 30.67569 11.76116 5.98E-18
B 0.59 0.07315 8.038061 2.05E-11
C 0.31 0.021489 14.33313 4.35E-22
Table 13.2 Example of estimated covariance matrix

for parameters of SEPD decline curve.
Parameter a b c
A 940.9979 2.230323 -0.64633
B 2.230323 0.005351 -0.00156
C -0.64633 -0.00156 0.000462
Table 13.3 Monte Carlo simulated coefficients of

the SEPD model. Mean 10,000: mean over 10,000
simulated coefficients.
a b c
1 353.46 0.56 0.32
2 361.47 0.58 0.31
3 377.53 0.64 0.29
4 409.62 0.7 0.28
5 380.53 0.64 0.29
6 325.66 0.5 0.33
7 391.54 0.67 0.28
8 370.27 0.61 0.3
9 337.69 0.53 0.32
10 372.27 0.62 0.3
Mean 10,000 361.04 0.59 0.31
Deterministic 360.78 0.59 0.31
Figure 13.7 shows fifty Monte Carlo generated curves with parameters
derived from estimates of coefficient of approximated curve and covari-
ance matrix of estimated parameters. One can observe asymmetry in the
distribution of simulated curves and presence of outlier curves.
Figure 13.8 shows uncertainty regions defined by the Eq. 13.13 for
two pairs of quantiles calculated for a set of 10,000 Monte Carlo simu-
lated decline curves. Quantiles were calculated for a range of production
months of 100-500. Inasmuch as the uncertainty range for the quantile
pair is wider, the prediction of possible production values is less certain.
On the other hand, estimated probability for predicted decline curve to
be within uncertainty range limits is higher for the quantile pair (P=0.05,
0.95). The black line shifted to the upper quantile bound is the actual pro-
duction decline curve. According to Figure 13.8, the actual decline curve
is within the uncertainty region for both quantile pairs (p=0.05, 0.95) and
(P=0.10, 0.90) although it is close to the upper boundary of the quantile
range (P=0.10, 0.90).
50
40
Production (bbl/day)
30
20
10
50 100 150 200 250 300 350 400

Months
Figure 13.7 Example of fifty Monte Carlo generated SEPD curves.
Monte carlo quantile uncertainty

Quantiles: 0.05-0.95 Quantiles: 0.10-0.90
40
Daily production (bbl)
30
20
10
100 200 300 400 100 200 300 400

Production months
Figure 13.8 Two quantile ranges for Monte Carlo generated production curves.
Approximation segment - 1-100 months. Prediction range - 100-500 months. Both,
approximated decline curve and approximating model are SEPD.
13.10 Block Bootstrap Forecast and Analysis

of Forecast Uncertainty
Block bootstrap is the bootstrap version designed for the analysis of
uncertainty in parameter estimates and in the forecast when input data
are correlated. At each bootstrap run, blocks of consecutive observations
are drawn with replacement from a set of blocks. Blocks may be overlap-
ping. A new data set is formed as the union of multiple blocks. Division of
data into blocks preserves data correlation within each block. Data within
a combination of several blocks has correlation matrix similar to that of
input data if the size of individual blocks is comparable to the width of a
range of correlation within the input data.
At each bootstrap run a new input data set is formed and approximating
models with parameters Xb [1 ≤ b ≤ Nb] (Nb is total number of bootstrap
runs) are formed. As in Eq. 13.11 and 13.12, parameters of the approximat-
ing models Xb and bootstrap generated curves are used to build aggregated
bootstrap-based forecasted production curves:
Nb
declMeanBT (t ) =
1
Nb ∑u(t , X )b (13.16)
b =1
and
declMedianBT (t ) = median[u(t , X b )] (13.17)

u
Figure 13.9 shows quantile regions derived from 10,000 bootstrap-

generated curves in the two-step decline curve approximation. At first step
grid search was run. Parameters of the approximated decline curve found
in grid search were used as starting parameters for Levenberg-Marquard
minimization of nonlinear least squares. Resulting approximating decline
curve and residuals of approximation were the source for 10,000 bootstrap
runs.
Actual decline curve is the line within quantile uncertainty region closer
to its upper bound.
Figures 13.8 and 13.9 indicate that the quantile range produced by
Monte Carlo simulation is narrower compared to that produced by the
block bootstrap simulation. Another important property is that average
and median over multiple decline curves produced both by Monte Carlo
Bootstrap quantile uncertainty

Quantiles: 0.05-0.95 Quantiles: 0.10-0.90
40
Daily production (bbl)
30
20
10
100 200 300 400 100 200 300 400

Production months
Figure 13.9 Bootstrap quantile uncertainty regions for SEPD approximating curve.
Approximated curve is generated using SEPD model and disturbed by random noise.
and block bootstrap simulation are both close to the center of the quantile
region. They may be treated as predicted values of the declined production.
13.11 Comparative Analysis of Results of Monte

Carlo and Bootstrap Simulations
The goal of both Monte Carlo and bootstrap simulation is two-fold: (a)
to forecast future values of hydrocarbon production and (b) to evaluate
uncertainty of the forecasted values. Three types of forecast produced by
both methods are: (1) Forecasted individual decline curves generated by
Monte Carlo or bootstrap methods, (2) mean forecasted values calculated
as means across individual randomly generated decline curves, and (3)
median forecasted values calculated using Eq. 13.12 and 13.16.
Table 13.4 shows values of the forecasted mean and median produc-
tion values derived from Monte Carlo and bootstrap decline curves data
sets. One can observe small but systematic differences in the forecasted
values. Both mean and median production values are higher for Monte
Table 13.4 Forecasted mean and median production values derived from Monte
Carlo and bootstrap methods.
Forecasted mean production Forecasted median production
Months Monte Carlo Bootstrap Monte Carlo Bootstrap
100 30.783 24.883 30.357 24.849
150 21.467 18.726 21.012 18.67
200 16.132 15.096 15.674 14.985
250 12.686 12.672 12.234 12.521
300 10.29 10.928 9.854 10.747
350 8.539 9.608 8.119 9.389
400 7.212 8.572 6.814 8.329
Carlo simulated decline curves at smaller production months. The differ-

ence is within limits of 7 to 15%. Another feature illustrated by Table 13.4
is higher forecasted values for the mean compared to median forecast for
both Monte Carlo and bootstrap simulation.
Upper and lower quantiles for monte carlo

and bootstrap generated decline curves
50
Bootstrap quantiles
Monte carlo quantiles
40
Oil production (bbl/day)
20 30
10
50 100 150 200 250 300 350 400

Months
Figure 13.10 Upper and lower quantiles for Monte Carlo and bootstrap generated
decline curves.
Figure 13.10 shows the upper and the lower quantiles for the two sets
of Monte Carlo and bootstrap generated decline curves. The total num-
ber of decline curves in both data sets is 2500. The upper and lower
quantiles were calculated for probabilities 0.9 and 0.1. The area between
upper and lower quantiles shows forecast uncertainty area. Probability
for the decline curve to be outside the uncertainty area in both data sets
is 0.2.
There is significant similarity between two uncertainty areas. They over-
lap, although Monte Carlo generated uncertainty area is systematically
narrower and slightly shifted towards larger production values. The most
reliable strategy would be to use results of both methods of decline curves
simulation for analysis of uncertainty and forecast.
13.12 Conclusions
Conclusions can be summarized as follows:
New multi-step production trend forecast and uncertainty analysis is
presented in this paper. It includes multi-dimensional grid search as the
first step followed by iterative minimization of nonlinear least squares.
Estimated parameters of the approximating model are used then for Monte
Carlo or bootstrap simulation of multiple decline curves, calculation of
aggregated forecast values, and analysis of uncertainty of forecasted pro-
duction trends. Parameters of the approximating model are used for the
deterministic decline curve forecast.
The following decline curve models are tested and their forecast effi-
ciency is evaluated: Power-Exponential, SEPD, Duong, and Hyperbolic.
Both grid search and two-step iterative minimization proved to be flexible
and able to work with any model with similar efficiency. In some situations
however, due to inappropriate selection of a starting point, iterative min-
imization may be unstable. In order to produce stable results the iterative
minimization needs a grid search as a first step.
Three types of forecast of production decline curves are designed and
analyzed in this paper:
• Forecasted production trend derived from parameters of the

model minimizing nonlinear least squares error.
• Forecasted production trend derived from multiple Monte
Carlo simulated curves.
• Forecasted production trend derived from multiple block
bootstrap generated curves.
Uncertainty analysis of forecast of the decline curves relies on unified

approach of Monte Carlo and block bootstrap simulation. Uncertainty
forecast region is calculated as the region between upper and lower quan-
tiles. The quantile values are calculated for fixed time moments using sets
of production values generated by either Monte Carlo or bootstrap.
The new method of Monte Carlo simulation and analysis of uncer-
tainty in forecast of decline curves relies on the use of covariance matrix
of estimates of the approximating model parameters produced by the
Lenvenberg-Marquard minimization (nlsLM R function). The covariance
matrix is used to produce multiple versions of a decline curve parameters
which are then used to generate multiple decline curves.
Comparative analysis of Monte Carlo and block bootstrap simulation
indicates that these methods produce uncertainty regions of different
width that may be mutually shifted in the time-production plane. More
reliable estimate of the uncertainty region in the time-production plane
may be achieved by the joint analysis of results produced by both methods.
References
1. S. Darwis, B.N. Ruchjana, and A.K. Permadi, Robust decline curve analysis.
J. Indones. Math. Soc. (MIHMI), 15(2): 105–111 (2009).
2. B.F. Towler, and S. Bansal, Hyperbolic decline-curve analysis using linear
regression. J. Petrol. Sci. and Engr., 8 (4): 257–268 (1993).
3. Z-Sh. Lin, C-H. Chen, G.V. Chilingar, H.H. Rieke, and S. Pal, Uncertainty
analysis of production decline data. Energy Sources, 27: 463–483 (2005).
4. V. Jochen, S. Holditch, and J. Spivey, Probabilistic reserves estimation using
decline curve analysis with the bootstrap method. SPE Annual Technical
Conference and Exhibition, 6–9 Oct., Denver, Colorado (1996).
5. Y. Cheng, Y. Wang, D. McVay, and J. Lee, Practical application of a probabilis-
tic approach to estimate reserves using production decline data. SPE Annual
Technical Conference and Exhibition, 9–12 Oct., Dallas, Texas (2010).
6. J.J. Arps. Analysis of decline curves. Trans. AIME, 160: 228–247 (1945).
Agarwal et al., Analyzing well production data. SPE Reservoir Eval. & Eng.,
2(5): (1999).8. D.M. Bates and D.G. Watts, Nonlinear Regression Analysis and
Its Applications, Wiley: 359 pp. (1988).
7. M. Khulud, K.M. Rahuma, H. Mohamed, N. Hissein, and S. Giuma,
Prediction of reservoir performance applying decline curve analysis. Int. J.
Chem. Eng. Appl., 4 (2): 74–77 (2013).
8. G.A. Watson, (Ed.), The Levenberg-Marquardt algorithm: implementation
and theory, in Lecture Notes in Mathematics: Numerical Analysis, Springer-
Verlag: Berlin, pp. 105–116 (1978).
14
Oil and Gas Company Production,
Reserves, and Valuation
Mark J. Kaiser *
Center for Energy Studies, Louisiana State University, Baton Rouge, LA
Abstract
The sustainability of an oil and gas company depends upon its ability to replace
reserves at a faster pace than its production rate. The primary determinants
of the value of an oil and gas company are its cash flows and earnings, which
are dependent upon the quantity and quality of the hydrocarbons it produces,
along with commodity sales prices, production potential, which is described by
its reserves, reserves replacement rate, and its inventory of capital assets, equip-
ment, infrastructure, and acreage. The purpose of this paper is to establish the
relationship between the primary factors that influence value for a cross-section of
oil and gas companies for the year ending 2010. We construct regression models
for majors and a random sample of North American independents according to
production, reserves, technology application, and geographic diversification. We
show that reserves and production are strong indicators of market capitalization,
and for independents, production and total assets are better proxies of company
value than reserves. Multinational independents are valued higher than domestic
producers and companies producing primarily from conventional assets exhibit a
modest price premium to unconventional producers. We infer the effective cap-
italization for a sample of private companies and the National Oil Companies of
OPEC and compare model-predicted market caps for companies domiciled out-
side North America.
Keywords: Company valuation, effective capitalization, valuation modeling
289
14.1 Introduction
The value of an oil and gas company and/or property is intended to reflect
the worth of the company and/or property on the open market. According
to the 2005 International Valuation Standards, worth is defined as “the
value of property to a particular investor, or class of investors, for iden-
tified investment objectives.” The market value of a property is defined as
the “estimated amount for which a property should exchange on the date
of valuation between a willing buyer and a willing seller in an arms-length
transaction after proper marketing wherein the parties had each acted
knowledgeably, prudently, and without compulsion… reflecting the col-
lective perceptions and actions of a market …” [1].
The primary value of any company is derived from its cash flows and
earnings, which are dependent upon the quantity and quality of the
product that it provides, along with the sales price. Oil and gas compa-
nies derive their earnings from producing commodities that serve other
businesses and consumers in the economy. Production is derived from
reserves and the inventory of capital assets, production equipment, infra-
structure, and acreage. Reserves lie below the surface and have not yet
been produced but are economically and technically viable to extract. In
North America, the United States Securities and Exchange Commission
(SEC), the Ontario Security Commission (OSC), Toronto Stock Exchange
(TSE), and Canadian Security Administration (CSA) provide guidelines
on resource classifications and company requirements to list on their
stock exchanges.
Any member of society with enough money can buy shares of a public
company, but a private company has only a few owners whose shares are
not offered to the public. To estimate the value of a public company, there
are four basic valuation techniques commonly employed—book value of
assets, discounted cash flow, price earnings multiple, and market value—
which can vary considerably depending on the assumptions applied [2, 3].
For a private company, the first three methods are not an option because
detailed financial information is not publicly released.
The sustainability of oil and gas production in modern society is a cen-
tral theme of economists and policy makers alike, not to mention a concern
among the general public and government agencies, but unfortunately, the
basic premises are frequently misunderstood. Sustainable oil and gas pro-
duction is measured by the reserves to production ratio, R/P, reported by
public companies on their annual statements, and collected on a country
Oil and Gas Company Production, Reserves, and Valuation 291
basis and reported by BP’s Annual Statistical Report and related surveys.
The world R/P ratio for oil has held reasonably steady at 40 for the past 40
years, while for gas, R/P has varied around 60. The exact value for the ratio
is unknowable, of course, as countries with the most significant reserves do
not follow third-party accounting systems for reporting. Most public com-
panies have R/P ratios less than 15–20, dictated in large part by the trade-
offs between capital investment and rate of return thresholds involved in
building reserves. In recent years, R/P ratios for oil and gas companies have
increased because of significant new unconventional resource discoveries
and developments.
The purpose of this paper is to describe the primary factors that impact
the value of an oil and gas company and establish the relationships that
exist between market capitalization, reserves, production, and assets
for a cross-section of public companies. We fix the time of assessment
on December 31, 2010 to coincide with the release of production and
reserves data. By fixing the time of assessment, we eliminate the impact
of commodity price volatility on market capitalization, but the model
results are necessarily linked to a specific point in time1. Information on
†
company valuation and its relationship to firm-specific attribute data

is useful in understanding the structure of the industry [4, 5], perform-
ing company due diligence [6, 7], and revealing the connection between
investment requirements and production output. The outline of the paper
is as follows: We begin by describing the two most important determinants
of oil and gas company value, reserves and production, followed by a dis-
cussion of secondary factors. Secondary factors are often more difficult to
quantify and are subject to measurement errors, which limits their applica-
tion. Summary statistics are presented for independents and international
oil companies. The market capitalization functional is specified and the
results of correlations and regression modeling are depicted for various
categorizations. We infer the effective market cap for a sample of private
companies, the National Oil Companies (NOCs) of OPEC, and companies
domiciled outside North America. We conclude with the limitations of the
analysis and challenges in estimation.
1
†
This is not a significant drawback, however, since it is relatively easy to incorporate time
into the analysis with additional data collection.
14.2 Reserves
14.2.1 Proved Reserves
The primary assets of oil and gas companies are their entitlements to future
production from reserves. Proved reserves are defined as the estimated2 ‡
remaining quantities of oil and gas anticipated to be economically produc-

ible, as of a given date, by application of development projects to known
accumulations under existing economic and operating conditions [8].
In addition, there must exist, or there must be a reasonable expectation
that there will exist, the legal right to produce or a revenue interest in the
production, installed means of delivering oil and gas to market, and all
permits and financing required to implement the project. Proved reserves
estimates must be made with “reasonable certainty” and are defined con-
servatively in the sense that the reserves estimates are met. Engineering
and geological data are needed to make the estimates, and generally speak-
ing, the knowledge offered by greater amounts of engineering and geologi-
cal data will improve the quality of the estimation. However, uncertainty in
reserves estimation is not necessarily reduced as fields mature [9], and only
after a field is abandoned are reserves known precisely.
14.2.2 Proved Reserves Categories

Proved reserves may be developed or undeveloped and are classified into
Proved Developing Producing (PDP), Proved Developed NonProducing
(PDNP), and Proved Undeveloped (PUD) categories. PDP reserves are
expected to be recovered from completion intervals that are open and
producing. PDNP reserves are expected to be recovered from completion
intervals that are open at the time of the estimate but are not producing
(shut-in) or completion intervals that are not yet open but behind existing
wells (behind-pipe). PUD reserves are expected to be recovered from new
wells on undrilled acreage or existing wells in new formations that require
significant capital expenditures to turn into producing properties. PDP is
the least risky and the most certain proved reserves class; conversely, PUD
‡
2
The physical attributes of the asset class, located miles under the ground in rocks with
variable properties and uncertain boundaries, relying on indirect measurements that are
expensive to perform, in environments that range from benign to harsh, in both stable
and politically risky countries, with contract terms that are time dependent, means that
reserves estimates and deliverability are always uncertain. Because future production is
subject to variable production rates, unknown prices and cost, and is impacted by regula-
tory and fiscal uncertainty, the value of reserves is also uncertain.
is the most risky and least certain reserves class. Companies that produce a
majority of their production from unconventional formations tend to have
a large percent of unconventional PUD reserves.
14.2.3 Reserves Reporting

Government agencies and regulatory bodies in many countries have devel-
oped reserves reporting regulations prescribing how volumes are classified
and which volumes are reported. In the U.S., the SPE/WPC/AAPG/SPEE3 §
Petroleum Resources Management System (PRMS) is the industry stan-

dard for classifying and reporting oil and gas reserves and resources. The
Securities and Exchange Commission, the regulatory agency that oversees
the U.S. stock market and the manner publicly traded companies evalu-
ate and report their assets and cash flows, require that companies report
proved developed, undeveloped, and total proved reserves.
Reserves volumes and values for publicly traded U.S. companies are
attached to financial statements and disclosed annually on Form 10-K,
while foreign issuers trading as an American Depository Receipt file Form
20-F. Private companies have no reserves disclosure requirements. The
regulatory oversight and transparency in reserves reporting vary widely
throughout the world and reserves volumes should always be interpreted
cautiously. For some NOCs, reserves data is considered a state secret.
14.2.4 Probable and Possible Reserves

On December 29, 2008, the SEC published revised rules for oil and gas
reserves disclosures, that among other changes, allow operators to disclose
probable and possible reserves, and which now closely align with guide-
lines in PRMS [10]. The relative uncertainty of reserves is characterized
by reference to deterministic categories—proved, P1 (“much more likely
than not”); probable, P2 (“as likely than not”); and possible, P3 (“possi-
ble, but not likely”)—or in probabilistic terms. If probabilistic methods
are used, there should be at least a 90% probability that the quantities of
proved reserves actually recovered will equal or exceed the estimate. For
probable and possible reserves, the exceedance probabilities are 50% for
probable and 10% for possible reserves. For a review of the modernization
of reporting requirements and useful discussion on reserves overbooking,
see [11–16].
§
3
Society of Petroleum Engineers, World Petroleum Congress, Society of Petroleum
Evaluation Engineers, American Association of Petroleum Geologist.
14.2.5 Contractual Differences

Under leases and concessions, the host government grants the producing
company the right to explore for, develop, produce, transport, and mar-
ket hydrocarbons within a fixed area for a specific amount of time. The
company typically bears all of the risks and costs for exploration, develop-
ment, and production and holds title to all resources produced while the
agreement is in effect. The contractor reports reserves consistent with the
networking interest after deduction of royalties.
Under production sharing contracts, the contractor is given the oppor-
tunity to recover the investment from production subject to specific limits
and terms (cost oil), and also receives a specific share of the production
remaining after cost recovery (profit oil). Resource ownership is retained
by the host government, and reserves consistent with the cost recovery
plus profit oil are reported by the contractor [17].
A typical feature of production sharing contracts is that as oil prices rise,
the amount of oil and gas the company is entitled to take, as opposed to the
government’s share, falls. Thus, while rising energy prices are beneficial to
the oil companies in general, they cut into overseas production growth and
the amount of booked reserves. According to IHS Herold, about 20% of the
proven reserves of ConocoPhillips, Exxon, BP, and Chevron were subject
to production sharing contracts in 2010; Shell and Total had about a 30%
exposure [18].
14.3 Production
Production is the causal result of reserves and an important measure of
performance because it determines gross revenue, and when combined
with costs, the cash flow and profitability of a property. Production data
for publicly traded companies are compiled at various levels of aggrega-
tion and frequency depending on the state/country of operation. Wells,
reservoirs, properties, fields, and projects are typical evaluation units and
production is frequently reported to regulatory authorities on a monthly
basis. Wells are the preferred unit for performance analysis and fields are
the highest level for which reserves are estimated.
Operators produce at rates to maximize return on investment, but dif-
ferences arise in how oil and gas is produced depending on ownership,
product type, location, and prices. National Oil Companies have broader
constraints and obligations than public corporations and OPEC NOCs are
subject to restrictions on production via a quota system.
Companies may shut-in or curtail gas production to protect wellbore

stability or because of low prices, while oil wells are almost always operated
at full capacity. Associated gas is produced in conjunction with crude oil,
and because crude is produced at capacity, the quantity of associated gas
is linked to crude production. Non-associated gas is more responsive to
price fluctuations. For example, non-associated gas may be choked back
(reduced) in the summer season in anticipation of higher gas prices during
the winter season. This sort of operational flexibility is generally not pres-
ent with associated gas production.
Commodity price provides thresholds for decision making, and are also
key elements in production sharing contracts, where royalty rates and other
terms are frequently linked to sliding price scales, and as prices increase, a
smaller share of production (and subsequently, booked reserves) are allo-
cated to the contractor.
Production and reserves are not strongly correlated at the field level
because production depends on the life cycle stage of the asset. At a cor-
porate level, however, production and reserves are expected to be more
closely related because fields of many different sizes, types, locations, and
ages are aggregated, which integrate and normalize the production and life
cycle variations of development. At a corporate level, production is more
likely to serve as a proxy for a company’s reserves position.
14.4 Factors that Impact Company Value

14.4.1 Ownership
The oil and gas industry consists of hundreds of public and private com-
panies of different sizes, strategies, and capabilities. Public corporations
have broadly dispersed ownership, while companies with concentrated
ownership are privately owned. In the U.S., private companies are the most
numerous and tend to be much smaller than publicly traded companies.
Internationally, the number and type of oil companies reflects historic
trends, national policy, ownership and access issues.
14.4.1.1 International Oil Companies

Oil and gas companies that compete globally, and often in partnership
with NOCs, are referred to as International Oil Companies (IOCs). IOCs
are listed on stock exchanges with 100% of their stock owned by the pub-
lic, and they publish annual reports and abide by regulatory rules of the
stock exchange listing. IOCs pursue activities to grow their stock price
and are primarily concerned about production, cash flows, and booking
reserves. Several IOCs are integrated across the supply chain in the sense
that they not only explore for, develop, and produce oil and gas, but also
refine, transport, and market their petroleum products. Exxon, Shell, BP,
Chevron, Total, ConocoPhillips and Eni are integrated IOCs commonly
referred to as the majors or supermajors. Exxon was the largest gas pro-
ducer in the U.S. in 2011 and the world’s largest publicly traded oil com-
pany. Eni is publicly traded and also government owned. In 2012, Conoco
Phillips split into separate E&P and refining businesses and is no longer
classified as an integrated IOC.
14.4.1.2 National Oil Companies

NOCs are owned entirely by state government and in contrast to IOCs,
production, reserves, revenue, profitability, and other key variables are not
publicly disclosed. NOCs dominate the global energy economy and control
more than three-quarters of the world’s oil and gas reserves and one-third
of supply. The structure, function, and roles of NOCs vary widely depend-
ing on the country’s economic and political system, strategic objectives,
reserves base, degree of privatization, and other factors. Most NOCs were
formed to manage the country’s hydrocarbon resources and local resources
are developed with the aid of international service companies. Resource-
rich NOCs typically have limited joint venture activity; e.g., Saudi Aramco
(Saudi Arabia), Russia (Lukoil). Some NOCs, such as Nigeria National
Petroleum Company, KazMunaiGas (Kazakhstan), Petronas (Malaysia),
and China National Petroleum Company, are actively engaged with IOCs.
14.4.1.3 Government Sponsored Entities

NOCs privatize (e.g., sell state assets to private investors) to raise capital
and change their mission and strategy. This has lead to a third class of
company that has become increasingly popular: companies that are part
public and part government-owned, referred to as Government Sponsored
Enterprises (GSEs). As NOCs get larger and more global and list their
shares, the boundaries between IOCs and NOCs become less distinct.
Eni from Italy, Petrobras from Brazil, Statoil from Norway, and
PetroChina and Sinopec from China are examples of GSEs. Some GSEs
have shed all previous government ownership (e.g., BP, Total, Repsol,
PetroCanada), while other GSEs continue to have varying degree of state
ownership (e.g., Sinopec (76%)4, Eni (30%), PetroChina (90%), Statoil

¶
(63%), Petrobras (32%), GazProm (51%), Rosneft (89%), and ONGC

(84%)). The publicly owned portion of GSEs provide production and
financial data in annual reports and are frequently listed on international
stock exchanges.
GSEs structure themselves as corporations, but their motivations are
both commercial and political, and these partially privatized companies
may operate in ways that reflect the interest of their national governments.
GSEs actively seek investment funds and have a structure that allows them
to access capital markets. Some companies were founded with the inten-
tion to manage and control the country’s oil and gas development while
other companies (e.g., Lukoil, Statoil, Petrobras, PetroChina, CNPC, and
ONGC) pursue development internationally and increasingly compete
with IOCs and each other. Some GSEs are as technologically advanced as
the majors. Petrobras, for example, is a world leader in deepwater devel-
opment while Statoil is a leader in Arctic offshore and subsea technology.
Most GSEs are actively engaged in joint venture activities with IOCs and
each other.
14.4.1.4 Independents and Juniors

Independents are non-government owned companies that focus either
upstream or downstream. Some of the largest E&P companies include
Occidental, Canadian National, Apache, Devon, BHP Billiton (Australia),
and Woodside (Australia). Small upstream producers are referred to as
juniors. Junior oil and gas companies number in the hundreds and pro-
duce between 500 to 10,000 barrels per day (b/d) and are a critical element
of the industry for technology transfer.
14.4.2 Degree of Integration

IOCs, NOCs, and GSEs are integrated to various extents across upstream
and downstream activities depending on their economic and political sys-
tems, reserves base, business strategy, consumption needs, and related fac-
tors. Independents explore for, develop, and produce oil and gas, but do
not own transportation or refinery or marketing operations. Companies
that refine, transport, and market their petroleum products in addition
to their upstream activities are referred to as integrated companies. All of
the major multinational IOCs are integrated although different strategies
¶
4
Estimated state ownership in 2010 denoted in parenthesis.
are employed which change over time. Some companies, such as BP and
Chevron, balance their production and refinery capacity, while Exxon and
Shell have a much higher refinery capacity than their production business
can provide, which require they purchase additional crude for their refin-
ery runs. Major exporting countries and NOC production capacity usually
far exceeds domestic refining capacity.
14.4.3 Product Mix

Most oil wells produce natural gas (called associated gas) and many gas
wells produce liquid hydrocarbons (called condensate or natural gas liq-
uids) that fall out of the production stream or are stripped out at a gas
plant. Wells and fields are classified as primarily oil or gas depending on the
volume of produced gas (measured in cubic feet) per unit of produced oil
(measured in barrels). We apply the same convention to reserves holdings
and classify companies based on their relative mix of oil and gas reserves.
Oil and gas proved reserves are denoted by Ro and Rg and companies with
Rg/Ro < 10,000 cf/bbl are classified as oil companies; gas companies are
identified by Rg/Ro > 10,000 cf/bbl.
14.4.4 Commodity Price

Oil and gas companies are dependent upon commodity prices for their
earnings and value, and they are price takers, meaning that they have to
sell their output at the prevailing market price. Oil is a global commodity
and price is controlled by global supply and demand conditions, while gas
is priced according to local and regional considerations. In the U.S., oil and
gas prices are determined in separate spot markets, and historically, gas has
traded at a significant discount to oil on a heat-equivalent basis. Natural
gas liquids prices vary over time with local supply and demand conditions,
petrochemical requirements, imports and infrastructure conditions.
Changes in the local demand and supply balance of gas impact prices
and production decisions. Companies with a large amount of production
outside of the U.S. are exposed to Brent oil prices and gas prices typically
linked to oil, which will bolster earnings relative to companies with mainly
U.S. production. Since late 2009, Chevron, for example, has earned higher
net income per barrel than Exxon due in part to Chevron’s greater leverage
to global oil prices rather than U.S. gas prices [19]. For large established
oil and gas companies, earnings and cash flow are expected to track com-
modity prices, while for smaller companies, the relationship is expected to
be weaker [3]. Robust and predictable relations in earnings and cash flow
for large companies may translate into a value premium relative to small
companies.
14.4.5 Production Cost

Production cost is a factor in company value as cost is a primary determi-
nant of profitability. Production cost (also called lifting costs) are the costs
to operate and maintain wells and related equipment and facilities per bar-
rel of oil equivalent (boe)5 produced by those facilities after the hydrocar-
**
bons have been found, acquired, and developed. Direct lifting costs are
total production spending minus production taxes. Total lifting costs are
the sum of direct lifting costs and production taxes. Production taxes typ-
ically rise and fall with changes in the prices of oil and natural gas, and
lifting cost varies by region and time. In 2009, worldwide total lifting costs
for U.S. IOCs were approximately $10/boe [20].
Each oil and gas deposit has its own extraction, processing, and transpor-
tation cost. Extraction cost depends on the size of deposit, reservoir conti-
nuity, location, fluid flow characteristics, and other factors. Processing cost
depends on the type and mixture of oil and gas, while transportation cost
is incurred to deliver the product to market and is usually a small part of
the total cost. Production cost is often a small part of the commodity price
over most of the life cycle of a field, but when deposits deplete, the cost of
production rises as a deposit approaches its economic limit. Companies
that are listed on U.S. exchange markets are required to file data on capital
expenditures, and for PDP reserves, the per-unit costs of extraction, but
because cost are aggregated over regional basins, the ability to infer useful
information from reported data is constrained.
14.4.6 Finding Cost

Finding cost are the average costs of adding proved reserves via explora-
tion and development activities. These costs are measured on a combined
basis in dollars per boe, and a number of difficulties arise in their applica-
tion. Finding costs are frequently reported over a specified period of time
(usually three to five years) and measurement errors and reliable delin-
eations of exploration and development expenditures, reserves additions,
and asset purchases often make comparisons ambiguous. Finding costs are
reported by a number of different organizations, but caution should always
be used in their usage. In general, companies with a low finding cost are
**
5
Heat-equivalent reserves are computed using 6,000 cf = 1 boe.
likely to be valued higher than companies with a high finding cost because
of its impact on project profitability and perception of market participants.
14.4.7 Assets
Assets represent capital invested. The scale of assets required to produce
oil and gas is enormous, especially for complex and large field develop-
ments in harsh environments. A company’s balance sheet summarizes its
financial position at a point in time and is a list of all its assets, liabilities,
and equity. Total assets are the total fixed assets of the firm reduced by
the accumulated depreciation of those assets as they age. The book value
of assets reflects the historical costs rather than current market value or
replacement cost. Assets typically include surface facilities, equipment and
infrastructure, processing plants, offshore structures, refineries, loading
facilities, pipelines, property, and buildings. Capital equipment is often
custom-designed and permanently installed at specific fields and have little
or no alternative uses.
14.4.8 Capital Structure

The capital structure of a company refers to the combination of securi-
ties employed (debt and equity) to raise and maintain the capital assets
required to conduct business. Debt carries an obligation to repay the prin-
cipal by a specified date plus a given rate of interest. Default on debt obli-
gations can place the company in bankruptcy. Equity is in common and
preferred stock. The degree to which a firm is funded by loans is known
as leverage and is usually expressed as the debt equity ratio (D/E). Firms
operating in the same segments of the value chain typically possess similar
capital structures, but considerable variation exists in the capital structure
of oil and gas companies and their D/E ratio. IOCs and GSEs use cash flow
from operations as the predominant source of funds and are also active in
the capital markets. Independents rely on revolving credit, long-term debt
and equity securities, as well as cash generated by operations.
A company’s D/E ratio is influenced by their participation across the
supply chain, access to capital markets, business strategy and operational
performance. A high debt ratio signifies a high risk for payment problems,
which can impede business growth and lead to liquidation. High indebt-
edness may prevent a company from obtaining additional financing to
fund future capital expenditures, acquisitions, or other general corporate
requirements. Analyst and rating agencies usually express concern with
high debt, financial complexity, and off-balance sheet obligations because
these all present risk for shareholders. Aggressive-growth independents

leverage their financial strength to capture business opportunities and may
have D/E > 5, while more conservative companies with D/E < 0.5 oper-
ate within cash flows to avoid carrying interest-bearing debt. Most IOCs
maintain a balanced approach to leverage and company D/E ratios typi-
cally range between 0.5 and 1.5.
14.4.9 Geologic Diversification

Reserves and production are classified according to the geologic nature
of the resource and the manner of its extraction. Conventional produc-
tion occurs from a discrete accumulation or set of accumulations that are
bounded by a downdip water contact from which hydrocarbons in liquid
or gaseous forms are extracted from medium to high permeable sands
with conventional technology. Unconventional reservoir systems are often
synonymous with “continuous” accumulations and include shale gas, tight
oil, and coalbed methane reservoirs. Unconventional oil and gas resources
are widely abundant, but their development is sensitive to technologic
and geologic risk and commodity price variation. Unconventional fields
are more expensive and riskier to develop than conventional fields and are
more difficult to characterize.
U.S. gas shale development was launched by small, nimble independents,
where their decentralized approach facilitated developments requiring mul-
tiple wells and trial-and-error approaches to optimize production. Majors
have traditionally focused on conventional megaprojects, but in 2009,
Exxon purchased XTO Energy for $41 billion to gain experience in uncon-
ventional gas production. Chevron and other majors have purchased acre-
age positions in unconventional plays and are currently pursuing exposure
in the area. When unconventional technologies and evaluation techniques
become standard practice, unconventional resources will be classified as
conventional. For all things equal, companies with large resource holdings
or production in unconventional plays are expected to be valued at a dis-
count to conventional resource positions because of the greater uncertain-
ties involved and higher levels of PUD reserves classification.
14.4.10 Geographic Diversification

Oil and gas development occurs throughout the world in a variety of envi-
ronments in countries with different political risks. Political risk represents
the unexpected impact on operating cash flows that result from political
actions and events, ranging from nationalization to unilaterally changing
contract terms. These risks can be minimized through insurance, strategic

alliances and partnering, avoiding geographical concentration, and main-
taining close and economic links with the host country. Companies pursue
different strategies to balance and optimize their asset portfolio.
The areas/basins of operations describe a company’s geographic diver-
sification and political risks. Remote, harsh, and environmentally sensitive
regions present unique challenges and risks. Offshore exploration and pro-
duction is more expensive and often more difficult technologically than
onshore operations. Various authors have attempted to quantify a compa-
ny’s political and geographic risk using expert opinion and ordinal rank-
ings, but such systems do not allow a robust or reliable comparison because
the underlying political systems are so complex and diverse and unpredict-
able that analytic models cannot reliably capture the essential drivers. Our
approach is simplistic and merely identifies companies according to their
multinational or domestic base of operations.
Multinational independents are involved in several countries and
regions, while small independents often specialize in one region where they
have expert knowledge in geology and contracting terms. IOCs are glob-
ally diversified and most top-tier independents are active in several coun-
tries. Anadarko, for example, is one of the most diversified independents
with areas of operation in the deepwater Gulf of Mexico, Algeria, Alaska,
Brazil, China, Indonesia, Mozambique, and West Africa. Most NOCs have
a strong regional focus, while many GSEs are similar to IOCs in their geo-
graphic coverage. Domestic independents usually operate exclusively off-
shore or onshore, although a few companies have assets in both areas.
14.4.11 Unobservable Factors

Additional factors which may impact company value but are not directly
observable or are highly subjective or temporary in nature, include man-
agement quality, takeover threats6, exploration potential, human capital,
††
and technology specialization. Because our assessment occurs at a point

in time, individual companies may be subject to specific events that do not
††
For example, Argentina’s largest oil company, YPF SA, majority owned by Spain’s Repsol
6
YPF, was the target of a recent government takeover. Prior to the takeover, many of YPF’s
concessions were revoked, sending its market capitalization to $9 billion in April 2012,
from $17.5 billion in 2011 [21]. Some industry insiders theorize the Argentine govern-
ment encouraged provincial leaders to revoke YPF’s oil concessions to reduce its mar-
ket capitalization to lower the amount the government would have to pay in the event of
nationalization.
impact other companies. For small or specialized companies these factors

might be important, but for most companies they are not expected to play
a large role.
14.5 Summary Statistics

14.5.1 Sample
A random sample of 70 independents with headquarters in North America
were considered along with seven majors: Exxon, Shell, BP, Chevron, Total,
ConocoPhillips, and Eni. The independents were identified from the 2010
Oil and Gas Journal’s OGJ150 (Figure 14.1) and other public sources.
The complete list of independents is provided in Tables A.1 and A.2 in
Appendix A.
14.5.2 Variables
A summary of the model variables and data sources are provided in Table
14.1. By fixing the time of assessment on December 31, 2010, we eliminate
the impact of oil and gas price on company valuation, and match the pro-
duction and reserves reporting of companies.
The market capitalization of a publicly traded corporation is the num-
ber of shares issued and outstanding, multiplied by the per share price at a
point in time.
100000 80
10000 70
Cumulative Reserves (Bboe)
Proved Reserves (MMboe)
60
1000
50
100
40
10
30
1
20
0.1 10
0.01 0
1 11 21 31 41 51 61 71 81 91 101 111 121 131
Companies
Figure 14.1 Independents sampled from the 2010 OGJ150 are shown as dark lines.
Table 14.1 Model variables and definitions.

Variable Notation Definition Source
Market CAP Number of outstanding Scottrade and
Capitalization shares multiplied Finance-
by stock price on Yahoo,
12/31/2010 Financial
Times
Proved Reserves Ro, Rg, Oil, gas, and boe Company Form
Rboe proved reserves 10-K
reported on
12/31/2010
Production Po, Pg, Oil, gas, and boe Company Form
Pboe annual production 10-K
in 2010
Proved PUD/R Ratio of proved Company Form
Undeveloped to undeveloped 10-K
Total Reserves reserves to total
Ratio proved reserves on
12/31/2010
Reserves to R/P Ratio of proved
Production Ratio reserves to 2010
annual production
in boe
Debt to Equity D/E Ratio of debt to equity Scottrade
Ratio on 12/31/2010
Reserves reported by consolidated subsidiaries7 and equity-accounted

‡‡
entities are combined. All resource companies are exposed to commodity

price swings due to economic activity supply and demand conditions, and
related factors.
Reserves life is computed as proved reserves divided by annual pro-
duction and is known as the R/P ratio. Proved undeveloped reserves are
reported in accord with SEC guidelines. A company’s proved undevel-
oped to total proved reserves, PUD/R, indicates how much of its proved
reserves base is currently producing. Reserves that are undeveloped require
‡‡
7
Consolidated subsidiaries are the business entities whose voting stock is less than 50%
controlled by the company.
significant capital expenditures to convert into producing fields and cash

flow generating assets.
Equity is the capital the owners of a firm have in the business and is capi-
tal at risk. Profits that are not paid out to owners are retained and add to the
equity base of the company. Debt is the capital gained from outside parties
for a limited period of time at a defined cost. The debt equity ratio (D/E) is
a measure of financial risk and measures the equity of the company relative
to its debt position.
14.5.3 Data Source

Market capitalization and D/E data was determined on December 31, 2010
using Scottrade, Finance-Yahoo, and Financial Times Global 500 [22].
Proved reserves and production data are collected from companies’ annual
reports for the year ending 2010. PUD/R and R/P are defined in terms
of heat-equivalent reserves and production and were computed based on
reported data; PUD/R is described as a percentage and R/P is described in
years. Asset values were collected from annual reports and the Oil and Gas
Journal.
14.5.4 International Oil Companies

In 2010, the seven majors held reserves of 48.5 billion barrels (Bbbl) oil and
258 trillion cubic feet (Tcf) gas; produced 7.4 Bboe, and collectively had a
R/P ratio of 12.4. Their combined market capitalization was $1.2 trillion
with reported book value of assets of $1.4 trillion. The D/E ratio of majors
averaged 1.1 and 40% of their proved reserves were undeveloped.
Exxon had the largest market capitalization on December 31, 2010
among IOCs at $369 billion followed by Shell at $209 billion, Chevron at
$184 billion, BP at $137 billion, and Total at $125 billion (Table 14.2). All
IOCs are primarily oil producing companies with Shell (Rg/Ro = 7.7) and
Exxon (Rg/Ro = 6.8) having the greatest percentage of gas reserves.
Exxon, Shell, BP, and Chevron held more than 10 Bboe reserves and
produced more than 1 Bboe in 2010. PUD/R ratios ranged from 27%
(ConocoPhillips) to 53% (Total); R/P ratios ranged from 10.5 (Chevron)
to 14.8 (Exxon); D/E ratios ranged from 0.61 (Eni) to 1.87 (BP). The D/E
ratios of Chevron (0.76) and Exxon (1.06) fell below the group average,
while ConocoPhillips (1.28) exceeded the average.
Table 14.2 Integrated oil company reserves and financial data (2010).
Proved reserves
R /R g 0
Oil BOE Production PUD/R (Mcf/ R/P Market Cap
Company (Bbbl) Gas (Tcf) (Bboe) (MMboe) (%) bbl) (yr) D/E ($billion)
Exxon 11.67 78.82 24.81 1,673 31 6.8 14.8 1.06 369
Shell 6.15 47.14 14.27 1,157 48 7.7 12.3 1.15 209
BP 10.71 42.70 17.83 1,378 44 4.0 12.9 1.87 137

Chevron 6.50 24.25 10.54 1,008 37 3.7 10.5 0.76 184
Total 5.20 25.79 9.50 830 53 5.0 11.4 1.38 125
ConocoPhilips 4.69 21.72 8.31 733 27 4.6 11.3 1.28 100
Eni 3.62 17.88 6.84 621 42 4.9 11.0 0.61 88
All 48.54 258.29 92.10 7,399 40 5.3 12.4 1.11 1,211
Source: Companies’ 2010 Annual Report, Financial Times, Yahoo-Finance, Scottrade.
Table 14.3 Statistical summary of the independent oil and gas companies sampled (2010).
Proved reserves
Liquid Total Production PUD/R Market Cap
Product Class Sample (MMbbl) Gas (Bcf) (MMboe) (MMboe) R/P (yr) (%) D/E ($million)
All Large 29 359(383) 3,606(3,514) 960(828) 73(69) 13(6.7) 39(14) 1.2(0.56) 12,746(11,014)
a
Small 41 17(19) 171(230) 45(49) 3.8(4.4) 12(21) 48(24) 1.4(4.6) 783(866)
All 70 158(299) 1,594(2,830) 424(699) 33(56) 13(17) 40(20) 1.2(2.9) 5,739(9,243)
Gas Large 15 161(176) 4,204(3,499) 861(692) 60(52) 14(21) 45(13) 1.3(0.43) 10,189(6,518)
Small 15 10(16) 305(304) 61(61) 4.4(4.9) 14(5.1) 45(20) 1.1(0.81) 776(1,061)
All 30 85(146) 2,255(3,157) 461(634) 32(46) 14(16) 45(18) 1.3(0.65) 5,482(6,630)
Oil Large 14 571(428) 2,966(3,416) 1,065( 88(81) 12(6.7) 34(11) 1.1(0.56) 15,486(13,829)
940)
Small 26 21(20) 94(118) 36(37) 3.5(4) 10(21) 51(25) 1.7(4.6) 787(731)
All 40 213(365) 1,099(2,443) 396(743) 33(63) 12(18) 35(21) 1.1(3.7) 5,932(10,791)
a
Standard deviations are denoted in parenthesis.
14.5.5 Independents
The 70 independents sampled represented 29 large and 41 small cap com-
panies distributed across 40 oil and 30 gas producers (Table 14.3). In total,
the 70 independents held combined reserves of 30 Bboe, produced 2.3
Bboe, and had a combined market cap of $401 billion at the end of 2010.
Apache, Devon, and Anadarko are the largest oil independents with
each company’s proved reserves over 2 Bboe and more than 200 MMboe
annual production, with market caps ranging from $34 billion (Devon) to
$46 billion (Apache).
Chesapeake, EOG Resources, and Talisman Energy are the largest
North American gas producing companies with reserves greater than 8
Tcfe and market caps ranging from $16.9 billion (Chesapeake) to $23.2 bil-
lion (EOG Resources). Both Chesapeake and EOG Resources report large
proved undeveloped reserves in unconventional plays (47% and 48% of
their reserves base, respectively).
Most independents hold less than 200 MMboe reserves: independents
with more than 200 MMboe reserves are classified as large-cap companies,
and the top-tier large cap independents have more than 1 Bboe reserves.
Top-tier large cap independents include oil companies Apache, Devon,
Anadarko, Marathon Oil, and Pioneer, and gas companies Chesapeake,
EOG Resources, Talisman Energy, and Noble Energy.
The average large-cap oil producer had 1.1 Bboe reserves, 88 MMboe
production, and a market cap of $15.5 billion on December 31, 2010, while
the average small-cap oil producer had 36 MMboe reserves, 3.5 MMboe
production, and $787 million market cap. The large cap oil producers had
a slightly smaller D/E ratio (1.1 vs. 1.7) and a slightly larger R/P ratio (12
vs. 10) compared to the small producer group and a smaller PUD/R per-
centage (34% vs. 53%).
The average large-cap gas producer held 861 MMboe reserves, pro-
duced 60 MMboe, and had a market cap of $10.2 billion on December 31,
2010, while the average small-cap gas producer had 61 MMboe reserves,
4.4 MMboe production, and $776 million market cap. Gas producers were
similar in R/P and PUD/R ratios, but large producers had slightly higher
average D/E ratios (1.3 vs. 1.1).
Aggressive-growth independents (e.g., Callon and ATP Oil & Gas)
leverage their financial strength to capture business opportunities and
have D/E > 10, while more conservative companies (e.g., Evolution and
Houston American Energy) follow a “zero-debt” policy to avoid carrying
interest-bearing debt and have D/E < 0.2. The independents had a group
average D/E ratio of 1.1 similar to the majors but with a much higher stan-
dard deviation reflecting a larger variation in financing structure.
14.6 Market Capitalization

14.6.1 Functional Specification
Market capitalization representing the value of the number of shares out-
standing multiplied by the share price on December 31, 2010 was regressed
by company class and primary product against oil and gas reserves, reserves
to production ratio, and debt equity ratio using the following relations: ·
CAP = a + b ⋅R boe + c ⋅R / P + d ⋅D / E
CAP = a+ b ⋅ R o + c ⋅R g +d ⋅ R / P + e ⋅ D / E.

where CAP is the market capitalization ($); Ro, Rg, Rboe represent the vol-
umes of oil (bbl), gas (cf) and heat-equivalent (boe) reserves, respectively;
R/P is the reserves to production ratio (yr); and D/E is the debt equity ratio.
14.6.2 Expectations
Proved reserves are expected to be a primary indicator of company value
and to be positively correlated with market capitalization at a given point in
time. Reserves represent the inventory of the company, and larger reserves
are expected to be associated with higher valuation and longer produc-
tion life. For all other things equal, high R/P ratios indicate that produc-
tion and cash flows are weighted to the future, and thus, high R/P ratios
could be negatively correlated with market value, but if large inventories
are perceived by the market as valuable, high R/P ratios could be positively
interpreted. Hence, the “expected” sign is ambiguous and the market is the
determining force in its realization.
Highly leveraged firms have higher fixed charges in the form of inter-
est payments relative to discretionary outlays, such as dividend payments.
The higher the D/E ratio, the greater the financial risk, and the lower the
expected market cap. The market cap for oil producers are expected to
dominate gas producers because of the price premium of liquid hydrocar-
bons, but obviously, many other factors also occur. Companies that are less
geologically and geographically diversified, or have significant percentage
of unconventional resources, also point to greater risk and may contribute

to a reduction in market cap.
14.7 International Oil Companies

Market capitalization for international and integrated oil companies are
correlated with both reserves (Figure 14.2) and production (Figure 14.3).
BP is valued significantly below the market average because of the liabili-
ties associated with Macondo oil spill which at the time of the assessment
were unsettled. On December 31, 2010, the market was valuing future
400
Exxon
y = 0.0132x–0.3031
Market Cap ($billion)
300
R2 = 0.7524
Shell
200 Chevron
BP
Eni Total
100
ConocoPhillips
0
0 5,000 10,000 15,000 20,000 25,000 30,000
Figure 14.2 Market capitalization and proved reserves – integrated oil companies (2010).
400
Exxon
y = 0.2183x–57.829
300
R2 = 0.7188
Shell
200 Chevron
BP
Eni Total
100
ConocoPhillips
0
400 600 800 1,000 1,200 1,400 1,600 1,800
Production (MMboe)
Figure 14.3 Market capitalization and production – integrated oil companies (2010).
liability at approximately $100 billion. In April 2011, the known liabilities

were estimated at $40 billion [23]. In July 2013, some observers tagged BP’s
total liability up to $125 billion [24, 25].
Assets are only weakly correlated with market capitalization because
they are reported at book value and the relative amount of refining, trans-
portation, and related infrastructure will vary significantly with the degree
of integration (Figure 14.4). We do not expect nor do we observe a strong
correspondence between market cap and asset value for IOCs.
Reserves provide the highest correlation and the greatest predictive power
of IOC value. Since only seven companies comprise the sample set, the use of
two or more variables in regression will over specify the model. Nonetheless,
the results of multi-factor models are reported as shown in Table 14.4. All the
coefficients are of the expected sign but R/P and D/E are not always statis-
tical significant.
500
Exxon
400
y = 1.3696x–88.798
R2 = 0.3468
300
Shell
200
Chevron BP
Total
100
ConocoPhillips Eni
0
100 150 200 250 300 350
Total Assets ($billion)
Figure 14.4 Market capitalization and total assets – integrated oil companies (2010).
Table 14.4 Regression results of integrated oil companies (2010).

CAP ($billion) = a + b ∙ Ro (Bbbl) + c ∙ Rg (Tcf) + d ∙ R/P (yr) + e ∙ D/E
a b c d e R2
527(1.8) 3.7(0.45) 7.1(3.1) -49(-1.5) -47(-1.2) 0.92
CAP ($billion) = a + b·Rboe (Bboe) + c·R/P (yr) + d·D/E
a b c d R2
141(0.46) 17(2.1) –6.1(– –98(–2.3) 0.83
0.18)
14.8 U.S. Independents

14.8.1 Large vs. Small Cap, Oil vs. Gas
Market capitalization for large independents are strongly correlated with
reserves for oil producers and less so for gas producers (Figure 14.5), and for
small-cap companies the correlations deteriorate significantly across both
groups (Figure 14.6). Large cap independents are more uniform as a group
50
y = 0.014x + 0.445
40 R2 = 0.9124
30
y = 0.0072x + 4.142
R2 = 0.6303
20
10
0
0 500 1,000 1,500 2,000 2,500 3,000 3,500
Gas Oil
Small-Cap Oil Producer Group Small-Cap Gas Producer Group
Linear (Gas) Linear (Oil)
Figure 14.5 Market capitalization and proved reserves — large-cap independents (2010).
4
3
y = 0.012x + 0.366
R2 = 0.3537
2
y = 0.011x + 0.0768
R2 = 0.4388
1
0
0 50 100 150 200
Gas Oil Linear (Gas) Linear (Oil)
Figure 14.6 Market capitalization and proved reserves — small-cap independents (2010).
in terms of production and financial metrics, whereas small-cap indepen-

dents are more heterogeneous which in part are reflected in low correlations.
Oil companies tend to have a greater market cap than gas companies
for both large and small producers which is probably the result of the price
premium afforded by liquid hydrocarbons and might also reflect the low-
cost high-margin attributes of the conventional resource class. Gas com-
panies with large positions in unconventional plays typically have a large
portion of proved undeveloped reserves on the books.
There is a strong and significant correlation between market cap and the
asset position of independents because assets are used in direct support
of oil and gas production activity and there are limited ancillary business
segments (Figure 14.7). This contrasts sharply with IOCs because of the
broader business segments over which IOCs operate.
Proved reserves and market capitalization for independents are also
more strongly correlated than IOCs (Figure 14.8). Reserves, production,
and assets are all strongly co-linear and are the primary components of
market capitalization of independents. Companies with small market cap-
italization tend to exhibit higher sensitivity to commodity price changes,
and although we did not include this aspect within the model, it is well
known that the net operating income of established integrated companies
correlates closely with commodity prices.
50
y = 0.8053x + 1.0413
R2 = 0.8088
40
30
y = 0.8301x + 1.7071
R2 = 0.7654
20
10
0
0 10 20 30 40 50 60
Total Assets ($billion)
Figure 14.7 Market capitalization and total assets – independents (2010).

50
y = 0.014x + 0.308
R2 = 0.9476
40
30
20
y = 0.0091x + 1.419
R2 = 0.762
10
0
0 500 1,000 1,500 2,000 2,500 3,000 3,500
Figure14. 8 Market capitalization and proved reserves – independents (2010).
14.8.2 Consolidated Small-Caps

Small-cap companies are expected to be valued on a proportional basis
similar to large-cap companies. In Figure 14.5, we consolidated the small-
cap oil and gas producers of the sample in terms of their reserves and
market values. The small-cap gas producer group nearly coincides with
the large-cap regression line as we would expect, whereas the small-cap
oil producer group is valued significantly higher relative to the large-cap
producers.
These differentials may be a general characteristic of small-cap compa-
nies or result from differences specific to the sample. Market premiums
reflect the market expectation that the company might make new discov-
eries, increase their production and reserves base, have greater future earn-
ings potential, or have an advantageous position in acreage or technology
application. Small-cap oil producers may be overvalued or reflect benefits,
real or perceived, of market participants.
14.8.3 Multinational vs. Domestic

Multinational independents have a higher valuation than domestic produc-
ers, which may be due to the profitability and size of fields, growth potential,
acreage positions, favorable contractual arrangements, and the premium
afforded by geographic diversification (Figure 14.9). Multinational inde-
pendents are the largest independents and exhibit higher premiums rela-
tive to independents with only domestic operations. We did not specify the
50
y = 0.014x + 0.5704
R2 = 0.9424
40
30
20
y = 0.0073x + 1.1748
10 R2 = 0.7119
0
0 500 1,000 1,500 2,000 2,500 3,000 3,500
Multinational Domestic Linear (Multinational) Linear (Domestic)
Figure 14.9 Multinational vs. domestic independents market capitalization (2010).
international and domestic portion of production or reserves since we do

not believe a finer delineation is suitable in the analysis.
14.8.4 Conventional vs. Unconventional

Producers with mostly conventional assets exhibit a modest premium rel-
ative to unconventional producers (Figure 14.10). In our classification, we
inferred technology application from annual reports and operating basins/
12
10
y =0.016x + 0.200
8
R2 = 0.7512
y = 0.011x + 0.467
6
R2 = 0.746
4
0
0 100 200 300 400 500 600 700 800
Conventional Unconventional Linear (Conventional) Linear (Unconventional)
Figure 14.10 Conventional vs. unconventional independents market capitalization (2010).

regions in a non-quantitative manner. We observe that the correlative rela-

tions are similar and the slope of the relation for companies with primarily
conventional production is only slightly greater than the unconventional
group. Unconventional production is subject to challenging econom-
ics and significant risk, and a high proportion of PUD reserves reflects
delayed and uncertain future cash flows, but the model did not distinguish
any significant differences between the resource types.
14.8.5 Production and Reserves

Production and reserves are strongly correlated for independents and pro-
duction is a better indicator of market cap than reserves (Figure 14.11;
Table 14.5). IOCs and independents are often considered to value reserves
over production, but the empirical evidence suggests that production is
more closely correlated with the market value of independents. Production
and reserves are primary factors in company valuation.
14.8.6 Regression Models

In Table 14.6 and Table 14.7, regression results for independents are depicted
using three and four variable model specifications. Oil and gas volume
models are superior to heat-equivalent reserves, and as expected, large-cap
producers yield stronger correlations than small-cap producers because
of the greater homogeneity of the companies. Reserves are the most
300
250
y = 0.0826x + 0.1143
Production (MMboe)
R² = 0.9614
200
150
y = 0.0692x + 1.2049
R² = 0.9206
100
50
0
0 500 1,000 1,500 2,000 2,500 3,000 3,500
Figure 14.11 Proved reserves and annual production - independents (2010).

Table 14.5 Regression results of independent oil companies (2010). CAP.

CAP ($million) = a + b · Pboe (MMboe)
Class Product a b R2
All Gas 1,107(2.18) 136(15.00) 0.89
Oil 313(1.04) 171(40.00) 0.98
All 501(1.70) 161(35.27) 0.95

Large-Cap Gas 3,217(2.99) 116(8.50) 0.84
Oil 928(0.92) 168(19.51) 0.97
All 1,476(1.77) 154(18.42) 0.92

Small-Cap Gas 37(0.14) 168(4.33) 0.59
Oil 485(2.77) 86(2.61) 0.22
All 302(2.06) 130(4.92) 0.37
Table 14.6 Regression results of independent oil companies – three independent

variables (2010).
CAP ($million) = a + b · Rboe (MMboe) + c · R/P (yr) + d · D/E
Class Product a b c d R2
All Gas 2,614(1.69)a 8.5(10.21) -29 -526(-0.55) 0.77
(-0.72)
Oil 659(1.01) 14(26.40) -20(-0.86) -57(-0.54) 0.95
All 1,276(1.72) 12(19.77) -31(-1.18) -85(-0.55) 0.85
Large- Gas 11,184(2.84) 6.2(5.01) -70(-1.41) -3,785 0.69
Cap (-1.60)
Oil 8,687(1.93) 13(10.53) -343 -1,604 0.93
(-1.67) (-0.56)
All 10,087(2.56) 10(8.45) -84(-1.32) -4,476 0.78
(-1.92)
Small- Gas 507(0.74) 13(2.98) -22(-0.42) -183(-0.61) 0.33
Cap
Oil 509(2.38) 12(3.43) -4.7 -32(-1.16) 0.30
(-0.80)
All 405(2.09) 11(4.73) -4.2 -29(-0.97) 0.34
(-0.64)
t-statistics are denoted in parenthesis.
a
Table 14.7 Regression results of independent oil companies - four independent

variables (2010).
CAP ($million) = a + b · Ro (MMbbl) + c · Rg (Bcf) + d · R/P (yr) + e · D/E
Class Product a b c d e R2
All Gas 1,872 27(6.85) 1.0(6.83) -7.3 -502 0.87
(1.61)a (-0.24) (-0.70)
Oil 555(0.84) 16(7.70) 2.1(6.84) -19(-0.81) -56(-0.53) 0.95
All 851 21(16.06) 1.2(10.86) -18(-0.92) -80(-0.70) 0.92
(1.55)
Large- Gas 8,958 22(4.61) 0.75(4.29) -42(-1.13) -3,278 0.83
Cap (3.01) (-1.88)
Oil 8,540 14(2.95) 2.2(4.00) -336 -1,658 0.92
(1.68) (-1.41) (-0.53)
All 7,702 19(9.1) 1.1(6.02) -46(-0.97) -3,849 0.88
(2.61) (-2.24)
Small- Gas 405(0.75) 50(3.62) 0.57(0.70) -18(-0.43) -53(-0.22) 0.58
Cap Oil 490(2.31) 23(2.31) 0.028 -6.1 -36(-1.32) 0.31
(0.02) (-1.02)
All 343(1.91) 26(4.53) 0.89(1.86) -5.0 -44(-1.55) 0.44
(-0.83)
t-statistics are denoted in parenthesis.
a
significant explanatory variable, and while the R/P and D/E coefficients
are both negative, they are not statistically significant.
14.9 Private Companies

There are a large number of private companies in the oil and gas indus-
try. If a private company “went public,” what market capitalization would
we expect? Using the oil and gas production levels for private companies
reported in the OGJP100, we estimate reserves and assets using the regres-
sion models for independents as shown in Table 14.8. Aera Energy, for
example, produced 43 MMboe in 2010, and using the regression models
derived for independents as a proxy, we would expect that Aera’s reserves
position to be approximately 525 MMboe based on $7.9 billion in assets.
Table 14.8 Estimated reserves and assets of private companies (2010).
Productiona
BOE Gas Oil Primary Estimated reserves Estimated assets
Rank Company (MMboe) (Bcf) (MMbbl) product (MMboe) ($billion)
1 Aera Energy 43 15 41 Oil 525 7.91
2 Merit Energy 24 98 8 Gas 331 3.35
3 Samson Lone Star 23 117 3 Gas 311 3.15
4 Hilcorp Energy 20 68 9 Oil 241 3.82
5 Yates Petroleum 17 92 1 Gas 227 2.28
6 Samson Resources 13 71 1 Gas 171 1.71
7 Walter Oil & Gas 12 44 4 Gas 151 1.51
8 EnerVest Operating 11 57 2 Gas 146 1.46
9 Endeavor Energy 10 24 6 Oil 122 2.10
Resources
10 Citation Oil & Gas 10 11 8 Oil 115 2.01
a
Source: Oil&Gas Financial Journal.
14.10 National Oil Companies of OPEC

The NOCs of OPEC are the largest oil and gas companies in the world.
Saudi Aramco is the world’s largest oil company with the largest reserves,
and along with the National Iranian Oil Company (NIOC), Iraq National
Oil Company (INOC), and Kuwait Oil Company (KOC), are descended
from a consortium set up by western IOCs before World War II. In 2010,
Aramco had daily oil and gas production capacity of about 12 MMboe/d
and combined reserves of 313 Bboe; NIOC had production capacity of
6 MMboe/d and combined reserves of 312 Bboe; INOC had production
capacity of 2 MMboe/d and combined reserves of 134 Bboe; and KOI had
production capacity of 2.6 MMboe/d and combined reserves of 112 Bboe.
The NOCs of OPEC are owned and operated entirely by state govern-
ment. If OPEC NOCs were privatized and listed on a stock exchange, what
market capitalizations would we expect? If oil and gas reserves volumes for
OPEC NOCs are accurate as shown in Table 14.9, NOC valuations would be
roughly similar to IOC valuations relative to their reserves position, and the
effective capitalizations can be estimated using the derived IOC relations.
Saudi Aramco is estimated to have the largest market capitalization
at $6.7 trillion, about four times the combined market cap of the seven
majors. National Iranian Oil Company would be worth $4.9 trillion. NOCs
in UAE, Venezuela, Iraq, and Kuwait would each have market caps greater
than $2 trillion. In total, the 12 NOCs of OPEC would have a combined
market cap of $28 trillion in 2010.
14.11 Government Sponsored Enterprises and Other

International Companies
In Table 14.10, the market cap for a sample of companies headquartered in
Australia (Woodside), Brazil (Petrobras), China (PetroChina, CNOOC),
and the UK (BG Group) are depicted. All of these companies are listed
on the U.S. stock exchange and are subject to SEC reserves disclosures.
PetroChina and Petrobras are partially state-owned and each has more
than 10 Bboe reserves and would be ranked among the top four IOCs if
they were privatized. Petrobras has a reputation for being a professional
deepwater field developer and operator but is subject to government influ-
ence. PetroChina is one of the largest companies in the world in terms
of the number of employees, and in 1999, its major domestic assets were
listed. CNOOC, BG Group, and Woodside hold reserves volumes similar
to large-cap U.S. independents.
Table 14.9 Effective market capitalization of National Oil Companies in OPEC (2010).
Liquids Reserves Production Effective market
National oil company (Bbbl) gas Total (MMboe) cap ($billion)
Country Sonatrach 12 (Tcf) (Bboe) 1,890 R/P(yr) 20 514
Algeria 39
Angola Sonangol 10 10 11 786 14 215
Ecuador Petroleos del Ecuador 7 0.3 7 185 35 128
Iran National Iranian Oil 138 1,046 312 2,847 110 4,929
Iraq Iraq National Oil 115 112 134 978 137 2,917
Kuwait Kuwait Petroleum 104 64 115 965 119 2,576
Libya National Oil 44 54 53 831 64 1,123
Nigeria Nigerian National Pet. 37 185 68 1,320 52 1,158
Qatar Qatar Petroleum 25 899 175 1,168 150 1,996
Saudi Arabia Saudi Aramco 262 264 306 4,411 69 6,706
U.A.E. Abu Dhabi National Oil 98 214 134 1,496 89 2,664
Venezuela Petroleos de Venezuela 99 176 129 1,285 100 2,641
OPEC 951 3,183 1,482 18,164 82 27,891
Source: EIA, OPEC.
Table 14.10 Actual and computed market capitalization of companies outside North America (2010).
Reserves Estimated market cap
Major Independent
Liquid Gas Total Production R/P Market cap model model
Company Country (Bbbl) (Tcf) (Bboe) (MMboe) (yr) D/E ($billion) ($billion) ($billion)
PetroChina China 11.28 65.50 22.19 1,228 18.1 0.8 303 506 290
Petrobras Brazil 10.76 11.95 12.76 878 14.5 0.7 229 124 222
CNOOC China 1.92 6.46 2.99 263 11.4 0.5 106 86 34
BG Group U.K. 0.95 11.69 2.89 236 12.3 0.9 69 53 49
Woodside Australia 0.18 6.45 1.25 69 18.0 0.8 34 46 14
400
Exxon
PetroChina
300
Petrobras y = 0.0132x–0.3031
R2 = 0.7524
200 Chevron
Shell
BP
CNOOC Total
100
Eni ConocoPhillips
BG Group
Woodside
0
0 5,000 10,000 15,000 20,000 25,000 30,000
IOC State-Owned Major International Independent
Figure 14.12 Market capitalization and proved reserves – state-owned majors and other
independents (2010).
Using only reserves as an indicator of market capitalization, PetroChina,

BG Group, and Woodside all fall above the IOCs market cap relation,
whereas CNOOC and Petrobras fall above the group average, indicating a
premium relative to reserves volumes (Figure 14.12). Using information on
company R/P and D/E data, the model-estimated capitalizations diverge
from the market data. PetroChina is estimated to have a market cap of $506
billion relative to its actual cap of $303 billion, while Petrobras is valued at
$124 billion relative to its cap of $229 billion. Interestingly, it appears that
the market is valuing PetroChina and Petrobras as if they were large-cap
independents, discounting their value because of state ownership, while
for CNOOC, BG Group and Woodside, the opposite is occurring and the
market is valuing these companies as if they were majors.
14.12 Conclusions
Constructing robust market capitalization models of oil and gas compa-
nies is subject to a number of estimation issues. In this analysis, we fixed
the time of assessment to coincide with the release of reserves and produc-
tion data, and it is clear that the model results will change over time and
are only valid on a relative basis. However, the relative positions of com-
panies and model results for most companies will not change significantly
on a year-to-year basis, unless exceptional events occur (e.g., BP and the
Macondo oil spill, merger, and acquisition activity) because finding and
developing reserves is a relatively slow and sequential process.
A large number of factors may impact capitalization of a company,
including the cost of production, earnings, management quality, and explo-
ration potential, etc. Some of these factors can be quantified but many can-
not. Factors that are not reported or disclosed limit the use of regression
models, but this limitation is relevant only if the excluded factors are sig-
nificant. In most cases, these factors are believed to be secondary, and thus
their exclusion is not expected to significantly impact the model results.
All of the major oil and gas companies in North America were enu-
merated and a large portion of independents was sampled. Large samples
reduce selection bias but do not eliminate its impact, and because all of the
measured factors are reported according to U.S. GAPP, the data sources are
believed to be reasonably consistent and accurate.
There are only a few majors and the use of multi-factor models are
over-specified. For independents, the model fits were generally robust, but
several coefficients of interest were not individually statistically significant.
As long as the coefficients are jointly significant, however, the estimating
equation can be utilized with benefit. The purpose of the regression model
is to estimate market value and is not necessarily concerned about the
direct significant of each control variable.
Because only a small number of transactions occur in the market place,
comparisons for private companies are difficult to perform, while for
the NOCs of OPEC, the values are speculative and based entirely on the
reserves volumes of the companies. We demonstrated comparisons for
GSEs and international companies listed on the U.S. stock exchange with
the model results.
References
1. International Valuation Standard, 7th Edn., International Valuation Standard
Committee, Washington, DC (2005).
2. J.B. Abrams: Quantitative Business Valuation. Wiley, New York (2010).
3. A. Damodaran: Investment Valuation. Wiley, New York (2002).
4. V.A. Van Vactor: Introduction to the Global Oil and Gas Business. PennWell,
Tulsa, OK (2010).
5. D. Chapman: Energy Resources and Energy Corporations. Cornell University
Press, Ithaca, NY (1983).
6. N. Antill and R. Arnott, Valuing Oil and Gas Companies. Woodlands Publishing
Ltd, UK (2000).
7. A.W. Howard and A.B. Harp, Oil and gas company valuations. Bus. Val. Rev.
28, 30–35 (2009).
8. SPE/WPC/AAPG/SPEE, Petroleum Resources Management System,
http://www.spe.org/industry/docs/Petroleum_Resources_Management_
System_2007. pdf#redirected_from=/industry/reserves/prms.php (2007).
9. F. Demirmen, Reserves estimation: the challenge to the industry. Journal of
Petroleum Technology, 5: 80–88 (2007).
10. US Securities and Exchange Commission (US SEC). Modernization of the
Oil and Gas Reporting Requirements. Confirming version (proposed rule),
17 CFR Parts 210, 211, 229, and 249, Release Nos. 33-8995; 34-59192; FR-78;
File No. S7-15-08, RIN 3235-AK00, http://www.sec.gov/rules/final/2008/33-
8995.pdf (2008).
11. R. Harrel and C. Cronquist, Estimation of Primary Reserves of Crude Oil,
Natural Gas, and Condensate, Chapter 16, in Reservoir Engineering and
Petrophysics, Vol V(B), E. Holstein, (Ed.), pp. 1480–1570, SPE, Richardson,
TX (2007).
12. B.G. Dharan, Improving the relevance and reliability of oil and gas reserves
disclosures. Prepared/Written testimony for the US House Committee on
Financial Services. 108th Congress, Second Session, 21 July 2004, http://
archives.financialservices.house.gov/media/pdf/072104bd.pdf (2004).
13. J.R. Etherington, Managing your business using integrated PRMS and SEC
standards, SPE 124938. SPE Annual Technical Conference and Exhibition, 4–7
October 2009, New Orleans, Louisiana, (2009).
14. W.J. Lee, Modernization of the SEC oil and gas reserves reporting require-
ments, SPE 123793. SPE Annual Technical Conference and Exhibition, New
Orleans, Louisiana, 4–7 October 2009, (2009).
15. M.A. McLane, Reserve overlooking – the problem no one wants to talk about,
Paper SPE 68580. SPE Hydrocarbon Economics and Evaluation Symposium,
Dallas, Texas, 2–3 April 2001, (2001).
16. G.T. Olsen, W.J. Lee, and T.A. Blasingame, Reserves overbooking: the prob-
lem we’re finally going to talk about, SPE 134014. SPE Annual Technical
Conference and Exhibition, Florence, Italy, 20–22 September 2010, (2010).
17. E.D. Young and C.L. McMichael, Evaluation of fiscal terms, reporting
reserves, and financials. SPE 77510. SPE Annual Technical Conference and
Exhibition, San Antonio, TX, September 29–October 2 2001, (2002).
18. L. Denning, Nailing big oil’s moving production target. Wall Street Journal,
July 3 (2012).
19. L. Denning, Chevron battles Exxon for big oil’s crown. Wall Street Journal,
Aug 15 (2012).
20. Performance Profiles of Major Energy Producers. U.S. Energy Information
Administration, Office of Energy Statistics, Washington, DC (2009).
21. T. Turner and M. Moffett, Behind the battle for Argentina’s oil. Wall Street
Journal, May 3 (2012).
22. Financial Times, FT Global 500, http://media.ft.com/cms/253867ca-1a60-

11e0-b003-00144feab49a.pdf (2010).
23. T. Fowler, BP spill saga far from over. Wall Street Journal, Oct 13 (2012).
24. P. Barrett, How BP got screwed on Gulf oil spill claims. Bloomberg
Businessweek, June 27 (2013).
25. P. Barrett, The probe into BP oil-spill payouts could be bigger than you think.
Boomberg Businessweek, July 3 (2013).
Table A.l Sample of independent oil producers3 (2010).
Reservesb
PUD/ GOR
Liquid Gas Total Production P1 R/P Market Capc (Mcf/
Company (MMbbl) (Bcf) (MMboe) (MMboe) (%) (yr) D/Ec ($million) bbl)
Apache 1,309 9,867 2,954 240 33 12.3 0.8 45,546 7.5
Devon 1,160 10,283 2,784 228 29 12.6 0.7 33,916 8.9
Anadarko 1,010 7,764 2,304 211 33 10.9 1.5 37,775 7.7
Marathon Oil 1,202 2,617 1,638 150 25 10.9 1.1 26,291 2.2
Pioneer Nat. 565 2,675 1,011 45 43 22.6 1.3 9,984 4.7
Res.
Nexen 905 494 987 90 52 11.0 1.6 12,045 0.5
SandRidge 252 1,763 546 20 59 27.2 2.4 2,972 7.0
Murphy 308 883 455 68 24 6.7 0.7 14,388 2.9
LINN Energy 227 1,233 433 16 36 26.8 1.1 5,961 5.4
Plains E&P 223 1,157 416 33 43 12.8 1.6 4,500 5.2
Denbury 338 358 398 27 40 14.9 1.1 7,636 1.1
Resources
(Continued)
Table A.l Sample of independent oil producers3 (2010). (Continued)
Reservesb
PUD/ GOR
Berry 166 630 271 12 51 22.7 1.8 2,316 3.8
Petroleum
Energen 103 954 262 19 23 13.9 1.0 3,475 9.2
Continental 225 840 365 68 23 5.4 0.0 10,005 3.7
Res.
Swift Energy 62 423 133 8 55 15.9 1.0 1,644 6.8

ATP Oil & 75 308 126 8 81 16.5 12.1 854 4.1
Gas
W&T 38 256 81 15 19 5.6 2.4 1,322 6.7
Offshore
Rosetta 32 289 80 8 49 9.5 0.9 1,957 9.1
Resources
Stone Energy 33 275 79 13 31 6.2 2.9 1,070 8.3
Brigham Exp. 52 88 67 3 65 22.5 0.8 3,160 1.7
(Continued)
Reservesb
PUD/ GOR
Helix Energy 25 227 63 8 61 8.4 1.8 1,287 9.1
Harvest Nat. 45 57 55 2 69 25.8 0.6 414 1.3
Res.
Clayton 38 79 51 5 32 9.4 2.6 1,008 2.1
Williams
QR Energy 21 58 30 5 32 6.1 1.4 664 2.8
Energy 17 61 27 5 7 5.6 0.3 594 3.6
Partners
Abraxas 10 85 24 2 76 13.1 -13.2d 347 8.7
Petroleum
Gulfport 20 16 22 2 63 11.3 1.9 976 0.8
Energy
GeoResources 11 39 18 2 35 9.6 0.8 444 3.5
PrimeEnergy 8 53 17 2 27 10.4 7.7 54 6.8
(Continued)
Reservesb
PUD/ GOR
Callon 8 33 14 1 52 15.8 12.6 172 4.1
Petroleum
Magnum 7 39 13 0.5 49 28.2 0.4 533 5.8
Hunter
Evolution 11 7 12 0.1 84 98.9 0.2 176 0.6

Kodiak Oil & 10 9 12 1 62 20.2 0.2 1,175 0.9
Gas
Isramco 3 24 7 1 0 8.6 8.1 229 7.1
Northern 6 4 6 1 59 7.2 0.2 1,687 0.6
O&G
(Continued)
Reservesb
PUD/ GOR
Lucas Energy 3 1 3 0.04 96 74.3 0.2 40 0.3
HKN 2 2 2 0.2 11 12.2 0.3 35 1.3
FieldPoint 1 3 2 0.1 22 15.6 1.2 33 2.3
Pyramid Oil 1 0.04 1 0.1 23 8.9 0.3 24 0.1
Houston 0.1 0.1 0.1 0.3 54 0.3 0.1 561 1.2
American
All 8533 43,952 15,858 1,317 35 12.0 1.1 236,709 5.2
a
Data collected from companies’ 2010 Annual Report, Financial Times, Yahoo-Finance, Scottrade.
b
Reported proved reserves on 2010 Annual Reports.
c
As of December 31, 2011.
d
Abraxas Petroleum’s book value of equity is negative because its total liability exceeded book value of its total assets.
Table A.2 Sample of independent gas producersa (2010).
Reservesb
Rg/Ro
Liquid Gas Total Production PUD/R Market Capc (Mcf/
Company (MMbbl) (Bcf) (Bcfe) (Bcfe) (%) R/P (yr) D/Ec ($million) bbl)
Chesapeake 285 15,390 17,100 1,035 47 16.5 1.4 16,945 54
EOG 538 8,472 11,700 864 48 13.5 1.1 23,218 16
Resources
Talisman 510 5,237 8,295 913 34 9.1 1.4 22,612 10
Energy
Noble Energy 365 4,361 6,551 473 54 13.8 0.9 15,236 12

Southwestern 1 4,930 4,936 405 45 12.2 1.0 13,026 4,930
Energy
Range 146 3,567 4,442 181 51 24.6 1.5 7,197 24
Resources
Ultra 32 4,200 4,390 214 60 20.6 2.2 7,309 133
Petroleum
Newfield 204 2,492 3,716 283 42 13.1 1.2 9,663 12
PetroHawk 47 3,110 3,392 246 65 13.8 1.2 11,261 66
(Continued)
Table A.2 Sample of independent gas producersa (2010). (Continued)
Reservesb
Rg/Ro
Contango Oil 11 2,644 2,712 28 80 96.9 1.1 3,936 233
& Gas
Cabot Oil & 9 2,644 2,701 131 36 20.7 1.1 3,936 279
Gas
Forest Oil 82 1,752 2,244 165 40 13.6 1.8 4,329 21
Quicksilver 68 1,556 1,962 130 48 15.1 2.3 2,506 23
Res.
Cimarex 105 1,254 1,884 218 23 8.7 0.7 7,525 12
Energy
EXCO 7 1,455 1,499 112 45 13.4 1.3 4,136 199
Resources
Comstock 4 1,026 1,051 73 50 14.3 0.8 1,179 243
Resources
SM Energy 57 640 984 110 30 8.9 1.3 3,713 11
(Continued)
Reservesb
Rg/Ro
Petroleum 34 657 861 38 65 22.9 1.2 972 19
Develop.
Carrizo 26 683 841 37 52 22.8 1.5 1,345 26
Goodrich 2 454 464 34 55 13.8 2.6 670 281
Petroleum
GMX 1 312 319 17 49 18.3 3.4 171 260
Resources
McMoRan 15 192 280 15 20 18.6 0.7 2,691 13
Exp.
Petroquest 2 175 184 28 31 6.5 1.1 467 108
Energy
Crimson 5 135 166 13 52 12.9 1.2 192 26
Exploration
Delta 2 123 134 15 8 9.1 1.0 22 64
Petroleum
(Continued)
Reservesb
Rg/Ro
Double Eagle 0.4 113 115 9 35 12.6 0.7 54 296
Pet.
Gasco Energy 0.5 40 43 4 0 9.8 0.9 42 85
CREDO 0.5 9 12 2 4 7.4 0.1 81 18
Petroleum
Mexco Energy 0.2 8 10 1 40 15.1 0.2 15 35
Royale Energy 0.02 6 6 1 12 9.1 1.2 22 337
All 2,560 67,637 82,995 5,793 45 14.3 1.3 164,471 26
a
Data collected from companies’ 2010 Annual Report, Financial Times, Yahoo-Finance, Scottrade.
b
Reported proved reserves on 2010 Annual Reports.
c
As of December 31, 2010.
Part 5
UNCONVENTIONAL RESERVOIRS
15
An Analytical Thermal-Model for
Optimization of Gas-Drilling in
Unconventional Tight-Sand Reservoirs
Boyun Guo1*, Gao Li2 and Jinze Song1
Petroleum Engineering Department, University of Louisiana at Lafayette

1
2
Petroleum Engineering Department, Southwest Petroleum University
Abstract
Drilling wells with gas, or gas-drilling, has recently been adopted to drill uncon-
ventional oil and gas reservoirs including tight sands, gas shale and oil shale.
However, the performance of gas-drilling is very unpredictable in many areas due
to lack of optimization of drilling parameters. Because gas properties are greatly
affected by temperature, a reliable thermal model is required for gas-drilling opti-
mization. Such a model is not available and this paper fills the gap. An analytical
thermal-model was derived in this study for predicting bottom hole gas tempera-
ture under various flow conditions. The result given by the analytical model was
compared with that by an existing approximation, showing an accuracy improve-
ment of 14%. Sensitivity analysis with the new model indicates that formation
fluid influx dominates the temperature profile inside and outside the drill string,
while Joule-Thomson cooling and drill cutting’s intrusion affect temperature
only at bottom hole. Applications of the model are demonstrated in an example.
This paper provides an analytical tool to drilling engineers for optimizing their
gas-drilling operations.
Keywords: Gas-drilling, temperature, prediction, unconventional, reservoirs
339
15.1 Introduction
Drilling optimization is a process of using optimal drilling parameters
to increase rate of penetration and enhance drilling performance. Such
parameters include weight on bit, rotary speed and fluid flow rate. Drilling
optimization is particularly important in drilling unconventional oil and
gas reservoirs including tight sands, gas shale and oil shale where the rate
of penetration is usually very low. The process of drilling wells with gas,
or gas-drilling, is difficult to optimize because gas properties are greatly
affected by temperature. A reliable thermal model is required for predict-
ing bottom hole temperature in optimizing weight on bit and rotary speed.
Such a model is not available from literature.
Gas-drilling (drilling with air, nitrogen, etc.) has been traditionally used
for increasing rate of penetration (ROP) in hard rock formations [1]. This
technology has recently gained a strong momentum in unconventional
oil and gas field development in North America and other regions. Holt
et al. [2] reported that this technology reduced drilling time by 50% in
hard rock drilling in China. George et al. [3] illustrated significant cost sav-
ings achieved through the use of PDC bits with air/foam drilling. Lays and
Grayson [4] documented the performance gains realized with PDC bits in
air/foam drilling in the Appalachian Basin. Wilhide et al. [5] reported the
first rotary steerable system drilling with dry air used to improve low cost
development of an Unconventional gas reservoir. Zreik et al. [6] showed
improved surface hole air-foam drilling performance in the Karstifed
Limestone (Papua New Guinea). Pletcher et al. [7] demonstrated that the
application of air drilling improved drilling efficiency in horizontal sand-
stone wells. The first application of gas-drilling in shale gas field develop-
ment was reported by Ford et al. [8]. It reduced drilling costs and improved
efficiency of developing the Fayetteville shale gas reserves on the northern
Arkansas side of the Arkoma Basin. The advantages of drilling with air as
the circulating medium rather than mud were found significant. Air drill-
ing delivered faster ROP compared and reduced mud costs and incidence
of lost circulation. Utilizing this technology enabled the operator to drill
approximately 58% more footage at high ROP. The increased durability has
reduced the number of bits/trips required to complete the interval. Maranuk
et al. [9] reported a unique system for Underbalanced drilling using air in
the Marcellus shale. They described a drilling system and drilling param-
eters highlighting the differences between mud and air drilling. They dis-
cussed the modifications to the bottom hole assembly (BHA) to increase
reliability and drilling fluids and their effects on various pressure regimes.
Analytical Thermal-Model for Optimization of Gas-Drilling 341
They identified the major disadvantages of using air for drilling as its lim-
itation to handle fluid influx, the reduction of carrying capacity compared
to foam and other normal mud regimes, and the increased flow velocities
required to ensure adequate cuttings removal.
A few researchers have investigated the methods for predicting fluid
temperature profiles in drilling circulation systems. Among them are
Zhang et al. [10], Hasan and Kabir [11], Eppelbaum et al. [12], and Kutasov
and Eppelbaum [13]. Unfortunately, all these methods were developed
for liquid-drilling, not for gas-drilling. The only method for gas-drilling
is the numerical simulator developed by Wang et al. [14]. The paper was
published in a Chinese journal and the simulator is not accessible to the
authors. Li et al. [15] presented an approximate mathematical model for
predicting bottom hole gas temperature in gas-drilling unconventional
tight reservoirs where the drilling annulus is treated as an insulator.
A new analytical solution for predicting gas temperature profiles inside
drill string and in the annulus was derived in this study for gas-drilling,
considering the effects of formation fluid influx, Joule-Thomson cooling,
and entrained drill cuttings. The bottom hole temperature given by the
new analytical solution are found significantly higher than that given by
Li et al.’s [15] model. Results of sensitivity analyses show that formation
fluid influx can significantly increase the temperature profiles in both the
drill string and the annulus. The Joule-Thomson cooling effect lowers the
temperature in the annulus only in a short interval near the bottom hole.
The drill cuttings entrained at the bottom hole can slightly increase the
temperature profile in the annulus.
15.2 Mathematical Model

Accurate prediction of gas temperature at bottom hole depends on the ability
of calculating the heat transfer during the gas counter-current flow inside
and outside the drill string. An analytical model for steady heat transfer is
derived in Appendix A. The model is briefly summarized in this section.
Major assumptions made in the model derivation include:
1. The thermal conductivity of casing is assumed to be infinitive.

2. Heat capacity of fluid is constant.
3. Friction-induced heat is negligible.
The thermal conductivities of steel casing, cement concrete, sandstone

rock, and air at 50 °C are 43 W/m- °C, 1.7 W/m- °C, 3 W/m- °C, and 0.03
W/m- °C, respectively. The high contrast (>50) in the thermal conductivity
values makes the casing a super conductor for the heat conduction in the
radial direction. Therefore, the first assumption is valid.
Heat capacity of gas is a function of temperature and pressure [16]. In
the temperature range between 0 °C and 100 °C at atmospheric pressure,
the heat capacity of air varies between 1,005 J/kg-C and 1,009 J/kg-C, or
within 0.40%. In gas-drilling operations the gas pressure in the drill string
is in a narrow range between 7 MPa and 10 MPa. The heat capacity of
air varies between 1,016.2 J/kg-C and 1,021.6 J/kg-C, or within 0.53%, in
this pressure range [17]. Considering the extreme condition of 0 °C and
10MPa, the heat capacity of air varies between 1,005 J/kg-C and 1,021.6 J/
kg-C, or within 1.65%, which justifies the second assumption.
All gases used in gas-drilling are dilute gases in the above-mentioned
ranges of pressure and temperature. Gas density varies from 1 to 100 kg/
m3 and gas viscosity changes from 13.3 × 10–6 m2/s to 22.1 × 10–6 m2/s
[17]. The friction pressure drop in the whole circulation system is 15 MPa
at most, with the friction pressure drop inside the drill string being less
than 5 MPa over a few thousand meters of length. This low pressure drop
is not expected to generate significant amount of heat, and thus the third
assumption is valid.
When gas is injected into a drill string, the heat brought to the inside of
string is proportional to the product of fluid heat capacity C and mass flow
rate ṁp where
ṁp = ρpQp (15.1)
where ρp and Qp are gas density and volumetric flow rate inside drill string.
As the gas fows down the drill string, the rate of heat transfer through
drill string is proportional to the thermal conductivity of string Kp. When
the drilling fluid expands below the bit orifices, its temperature drops due
to Joule-Thomson cooling effect. The downstream temperature can be
expressed as [18]:
k −1
P  k
Tdn = Tup  dn  (15.2)
P 
up

where Tdn and Tup are the absolute temperatures in the downstream and
upstream of bit orifices, respectively, Pdn and Pup are the absolute pressures
in the downstream and upstream of bit orifices, respectively, and k is the
specific heat ratio of gas (k = 1.3 for gas). The temperature drop at the bit
is expressed as:
ΔTj = Tup − Tdn (15.3)
The gas receives heat from the entrained drill cuttings and formation oil
influx. Assuming all formation fluid influx occurs at bottom hole, the fluid
temperature should change at bottom hole in the annulus by
ΔTb = − ΔTj + ΔTc + ΔTf (15.4)
where ∆Tc and ∆Tf are temperature changes due to added drill cutting
and fluid influx, respectively. It can be shown that in the practical drilling
conditions where the rate of penetration is less than 50 m/hour, the term
∆Tc is negligible. The annular temperature at the bottom hole Tbh can be
expressed as:
C pm p (Tp − ∆Tj ) + C f m f Tmax

Tbh (15.5)
C pm p + C f m f

where Tp is the temperature of fluid inside the drill string, Cf is the heat
capacity of the fluid influx, mf is the mass flow rate of fluid influx, and Tmax
is the geo-temperature at the bottom hole depth. Therefore the tempera-
ture change at the bottom hole can be expressed as:
C pm p (Tp − ∆Tj ) + C f m f Tmax

∆Tb = Tbh − Tp − Tp (15.6)
C pm p + C f m f

The heat transfer in the annulus depends on the product of mixture heat
capacity Ca and mixture mass fow rate ṁc where
Caṁa = Cpṁp + Csṁs + Cfṁf (15.7)
where the product of heat capacity and mass fow rate of solid cuttings Csṁs
is further expressed in two terms:
Csṁs = Chṁh + Crṁr (15.8)

where Ch and Cr are the heat capacities of hydrocarbon and dry rock in the
cuttings, respectively. The mass flow rates of the hydrocarbon and rock in
the cuttings are respectively expressed as:
π 2
m h = Db R pϕρ h (15.9)
4
and
π 2
m r = Db R p (1− ϕ ) ρ r (15.10)
4
where Db, Rp, φ, ρh, and ρr are drill bit diameter, rate of penetration, rock
porosity, density of hydrocarbon, and density of rock, respectively. The Cf
in (15.4) is heat capacity of formation influx fluid (usually oil) and mass
flow rate of formation fluid influx is expressed as:
ṁf = ρfQf (15.11)
where ρf and Qf are density of fluid influx and flow rate of fluid influx,
respectively. As the fluid mixture flows up the annulus, the rates of heat
transfer through drill string and cement sheath are proportional to the
thermal conductivities of drill string Kp and cement sheath Kc, respectively
(the thermal conductivity of casing is assumed to be infinity compared to
that of cement sheath).
The gas temperatures inside the drill string Tp and in the annulus Ta
take the following forms respectively (derivation of solution is given in
Appendix):
AG + ABTg 0 − G( B + E )
Tp = C1 Ae r1L + C2 Ae r2 L + GL + (15.12)
AB
and
AG + ABTg 0 − EG
Ta = C1 ( A + r1 )e r1L + C2 ( A + r2 )e r2 L + GL +
AB
(15.13)
where
AB( A∆Tb − G ) −  ABTp 0 − ABTg 0 − AG +G( B + E ) r2e r2 Lmax

C1 =
A2 B(r1e r1Lmax − r2e r2 Lmax )
(15.14)
− AB( A∆Tb − G ) +  ABTp 0 − ABTg 0 − AG +G( B + E ) r1e r1 Lmax

C2 =
(15.15)
B + E − A + ( B + E − A)2 + 4 AB
r1 = (15.16)
2
B + E − A − ( B + E − A)2 + 4 AB (15.17)
r2 =
2
where
π dpK p
A= (15.18)
C pm pt p

π dc K c (15.19)
B=
Cam atc
π dpK p
E= (15.20)
Cam at p

where all symbols are defined in the Nomenclature section.

Because (15.12) involves the in-string temperature Tp at bottom depth,
it is necessary to solve (15.12) and (15.13) simultaneously with a numerical
method such as Newton-Raphson iteration. These equations were solved
in a spreadsheet program using the Goal Seek tool in MS Excel.
15.3 Model Comparison

The newly derived analytical solution was coded in an MS Excel spread-
sheet to compare with other models. A number of numerical models have
been presented for fluid temperature prediction, including Keller et al. [19],
Table 15.1 Data used in comparison of analytical solutions.
Factor Value Unit
Depth (L) 2000 M
Bit diameter (Db) 0.201 M
Inner diameter of cement (dc) 0.245 M
Outer diameter of cement (Dc) 0.311 M
Outer diameter of drill string (Dp) 0.114 M
Inner diameter of drill string (dp) 0.098 M
Geothermal temp at surface (Tg0) 20 °C
Geothermal gradient (G) 0.0245 C/m
Thermal conductivity of cement (Kc) 1.7 W/m-°C
Thermal conductivity of annulus (Kp) 0.0271 W/m-°C
Injection rate (Qp) 1.08 m3/s
Temperature of injected fuid (Tp0) 40 °C
Heat capacity of fuid inside pipe (Cp) 1005 J/kg-°C
Heat capacity of rock (Cr) 920 J/kg-°C
Heat capacity of formation fuid (Cf) 1880 J/kg-°C
Porosity (φ) 0.3
Rate of penetration (Rp) 0 m/h
Temperature drop at bit (∆Tb) 0 °C
Density of rock (ρr) 2650 kg/m3
Density of formation fuid (ρf) 1000 kg/m3
Density of injected fuid (ρp) 1.127 kg/m3
Formation fuid infux rate (Qf) 0 m3/s
Wooley [20], Marshall and Bentsen [21], Kabir et al. [22] and Hasan and
Kabir [11]. Unfortunately all these models were developed for liquid or
multi-phase flow. They are not applicable to gas flow in gas-drilling. The
analytical model for gas-drilling presented by Li et al. [15] was used for
comparison. The data used in the models are provided in Table 15.1.
Gas is the major component of air (>78%). Heat capacity of gas is a func-
tion of temperature and pressure [16]. In the temperature range between
0 °C and 100 °C at atmospheric pressure, the heat capacity of air varies
between 1,005 J/kg-C and 1,009 J/kg-C, or within 0.40%. In gas-drilling
operations the gas pressure in the drill string is in a narrow range between
7 MPa and 10 MPa. The heat capacity of gas varies between 1,016.2 J/
kg-C and 1,021.6 J/kg-C, or within 0.53%, in this pressure range [17].
Considering the extreme condition of 0 °C and 10MPa, the heat capacity
of gas varies between 1,005 J/kg-C and 1,021.6 J/kg-C, or within 1.65%.
Therefore, the heat capacity of gas was assumed to be constant in this study.
Figure 15.1 indicates that the injected gas is cooled down in the upper
section of drill string by the geothermal gradient. Gas is then heated up by
the geothermal gradient in the lower section of drill string. After arrival
in the annulus, the gas is quickly heated up by the geothermal gradient in
the lower section of the annulus. Eventually the gas is cooled down in the
upper section of the annulus by the geothermal gradient.
Also presented in Figure 15.1 is the gas temperature profile inside the
drill string given by Li et al.’s [15] analytical model. It is shown that the
bottom hole temperature given by the new analytical model is 7 °C higher
80
70
60
Temperature (C)
50
40
30 Temperature in the annulus
20 Geothermal temperature
Temperature inside drill string
10 Temperature inside drill string by Li et al. (2015)
0
0 500 1000 1500 2000 2500
Measured depth (m)
Figure 15.1 A comparison of temperature profiles given by different analytical models.

than that given by Li et al.’s [15] model. If the new model is considered to
be accurate, Li et al.’s [15] model is expected to underestimate bottom hole
temperature by 14%.
15.4 Sensitivity Analysis

Previous models do not consider the effects of formation fluid influx,
Joule-Thomson cooling, and entrained drill cuttings at bottom hole on the
temperature profiles inside drill string and in the annulus. These effects
were analyzed with the new model in this study. Figure 15.2 demonstrates
the effect of formation fluid influx on the temperature profiles. It shows
that the formation fluid influx can significantly increase the temperature
profiles in both the drill string and the annulus. Figure 15.3 illustrates the
effect of Joule-Thomson cooling on the temperature profiles. It shows that
the Joule-Thomson cooling effect lowers the temperature in the annulus
only at the bottom hole. It diminishes quickly in a very short interval when
the drilling fluid moves up the annulus. Figure 15.4 shows the effect of
entrained drill cuttings on the temperature profiles. It indicates that the
drill cuttings can slightly increase the temperature profile in the annulus
even at a very high rate of penetration of up to 60 m/hour.
80
70
60
Temperature (C)
50
40
30 Temperature inside drill string - no influx

Temperature in annulus - no influx
20 Temperature inside drill string - influx 0.005 m3/s
Temperature in annulus - influx 0.005 m3/s
Temperature inside drill string - influx 0.01 m3/s
10
Temperature in annulus - influx 0.01 m3/s
0
0 500 1000 1500 2000 2500
Measured depth (m)
Figure 15.2 Effect of formation fluid influx on the temperature profiles.

70
60
50
Temperature (C)
40
30
Temperature inside drill string - no Joule-Thomson cooling

20 Temperature in annulus - no Joule-Thomson cooling
Temperature inside drill string - Joule-Thomson cooling 5 C
10 Temperature in annulus - Joule-Thomson cooling 5 C
Temperature inside drill string - Joule-Thomson cooling 10 C
Temperature in annulus - Joule-Thomson cooling 10 C
0
0 500 1000 1500 2000 2500
Measured depth (m)
Figure 15.3 Effect of Joule-Thomson cooling on the temperature profiles.
70
60
50
Temperature (C)
40
30
Temperature inside drill string - no rate of penetration

20 Temperature in annulus - no rate of penetration
Temperature inside drill string - rate of penetration 30 m/h
10 Temperature in annulus - rate of penetration 30 m/h
Temperature inside drill string - rate of penetration 60 m/h
Temperature in annulus - rate of penetration 60 m/h
0
0 500 1000 1500 2000 2500
Measured depth (m)
Figure 15.4 Effect of entrained drill cuttings on the temperature profiles.
15.5 Model Applications

It was planned to increase rate of penetration (ROP) using the low-
temperature effect in the XYZ unconventional tight reservoir. Drilling con-
ditions are similar to that described in Table 15.1. Three bit orifices with
0.01 m diameter each were proposed at well depth 2,000 ft. The pressures
in the downstream and upstream of the bit were predicted by a hydraulics

model to be 1.28 MPa and 2.10 MPa, respectively. The potential of increas-
ing ROP was estimated using the presented analytical model.
Figure 15.1 shows that the gas temperature inside the drill string,
in the upstream of the bit orifices, is 49 °C. The gas temperature in the
downstream of the bit orifices (bottom hole) is predicted based on Joule-
Thomson cooling at bit [18]:
k −1
P  k
Tdn = Tup  dn  (15.21)
P  up

upstream of bit orifces, respectively, pdn and pup are the absolute pressures
in the downstream and upstream of bit orifces, respectively, and k is the
specific heat ratio of gas (k = 1.3 for air). Equation (15.21) gives in this case:
1.3−1
 1.28  1.3
Tdn = (49 + 273.15) = 287°K = 14.2°C
 2.10 
The fold of increase in weight on bit (WOB) of PDC bit can be estimated
based on a simple relation derived from Glowka and Stone’s [23] thermal
model for bit cutters:
tcr − Tdn , C (15.22)

FWOB =
tcr − Tdn , N
where
FWOB = fold of increase in WOB by Joule-Thomson cooling
tcr = critical temperature of PDC cutter, 350 °C
tdn,C = gas temperature in the downstream of bit after cooling, °C
tdn,N = gas temperature in the downstream of bit before cooling, °C
In this case Eq. (15.22) predicts:
350 − 14.2
FWOB = = 1.12 or 12% increase.
350 − 49
Since the rate of penetration (ROP) is approximately proportional to

the weight on bit [24], it is expected that the 12% increase in weight on bit
should increase ROP by about 12%.
15.6 Conclusions
A new closed-form analytical solution for predicting gas temperature pro-
files inside drill strings and in the annulus was derived in this study for
gas-drilling. The new solution has advantages over existing solutions in
that it can handle formation fluid influx, Joule-Thomson cooling effect,
and entrained drill cuttings. The following conclusions are drawn from this
study:
1. An example calculation shows that the injected hot gas is

cooled down in the upper section of drill string by the geo-
thermal gradient. Gas is then heated up by the geothermal
gradient in the lower section of drill string. After arrival in
the annulus, the gas is quickly heated up by the geo-thermal
gradient in the lower section of the annulus. Eventually, the
gas is cooled down in the upper section of the annulus by the
geothermal gradient.
2. The temperature profiles given by the new analytical solu-
tion are significantly different from that given by Li et al.’s
analytical model. If the new model is considered to be accu-
rate, Li et al.’s [15] model is expected to underestimate bot-
tom hole temperature by 14%.
3. Results of sensitivity analyses show that formation fluid
influx can significantly increase the temperature profiles in
both the drill string and the annulus. The Joule-Thomson
cooling effect lowers the temperature in the annulus only
at the bottom hole. The drill cuttings entrained at the bot-
tom hole can slightly increase the temperature profile in the
annulus.
Direct measurement of temperature profiles are required to further

validate the new analytical solution. Once validated the new solution can
replace numerical simulators that are not readily available to field engi-
neers in general.
Nomenclature
Aa = cross-sectional area of annulus (m2)
Ap = cross-sectional area of drill pipe (m2)
Ca = heat capacity of fluid in the annulus (J/kg-°C)
Cf = heat capacity of formation fluid influx (J/kg-°C)
Ch = heat capacity of hydrocarbons in cuttings (J/kg-°C)
Cp = heat capacity of fluid inside drill pipe (J/kg-°C)
Cr = heat capacity of rock (J/kg-°C)
Cs = heat capacity of solid in the annulus (J/kg-°C)
Db = bit-diameter (m)
dc = inner-diameter of cement sheath (m)
Dc = outer-diameter of cement sheath (m)
dp = inner-diameter of drill pipe (m)
Dp = outer-diameter of drill pipe (m)
G = geothermal gradient (°C/m)
k = specific heat ratio of gas (dimensionless)
Kc = thermal conductivity of cement (W/m- °C)
Kp = thermal conductivity of drill pipe (W/m-°C)
L = wellbore depth along the drill string (m)
Lmax = the maximum hole depth (m)
ṁa = mass flow rate in the annulus (kg/s)
ṁf = mass flow rate of formation fluid influx (kg/s)
ṁh = mass flow rate of hydrocarbons in cuttings (kg/s)
ṁp = mass flow rate inside the drill pipe (kg/s)
ṁs = mass flow rate of solid cuttings in the annulus (kg/s)
ṁr = mass flow rate of rock (kg/s)
Pdown = absolute pressure in the downstream (psi)
Pup = absolute pressure in the upstream (psi)
Qf = formation fluid influx rate (m3/s)
Qp = fluid injection flow rate (m3/s)
RP = rate of penetration (m/s)
tc = thickness of cement sheath (m)
tp = wall thickness of drill pipe (m)
Ta = temperature of annular fluid (°C)
∆Tb = temperature change at drill bit (°C)
Tdn = absolute temperature in the downstream (°C)
Tg = geothermal temperature at depth (°C)
Tg0 = geothermal temperature at surface (°C)
Tp = temperature of fluid inside drill pipe at depth (°C)
Tp0 = temperature of fluid inside drill pipe at surface (°C)
Tup = absolute temperature in the upstream (°C)
Greeks
φ = porosity of rock (dimensionless)
ρa = fluid mixture density in the annulus (kg/m3)
ρf = density of formation fluid (kg/m3)
ρh = density of hydrocarbons in cuttings (kg/m3)
ρp = fluid density inside drill pipe (kg/m3)
ρr = density of dry rock (kg/m3)
ρs = solid density in the annulus (kg/m3)
Acknowledgements
This research was supported by the China National Natural Science
Foundation Founding No. 51274220, No. 51134004, No. 51221003,
51274045, 51274221, and No. 51334003.
References
1. W.C. Lyons, B. Guo, and F.A. Seidel, Air and Gas Drilling Manual, 2nd edition,
McGraw-Hill, New York (2001).
2. C. Holt, S.W. Nas, C. Shen, and X. Niu, Managed pressure drilling reduces
China Hard Rock Drilling by Half, Paper SPE 105490 presented at the SPE/
IADC Drilling Conference held in Amsterdam, 20–22 February, 2007.
3. B. George, E.S. Grayson, R. Lays, F.C. Felderhoff, M.L. Doster, and M.A.
Holmes, Significant cost savings achieved through the use of PDC bits in
compressed air/foam applications, Paper SPE 116118 presented at the SPE
Annual Technical Conference and Exhibition held in Colorado, Denver,
21–24 September, 2008.
4. R. Lays and E.S. Grayson, Performance gains realized with PDC bits in Air/
Foam applications in the Appalachian Basin, Paper SPE 117695 presented
at the SPE Eastern Regional/AAPG Eastern Section Joint Meeting held in
Pittsburgh, 11–15 October, 2008.
5. S. Wilhide, J. Smith, D. Doebereiner, B. Raymond, D. Weisbeck, and B.
Ziemke, First rotary steerable system drilling with dry air is used to further
improve low cost development of an unconventional gas reservoir, Paper SPE
135471 presented at the SPE Annual Technical Conference and Exhibition
held in Florence, Italy, 19–22 September, 2010.
6. A. Zreik, R.H.J. Pinkstone, and A.I. Ross, Improving surface hole air-foam
drilling performance in Karstifed Limestone (Papua New Guinea), Paper
SPE 135859 presented at the IADC/SPE Asia Pacific Drilling Technology
Conference and Exhibition held in Ho Chi Minh City, Vietnam, 1–3

November, 2010.
7. J.P. Pletcher II, A. Scarr, J. Smith, S. Swadi, and C. Rogers, Application of air
percussion drilling improves drilling effciency in horizontal sandstone wells,
Paper SPE 135308 presented at the SPE Annual Technical Conference and
Exhibition held in Florence, Italy, 19–22 September, 2010.
8. R. Ford, A. Stone, A. Spedale, R. Slaughter, S. Swadi, and C. Dewey, Efficiently
developing Fayetteville Shale Gas Reserves: Percussion drilling solves appli-
cation challenges/reduces drilling costs, Paper SPE 148828 presented at the
SPE Eastern Regional Meeting held in Columbus 17–19 August, 2011.
9. C. Maranuk, A. Rodriguez, J. Trapasso, and J. Watson, Unique system for
underbalanced drilling using air in the Marcellus Shale, Paper SPE 171024
presented at the SPE Eastern Regional Meeting in Charleston, 21–23 October
2014. JPT 67(2), 87–89 (February 2015).
10. Y. Zhang, L. Pan, K. Pruess, and S. Finsterle, A time-convolution approach
for modeling heat exchange between a Wellbore and surrounding formation.
Geothermics 40, 261–266 (2011).
11. A.R. Hasan and C.S. Kabir, Wellbore heat-transfer modeling and applica-
tions. J. Petrol. Sci. Eng. 86–87(2012), 127–136 (2012).
12. L.V. Eppelbaum, I.M. Kutasov, and A.N. Pilchin, Applied Geothermics, p. 751,
Springer, New York (2014).
13. I.M. Kutasov and L.V. Eppelbaum, Wellbore and formation temperatures
during drilling, cementing of casing and shut-in, Proceedings of the World
Geothermal Congress 2015 held in Melbourne, Australia, 19–25 April 2015.
14. C. Wang, Y. Meng, and W. Jiang, Variation in Wellbore temperature and its
effects on gas injection rate. Natural Gas Industry Press (in Chinese) 27(10),
67–69 (2007).
15. J. Li, B. Guo, and B. Li, A closed form mathematical model for predicting gas
temperature in gas-drilling unconventional tight reservoirs. J. Nat. Gas Sci.
Eng. 27, 284–289 (2015).
16. M.M. Abbott and H.C. van Ness, Thermodynamics, 2nd edition, pp. 100-101,
McGraw-Hill Publishing Company, New York (1989).
17. K. Kadoya, N. Matsunaga, and A. Nagashima, Viscosity and thermal conduc-
tivity of dry air in the gaseous phase. J. Phys. Chem. Ref. Data 14(4), 947–970
(1985).
18. B. Guo and G. Liu, Applied Drilling Circulation Systems, pp. 145–150, Elsevier,
Oxford (2011).
19. H.H. Keller, E.J. Couch, and P.M. Berry, Temperature distribution in circu-
lating mud columns. SPE Journal 13(1), 23–30 (1973).
20. G.R. Wooley, Computing downhole temperatures in circulation, injection,
and production wells. J. Petrol. Technol. 32(19), 1509–1522 (Sept. 1980).
21. D.W. Marshall and R.G. Bentsen, A computer model to determine the tem-
perature distributions in a Wellbore. J. Can. Technol. 21(1), 63–75 (Jan-Feb.
1982).
22. C.S. Kabir, A.R. Hasan, G.E. Kouba, and N.M. Ameen, Determining circu-
lating fluid temperature in drilling, workover, and well-control operations.
SPE Drilling & Completion 11(2), 74–79 (June 1996).
23. D.A. Glowka and C.M. Stone, Thermal response of polycrystalline diamond
compact cutters under simulated downhole conditions. SPE Journal (April
1985), 143–156 (1985).
24. P. L. Moore, Five factor that affect drilling rate. Oil Gas J. 56(40), 141–162
(1958).
Appendix A: Steady Heat Transfer Solution for Fluid

Temperature in Counter-Current Flow
Assumptions
The following assumptions are made in model formulation:
a. The thermal conductivities of casing are assumed to be

infinitive.
b. The geothermal gradient behind the annulus is not affected
by borehole fluid.
c. Heat capacity of fluid is constant.
d. Friction-induced heat is negligible.
Governing Equation
Figure 15.5 depicts a small element of a borehole section with a drill string
at center. Consider the heat flow inside the drill pipe during a time period
of ∆t. Heat balance is given by
Qp,in – Qp,out – qp = Qp,chng (A.1)
where
Qp,in = heat energy brought into the drill pipe element by fluid due to
convection, J
Qp,out = heat energy carried away the drill pipe element by fluid due to
convection, J
qp = heat transfer through the drill pipe due to conduction, J
Qp,chng = change of heat energy in the fluid, J.
These terms can be further formulated as
Qp,in = CpṁpTp,LΔt (A.2)

Qp, in Qa, out

Casing Drill pipe Drill pipe Casing
Cement
Cement Annulus
Inside
Annulus
drill pipe
ΔL qp qa
Qp, out Qa, in
Figure 15.5 Sketch illustrating heat transfer in a borehole section.
Qp,out = CpṁpTp,L+ΔLΔt (A.3)
 ∂Tp 
q p = π d p K p∆L  − ∆t (A.4)
 ∂r 
Qp,chng = CpρpApΔLΔTp (A.5)
Substituting (A.2) through (A.5) into (A.1) gives
 ∂ Tp 
C pm p∆t (Tp , L − Tp , L +∆L ) + π d p K p∆L  t = ρpC p Ap∆L∆Tp
 ∂ r 
(A.6)
Dividing all the terms of this equation by ∆L∆t yields
(Tp,L − Tp,L +∆L ) ∂ Tp ∆Tp

C pm p + π dpK p = ρpC p Ap (A.7)
∆L ∂r ∆t
For infinitesimal of ∆L and ∆t, this equation becomes

∂ Tp ρ p Ap ∂ Tp π d p K p ∂ Tp
+ = (A.8)

∂L m p ∂ t C pm p ∂ r

The radial-temperature gradient in the insulation layer can be formu-

lated as
∂ Tp Τa − Tp
= (A.9)

∂r tp

Substituting Eq. (A.9) into Eq. (A.8) yields
∂ Tp ∂ Tp
+ λp + α p (Tp − Ta ) = 0 (A.10)
∂L ∂t
where
ρ p Ap
λp = (A.11)
m p

π dpK p
αp = (A.12)
C pm pt p

Consider the heat flow in the annulus during a time period of ∆t. Heat
balance is given by
Qa,in – Qa,out + qp – qa = Qa,chng (A.13)
where
Qa,in = heat energy brought into the drill pipe element by fuid due to
convection, J
Qa,out = heat energy carried away the drill pipe element by fuid due to
convection, J
qa = heat transfer through casing and cement due to conduction, J
Qa,chng = change of heat energy in the fuid, J.
Qa,in = CaṁaTa,L + ΔLΔt (A.14)
Qa,out = CaṁaTa,LΔt (A.15)
 ∂T 
qa = π dc K c ∆L  − a  ∆t (A.16)
 ∂r 

Qa,chng = CaρaAaΔLΔTa (A.17)
Substituting (A.14) through (A.17) into (A.13) gives
 ∂ Tp 
Cam a∆t (Ta,L +∆L − Ta,L ) + π d p K p∆L  ∆t
 ∂ r 
 ∂T 
+ π da K a∆L  a  ∆t = ρaCa Aa∆L∆Ta
 ∂r 
(A.18)
Dividing all the terms of this equation by ∆L∆t yields
(Ta,L +∆L − Ta,L )  ∂ Tp   ∂T  ∆T

Cam a + π dpK p   + π dc K c  a  = ρaCa Aa a
∆L  ∂r   ∂r  ∆t

(A.19)
For infnitesimal of ∆L and ∆t, this equation becomes
∂ Ta ∂T  ∂ Tp   ∂T 
Cam a − ρaCa Aa a − π d p K p   + π dc K c  a  = 0
∂t ∂t  ∂r   ∂r 

(A.20)

lated as
∂ Tp Τa − Tp
= (A.21)

∂r tp

and
∂ Ta Τ g − Ta
= (A.22)
∂r tc
Substituting (A.21) and (A.22) into (A.20) yields
∂Ta ∂T
∂L ∂t
( )
− λ a a + β a Tp − Ta − α a Ta − Tg = 0

( ) (A.23)
where
ρa Aa
λa = (A.24)
m a
π dpK p
βa = (A.25)
Cam at p

π dc K c
αa = (A.26)
Cam atc
The temperatures Tp and Ta at any given depth can be solved numeri-

cally from (A.10) and (A.23).
For steady heat flow, (A.10) and (A.23) can be written as:
∂ Tp
∂L
(
+ α p Tp − Ta = 0 )
(A.27)
∂ Ta
∂L
( ) (
+ β a Tp − Ta − α a Ta − Tg = 0

) (A.28)
where the geo-temperature can be expressed as:
Tg = Tg0 + GL. (A.29)
Boundary Conditions
The boundary conditions for solving (A.27) and (A.28) are expressed as
Tp = Tp0 at L = 0 (A.30)
Ta = Tp +ΔTb at L = Lmax (A.31)
Solution
The governing (A.27) and (A.28) subjected to the boundary conditions
(A.30) and (A.31) were solved with the method of characteristics. The
solutions take the following form:
Aa + ABb − a( B + E )
Tp = C1 Ae r1L + C2 Aer2 L + aL + (A.32)
AB
Aa + ABb − aE
Ta = C1 ( A + r1 )e r1L + C2 ( A + r2 )e r2 L + aL + (A.33)
AB
where
AB( AD − a) − [ ABC − ABb − Aa + a( B + E )]r2e r2 Lmax

C1 =
A 2 B(r1e r1Lmax − r2e r2 Lmax )
(A.34)
− AB( AD − a) + [ ABC − ABb − Aa + a( B + E )]r1e r1Lmax

C2 =
(A.35)
B + E − A + ( B + E − A)2 + 4 AB
r1 = (A.36)
2
B + E − A − ( B + E − A)2 + 4 AB
r2 = (A.37)
2
where A = αp, B = αa, C = Tp0, D = ∆Tb, E = βa, a = G, and b = Tg0.

16
Development of an Analytical Model for
Predicting the Fluid Temperature Profile
in Drilling Gas Hydrates Reservoirs
Liqun Shan, Boyun Guo* and Xiao Cai
University of Louisiana at Lafayette, Petroleum Engineering, Lafayette, LA, USA
Abstract
Production of natural gas from unconventional gas-hydrate reservoirs faces sev-
eral challenges. One of them is the well control issue due to the natural gas released
from gas hydrates during well drilling. It is very important for drilling engineers
to know if the temperature of drilling fluid in the borehole is lower than the crit-
ical temperature that causes hydrate dissociation. However, there is no reliable
method to predict the fluid temperature for designing drilling hydraulics. This
paper fills the gap. An analytical model was developed in this study for predict-
ing temperature profile in drilling gas-hydrate deposits. A case study indicates a
good consistency between model-implications and field observations. Sensitivity
analyses with the model show that the bottom-hole temperature in gas-hydrate
drilling is dominated by the temperature and flow rate of the injected drilling fluid.
The temperature of drilling fluid in the annulus can become greater than the geo-
temperature at the same depth at high fluid flow rate. The Joule-Thomason cooling
effect below the drill bit nozzles rapidly diminishes in a short interval above the
bottom hole due to the heating effect of geo-thermal gradient. The rate of pene-
tration of drill bit has a negligible effect on the fluid temperature profile due to the
low percentage of heat flow contributed by the drill cuttings. This paper provides
drilling engineers a rigorous method for predicting wellbore temperature profile
during drilling gas-hydrates reservoirs.
Keywords: Gas, hydrates, temperature, analytical, model
363
16.1 Introduction
The amount of energy trapped in the naturally occurring gas hydrates has
been found in the world to be about twice the amount of energy found in all
recoverable fossil fuels. Due to the large amount of natural gas stored in the
huge deposits of gas hydrates, gas hydrates have been considered to be the
future clean energy resources [1]. Production of natural gas from gas hydrates
faces great challenges, uncertainties, and special issues [2, 3]. Challenges
encountered during drilling gas hydrates wells are complications due to the
gas released from hydrates. The in-situ gas hydrates decompose when pres-
sure is reduced and temperature is increased due to the circulation of drilling
fluid in the borehole. The release of gas from the hydrates can cause borehole
stability [4] and well control problems [5, 6]. Therefore, accurately predicting
the temperature in the wellbore is critical to drilling gas hydrates wells.
Some mathematical models are available for estimating fluid tempera-
ture during drilling conventional wells. An approximate solution that cou-
ples a pseudo-steady heat transfer in the wellbore to transient heat flow in
the formation was presented by Ramey [7]. Raymond [8] derived complete
heat transfer equations for the well-bore-reservoir system. A number of
further investigations have been conducted since then. Holmes and Swift
[9] presented a method to calculate circulating mud temperatures under
pseudo-steady flow conditions. Sump and Williams [10] used a mathemat-
ical model to predict wellbore temperature during mud circulation and
cementing operations. All of these early analytical solutions do not con-
sider the rate of penetration and heat-transport due to mixing of mud with
drill cuttings. Keller et al. [11] provided a numerical transient heat transfer
model for predicting temperature distribution in circulating mud columns.
Wooley [12] applied finite difference method to computing downhole tem-
peratures in circulation, injection, and production wells. His model predic-
tions were found in agreement with field temperature data. Following Keller
et al.’s approach Marshall and Bentsen [13] developed a computer model to
calculate the temperature distribution in a well-bore using finite difference
method with an improved solution procedure. Arnold [14] calculated tem-
perature variation in circulating wellbore fluid. Kabir et al. [15] presented
mathematical models for determining the circulating fluid temperature in
drilling, workover, and well-control operations. Results from their models
do not agree with that from Wooley’s (1980) [12] model near the bottom
hole. This is attributed to the fact that Kabir et al. made an assumption
of stationary temperature at the bottom hole, i.e., the temperature gra-
dient in the axial direction was assumed to be zero at the bottom hole.
Development of Analytical Model for Predicting Fluid Profile 365
Nguyen et al. [16] developed a temperature model for modeling thermal

effects on wellbore stability accounting for frictional heat. Zhang et al.
[17] presented a time-convolution approach for modeling heat exchange
between a wellbore and its surrounding formation. Wu et al. [18] demon-
strated a coupled numerical model for wellbore/reservoir temperature pre-
diction and stress analysis during fluid circulation. Hasan and Kabir [19]
presented a unified approach to modeling wellbore heat transfer under
diverse situations. Analytical fluid temperature expressions for both sin-
gle and multiple flow conduits were provided. Their transient expressions
allow coupling with fluid flow models for handling many different types
of flow problems, both steady- and unsteady-state, encountered in hydro-
carbon production systems. Most recently, Kutasov and Eppelbaum [20]
presented a transient thermal model to predict well-bore and formation
temperatures during drilling, cementing of casing and shut-in.
In summary, a number of analytical and numerical models have been
developed in the past five decades for predicting wellbore fluid tempera-
ture profile under steady and transient heat flow conditions. Although they
have been used in the oil and gas industry with success, three facts were not
taken into consideration in model development. They are 1) the tempera-
ture of fluid in the annular space is affected by the drill cuttings entrained
at the bottom hole; 2) the thermal properties (heat capacity and thermal
conductivity) of the fluid in the annulus are different from that inside the
drill string due to the addition of drill cuttings to the system; and 3) the
Joule-Thomason effect at the drill bit nozzles reduces fluid temperature.
These facts are especially pronounced in drilling gas hydrates deposits.
This paper presents a new analytical solution to the fluid temperature
profiles inside and outside the drill string, considering the effect of gas
hydrates cuttings entrained at the bottom hole and the temperature drop
at bit due to Joule-Thomason cooling effect. Case studies and sensitivity
analyses with the new model are presented.
16.2 Mathematical Model

This section describes the new analytical model developed for predicting
the fluid temperature profiles inside and outside the drill string during gas
hydrates drilling. Definitions of symbols are shown in the Nomenclature
section. The following assumptions are made in model formulation:
a. The thermal conductivities of casing are assumed to be

infinitive.
b. The geothermal gradient behind the annulus is not affected

by borehole fluid.
c. Heat capacity of fluid is constant.
d. Friction-induced heat is negligible.
Figure 16.1 depicts a small element of a borehole section with a drill

string at center.
Consider the heat flow inside the drill pipe during a time period of Δt.
Heat balance is given by
Qp,in – Qp,out – qp = Qp,chng (16.1)
where
Qp,in = heat energy brought into the drill pipe element by fluid due to
convection, J
Qp,out = heat energy carried away the drill pipe element by fluid due to
convection, J
qp = heat transfer through the drill pipe due to conduction, J
Qp,chng = change of heat energy in the fluid, J.
Qp,in = Cpm. pTp,LΔt (16.2)
Qp, in Qa, out

Casing Drill pipe Drill pipe Casing
Cement
Cement
Inside
Annulus drill pipe Annulus
ΔL qp qa
Qp, out Qa, in
Figure 16.1 Sketch illustrating heat transfer in a borehole section.

Qp,out = Cpm. pTp,L+ΔLΔt (16.3)
 ∂ Tp 
q p = π d p K p∆L  − ∆t (16.4)
 ∂ r 
Qp,chng = CpρpApΔLΔTp (16.5)
Substituting Eqs. (16.2) through (16.5) into Eq. (16.1) gives
 ∂ Tp 
C pm p∆t (Tp , L − Tp , L +∆L ) + π d p K p∆L   ∆t = ρ pC p Ap∆L∆Tp (16.6)
 ∂ r
Dividing all the terms of this equation by ΔLΔt yields
(Tp ,L − Tp ,L +∆L ) ∂ Tp ∆Tp

C pm p + π dpK p = ρ pC p A p (16.7)
∆L ∂r ∆t
For infinitesimal of ΔL and Δt, this equation becomes
∂ Tp ρ p Ap ∂ Tp π d p K p ∂ Tp
+ = (16.8)

∂L m p ∂ t C pm p ∂ r

lated as
∂ Tp Ta − Tp
= (16.9)

∂r tp

Substituting Eq. (16.9) into Eq. (16.8) yields
∂ Tp ∂ Tp
+ λp + a p (Tp − Ta ) = 0 (16.10)
∂L ∂t
where
ρ p Ap
λp = (16.11)
m p

π dpK p .
ap = (16.12)
C pm pt p

Consider the heat flow in the annulus during a time period of Δt. Heat
balance is given by
Qa,in − Qa,out + qp – qa = Qa,chng (16.13)
where
Qa,in = heat energy brought into the drill pipe element by fluid due to
convection, J
Qa,out = heat energy carried away through the drill pipe element by fluid
due to convection, J
qa = heat transfer through casing and cement due to conduction, J
Qa,chng = change of heat energy in the fluid, J.
Qa,in = Cam. aTa,L+ΔLΔt (16.14)
Qa,out = Cam. aTa,LΔt (16.15)
 ∂T 
qa = π dc K c ∆L  − a  ∆t (16.16)
 ∂r 

Qa,chng = CaρaAaΔLΔTa (16.17)
Substituting Eqs. (16.14) through (16.17) into Eq. (16.13) gives
 ∂ Tp   ∂T 
Cam a∆t (Ta,L +∆L − Ta,L ) − π d p K p∆L   ∆t + π dc K c ∆L  a  ∆t
 ∂r   ∂r 
= ρaCa Aa∆L∆Ta (16.18)

Dividing all the terms of this equation by ΔLΔt yields
(Ta,L +∆L − Ta,L )  ∂ Tp   ∂T  ∆T

Cam a − π dpK p   + π dc K c  a  = ρaCa Aa a
∆L  ∂r   ∂r  ∆t
(16.19)
For infinitesimal of ΔL and Δt, this equation becomes
∂ Ta ∆T  ∂ Tp   ∂T 
Cam a − ρaCa Aa a − π d p K p   + π dc K c  a  = 0
∂L ∆t  ∂r   ∂r 
(16.20)

lated as
∂ Tp Ta − Tp
= (16.21)

∂r tp

and
∂ Ta Tg − Ta (16.22)
=
∂r tc
Substituting Eqs. (16.21) and (16.22) into Eq. (16.20) yields
∂Ta ∂T
− λa a + βa (Tp − Ta ) − aa (Ta − Tg ) = 0 (16.23)
∂L ∂t

where
ρa Aa
λa = (16.24)
m a
π dpK p
βa = (16.25)
Cam at p

π dc K c
aa = (16.26)
Cam atc
The temperatures Tp and Ta at any given depth can be solved numeri-

cally from Eqs. (16.10) and (16.23).
For steady heat flow, Eqs. (16.10) and (16.23) can be written as:
∂ Tp
+ a p (Tp − Ta ) = 0 (16.27)
∂L
∂ Ta
+ βa (Tp − Ta ) − aa (Ta − Tg ) = 0 (16.28)
∂L
where the geo-temperature can be expressed as:
Tg = Tg0 + GL (16.29)
It is generally recognized in drilling practice that the cooling effect of

sea water around the steel riser and drill string is so significant that the
temperature of the drilling mud inside the drill string at the sea floor depth
reaches the known temperature of sea water at the same depth. Due to
the high thermal conductivities of steal drill pipe and water-base drilling
mud, the mud temperature at sea floor does not change notably with time.
Therefore, the boundary conditions for solving Eqs. (16.27) and (16.28) are
expressed as
Tp = Tp0 at L = 0 (16.30)
Ta = Tp + Tb0 at L = Lmax (16.31)
The governing equations (16.27) and (16.28) subjected to the boundary

conditions (16.30) and (16.31) were solved with the method of character-
istics. The solutions take the following form:
AG + ABTg 0 − G( B + E )
Tp = C1 Ae r1L + C2 Ae r2 L + GL + (16.32)
AB
and
AG + ABTg 0 − EG
Ta = C1 ( A + r1 )e r1L + C2 ( A + r2 )e r2 L + GL +
AB
(16.33)
where
AB( ATb 0 − G ) − [ ABTp 0 − ABTg 0 − AG + G( B + E )]r2e r2 Lmax

C1 =
(16.34)
− AB( ATb 0 − G ) + [ ABTp 0 − ABTg 0 − AG + G( B + E )]r1e r1Lmax

C2 =
(16.35)
B + E − A + ( B + E − A)2 + 4 AB (16.36)
r1 =
2
B + E − A − ( B + E − A)2 + 4 AB (16.37)
r2 =
2
where
π dpK p
A= (16.38)
C pm pt p

π dc K c
B= (16.39)
Cam atc
π dpK p .
E= (16.40)
Cam at p

This mathematical model has a unique advantage over the existing

models in that it handles the mass and heat of gas hydrates added to the
system at the bottom hole. The product of heat capacity and mass flow rate
.
in the annulus Cam a is expressed as
Cam a = C pm p + Csm s (16.41)

where the mass flow rate inside the drill string is expressed as
.
m p = ρpQp. (16.42)
where ρp and Qp are density of fluid inside the pipe and fluid flow rate,
.
respectively The product of heat capacity and mass flow rate of solid Csm s
is further expressed in two terms:
Csm. s = Chm. h + Crm.r (16.43)
where
π 2
m h = Db R pϕρh (16.44)
4
and
π 2
m r = Db R p (1 − ϕ )ρr (16.45)
4
where Db, Rp, ϕ, and ρr are drill bit diameter, rate of penetration, rock
porosity, and density of rock, respectively.
It is understood that the temperature change at drill bit due to Joule-
Thomason effect (Tb0) depends on fluid properties. For compressive flu-
ids the temperature in the downstream of the bit orifices (bottom hole) is
lower than the temperature in the upstream and can be predicted based on
isentropic process [21]:
k −1
P  k
Tdn = Tup  dn  (16.46)
P  up

upstream of bit orifices, respectively, Pdn and Pup are the absolute pressures
in the down-stream and upstream of bit orifices, respectively, and k is the
specific heat ratio of fluid (k = 1.3 for air). For incompressive fluids, such as
water, the temperature in the downstream of the bit orifices (bottom hole)
has been found to be higher than the temperature in the upstream due
to the negative Joule-Thompson coefficient [22]. Because this temperature
change is fluid-dependent, the value of Tb0 in the analytical model is left
user-specified.
16.3 Case Study

The National Gas Hydrate Expedition Programme (NGHP-01) in India
found the occurrence of natural gas hydrates (NGH) along the passive con-
tinental margins of the Indian peninsula and in the Andaman convergent
margin [23]. Drilling result from well NGHP-01-17A in the Andaman
Islands indicates gas hydrate saturations in two zones at intervals of 547m to
570m and 586m to 602m below the sea floor to be 10% and 20 % respectively.
The geo-temperature of the hydrate-bearing layer is about 10.6 °C [24, 25].
Table 16.1 shows well configuration data and operation parameter val-
ues used in drilling the gas hydrate layer. Estimated material properties
from Incropera and Dewitt [26] are presented in Table 16.2. Figure 16.2
shows temperature profiles calculated by Eqs. (16.32) and (16.33). This fig-
ure illustrates that the drilling mud inside the drill string is heated up while
it flows down the drill string. But the mud temperature does not rise close
to the geothermal temperature at the bottom hole. It also indicates that
the upward flowing fluid mixture (mud and rock cuttings with hydrates)
in the annulus is heated to the geo-temperature in a short interval above
the bottom hole. But it goes beyond the geothermal temperature at the
same depth near the sea floor. The mud pressure at the top of the hydrate-
bearing layer (1,891 m) is about 18.9 MPa. The hydrate dissociation
temperature at this pressure is about 18.3 °C. Because the geothermal
Table 16.1 Well and operation parameter values in drilling well

NGHP-01-17A.
Total depth below see floor 602 m
Bit diameter 0.2032 m
Outer diameter of drill pipe 0.127 m
Inner diameter of drill pipe 0.108 m
Geothermal temperature at sea floor 1.5 °C
Geothermal gradient 0.015 °C
Mud injection rate 0.025 m3/s
Mud density 1100 kg/m3
Mud temperature inside drill pipe at sea floor 2 °C
Rate of penetration 0.008 m/s
Table 16.2 Estimated material properties in drilling well

NGHP-01-17A.
Thermal conductivity of drill pipe 43 W/m-C
Heat capacity of mud inside drill pipe 4210 J/kg-C
Heat capacity of rock 920 J/kg-C
Heat capacity of hydrates 2050 J/kg-C
Rock porosity 0.6
Density of rock 2650 kg/m3
Density of hydrates 985 kg/m3
12
10
8
Temperature (C)
4
Temperature in the annulus
0
0 100 200 300 400 500 600 700
Depth below sea floor (m)
Figure 16.2 Model-predicted temperature profiles in drilling well NGHP-01-17A.
temperature of the hydrate-bearing layer is only 10.6 °C, it is expected that

the fluid temperature at the hydrate-depth will never reach the hydrate
dissociation temperature. Therefore the gas cut in the upper section of
borehole will be from only the drill cuttings created by the drill bit, which
should not be a threat to drilling operations if degassers work properly. In
fact, this well was drilled successfully without well control issues reported.
16.4 Sensitivity Analysis

For well control concerns, it is desirable to explore major factors affecting
annular temperature profile during gas-hydrate drilling. Previous studies
suggest that the annular temperature increases with the temperature of

injected fluid and fluid flow rate [15, 19, 20]. Joule-Thomason cooling
effect below the drill bit nozzles affects the temperature profile in the annu-
lus [21, 27]. Rate of penetration (ROP) may also affect annular temperature
because drill cuttings feed the annulus with heat in addition to mass. These
factors were further investigated with the derived analytical model using
the same data presented in Table 16.1 with one parameter value changed
at a time.
Figure 16.3 presents temperature profiles calculated by Eqs. (16.32)
and (16.33) with a mud temperature at sea floor depth being equal to the
hydrate reservoir temperature 10.6 °C. It indicates that the mud inside the
drill string is cooled down at the shallow depth and then heated up while it
flows along the drill string. But the mud temperature does not rise close to
the geothermal temperature at the bottom hole. It also demonstrates that
the upward flowing fluid mixture (mud and rock cuttings with hydrates)
in the annulus is heated to the geo-temperature in a short interval above
the bottom hole. But at the depth of sea floor, it becomes 1.8 °C which is
slightly greater than the geothermal temperature at the same depth. Since
this annular temperature is much lower than the hydrate dissociation
temperature 18.3 °C, it is predicted that the hydrates in the drill cuttings
will not dissociate at the depth of sea floor, i.e., there should be no free gas
in the mud at this depth.
Figure 16.4 illustrates temperature profiles calculated by Eqs. (16.32)
and (16.33) with a mud flow rate of 0.05 m3/s. A comparison between
12
10
Temperature (C)
4
0
0 100 200 300 400 500 600 700
Figure 16.3 Model-predicted temperature profiles with injection mud temperature

10.6 °C.
12
10
Temperature (C)
4
0
0 100 200 300 400 500 600 700
Figure 16.4 Model-predicted temperature profiles with mud flow rate 0.05 m3/s.
12
10
8
Temperature (C)

0
0 100 200 300 400 500 600 700
Figure 16.5 Model-predicted temperature profiles with 3 °C-temperature drop at the drill
bit due to Joule-Thomson effect.
Figures 16.1 and 16.3 shows that the fast injection of cold mud reduces bot-
tom hole temperature. The upward flowing fluid mixture (mud and rock
cuttings with hydrates) in the annulus is heated to the geo-temperature in
a short interval above the bottom hole. Also the temperature of the upward
flowing fluid mixture in the annulus is slightly elevated due to the fast con-
vection of heat from the deeper section.
Figure 16.5 demonstrates temperature profiles calculated by Eqs. (16.32)
and (16.33), considering a 3 °C temperature drop due to Joule-Thomson
effect below the drill bit nozzles. This assumption is made for oil-base mud
with a positive Joule–Thomason coefficient. Oil-base muds are often utilized
in drilling hydrate-bearing formations to prevent hole enlargement due to
decomposition of gas hydrates. A comparison between Figures 16.1 and 16.4
shows that the Joule-Thomason effect rapidly diminishes due to the heating
effect of geo-thermal gradient in a short interval above the bottom hole.
The effect of ROP on annular temperature was investigated with Eq.
(16.33). When the ROP is doubled, its effect was found not noticeable. This
is because the heat added by the drilling cuttings is less than 0.5% of the
total heat flow at the total well depth.
16.5 Conclusions
An analytical model was developed in this study for predicting tempera-
ture profiles in drilling gas hydrates deposits. A case study with the model
was performed. Parameter sensitivity analyses were carried out with the
model. The following conclusions are drawn:
1. For the real case studied, the drilling fluid temperature at

the top of the hydrate-bearing interval was predicted by the
model to be lower than the hydrate dissociation tempera-
ture at the same depth, and thus the gas hydrates were not
expected to decompose during the well drilling process. This
result is consistent with the real situation where no bore-
hole enlargement was found due to hydrate decomposition.
However, due to the lack of temperature measurement, no
direct comparison between the model-predicted and mea-
sured temperatures was made. The consistency between
field observation and model implication is evidenced by the
resultant trouble-free drilling operation.
2. The injected drilling fluid is rapidly heated up by the
geo-thermal gradient after entering the annular space. The
temperature of the drilling fluid in the shallow annulus can
become greater than the geo-temperature at the same depth,
depending on the fluid flow rate.
3. The bottom hole temperature is dominated by the flow rate
of the injected drilling fluid and is always below the geo-
temperature at the same depth due to the Joule-Thomason
cooling effect at the drill bit nozzles. But this effect rapidly
diminishes in a short interval above the bottom hole due to

the heating effect of the geo-thermal gradient.
4. The effect of temperature of the injected drilling fluid on
annular temperature diminishes as depth increases due to
the heating effect of the geo-thermal gradient.
5. The rate of penetration has a negligible effect on the fluid
temperature profile due to the low percentage of heat con-
tributed by the drill cuttings.
Acknowledgements
This research was supported by the China National Natural Science
Foundation Founding No. 51134004, No. 51274220, and No. 51334003.
Nomenclature
Aa = cross-sectional area of annulus open for fluid flow (m2)
Ap = cross-sectional area of drill pipe open for fluid flow (m2)
Ca = heat capacity of fluid in the annulus (J/kg- °C)
Cp = heat capacity of fluid inside drill pipe (J/kg- °C)
Cs = heat capacity of solid in the annulus (J/kg- °C)
Ch = heat capacity of gas hydrates (J/kg- °C)
Cr = heat capacity of rock (J/kg- °C)
Db = bit-diameter (m)
dc = inner-diameter of cement sheath (m)
Dc = outer-diameter of cement sheath (m)
dp = inner-diameter of drill pipe (m)
Dp = outer-diameter of drill pipe (m)
k = specific heat ratio of fluid, dimensionless
Kc = thermal conductivity of cement (W/m- °C)
Kp = thermal conductivity of drill pipe (W/m- °C)
L = wellbore depth along the drill string (m)
Lmax = the maximum hole depth (m)
.
m a = mass flow rate in the annulus (kg/s)
.
m p = mass flow rate inside the drill pipe (kg/s)
.
m s = mass flow rate of solid cuttings in the annulus (kg/s)
.
m h = mass flow rate of gas hydrates (kg/s)
.
m r = mass flow rate of rock (kg/s)
Qp = fluid flow rate (m3/s)
Pdn = absolute pressure in the downstream (MPa)

Pup = absolute pressure in the upstream (MPa)
RP = rate of penetration (m/s)
tc = thickness of cement sheath (m)
tp = wall thickness of drill pipe (m)
Ta = temperature of annular fluid (°C)
Tb0 = temperature change at drill bit due to Joule-Thomason effect (°C)
Tdn = absolute temperature in the downstream (°K)
Tg = geothermal temperature at depth (°C)
Tg0 = geothermal temperature at surface (°C)
Tp = temperature of fluid inside drill pipe at depth (°C)
Tp0 = temperature of fluid inside drill pipe at surface (°C)
Tup = absolute temperature in the upstream (°K)
Greeks
ϕ = porosity of rock
ρa = fluid density in the annulus (kg/m3)
ρh = density of hydrates (kg/m3)
ρp = fluid density inside drill pipe (kg/m3)
ρr = density of dry rock (kg/m3)
ρs = solid density in the annulus (kg/m3)
References
1. Y.F Makogon, Natural gas hydrates–A promising source of energy. J. Nat.
Gas Sci. and Eng. 2, 49–59 (2010).
2. G.J. Moridis, T.S. Collett, R. Boswell, M. Kurihara, M.T. Reagan, C. Koh, and
E.D. Sloan, Toward Production from gas Hydrates: Current status, assess-
ment of resources, and simulation-based evaluation of technology and
potential. SPE J. 12(5), 745–771 (2009).
3. G.J. Moridis, T.S. Collett, M. Pooladi-Darvish, S. Hancock, C. Santamarina, R.
Boswell, T. Kneafsey, J. Rutqvist, M.B. Kowalsky, M. Reagan, E.D. Sloan, A.K.
Sum, and C.A. Koh, Challenges, uncertainties, and issues facing gas produc-
tion from Gas-Hydrate deposits. SPE Reserv. Eval. Eng. 14, 76–112 (2011).
4. T. Khabibullin, G. Falcone, and C. Teodoriu, Drilling through gas-hydrate
sediments: Managing Wellbore-Stability risks. SPE Drilling & Completion
26, 287–294 (June 2011).
5. D.M. Hannegan, R.J. Todd, D.M. Pritchard, and B. Jonasson, MPD–uniquely
applicable to methane hydrate drilling, SPE/IADC Underbalanced Technology
Conference and Exhibition, SPE/IADC 91560 (2004).
6. D.M. Hannegan, Methane hydrate drilling technology, Offshore Technology

Conference, OTC 17448 (2012).
7. H.J. Ramey, Wellbore heat transmission. J. Pet. Technol. 14, 427–435 (1962).
8. L.R. Raymond, Temperature distribution in a circulating drilling fluid. SPE–
AIME 21, 333–342 (1969).
9. C.S. Holmes and S.C. Swift, Calculations of Circulating Mud Temperatures.
J. Petrol. Technol. 22(6), 670–674 (1970).
10. G.D. Sump and B.B. Williams, Prediction of wellbore temperature during
mud circulation and cementing operations, J. Eng. Ind. 95, 1083–1092 (1973).
11. H.H. Keller, E.J. Couch, and P.M. Berry, Temperature distribution in circu-
lating mud columns. SPE J. 13(1), 23–30 (1973).
12. G.R. Wooley, Computing downhole temperatures in circulation, injection,
and production wells. JPT 32, 1509–1522 (1980).
13. D.W. Marshall and R.G. Bentsen, A computer model to determine the tem-
perature distributions in a wellbore. J. Can. Technol. 21, 63–75 (1982).
14. F.C. Arnold, Temperature variation in a circulating wellbore fluid. ASME J.
Energy Res. Technol. 112, 79–83 (1990).
15. C.S. Kabir, A.R. Hasan, G.E. Kouba, and N.M. Ameen, Determining circulat-
ing fluid temperature in drilling, workover, and well-control operations. SPE
Drilling & Completion 11, 74–79 (1996).
16. D. Nguyen, S. Miska, M. Yu, and A. Saasen, Modeling thermal effects on
wellbore Stability, SPE 133428 presented at the SPE Trinidad and Tobago
Energy Resources Conference (2010).
17. Y. Zhang, L. Pan, K. Pruess, and S. Finsterle, A time-convolution approach
for modeling heat exchange between a wellbore and surrounding formation.
Geothermics 40, 261–266 (2011).
18. B. Wu, X. Zhang, B. Wu, R.G. Jeffrey, and A.P. Bunger, A coupled model for well-
bore/ reservoir temperature prediction and stress analysis during fluid circula-
tion, 46th US Rock Mechanics/Geomechanics Symposium, pp. 24–27 (2012).
19. A.R. Hasan and C.S. Kabir, Wellbore heat-transfer modeling and applica-
tions. J. Petrol. Sci. Eng. 86, 127–136 (2012).
20. I.M. Kutasov and L.V. Eppelbaum, Wellbore and formation temperatures
during drilling, cementing of casing and shut-in, World Geothermal Congress,
pp. 19–25 (2015).
21. B. Guo and G. Liu, Applied Drilling Circulation Systems, Elsevier, Oxford (2011).
22. P.H. Stauffer, K.C. Lewis, J.S. Stein, B.J. Travis, P. Lichtner, and G. Zyvoloski,
Joule-Thomson effects on the fl ow of liquid water. Transport in Porous Media
105(3), 471–485 (2014).
23. T.S. Collett, R. Boswell, J.R. Cochran, P. Kumar, M. Lall, A. Mazumdar, M.V.
Ramana, T. Ramprasad, M. Riedel, K. Sain, A.V. Sathe, and K. Vishwanath,
Geologic implications of gas hydrates in the offshore of India: Results of the
national gas hydrate program expedition 01. Mar. Petrol. Geol. 58A, 3–28 (2014).
24. U. Shankar and M. Riedel, Heat fl ow and gas hydrate saturation estimates
from andaman Sea. India. Mar. Petrol. Geol. 43, 434–449 (2013).
25. W.J. Winters, Physical and geotechnical properties of gas hydrate bearing
sediment from offshore India and northern cascadia margin compared
to other hydrate reservoirs, 7th International Conference on Gas Hydrates,
pp. 1–22 (2011).
26. F. Incropera and D. Dewitt, Heat Transfer, John Wiley and Sons Inc. New
York (1985).
27. J. Li, B. Guo, and B. Li, A closed form mathematical model for predicting gas
temperature in gas-drilling unconventional tight reservoirs. J. Nat. Gas Sci.
Eng. 27, 284–289 (2015).
17
Distinguishing Between Brine-
Saturated and Gas-Saturated Shaly
Formations with a Monte-Carlo
Simulation of Seismic Velocities
Simon Katz*, George Chilingar and Leonid Khilyuk
Russian Academy of Natural Sciences, US Branch,

Abstract
The Monte-Carlo method was used to model seismic velocities in shaly sands
defined by a pair of linear equations with randomized coefficients. The starting
point for this research was a deterministic model with linear equations defining
compressional (Vp) and shear waves (Vs) velocities in gas or brine-filled shaly
deposits in the North Sea. We introduced two levels of randomization. At the first
level, porosity and clay content were treated as random variables. At the second
level, in addition to randomized porosity and clay content, coefficients of equa-
tions were randomized and random errors in seismic velocities were added. The
effects of errors in measured seismic velocities and random coefficients were not
separable. Therefore, we introduced a summary measurement of the randomiza-
tion effect that includes the effects of all random parameters.
The goal of the Monte Carlo simulation was to evaluate the efficiency of iden-
tification of gas-saturated deposits using seismic velocity data. We tested three
methods for the identification of gas-filled formations: k-nearest neighbor, recur-
sive partitioning, and linear discriminant analysis. The efficiency of identification
was characterized by two conditional probabilities: (1) The Probability of True
Discovery, which is defined as the probability of identification of a gas-saturated
formation as gas saturated, and (2) The Probability of False Discovery which is
defined as the false identification of a brine-saturated formation as gas saturated.
Both probabilities were evaluated as functions of the width of distribution of
383
random components in the equations for Vp and Vs velocities in the formations of

two types and as a function of the summary randomization parameter. The results
of identifying gas-saturated formations using only Vp were negative. On the other
hand, the combination of parameters derived from the Vp and Vs velocities was a
strong indicator of a gas-saturated formation.
Keywords: Monte carlo, classification, probability of true discovery, seismic

velocities, brine, gas, shales
17.1 Introduction
The goal of this paper is to analyze the efficiency of classification of gas
and brine-filled shaly formations using the Monte Carlo technique. The
economic success of shale gas in the US over the last ten years has led to
the rapid development of shale gas in the world, thus improving the sus-
tainability of worldwide energy resources.
The Monte Carlo method relies on repeated random sampling followed
by the calculation of statistical attributes defined by a set of randomized
equations. We generated up to 10000 random samples of porosity, clay
content and coefficients of model equations to calculate multiple replicas
of Vp and Vs velocities.
The results of the Monte Carlo simulations were used to evaluate the
efficiency of identifying gas-filled formations against brine-filled ones. The
identification efficiency is characterized by the following two conditional
probabilities:
1. Probability of classifying gas-filled formations as gas-filled.

This probability is called the probability of true discovery.
2. Probability of classifying brine-filled formations as gas-filled.
This probability is called the probability of false discovery.
Both classification probabilities are evaluated by three classification

techniques - K-Nearest Neighbor (KNN) [1], Recursive partitioning
(RPART) [2], and Linear Discriminant Analysis (LDA) [3]. All three meth-
ods belong to the class of supervised methods and rely on the use of train-
ing and test sets. The training set is used to optimize the parameters of the
classification procedure. In all three classification methods we use cross
validation as an optimization tool. As soon as the classification procedure
is optimized it may be used to classify new data from the test set.
We demonstrate that both classification probabilities strongly depend
on the size of the error of estimation of Vp and Vs velocities from seismic
Distinguishing Brine-Saturated and Gas-Saturated Formations 385
data and on the uncertainty of the coefficients in the Vp, Vs models. Another
factor that affects the classification probabilities is the set of predictors used
for classification.
The Monte Carlo simulation and all statistical analysis is done with
statistical software R [4–6]. The general principles of statistical classifica-
tion techniques have been studied [1, 6, 7]. Some approaches to classifi-
cation techniques and machine learning methods in geological problems
are presented in academic research [8]. The Monte Carlo techniques were
used in seismic inversion and geological research in various publications
[9–11]. Deterministic equations that link seismic velocities with porosity,
clay content and the effect of pore fluid are discussed in studies by Schon
and Buryakovsky et al. [12, 13]. The starting point for the analysis done in
this paper was a model for Vp and Vs velocities expressed as linear functions
of porosity and clay content. These types of models were built as a result of
linear regression for the data inside a limited range of porosity variations,
De-hua Han*, A. Nur (1986).
17.2 Random Models for Seismic Velocities

We start with deterministic equations for Vp and Vs velocities in North Sea
shaly brine or gas saturated reservoir formations [14].
Gas: Vp = ag,0 – ag,1 * ϕ – ag,2 * C (17.1)
Vs = bg,0 – bg,1 * ϕ – bg,2 * C (17.2)
Brine: Vp = ab,0 – ab,1 * ϕ – ab,2 * C (17.3)
Vs = bb,0 – bb,1 * ϕ – bb,2 * C (17.4)
where ϕ-porosity, C-clay content as fraction, Vp and Vs-compressional and

shear waves velocities, respectively.
The coefficients in equations 17.1 to 17.4 have the following values:
ag,0 = 4.82; ag,1 = 5.02; ag,2 = 0.597; bg,0 = 3.26 ; bg,1 = 3.03; bg,2 = 0.892
ab,0 = 5.46; ab,1 = 6.29 ; ab,2 = 1.1; bb,0 = 3.32 ; bb,1 = 3.62; bb,2 = 0.952
We assume that equations 17.1–17.4 define the mean values of the sto-
chastic Vp and Vs velocities. For purposes of the Monte-Carlo simulation
the equations 17.1–17.4 are modified to incorporate the assumption of

porosity and clay content variability, coefficient uncertainty and errors in
Vp, Vs estimations.
At the first randomization level, velocities Vp and Vs are described by the
set of equations 17.1–17.4 with porosity and clay content treated as random
variables.
At the second randomization level we used two types of modified equations:
a. Equations 17.1–17.4 with an additional random component
Gas: Vp = ag,0 – ag,1 * ϕ – ag,2 * C+ηg,1 (17.5)
Vs = bg,0 – bg,1 * ϕ – bg,2 * C+ηg,2 (17.6)
Brine: Vp = ab,0 – ab,1 * ϕ – ab,2 * C+ηb,1 (17.7)
Vs = bb,0 – bb,1 * ϕ – bb,2 * C+ηb,2 (17.8)
Random components hg,1, hg,2, hb,1, and hb,2 simulate errors in the Vp, Vs
velocities estimated from seismic data and the imprecise nature of coeffi-
cients ag,0, bg,0, ab,0, bb,0.
b. Equations 17.1–17.4 with randomized coefficients in Equations 17.1–17.4.
Gas: Vp = ag,0 + δg,1 – (ag,1 + δg,1)* ϕ – (ag,2 + μb,2)* C (17.9)
Vs = bg,0 + μg,0 – (bg,1 + μg,1)* ϕ – (bg,2 + μg,2)* C (17.10)
Brine: Vp = ab,0 + δb,0 – (ab,1 + δb,1)* ϕ – ab,2 *(1 + δb,2)* C (17.11)
Vs = bb,0 + μb,0 – (bb,1 + μb,1)* ϕ – (bb,2 + μb,2)* C (17.12)
Random variables δg,0, δg,1, μg,2, μg,0, μg,1, μg,2, δb,0, δg,1, μb,2, μb,0, μb,1, δg,0
model the imprecise nature of equations 17.1–17.4. It is important to note
that the effect of random variables δg,0, μg,0, δb,0, μg,0 is identical to the effect
of the errors in Vp, Vs estimated from seismic data.
We assume that the random variables in equations 17.9–17.12 have zero
mean, mutually independent, and their probability distributions are uni-
form in a predefined interval symmetric around zero.
17.3 Variability of Seismic Velocities Predicted

by Random Models
The uncertainty of the modeling results is defined by the combination of
errors in recorded seismic velocities and the uncertainty due to random
variations of the coefficients in the model.
Velocities Vp and Vs, defined by equations 17.9–17.12, are treated as
random variables with mean values defined by equations 17.1–17.4. To
evaluate the variability of Vp and Vs we generated multiple replicas of the
Vp and Vs velocities from equations 17.9–17.13. Using this set of velocity
replicas we estimated the standard deviation. We assumed that the porosity
and clay content were uniformly distributed in the intervals (0.1, 0.16) and
(0.3, 0.4), respectively. Random samples of the variables δg,0, δg,0, μg,0, μg,0,
δg,1, δg,2, μg,1, μg,1, δg,2, δb,2, μb,2, μb,2 were generated under the assumption of
uniform distribution in the interval (-a, a) with the probability of density
distribution:
 1
 ; x <= a (17.13)
P( x ) =  2a
 0; x > a

Figures 17.1 and 17.2 show standard deviations calculated from two sets
of simulated Vp, Vs velocities as functions of half the width of the interval of
distribution (parameter a in eq. 17.13). For each value of parameter a, ten
thousand Vp(k) and Vs(k), 1<=k<=10000, samples are calculated. These
samples are used to calculate standard deviations.
According to Figures 17.1 and 17.2, the standard deviations of the Vp,
Vs and Vp/Vs ratios are strongly dependent on the variability of random
components in equations 17.9–17.12 and monotonically increase with an
increase of half the width of the distribution interval.
One can observe the strong effect of the width of distribution interval
of random components in eqs. 17.9–17.12 on the variability of Vp, Vs and
their ratio. The Vp standard deviation is somewhat higher than the stan-
dard deviation of Vs. The standard deviation of the Vp/Vs ratio is about the
same for gas-filled and brine-filled formations.
Variability of predicted velocity due to model uncertainty
0.14
Velocity STD
0.10
std(Vp)
0.06
std(Vs)
0.05 0.10 0.15 0.20

Half width of the distribution interval
Figure 17.1 Gas-filled formations. The standard deviations of calculated Vp, Vs serve
as unctions of the half-width of the distribution interval for random components in
eqs. 17.10–17.13. The standard deviation is calculated using 10000 Vp and Vs samples
for a range of porosity from 0.1 to 0.16 and clay content in the range of 0.3 to 0.4. Standard
deviation is given in km/sec.
Variability of predicted Vp/Vs ratio due to model uncertainty
0.08
STD of Vp/Vs ratio
0.04
GAS
BRINE
0.00
0.05 0.10 0.15 0.20
Half width of the distribution interval
Figure 17.2 Standard deviations of the Vp/Vs ratio for gas and brine saturated formations.
Vp and Vs velocities were calculated for 10000 samples of random components in eqs.
17.10–17.14.
17.4 The Separability of (Vp, Vs) Clusters for Gas-

and Brine-Saturated Formations
Figure 17.3 shows two clusters of (Vp, Vs) velocities calculated for two val-
ues of the parameter a in eq. 17.13. Modeling was done for a pore range of
(0.10, 0.16) and a clay range of (0.3, 0.4).
Uncertainty = 0 Uncertainty = 0.2
4.5 4.5
4.0 4.0
Vp
Vp
3.5 3.5
GAS GAS
3.0 Brine Brine
3.0
2.50 2.65 2.3 2.6 2.9

Vs Vs
Figure 17.3 Clusters of (Vp,Vs) velocities for gas-filled and brine-filled formations.
Two velocity clusters are perfectly separable in the case of the first level
randomization of the model. The velocity clusters partially overlap if the
models are randomized with strong random components in the coeffi-
cients of equations 17.9–17.12.
17.5 Reliability Analysis of Identifying Gas-Filled

Formations
According to equations 17.9–17.12, the effects of the errors in the Vp, Vs
velocities, estimated from seismic data and uncertainty in the coefficients
of equations, are inseparable. We introduced a summary attribute named
‘Total Uncertainty’ to characterize the effect of random parameters on the
efficiency of identifying gas-filled formations:
TU = V.std + a (17.14)
Where V.std is the standard deviation of the Vs and Vp velocities esti-

mated from external data.
a - The half width of the distribution interval for uniformly distributed
random components in equations 17.9–17.12.
We characterized the reliability of identifying gas-filled formations
by two conditional probabilities – the probability of true discovery and
the probability of false discovery. The probability of true discovery is
defined as the probability of classifying gas-filled formations as gas-

filled. The probability of false discovery is the probability of classifying
brine-filled formation as gas-filled. We used the Monte Carlo simula-
tion to evaluate these probabilities for three classification techniques -
k-nearest neighbor, recursive partition, and linear discriminant
analysis.
For each value of total uncertainty and each classification technique,
we tested the efficiency of classifi cation with the following sets of pre-
dictors (Vp), (Vp/Vs), (Vp, Vs, Vp/Vs), (Vp, Vs, Vp/Vs, Vp* Vs). For each
value of total uncertainty and each set of predictors we built 10 pairs of
training and test sets. Each set included 100 records for gas-filled forma-
tions and 100 records for brine-filled formations. The results presented
in the three following sections are an average over 10 pairs of training
and test sets.
All three classification techniques relied on the following classification
model:
Class ~ predictors (17.15)
 GAS 
where Class = 
 BRINE 
(17.16)
and the predictor component in eq. 17.15 has one of the following forms:
predictors = Vp (17.17)

predictors = Vp/Vs (17.18)
predictors = Vp + Vs+ Vp/Vs (17.19)
predictors = Vp + Vs+ Vp/Vs+ Vp * Vs (17.20)
Each record in the training set has the form (class, predictors).
Estimates of the probability of true and false discoveries were calculated
as follows:
N test (G | G ))
P .true.discovery = (17.21)
N test (G )
N test (G | B))
P . false.discovery = (17.22)
N test ( B)
where
Ntest (G|G)) - number of records in the test set with Class variable equal
GAS that is classified as GAS.
Ntest (G|B)) - number of records in the test set with class variable equal
BRINE that is classified as GAS.
Ntest (G)) - number of records in the test set with Class variable equal
GAS.
Ntest (B)) number of records in the test set with Class variable equal
BRINE.
Additional useful parameter – the probability of missed discovery is
defined as follows:
N test (G | G )) (17.23)
P .missed .discovery = 1 −
N test (G )
17.5.1 Classification with K-Nearest Neighbor

The K-nearest neighbor algorithm starts with a calculation of distances
between a classified record and records in both classes from the train-
ing set. K-nearest records are selected and the number of records from
both classes within the K-nearest set is calculated. A classified record is
assigned to the class with the largest number of records in the K-nearest
neighborhood. Optimization includes selection of the size K of the nearest
neighborhood.
According to Figure 17.4, three predictor sets (eqs. 17.18–17.20) guaran-
tee the highest probability of true discovery at small values of total uncer-
tainty. For these predictors, the probability of true discovery is larger than
0.9 in a range of total uncertainty from 0 to 0.1. For high total uncertainty,
two predictor sets (eqs. 17.18 and 17.20) guarantee that the probability of
true discovery is higher than 0.7. Both single predictors, Vp and Vp/Vs, are
inferior to a group of predictors.
According to Figure 17.5, classification with a single predictor Vp leads to
a comparatively high probability of false discovery even at small values of
total uncertainty. Three other sets of predictors give the probability of false
discovery as smaller than 0.1 in the range of total uncertainty from 0 to 0.1.
1.0
0.8
Probability
0.6 Predictors
Vp
Vp+Vs+Vp/Vs
Vp+Vs+Vp/Vs +Vp*Vs
0.4
Vp/Vs
0.1 0.2 0.3 0.4

Total uncertainty
Figure 17.4 K-nearest neighbor and the probability of true discovery.
0.4
0.3
Probability
0.2
Predictors
0.1 Vp
Vp+Vs+Vp/Vs
Vp+Vs+Vp/Vs +Vp*Vs
0.0 Vp/Vs
0.1 0.2 0.3 0.4

Total uncertainty
Figure 17.5 K-nearest neighbor and the probability of false discovery.
When total uncertainty is as high as 0.4, two sets of predictors (eqs. 17.19,
17.20) lead to the probability of false discovery being no higher than 0.3.
17.5.2 Classification with Recursive Partitioning

The recursive partitioning algorithm splits the data on one of the predictors
into two sets such that the maximum separation of two classes is achieved.
This creates a branch in the decision tree. The partitioning algorithm is
then run recursively on all the predictors. The optimization of the recur-
sive partitioning includes limiting the number of branches to avoid an over
fit of the data.
According to Figure 17.6, two predictor sets (eqs. 17.19 and 17.20)
guarantee the highest probability of true discovery at small values of total
uncertainty. For these predictors, the probability of true discovery is larger
than 0.9 in the range of total uncertainty (0, 0.1). In the case of high total
uncertainty ~ 0.4 , two predictor sets (eqs. 17.19 and 17.20) guarantee that
the probability of true discovery is around 0.7. As in the case of KNN clas-
sification, single predictors produce inferior results compared to results
with the group of predictors.
According to Figure 17.7, classification with a single predictor Vp leads
to a comparatively high probability of false discovery even when total
1.0
0.8
Probability
0.6 Predictors
Vp
Vp+Vs+Vp/Vs
Vp+Vs+Vp/Vs +Vp*Vs
0.4
Vp/Vs
0.1 0.2 0.3 0.4

Total uncertainty
Figure 17.6 Recursive partitioning and the probability of true discovery.
Probability of false discovery

0.4
0.3
Probability
0.2
Predictors
0.1 Vp
Vp+Vs+Vp/Vs
Vp+Vs+Vp/Vs +Vp*Vs
0.0 Vp/Vs
0.1 0.2 0.3 0.4

Total uncertainty
Figure 17.7 Recursive partition.

uncertainty is around zero. Three other sets of predictors give the proba-
bility of false discovery as smaller than 0.1 in a range of total uncertainty of
(0, 0.1). When total uncertainty is as high as 0.4, two sets of predictors (eqs.
17.19, 17.20) lead to a probability of false discovery of around 0.33. Vp/Vs as
a predictor gives a false discovery rate as high as 0.4.
17.5.3 Classification with Linear Discriminant Analysis

Linear Discriminant Analysis (LDA) (Figures 17.8 and 17.9) involves the
maximization of the ratio between class variance to the sum of within
classes’ variances and builds a decision region between the classes.
Maximizing this ratio leads to maximum separability between two classes.
As soon as a decision region is constructed on the training set, the method
may be applied to the test set with an undefined class variable.
According to Figure 17.8, three predictor sets (eqs. 17.18, 17.19, and
17.20) guarantee the highest probability of true discovery at small values
of total uncertainty. For these predictors, the probability of true discovery
is larger than 0.9 in a range of total uncertainty from 0 to 0.1. In case of a
high total uncertainty around 0.4, two predictor sets (eqs. 17.19 and 17.20)
guarantee the probability of true discovery to be around 0.77.
Figure 17.9 shows that classification with a single predictor Vp leads to
a comparatively high probability of false discovery even when the total
uncertainty is around zero. Three other sets of predictors give the probabil-
ity of false discovery as smaller than 0.1 in a range of total uncertainty from
0 to 0.1. When total uncertainty is as high as 0.4, two sets of predictors
(eqs. 17.19, 17.20) lead to a probability of false discovery of around 0.27.
1.0
0.8
Probability
0.6 Predictors
Vp
Vp+Vs+Vp/Vs
Vp+Vs+Vp/Vs +Vp*Vs
0.4
Vp/Vs
0.1 0.2 0.3 0.4

Total uncertainty
Figure 17.8 Linear discriminant analysis and the probability of true discovery.
0.4
0.3
Probability
0.2
Predictors
0.1 Vp
Vp+Vs+Vp/Vs
Vp+Vs+Vp/Vs +Vp*Vs
0.0 Vp/Vs
0.1 0.2 0.3 0.4

Total uncertainty
Figure 17.9 Linear discriminant analysis and the probability of false discovery.
17.5.4 Comparison of the Three Classification Techniques

Figure 17.10 shows the probability of true discovery for the three classifica-
tion techniques with predictors defined by equation 17.20. This set of pre-
dictors outperforms other predictor combinations for all three techniques.
All three methods produced excellent results in the case of small total
uncertainty. In the range of total uncertainty (0, 0.1) the probability
of true discovery is higher than 0.9 for all three methods. When total
uncertainty is as high as 0.4, the two methods, K-nearest neighbor and
linear discriminant analysis, produce a high probability of true discov-
ery of around 0.75.
1.0
0.9
Probability
0.8
0.7
KNN
RPART
LDA
0.6
0.1 0.2 0.3 0.4
Total uncertainty
Figure 17.10 The probability of true discovery and the three classification techniques
Classification model: class =Vp + Vs + Vp/Vs + Vp *Vs.
0.30
0.20
Probability
0.10
KNN
0.00 RPART
LDA
0.1 0.2 0.3 0.4

Total uncertainty
Figure 17.11 The probability of false discovery.
Classification model: predictors =Vp + Vs + Vp/Vs + Vp *Vs
According to Figure 17.11, all three methods produced a low probability of

false discovery, smaller than 0.1 if the total uncertainty does not exceed 0.1.
For a high total uncertainty in the area around 0.4, K-nearest neighbor and
linear discriminant analysis produced false discovery rates not exceeding 0.3.
17.6 Conclusions
1. The Monte-Carlo method was used to model seismic velocities
defined by a pair of linear equations with randomized coefficients.
Deterministic versions of these equations were designed by [14]
to model Vp and Vs velocities in gas or brine-filled shaly deposits
in the North Sea.
2. We introduced two levels of randomization. At the first level,
porosity and clay content were treated as random variables. At
the second level, in addition to randomized porosity and clay
content, coefficients of equations were randomized and random
errors in seismic velocities were included in the model. The effects
of errors in measured seismic velocities and random coefficients
in the equations were not separable. Therefore, we introduced a
summary measurement of randomization effects that included
the effects of all random parameters.
3. A summary attribute named ‘total uncertainty’ was constructed
to characterize the effect of random parameters and errors in
recorded seismic velocities on the efficiency of identification of

gas-filled formations.
4. The results of the Monte Carlo simulations were used to evaluate
the efficiency of identifying gas-filled formations vs. brine-filled
ones. Classifier attributes were derived from a pair of Vp and Vs
velocities and in addition to Vp and Vs included the Vp/Vs ratio and
the product of Vp and Vs.
5. Three classification techniques - K-nearest neighbor, recursive
partitioning, and linear discriminant analysis were tested using
the Monte Carlo simulated data. All three methods included uti-
lization of the training set which is used to optimize a classifica-
tion technique. The optimized technique was applied to the test
data set.
6. We tested following combinations of predictors: (Vp), (Vp/Vs), (Vp
Vs, Vp/Vs), (Vp, Vs, Vp/Vs, Vp*Vs). The best classification result for
all three methods were obtained for the (Vp Vs, Vp/Vs) and (Vp, Vs,
Vp/Vs, Vp*Vs) sets of predictors.
7. The efficiency of identifying gas-filled formations was character-
ized by two conditional probabilities - probabilities of true and
false discovery. The probability of true discovery declines while
the probability of false discovery increases with increased total
uncertainty of the model. Still, when total model uncertainty is as
high as 0.4, two techniques - KNN and LDA had probabilities of
true and false discovery around 0.75 and 0.28 respectively.
8. The analysis presented in this paper is done for the models of
shaly rocks with porosity and clay content uniformly distributed
in the intervals (01, 0.16) and (0.3, 0.4), respectively. The meth-
odology presented in the paper is applicable to other distribu-
tions of porosity and clay content and to different classification
problems. For example, it may be used to study the reliability of
different lithophacies classification.
References
1. R. Duda, E. Peter, E. Hart, and. D. Stork, Pattern Classification (2nd Edition),
pp. 637, Whiley Interscience Publication (2000).
2. L. Breiman, J. Freidman, and R. Olsen, C. Stone, Classification and Regression
Trees, p. 368, Chapman & Hall/CRC, Boca Raton (1984).
3. G.J. McLachlan, Discriminant Analysis and Statistical Pattern Recognition,
pp. 530, Wiley Interscience (2004).
4. R. Christian, and G. Casella, Introducing Monte Carlo Methods with R (Use

R!), Springer (2009).
5. O. Jones, R. Mallardet, and A. Robinson, Introduction to Scientific
Programming and Simulation Using R. pp. 453, Chapman & Hall/CRC The R
Series (2009).
6. N. Matloff, The Art of R Programming: A Tour of Statistical Software Design,
pp. 373 (2011).
7. F. Provost, and T. Fawcet, Robust classification for imprecise environments.
Mach. Learn. 203–231 pp. (1995).
8. P. Avseth, T. Mukerji, and G. Mavko. Quantitative Seismic Interpretation.
Applying Rock Physics Tools to Reduce Interpretation Risk, p. 376,
Cambridge University Press (2010).
9. R. Bachrac, On of porosity and saturation using stochastic rock physics mod-
eling. Geophysics 71, pp. 53–63, (2006).
10. A. Mewes, B. Kulessa, J.D. McKinley, and A.M. Binley et.al, Anisotropic seis-
mic inversion using a multigrid Monte Carlo approach. Geophys. J. Int. 183,
pp. 267–276 (2010).
11. R. Mireault, and L. Dean. Reservoir Engineering For Geologists, Part 8b:
Monte Carlo Simulation/Risk Assessment. Can. Soc. Petrol. Reservoir
Geologists, 35, 34–45 pp. (2008).
12. J. Schon, Physical properties of Rocks. A Workbook, Volume 8, Series J. Cubbit
(Ed.), pp. 481, Elsevier (2011).
13. L. Buryakovsky, G. Chilingar, H. Rieke, and S. Shin, Fundamentals of the
Petrophysics of Oil and Gas Reservoirs,. pp. 374, Wiley, (2012). 374 pp.
14. D. Marion, and D. Jizba, Acoustic properties and their dependence on
porosity, mineralogy, and saturation; application to field-scale measure-
ment. in Advances in Core Evaluation - Reservoir Management, Washington,
C. Chardair-Rivere, P.F. Worthington (Eds.),. Advances in core evaluation -
reservoir management. Proc. 3rd Core Analysis Symposium, Washington,
Paris. (1992).
18
Shale Mechanical Properties Influence
Factors Overview and Experimental
Investigation on Water Content Effects
Hui Li1*, Bitao Lai2 and Shuhua Lin1
1
Department of Petroleum Engineering, University of Louisiana at Lafayette,
Lafayette, LA, USA
2
Aramco Services Company, Houston Research Center, Park Row Dr,
Houston, TX, USA
Abstract
In the last decade, the unconventional resource has changed the global energy
and petroleum industry. The technological innovations of hydraulic fracturing
and horizontal drilling techniques, which require a good understanding of the
rock mechanical properties of shale, have driven the economical production of
shale reservoirs. However, the intrinsic rock-mechanical properties of shale rocks
are extremely complex; they are affected by many factors, including confining
pressure, water content, water salinity, TOC (total organic carbon), clay content,
bedding plane orientation, mineralogy, anisotropy and others. Some factors (i.e.,
water content, TOC and clay content etc.) can significantly decrease the strength
of shale rocks, and this weakening mechanism needs to be considered in hydraulic
fracture designing and wellbore stability.
This paper systematically investigates various key factors that impact the rock-
mechanical properties especially the mechanical properties of shale rocks, including
confining pressure, water content, TOC (Total organic carbon), clay content, bed-
ding plane orientation, porosity, anisotropy, and temperature effects. We experimen-
tally investigated the impact of water saturation on Barnett shale’s rock-mechanical
properties (Young’s modulus, E and Poisson’s ratio, ν and uniaxial unconfined com-
pressive strength (UCS)) using a Material Testing System (MTS-810). Experimental
results showed that: 1) water saturation has a significant impact on Young’s modulus
and UCS. Young’s modulus decreases 6.1% with an increase of a water saturation
399
(1%) of Barnett shale core corresponding to the value of dry shale’s Young’s modu-
lus; 2) Young’s modulus and UCS decrease linearly with increasing water saturation;
And 3) the relationship between water saturation and Poisson’s ratio is not obvious.
Keywords: Elastic properties, water content, shale rocks
18.1 Introduction
In the last decade, the unconventional resource has changed the global
energy and petroleum industry. According to the International Energy
Agency (IEA), the volume of unconventional gas resources is estimated at
380,000 billion cubic meters (Gm3), equivalent to about 50% of global gas
resources, and shale gas accounts for the biggest share of these resources
(Unconventional Gas, 2012) [1]. In the U.S.A., Shale gas has been produced
for many years, and accounts for about 8% of total natural gas production
(Warlick, 2006) [2]. Gas shales in the U.S.A. are predicted (AEO, 2011)
to become the source of 45% of all gas production by 2035 [3], especially
considering that other fossil energy resources constitute larger threats for
climatic shift, environmental pollution and potential risks for production/
exploration. The increasing significance of shale gas plays has led to the
need for deeper understanding of shale behaviors.
Shale is a sedimentary rock formation which is generally composed
of clay, quartz and other minerals (Survey of Energy, 2010) [4]. Shale has
extremely low permeability (ranging from 10–6 md to 10–2 md) and nano-
pores in shale matrix (Liu et al., 2012) [5]. Therefore, Shale gas cannot be
produced at economical rates without stimulation, typically hydraulic frac-
turing, and horizontal wells (Tran, 2009) [6]. Some of the most critical factors
that affect hydraulic fracture propagation, wellbore stability, and the produc-
tivity of shale gas wells are the rock-mechanical properties of shale rocks
(i.e., Young’s modulus, E, Poisson’s ratio, ν, and the strength of shale rock)
(Sharma et al., 2004) [7]. Meanwhile, mechanical properties of gas shale are
affected by many factors, including confining stress, water content, TOC
(Total organic carbon), bedding plane orientation, anisotropy, temperature
and so on. Accordingly, the development of shale reservoir should consider
the effects of these factors and shale gas development nowadays becomes
one of the thorniest technical challenges faced by the petroleum industry.
18.2 Influence Factors

Mechanical properties of gas shale are influenced by a number of factors,
including effective pressure, water content, temperature, clay content,
Shale Mechanical Properties Influence Factors 401
bedding plane orientation, total organic content and so on. Each factor is
reviewed as follows.
18.2.1 Effective Pressure

Effective pressure has a strong impact on shale’s mechanical properties,
which is defined as
Pe = Pc − nPp (18.1)
Where Pe is effective pressure, Pc is confining pressure, Pp is pore

pressure, and n is effective pressure coefficient. If n < 1, the confining
pressure effect is more important than pore pressure. If n > 1, the con-
fining pressure is less important than pore pressure. If n = 1, confining
pressure and pore pressure have equal effects (Tinni et al., 2011). Zhang
and Bentley (2000) conducted velocity measurement on sandstones, and
observed that compressional velocity increased with effective pressure,
especially at low effective pressure range (Figure 18.1). They also con-
cluded that both bulk and shear moduli increase with effective pressure,
and the explanation is the closing of cracks and pores (Toksoz et al.,
1976). Dewhurst et al., (2008) measured ultrasonic velocities of Officer
Basin shale samples and concluded that compressional- and shear-
wave velocities increased linearly with mean effective stress. Tahini and
Abousleiman (2010) conducted uniaxial compressive tests on sandstone
samples and measured the compressional- and shear-wave velocities
simultaneously; they observed that both compressional- and shear-wave
velocities increase with uniaxial stress [8–12].
In the reservoir, the rocks are subjected to a combination of stresses
(a combination of compressive, shear and tensile stresses). The combined
6.0
5.0
VP (km/s)
4.0
3.0
2.0
0 10 20 30 40 50
Effective pressure (MPa)
Figure 18.1 Vp versus effective pressure for sandstone samples [9] (Zhang and Bentley,
2000).
Rupture
line
Sheor
o
d
c
•
Tension 0 Compression
f 'c f 'c
Figure 18.2 Typical Mohr rupture diagram for concrete [13] (Wuerker, 1959).
stress effects and confinement can be treated together with Mohr’s theory
(Wuerker, 1959). Mohr’s theory of failure (Mohr’s rupture line) has been
found to be one of most useful tools in analyzing the stress of rocks (Chong
et al., 1984; LaI, 1999). Coates and Denoo (1981) applied Mohr’s Circle to
determine a sand’s ultimate strength. Zhang et al., (2006) and Islam et al.,
(2010) conducted shale wellbore stability analysis using Mohr’s theory.
Rocks will not fail if the stress is within the Mohr’s rupture line. Figure 18.2
shows the typical Mohr rupture diagram made by Wuerker (1959) [13–18].
18.2.2 Porosity
The influence of porosity on shale’s rock-mechanical properties has been
well investigated. Spikes and Dvorkin (2004) studied the relation between
Poisson’s ratio and porosity using synthetic seismic modeling. Based on
the simulation and well logging analysis, Poisson’s ratio increases with the
increment of porosity (see Figure 18.3) [19].
Kumar et al., (2012) studied porosity dependent elastic modulus of
shale by using nano-indentation technique. Based on their research,
Young’s modulus decreases with an increase of shale porosity (Figure
18.4). Furthermore, Kumar et al. (2012) concluded that the development of
porosity took place during the thermal maturation of organic matter, and
localized presence of porosity played a crucial role in mechanical proper-
ties of organics in shale [20, 21].
0.38 130
1.0 0.8
0.36 120
110
0.34
Poisson’s ratio
100
0.32 90
0.30 80
0.0 70
0.28
60
0.26 50
0.6 0.4 0.2
0.24 Clay content 40
0.00 0.05 0.10 0.15 0.20 0.25 0.30 GR
Porosity
Figure 18.3 Poisson’s ratio versus total porosity [19] (Spikes and Dvorkin, 2004).
100 Woodford
Ei = 67–3.8*φ Barnett
80 R2 = 0.4
Haynesville
60 Eagle ford
Ei, GPa
40
20
0
0 5 10 15
φ, %
Figure 18.4 The relationship of Young’s modulus and porosity [20] (Kumar et al., 2012).
18.2.3 Water Content

Water content weakening rock strength has been studied for years. Several
mechanisms are proposed to explain the water weakening effects. Meng
et al. (2005) proposed that with the increase of water content, the activity
of water molecules increases, thus the pore pressure increases and effec-
tive stress decreases. As a result, the elastic yield limit of the rock reduces,
and the shear strength decreases, which might lead to a shear failure.
Hadizadeh and Law (1991) proposed the stress corrosion effects based
on the way highly stressed silicon-oxygen bonds at micro-crack tips may
be hydrolyzed and weakened, permitting the crack to propagate. Delage
et al., (1996) and Lord et al. (2005) believed that capillary pressure
decrease is the reason for water weakening effects. The water menisci will
pull the grains together and thereby contribute to cohesion. Increasing
water content will make the menisci become weaker or disappear, and
the rock will become weaker [22–25].
Van Eeckhout (1976) studied the effect of humidity on the strength and
elastic properties of a few samples obtained from coal mine shale. Their
experimental work showed that with an increase of moisture content (rel-
ative humilities of 0.48 and 100%), there was a reduction in compressive
strength and Young’s modulus, and an increase in Poisson’s ratio [26].
Lashkaripour and Passaris (1993) compiled a database with selected
values of shale rock properties and concluded that there was also a sig-
nificant negative correlation between water content and compressive
strength in coal mines shale. Uniaxial tests were widely used to study
the weakening effect of water content on shale’s rock-mechanical proper-
ties. Steiger and Leung (1989) performed uniaxial compressive strength
(UCS) tests on three categories of shale rocks under both dry and wet
conditions. The testing results demonstrated that a dryer sample (lower
water content) has a higher unconfined compressive strength. Table 18.1
lists the uniaxial compressive strengths measured with both wet and dry
samples. As Table 18.1 shows, the uniaxial compressive strengths mea-
sured on dry samples can be 2 to 10 times larger than that of the wet
samples [27, 28].
The elastic properties changed remarkably during desaturation and
resaturation tests (Pham et al., 2005). Pham et al. (2005) carried out
desaturation and resaturation uniaxial compressive tests on Bure shale to
observe the influence of water content or RH (relative humidity) on elastic
properties of shale specimens. The UCS and Young’s modulus increased
when the specimens were desaturated and decreased during resaturation
(Figure 18.5 (a) and (b)) [29].
Lin and Lai (2013) conducted laboratory tests on water satura-
tion effects on Barnett shale’s geomechanical properties using uniaxial
strength testing method.
The results showed that Young’s modulus and unconfining compressive
strength decreased linearly with increasing water saturation. Poisson’s ratio
showed insensitivity to the change of water content [30].
Table 18.1 Uniaxial compressive strength of typical shales [28] (Steiger and
Leung, 1989).
Sample condition Unconfined compressive strength (1000 psi)
Shale E Shale F Shale G
Wet 6.5 4 0.5
Dry 14 12 5
12000
11000
The Young modulus (MPa)

10000
9000
8000
7000
6000
5000
4000
30 40 50 60 70 80 90 100
(a) Relative humidity (%)
60
55
Compressive strength (MPa)
50
45
40
35
30
25
30 40 50 60 70 80 90 100
(b) Relative humidity (%)
Figure 18.5 (a) The variation of Young’s modulus with RH [29] (Pham et al., 2005) and
(b) The variation compressive strength with RH [29] (Pham et al., 2005).
18.2.4 Salt Solutions

Water salinity plays an important role in shale mechanical and petro-
physical properties. Shale fabric structure is altered after being exposed
to salt solutions (Zhang et al., 2006 and Gomez et al., 2014). The influ-
ence of different salt types on shale’s mechanical properties has been stud-
ied by many researchers. Zhang et al., (2006) conducted a biaxial test to
study the effects of different types of salt solutions (KCl, NaCl, CaCl2) on
shale strength using Pierre I shale samples and Arco shale samples, both
of which are rich of clay content (about 64%). For Pierre I, Zhang et al.,
(2006) found that the compressive strength of shale increased after inter-
action with ionic solutions. The higher the concentration, the higher is the
compressive strength. Among the three salts, the strength enhancement

for KCl is the largest. However, for Arco shale samples, Zhang et al. (2006)
found that compressive strength decreased after exposure to both NaCl
and CaCl2 solutions, but increased when exposure to KCl solutions. Akrad
et al. (2011) revealed that KCl solution had little effect on the change of
Young’s Modulus of Middle Bakken formation which is a clay-poor forma-
tion, but could reduce the decline of the Young’s Modulus of Barnet forma-
tion (rich in water-sensitive clays) comparing with fresh water. Therefore,
we believe that the complexity of the water salinity effects is likely due to
mineralogical composition especially the water-sensitive clay content of
the formation [31–33].
18.2.5 Total Organic Carbon (TOC)

The organic matter in shales has the features of softness, high resistivity
and having non-zero shear modulus. The TOC content has significant
weakening effects on the geomechanical behavior of shales (Vernik and
Liu 1997; Quirein et al., 2012 and Murphy et al., 2013) [34–36].
Vernik and Liu (1997) used ultrasonic velocity to investigate the TOC
effects, clay content, and shale’s anisotropy under a wide range of effective
pressure conditions. The experiment results showed that the elastic anisot-
ropy of shale increases significantly with compaction [34]. Additionally,
the matrix anisotropy of shale increased dramatically with kerogen con-
tent. Gupta et al., (2012) cross-plotted ultrasonic velocities and elastic
moduli with TOC for some Woodford shale samples. Figure 18.6 indicates
that TOC varies inversely with both Vp and Vs, and Young’s modulus (E) is
inversely proportional with TOC for E<60 GPa and TOC>4% [37].
Bandyopadhyay et al., (2012) used a theoretical rock-physics modeling
framework to study the influence of TOC on dynamic rock properties.
(a) (b) (c)

6 90
Young’s modulus (GPa)
3.5 80
VP –fast (km/s)
VS –fast (km/s)
5.5
70
5 3
60
4.5 50
2.5
4 40
30
0 2 4 6 8 10 12 0 2 4 6 8 10 12 0 2 4 6 8 10 12
TOC (%) TOC (%) TOC (%)
Figure 18.6 The relationship of Vp-fast & Vs-fast and Young’s modulus with TOC [37]
(Gupta et al., 2012).
100 Woodford
y = 58.5–3.3* TOC
R2 = 0.48 Barnett
80
Haynesville
60 Eagle Ford
Ei, GPa
40
20
0
0 5 10 15
TOC, wt%
Figure 18.7 The relationship of Young’s modulus and TOC [21] (Kumar et al., 2012).
Based on the experiment results, TOC reduced both AI (acoustic imped-

ance) and Vp/Vs, and caused elastic and resistivity anisotropy [38]. Kumar
et al. (2012) studied the effect of TOC on mechanical properties of shale
using nano-indentation technique. According to the study of Kumar et al.
(2012), Young’s modulus decreased substantially with the increasing TOC
(Figure 18.7). Futher, the mechanical properties of kerogen also related to
the maturation [20, 21].
Zargari et al. (2011) studied the effects of maturity on elastic properties
of organic rich shales. Ten Bakken shale samples were selected and synthet-
ically matured through hydrous pyrolysis technique. Zargari et al. (2011)
found out that naturally mature samples showed an increase in modulus
with maturity, while the change in the modulus of clay and kerogen of the
samples subject to pyrolysis depended on the initial maturity [39].
18.2.6 Clay Content

Clay is a key factor that influences the rock mechanical properties. Clay-
rich rocks display ductile mechanical properties and clay deficient rocks
have a more brittle character (Yao et al., 2010) [40]. Kumar et al., (2012)
applied Fourier Transform Infrared (FTIR) technique to analyse clay con-
tent of shale rocks and found that nano-indentation-based Young’s mod-
ulus decreases as total clay content increases (Figure 18.8) [21]. Sone and
Zoback (2011) conducted uniaxial compressive tests on shale specimens
and used X-ray diffraction (XRD) to measure clay content of the samples.
The results showed that the UCS (uniaxial compressive strength) decreased
significantly with increasing of clay content (see Figure 18.9) [41]. Similar
results have been found by Akrad et al., (2011) [33].
100 Woodford
90 Barnett
y = 63–0.45* clay
80 R2 = 0.47 Haynesville
70 Ordovician
60 Eagle ford
Ei, GPa
50
40
30
20
10
0
0 20 40 60 80 100
Clay, wt%
Figure 18.8 The relationship of Young’s modulus and clay content [21] (Kumar et al.,
2012).
250
200
UCS (MPa)
150
100 Barnett dark

Barnett light
Haynesville dark
50 Haynesville light
Fort St. John
0
0 10 20 30 40 50
Clay content (%)
Figure 18.9 The influence of clay content on UCS [41] (Sone and Zoback, 2011).
18.2.7 Bedding Plane Orientation

Shales are highly laminated and anisotropic rocks. The bedding plane orien-
tation can significantly impact the anisotropic behavior as well as geo- and
rock-mechanical properties. Willson et al., (2007) conducted uniaxial tests
and concluded that formation unconfined compressive strength was high-
est at orientations perpendicular to bedding (bedding plane angle equals
to 0°), lowest when bedding plane angle was 60°, and had an intermediate
value when bedding plane angle was 90° (see Figure 18.10) [42].
Abousleiman et al., (2008) performed direct shear tests on cores with
different bedding planes (perpendicular to bedding and parallel to bed-
ding). The experimental results are shown in Table 18.2 and Figure 18.11.
1.2
1
normalized to that measured
perpendicular to bedding
0.8
0.6
0.4
0.2
0
0 10 20 30 40 50 60 70 80 90
Sample orientation to bedding (0 = perpendicular, 90 = parallel)
Tournemire27 Siltshale28 Mudshale28
2 2
HRZ shale Kingak shale Pierre 1 shale
Figure 18.10 Experimental data showing strength anisotropy in shales [42] (Willson
et al., 2007).
Table 18.2 Summary of difference in strength for the upper &

middle woodford shale samples [43] (Abousleiman et al., 2008).
Conf. pressure Strength difference
Interval (MPa) (Horizontal-Vertical)
Upper Woodford 6.89 -7%
13.79 4%
Middle Woodford 6.89 12%
13.79 -8%
Due to the limited number of samples, the authors claimed that the effects
of bedding plane to the strength was unclear and needed more experimen-
tal supportive results [43].
Wu and Tan (2010) measured acoustic velocity on shale samples with
various bedding planes. They found that the Young’s modulus from paral-
lel to the bedding planes was about 37% higher than that perpendicular to
the bedding planes. According to the triaxial stress experiment conducted
by Li et al., (2012), the core samples perpendicular to the bedding plane
have the highest strength, then the samples paralleling bedding plane, and
100
Middle-vertical
Middle-horizontal
90 Upper-vertical
Deviatoric strength (MPa) Upper-horizontal
80
70
60
50
40
6 7 8 9 10 11 12 13 14
Figure 18.11 Strength of horizontally and vertically cored samples from Upper and
Middle Woodford [43] (Abousleiman et al., 2008).
the strength of the samples in the bedding plane orientation 60° is lowest
(see Figure 18.12) [44, 45].
Alqahtani et al. (2013) conducted triaxial tests and ultrasonic velocity
tests to study the effects of bedding plane orientation (0°, 45° and 90°) on
Green River, Eagle Ford and Mancos shales. The results demonstrated that
the maximum compressive strength, static elastic moduli, and compressive
velocity occurred at horizontal samples (see Figure 18.13) [46].
250
a
Compressive strength, MPa
200
a
150
100
50
0
0 20 40 60 80 100
Angle of coring and bedding planes normal, deg
Figure 18.12 The strength of shale core in different direction [45] (Li et al., 2012).
25000
20000
Compressive strength (psi)
15000
Vertical
45
10000 Horizontal
5000
0
Green river Mancos Eagle Ford, La Salle
Figure 18.13 Effect of the bedding angle on compressive failure strength [46] (Alqahtani
et al., 2013).
18.2.8 Mineralogy
Mineralogy is another key factor that affects shale mechanical properties.
To examine mineral compositions effects on elastic moduli, Marouby and
Heidari (2011) used a self-consistent approximation (Voigt-Reuss-Hill’s
approximation) method to calculate shale matrix bulk (Kmat) and shear
moduli (μmat) based on the interpretation of sonic, nuclear, and resistivity
logs. Table 18.3 shows different mineral constituents of three-layer shale
formation (Cq, Ccal, Cill, Cpy and Cker respectively denotes quartz, calcite,
Table 18.3 Different mineral constituents of three-layer formation [47]

(Marouby and Heidari, 2011).
Layer thickess (ft) Cq Ccal Cill Cpy Cker Sw φ
15 0.65 0.14 0.14 0.03 0.04 0.3 0.06
15 0.14 0.14 0.65 0.03 0.04 0.3 0.06
15 0.14 0.65 0.14 0.03 0.04 0.3 0.06
0 0
5 5
10 10
15 15
Depth (ft)
20 20
25 25
30 30
35 35
40 40
45 45
13 14 15 16 8 9 10
Ksat (GPa) µsat (GPa)
Figure 18.14 The bulk and shear moduli with different mineral concentrations of 3 layer
formation [47] (Marouby and Heidari, 2011).
100
y = 26.3 + 0.4* QC
80 R2 = 0.48
60
Ei, GPa
40
20
0
0 20 40 60 80 100
Quartz plus carbonate, wt%
Figure 18.15 The dependence of Young’s modulus on quartz and carbonate [20] (Kumar
et al., 2012).
illite, pyrite and kerogen concentration), and Figure 18.14 shows that dif-
ferent mineral concentrations lead to different bulk and shear moduli.
Based on the study of Marouby and Heidari (2011), mineral constituent is
a critical factor that infuences elastic properties of shale [47].
Akrad et al. (2011) using nano-indentation observed that elastic mod-

ulus decreased more significantly for calcite-rich shale rock samples than
quartz-rich shale [33]. Kumar et al., (2012) used the same technique to
study the effect of carbonate and quartz content on mechanical properties
of shale and concluded that an increase in carbonate and quartz content
results in an increase of Young’s modulus (Figure 18.15) [20].
18.2.9 Anisotropy
Shale generally is a highly anisotropic laminated sedimentary rock. The
degree of anisotropy of shale depends on many factors such as kerogen con-
tent (Vernik and Landis, 1996), TOC and clay content (Kumar et al., 2012).
The larger the kerogen, TOC and clay content, the higher is the anisotropy
of the shale (Vernik and Nur, 1992; Sondergeld et al., 2000; Sondergeld and
Rai, 2011; Kumar et al., 2012 and Zhu et al., 2012) [20, 48–52].
Abousleiman et al., (2008) studied the relation of elastic strength and
anisotropy of Woodford shale samples using an inclined direct shear test-
ing device (IDSTD). According to the study of Abousleiman et al., (2008),
anisotropy influences the elastic properties of shale specimens. But the
relationship of Woodshale strength and anisotropy is not yet clear due
to the limited number of tests. Dewhurst et al., (2008) conducted triaxial
tests using North Sea Shale samples and measured ultrasonic P-wave and
S-wave. They concluded that the shale speciments exhibited slightly larger
P-wave anisotropy and moderate S-wave anisotropy [11, 43].
18.2.10 Temperature
Temperature also can influence geomechanical properties of shale. High
temperatures will lead to kerogen decomposition, thereby weakening the
shale strength. Handin and Hager (1958) performed short-time triaxial
compression tests on shale samples to study the deformation of rocks at
elevated temperature conditions. The experimental results show that an
increase of temperature at constant pressure reduces the yield stress and
raises the ultimate strength by increasing the ductility of shale. Closmann
and Bradley (1979) investigated the temperature effects on mechanical
properties of oil shale. The testing results demonstrated that with tem-
perature increasing, Young’s modulus, tensile strength and compressive
strength all decrease. Similarly, Akrad et al., (2011) used nano-indentation
technique and found out that the Young’s modulus of shale decreased
significantly when they were heated to high temperatures (~300 °F) [33,
53, 54].
18.3 Experimental Investigation of Water Saturation

Effects on Shale’s Mechanical Properties
18.3.1 Experiment Description
In this work, we performed an experimental study to investigate the effects
of water saturation on shale mechanical properties. Nineteen Barnett shale
cores with geometry of 1 inch diameter and 2 inch length were partially
saturated with KCl solution. The core samples were saturated by the hour
(2 hours, 4 hours, 6 hours, 8 hours, 10 hours, and more) to let the cores
reach different water saturation. Uniaxial compressive tests were then per-
formed on these samples at room temperature.
The Barnett shale cores in this work were extracted from a depth of
7,500 ft in the Fort Worth and Permian Basin in Texas. The total organic
carbon (TOC) content by weight in the Barnett Shale is reported to average
3.3–4.5%. The average density is 2.376 g/cm3. ASTM standard on uniaxial
compressive strength testing was employed in this experimental work.
18.3.2 Results and Discussion

Nineteen (19) Barnett shale samples have been saturated to different water
saturations and tested in this work. Young’s modulus was determined
according to Average modulus, and Poisson’s ratio was calculated at 50%
UCS. The results are shown in Table 18.4. It can be seen that the water satu-
ration of the core samples was varied from 0% (totally dry) to 35%. Results
of Young’s modulus, Poisson’s ratio and UCS are listed in the table.
Based on the data from Table 18.4, we plot the relationship between
water saturation and Young’s modulus (see Figure 18.16). From this figure
it can be seen that Young’s modulus decreases with an increase of water
saturation. Curve fitting shows that the relationship between Young’s mod-
ulus and water saturation is linear and can be demonstrated by Equation
18.1. The matching coefficient is 0.69.
Young’s modulus decreases 6.1% with an increase of a water saturation
(1%) of Barnett shale core corresponding to the value of dry shale’s Young’s
modulus.
E = −22820 · Sw + 3758859 (18.1)
Plot Poisson’s ratio vs. water saturn Figure 18.17. It can be seen that
partially saturated Barnett shale’s Poisson’s ratio fluctuates around the val-
ues of the dry samples. In other words, there is no obvious relationship
between water saturation and Poisson’s ratio.
Table 18.4 The result of shale cores from uniaxial tests.
φ Sw E ν UCS
Test # % % (10 ) psi
6
psi
1 0 3.88 0.224 4,528
2 15.5 32.2 2.4 0.287 3,159
3 15.1 20.9 3.58 0.45 3,989
4 13.8 2.4 3.64 4,193
5 14.3 29.9 3.11 0.092 2,939
6 14.0 17.8 3.4 0.311 3,433
7 15.0 7.4 3.59 3,872
8 0 3.64 4,429
9 13.9 35.1 2.87 0.21 2,965
10 14.4 33.6 2.72 0.725 2,793
11 14.3 9.3 2.73 2,903
12 14.5 11.3 3.49 4,232
13 14.6 10.9 3.52 0.278 3,960
14 15.4 22.2 3.51 0.277 4,463
15 15.3 8.4 3.37 3,444
16 16.0 8.6 3.33 0.04 3,871
17 15.2 12.8 3.73 0.109 4,077
18 15.3 24.7 3.14 0.107 2,760
19 16.3 13.2 2.84 1,906
Where φ is porosity; Sw is water saturation; E is Young’s modulus, psi; ν is
Poisson’s ratio; and UCS is uniaxial compressive strength, psi.
Figure 18.18 demonstrates that the relationship between Uniaxial

Compressive Strength (UCS) and water saturation is linear. The UCS
decreases with the increase of the water saturation. Equation 18.2 indicates
the relationship between UCS and water saturation. The matching coeffi-
cient is 0.7627.
4000000
3800000 Young’s modulus vs. Sw
3600000
Young’s modulus (psi)
3400000
3200000
3000000
2800000 y = –22819.98x + 3758858.52
R2 = 0.69
2600000
2400000 Young’s modulus
2200000 Fitting line of Young’s modulus
2000000
0 10 20 30 40
Water saturation (%)
Figure 8.16 The relationship between Sw and Young’s modulus.
0.8
Poisson’s ratio vs. Sw
0.7
0.6
Poisson’s ratio
Poisson’s ratio
0.5
Fitting line of Poisson’s ratio
0.4
0.3
0.2
y = 0.0012x + 0.2166
0.1
R2 = 0.0203
0
0 10 20 30 40
Figure 18.17 The relationship between Sw and Poisson’s ratio.
UCS = − 43 · Sw + 4357 (18.2)
Previous work and our experimental results show that induced water
tends to decrease the strength and Young’s modulus. Shales are rich in clay
minerals, and in the presence of water, clays with expandable minerals
such as smectites will expand and make shale softer and weaker.
5000
UCS vs. Sw
4500
4000
3500
UCS (psi)
3000
2500
y = –43.252x + 4357.2
2000 UCS
Fitting line of UCS R2 = 0.7627
1500
0 10 20 30 40
Figure 18.18 The relationship between Sw and UCS.
18.3.3 Error Analysis of Experiments

Take the data of Young’s modulus with saturation as an example of how
an error analysis is performed (Mills and Chang, 2004) [55]. Assume the
measured water saturation data is xi, the measured Young’s modulus data
is yi, and the fitting value of Young’s modulus is Yi. To find the “best” line,
we need to calculate the slope m and the intercept C (Equation 18.3 &
Equation 18.4). Next, the standard error Sy for the curve-fit is calculated
(Equation 18.5 & Equation 18.6). Repeating the same calculation, the result
of the values of m, C and Sy for Young’s modulus and UCS are shown in
Table 18.5 and Figure 18.19. Figure 18.19 indicates difference between the
best line and the fitting line of Young’s modulus and UCS.
Table 18.5 The result of error analysis.

Parameters Young’s modulus UCS
m -24,359.42 -35.01
C 3,674,567.36 4,128.51
Sy 519,658.53 890.88
4000000 5000
Young’s modulus vs. Sw UCS vs. Sw
4500
3500000 y = –22820x + 4E+06

2
R = 0.6939 4000 y = –43.252x + 4357.2
2
R = 0.7627
3500
UCS (psi)
3000000
3000
Young’s modulus
2500 UCS
2500000 The best line The best line
Fitting line of Young’s modulus 2000
Fitting line of UCS
2000000 1500
0 10 20 30 40 50 0 10 20 30 40 50
(a) Water saturation (%) ( b) Water saturation (%)
Figure 18.19 (a) and (b) The fitting line and the “best” line for E and UCS.
m=
∑ x y − xy i i
∑ x − (∑ x ) / N
2 (18.3)
2
i

C = y − mx (18.4)

∑ D = ∑( y − Y )
i
2
i i
2
(18.5)

Sy =
1
N −2 ∑D i
2

(18.6)
Where xi is the water saturation value, yi is the measured Young’s modu-

lus, Yi is the fitting Young’s modulus, x is the average water saturation data,
y is the average measure Young’s modulus, N is the number of testing
points, m is the “best” line, C is the intercept of the “best” line, and Sy is the
standard error for the curve-ft.
To better understand the result of the experiment, we plot the upper and
lower limit of the data for Young’s modulus and UCS, shown in Figures
18.20 and 18.21. Also, the range of each point is plotted in the figures.
18.4 Conclusions
1.
Shale is a highly laminated anisotropic sedimentary rock. The
mechanical properties of shale are affected by many factors. After a
thorough review of key factors that affect shale’s mechanical proper-
ties, the following conclusions can be drawn.
4500000 Young’s modulus
2800000 3200000 3600000 4000000

Young’s modulus vs. Sw
4000000

3500000
3000000
2500000 y = –22819.98x + 3758858.52

R2 = 0.69
2000000
0 10 20 30 40 50 0 5 10 15 20 25 30 35
(a) Water saturation (%) (b) Water saturation (%)
Figure 18.20 (a) and (b) The data range analysis for E.
UCS
2500 3000 3500 4000 4500 5000
5000
UCS vs. Sw
4500
4000
UCS (psi)
3500
UCS (psi)
3000
2500
y = –43.252x + 4357.2
2000 R2 = 0.7627
1500
0 10 20 30 40 50 0 5 10 15 20 25 30 35
(a) Water saturation (%) (b) Water saturation (%)
Figure 18.21 (a) and (b) The data range analysis for UCS.
1. The compressive strength and Young’s modulus of shale

increase with the increasing confinement pressure. Poisson’s
ratio decreases slightly with increasing confining stress.
2. Porosity, water content, and temperature have similar effects
on shale. The UCS and Young’s modulus of shales decrease
remarkably with the increasing porosity, water content, and
temperature. Salt solutions have diverse influences on elastic
properties of shales. It appears that different salt types can
have very different effects on shale’s mechanical properties.
3. TOC is a crucial factor, which could increase anisotropy and
is inversely proportional to AI, Vp, Vs and Young’s modulus
of shale rocks. Shales with higher clay content are weaker
and the higher the clay content of the shale, the lower the
Young’s modulus and strength.
4. Bedding plane orientation can highly impact the mechanical

properties of gas shale. However, further study is needed to
draw from conclusions. Mineralogy also affects the elastic
moduli of shales; an increase in carbonate and quartz con-
tent tends to significantly increase Young’s modulus.
2. Nineteen Barnett shale samples with different water contents

(assumed proportional to water saturation) were tested and Young’s
modulus, Poisson’s ratio and UCS determined. Young’s modulus
and UCS decrease with increasing water content, and the relation-
ship between water content and Young’s modulus and UCS is lin-
ear. The relationship between the water content and Poisson’s ratio
is not obvious. The water content appears to have little effect on
Poisson’s ratio.
Specifc conclusions are:
1. The relationship between water saturation and Young’s mod-

ulus can be given as E = –22,820Sw + 3,758,859, and the data
matching coefficient is 0.69. Young’s modulus decreases
6.1% from the dry shale Young’s modulus with an increase
of 1% water saturation for the Barnett shale core samples.
2. The relationship between the water saturation and Poisson’s
ratio is not obvious. The water saturation has little effect on
Poisson’s ratio.
3. The relationship between water saturation and UCS is lin-
ear as UCS = –43Sw + 4,357. The data matching coefficient is
0.7627. UCS decreases with the increase of water saturation.
Acknowledgements
The authors are grateful to Aramco Service Company for the support and
the permission to publish this paper. We thank the Petroleum Engineering
Department of ULL for the support of this study.
References
1. Total: Unconventional Gas. Unconventional Gas Exploration & Production,
Paris, France (2012).
2. D. Warlick, Gas shale and CBM development in North America. Oil Gas
Financial Journal. 3(11), 1–5 (2006).
3. Annual Energy Outlook (AEO). National Energy Modeling System, Run Ref,
D120810C (2011).
4. Survey of Energy Resources: Focus on Shale Gas, Report of World Energy
Council, London, United Kingdom (2010).
5. C. L. Liu, R. Sharma, and A.Tutuncu, Shale Gas Resources: Energy Potential and
Associated Exploitation Challenges for Coupled Geomechanics and Transport
Characteristics, 46th U.S. Rock Mechanics/Geomechanics Symposium, Chicago,
Illinois, June 24–27, Paper No. ARMA-2012-209 (2012).
6. M. H. Tran, Geomechanics Field and Laboratory Characterization of
Woodford Shale, A thesis for the degree of Master of Science. University of
Oklahoma, Norman, Oklahoma (2009).
7. M. M. Sharma, P. B. Gadde, R. Sullivan, R. Sigal, R. Fielder, D. Copeland, L.
Griffn, and L. Weijers, Slick Water and Hybrid Fracs in the Bossier: Some
Lessons Learnt, SPE Annual Technical Conference and Exhibition, Houston,
TX, USA, September 26–29, Paper No. SPE 89876 (2004).
8. A. Tinni, C. H. Sondergeld, C. S. Rai, and H. Kouam, Effective Pressure
and Microstructure Control on Resistivity Formation Factor and Seismic
Waves Velocities, SPE Annual Technical Conference and Exhibition, Denver,
Colorado, USA, 30 October-2 November, Paper No. SPE-147432-MS
(2011).
9. J. J. Zhang and L. R. Bentley, Change of Elastic Moduli of Dry Sandstone
with Effective Pressure, SEG Annual Meeting, August 6–11, Calgary, Alberta,
Paper No. SEG-2000-1826 (2000).
10. M. N. Toksoz, C. H. Cheng, and A. Timur, Velocities of seismic waves in
porous rocks. Geophysics 41, 621–645 (1976).
11. D. N. Dewhurst, A. F. Siggins, U. Kuila, M. B. Clennell, M. D. Raven, and
H. M. Nordgard-Bolas, Elastic, Geomechanical and Petrophysical Properties
of Shales, 42nd US Rock Mechanics Symposium and 2nd U.S.-Canada Rock
Mechanics Symposium, San Francisco, USA, June 29- July 2, Paper No. ARME
08-208 (2008).
12. A. M. Al-Tahini and Y. N. Abousleiman, Pore-pressure-coeffcient anisot-
ropy measurements for intrinsic and induced anisotropy in sandstone. SPE
Reservoir Evaluation & Engineering 13(2), 265–274 (2010).
13. R. G. Wuerker, Infuence of Stress Rate and Other Factors on Strength and
Elastic Properties of Rocks, 3rd U.S. Symposium on Rock Mechanics (USRMS),
Golden, CO, April 20–22, Paper No. ARMA-59-0032 (1959).
14. K. Chong, J. Chen, G. Dana, and S. Sailor, Triaxial testing of Devonian oil
shale. J. Geotech. Engrg. 110(10), 1491–1497 (1984).
15. M. LaI, Shale Stability: Drilling Fluid Interaction and Shale Strength, SPE
Latin American and Caribbean Petroleum Engineering Conference, Caracas,
Venezuela, April 21–23, Paper No. SPE 54356 (1999).
16. G. R. Coates and S. A. Denoo, Mechanical Properties Program Using Borehole

Analysis and Mohr’s Circle, SPWLA 22nd Annual Logging Symposium,
Mexico City, Mexico, June 23–26, Paper No. SPWLA-1981-DD (1981).
17. J. G. Zhang, A. Talal, C. Martin, S. Mukul, C. David, and B. Saddok,
Compressive Strength and Acoustic Properties Changes in Shale with
Exposure to Water-Based Fluids, 41st U.S. Symposium on Rock Mechanics
(USRMS), Golden, CO, June 17–21, Paper No. ARMA-06-900 (2006).
18. M. A. Islam, P. Skalle, A. M. Al-Ajmi, and O. K. Soreide, Stability Analysis
in Shale Through Deviated Boreholes Using the Mohr And Mogi-Coulomb
Failure Criteria, 44th U.S. Rock Mechanics Symposium and 5th U.S.-Canada
Rock Mechanics Symposium, Salt Lake City, Utah, June 27–30, Paper No.
ARMA-10-432 (2010).
19. K. T. Spikes and J. P. Dvorkin, Reservoir and Elastic Property Prediction
Away from Well Control, Offshore Technology Conference, Houston, Texas,
May 3–6, Paper No. OTC-16922-MS (2004).
20. V. Kumar, C. H. Sondergeld, and C. S. Rai, Nano to Macro Mechanical
Characterization of Shale, SPE Annual Technical Conference and Exhibition,
San Antonio, Texas, USA, October 8–10, Paper No. SPE-159804-MS (2012).
21. V. Kumar, M. E. Curtis, N. Gupta, C. H. Sondergeld, and C. S. Rai, Estimation
of Elastic Properties of Organic Matter in Woodford Shale through
Nanoindentation Measurements, SPE Canadian Unconventional Resources
Conference, Calgary, Alberta, Canada, October 30–November 1, Paper No.
SPE-162778-MS (2012).
22. Z. P. Meng, Y. Wu, and J. Tiedemann, Analysis of Mechanical Properties of
Sedimentary Rocks of Coal Measures and Their Influencing Factors, 40th
U.S. Symposium on Rock Mechanics (USRMS), Anchorage, AK, June 25–29,
Paper No. ARMA 05-687 (2005).
23. J. Hadizadeh, and R. D. Law, Water weakening of sandstone and quartzite
deformed at various stress and strain rates. Int. J. Rock Mech. Min. Sci. 28,
431–440 (1991).
24. P. Delage, Y.J. Cui, and C. Schroeder, Subsidence and Capillary Effects in
Chalks, ISRM International Symposium, Turin, Italy, September 2–5, Paper
No. ISRM-EUROCK-1996-160 (1996).
25. D. L. Lord, A. H. Demond, and K. F. Hayes, Comparison of capillary pressure
relationships of organic liquid-water systems containing an organic acid or
base. J. Contam. Hydrol. 77(3), 195–208 (2005).
16. E. M. Van Eeckhout, The mechanisms of strength reduction due to moisture
in coal mine shale’s. Int. J. Rock Mech. Min. 13, 61–67 (1976).
27. G. R. Lashkaripour and E. K. Passaris, Correlations between Index Parameters
and Mechanical Properties of Shale, 8th International Conferfence on Rock
Mech, Tokyo, 1, 257–261 (1993).
28. R. P. Steiger and P. K. Leung, Lecture: Predictions of Wellbore Stability in
Shale Formations at Great Depth, ISRM International Symposium, Pau,
France, August 30-September 2, Paper No. ISRM-IS-1989-148 (1989).
29. Q. T. Pham, F. Vales, L. Malinsky, D. N. Minh, and H. Gharbi, Effect of

Desaturation-Resaturation on Mechanical Behaviour of Shale, 40th U.S.
Symposium on Rock Mechanics (USRMS), Anchorage, AK, June 25–29, Paper
No. ARMA-05-793 (2005).
30. S. H. Lin and B. T. Lai, Experimental Investigation of Water Saturation
Effects on Barnett Shale’s Geomechanical Behaviors, SPE Annual Technical
Conference and Exhibition, New Orleans, Louisiana, USA, September 30 -
October 2, Paper No. SPE-166234-MS (2013).
31. J. G. Zhang, A. Talal, C. Martin, S. Mukul, C. David and B. Saddok,
Compressive Strength and Acoustic Properties Changes in Shale with
Exposure to Water-Based Fluids, 41st U.S. Symposium on Rock Mechanics
(USRMS), Golden, CO, June 17–21, Paper No. ARMA-06-900 (2006).
32. S. L. Gomez, R. Leonard, D. Li and W. He, Shale-Fluid Interactions and
Drilling Fluid Designs, International Petroleum Technology Conference,
Doha, Qatar, January 19–22, Paper No. IPTC-17235-MS (2014).
33. O. M. Akrad, J. L. Miskimins, and M. Prasad, The Effects of Fracturing Fluids
on Shale Rock Mechanical Properties and Proppant Embedment, SPE Annual
Technical Conference and Exhibition, Denver, Colorado, USA, October 30 -
November 2, Paper No. SPE-146658-MS (2011).
34. L. Vernik and X. Liu, Velocity Anisotropy in shales: A petrophysical study.
Geophysics 62, 521–532 (1997).
35. J. A. Quirein, E. E. Murphy, G. Praznik, J. M. Witkowsky, S. Shannon, and D.
Buller, A Comparison of Core and Well Log Data to Evaluate Porosity, TOC,
and Hydrocarbon Volume in the Eagle Ford Shale, SPE Annual Technical
Conference and Exhibition, San Antonio, Texas, USA, October 8–10, Paper
No. SPE-159904-MS (2012).
36. E. E. Murphy, G. Praznik, J. Quirein, E. J. Galford, J. M. Witkowsky, and S.
Chen, A Workflow to Evaluate Mineralogy, Porosity, TOC, and Hydrocarbon
Volume in the Eagle Ford Shale, SPE Unconventional Resources Conference
and Exhibition-Asia Pacific, Brisbane, Australia, November 11–13, Paper No.
SPE-167012-MS (2013).
37. N. Gupta, C. S. Rai, and C. H. Sondergeld, Integrated Petrophysical
Characterization of the Woodford Shale in Oklahoma, SPWLA 53rd Annual
Logging Symposium, Cartagena, Colombia, June 16–20, Paper No. SPWLA-
2012-250 (2012).
38. K. Bandyopadhyay, R. Sain, E. Liu, C. Harris, A. Martinez, M. Payne, and
Y. P. Zhu, Rock Property Inversion in Organic-Rich Shale: Uncertainties,
Ambiguities, and Pitfalls, SEG Annual Meeting, Las Vegas, Nevada, November
4–9, Paper No. SEG-2012-0932 (2012).
39. S. Zargari, M. Prasad, K. C. Mba, and E. Mattson, Organic Maturity, Hydrous
Pyrolysis and Elastic Property in Shales, Canadian Unconventional Resources
Conference, Alberta, Canada, November 15–17, Paper No. SPE-149403-MS
(2011).
40. Y. Yao , S. V. Gosavi, K. H. Searles, and T. K. Ellison, Cohesive Fracture

Mechanics Based Analysis to Model Ductile Rock Fracture, 44th U.S. Rock
Mechanics Symposium and 5th U.S.-Canada Rock Mechanics Symposium, Salt
Lake City, Utah, June 27–30, Paper No. ARMA-10-140 (2010).
41. H. Sone and M. D. Zoback, Visco-plastic Properties of Shale Gas Reservoir
Rocks, 45th U.S. Rock Mechanics/Geomechanics Symposium, San Francisco,
California, June 26–29, Paper No. ARMA-11-417 (2011).
42. S. M. Willson, S. T. Edwards, A. J. Crook, A. Bere, D. Moos, P. Peska,
and N. C. Last, Assuring Stability in Extended Reach Wells-Analyses,
Practices and Mitigations, SPE/IADC Drilling Conference, Amsterdam, The
Netherlands, February 20–22, Paper No. SPE-105405-MS (2007).
43. Y. N. Abousleiman, M. H. Tran, and S. Hoang, Laboratory Characterization
of Anisotropy and Fluid Effects on Shale Mechanical Properties Using
Inclined Direct Shear Testing Device IDSTD™, 42nd U.S. Rock Mechanics
Symposium (USRMS), San Francisco, CA, June 29-July 2, Paper No. ARMA-
08-256 (2008).
44. B. Wu and C. P. Tan, Effect of Shale Bedding Plane Failure on Wellbore
Stability – Example from Analyzing Stuck-Pipe Wells, 44th US Rock
Mechanics Symposium and 5th US Canada Rock Mechanics Symposium, Salt
Lake City, UT, USA, June 27–30, Paper No. ARMA-10-350 (2010).
45. Y.F. Li, Y.Q. Fu, G. Tang, C.Y. She, J. Guo, and J.Y. Zhang, Effect of Weak
Bedding Planes on Wellbore Stability for Shale Gas Wells, IADC/SPE Asia
Pacifc Drilling Technology Conference and Exhibition, Tianjin, China, July
9–11, Paper No. SPE-155666-MS (2012).
46. A. A. Alqahtani, M. Mokhtari, A.N. Tutuncu, and S. Sonnenberg, Effect of
Mineralogy and Petrophysical Characteristics on Acoustic and Mechanical
Properties of Organic Rich Shale, Unconventional Resources Technology
Conference, Denver, Colorado, USA, August 12–14, Paper No. SPE-
168899-MS (2013).
47. P. Marouby and Z. Heidari, Assessment of In-Situ Elastic Properties of
Hydrocarbon-Bearing Shale with the Joint Quantitative Interpretation
of Sonic, Nuclear, and Resistivity Logs, SPWLA 52nd Annual Logging
Symposium, Colorado Springs, Colorado, May 14–18, Paper No. SPWLA-
2011-O presented (2011).
48. L. Vernik and C. Landis, Elastic Anisotropy of Source Rocks: Implications for
Hydrocarbon Generation and Primary Migration. B. Am. Assn. Petrol. Geol.
80, 531–544 (1996).
49. L. Vernik and A. Nur, Ultrasonic velocity and anisotropy of hydrocarbon
source rocks. Geophysics 57, 727–735 (1992).
50. C. H. Sondergeld, C. S. Rai, R. W. Margesson, and K. J. Whidden, Ultrasonic
Measurement of Anisotropy on Kimmeridge Shale, SEG 78th Annual
International Meeting, Calgary, Canada, August 6–11 (2000).
51. C. H. Sondergeld and C.S. Rai, Elastic anisotropy of shales. The Leading Edge
30, 324–331 (2011).
52. Y. P. Zhu, S. Y. Xu, M. Payne, A. Martinez, E. Liu, C. Harris and K.

Bandyopadhyay, Improved Rock-Physics Model for Shale Gas Reservoirs,
SEG Annual Meeting, Las Vegas, Nevada, November 4–9, Paper No. SEG-
2012-0927 (2012).
53. J. Handin and R. Hager, Experimental deformation of sedimentary rocks
under confning pressure: Tests at high temperature. B. Am. Assn. Petrol. Geol.
42(12), 2892–2934 (1958).
54. P. J. Closmann and W. B. Bradley, The effect of temperature on tensile and
compressive strengths and Young’s modulus of oil shale. SPE Journal 19(5),
301–312 (1979).
55. A. F. Mills and B. H. Chang, Error Analysis of Experiments-A Manual for
Engineering Students, University of California, Los Angeles, California
90095 (2004).
Part 6
ENHANCE OIL RECOVERY
19
A Numerical Investigation of Enhanced
Oil Recovery Using Hydrophilic Nanofuids
Yin Feng, Liyuan Cao* and Erxiu Shi
University of Louisiana at Lafayette, Lafayette, LA, USA
Abstract
In order to explore the potential applications of nanotechnology in oil industry to
enhance oil recovery, numerous studies regarding nanoparticle mobility in porous
media have been conducted in the past decade. Theoretical modeling was carried
out to investigate nanoparticle retention mechanism as well. However, most of the
simulations were either limited to laboratory scale or only accounted for transport
and retention behaviors. In this paper, the effect of hydrophilic nanofuids on oil
recovery was quantified at a reservoir scale. Simulations were performed using an
in-house simulator and three essential tasks were accomplished: (1) The in-house
simulation framework consists of two critical components: a two-phase displace-
ment module (accounts for waterflooding) and a transport/retention module
(accounts for Advection-Dispersion-Retention behavior exhibited by nanoparti-
cles). Those two modules were verified with two simulation tools (Eclipse from
Schlumberger and MNM1D by Tosco et al.), respectively in this study; (2) A solu-
tion algorithm bridging the above two modules was developed and implemented
to demonstrate the effect of nanoparticle retention on improving displacing effi-
ciency; (3) Different injection scenarios simulated include continuous injection,
slug injection, and postflush, and the effects of injection time, injection rate, and
slug size on oil recovery were studied. The results discovered that when nanofuids
flooding is used after waterflooding as a tertiary recovery method, early nano-
fuids injection improves higher oil recovery, but results in more nanoparticle loss.
Higher injection rate of nanofuids could help improve the flooding efficiency, but
not the ultimate oil recovery and cause more nanoparticle loss as well. Compared
to continuous injection, slug injection with postflush saves a lot of nanoparticles
but present a similar contribution in a Nano-EOR process.
429
Keywords: Nanoflooding, reservoir simulation, enhanced oil recovery,

nanoparticle retention
19.1 Introduction
Introduced by Feynman in 1959, nanotechnology was an idea to
precisely manipulate atoms and molecules and to create nanoscale
machines. Nanotechnology has shown its potential for revolution-
ary changes in oil and gas industry in the past few years. Engineered
nanoparticles have been proposed for many potential applications in
upstream oil and gas industry, especially as emulsion/foam stabilizers for
enhanced oil recovery (EOR) purpose or as nano-sensors for formation
evaluation. Zhang et al., (2009), and Espinosa et al., (2010) both have
found that surface-coated silica nanoparticles can stabilize emulsions
and foams for mobility control in EOR through experimental studies.
Functionalized nanoparticles can be injected into reservoir formations
as sensors to detect certain reservoir rock and/or fluids properties [1].
Other engineering applications in the improved oil recovery process
include injecting nanoparticle dispersion into reservoirs for improved
oil recovery [2–6]. Several studies have proved that SiO2 nanoparticles
(NPs) can alter the wettability of reservoir rock and reduce the interfa-
cial tension (IFT) between crude oil and brine phases by their adsorp-
tion onto the rock surface [7], (Onyekonwu and Ogolo 2010). Le et al.
(2009) studied synergistic blends of SiO2 nanoparticles and surfactants
for EOR in high-temperature reservoirs. They proved that some blends
have great potential in EOR application because of their thermostability
at 91°C, extremely good IFT reduction and very high oil displacement
efficiency. Hendraningrat et al., (2013) used Hydrophilic silica nanopar-
ticles for a coreflood experiment to investigate nanofuids for EOR. Their
results showed that oil recovery increased with nanofuids flooding in
high permeability sand stone core sample, but not for low permeabil-
ity one. For both environmental concern and reservoir application, it is
important to understand nanomaterial transport properties in porous
media with single- or two-phase flow.
Over the past few years, nanoparticles transport in porous media has
been studied. Both experiments and theoretical models were carried
out to evaluate the nanoparticle mobility and investigate the nanopar-
ticle retention mechanism. Systematical nanomaterial transport exper-
iments were carried out by Lecoanet et al. [8] to evaluate their mobility
Numerical Investigation of Enhanced Oil Recovery 431
in water-saturated glass-bead columns. Those nanomaterial diameters

were between 1.2 nm and 303 nm. Most of them penetrated the media
and their effluent concentration reached initial injection concentration
after long-term injection. Jeong and Kim [9] investigated copper oxide
nanoparticles transport in two-dimensional porous media. They visual-
ized nanoparticle aggregation in the pores and found flow velocity and
surfactant content in the nanoparticle dispersion both had important
effects on the nanoparticle deposition and aggregation. The mobility of
surface-modified iron nanoparticles through water-saturated columns
was evaluated under different flow conditions [10–12]. The effects of
porous medium structure, flow rate and ionic strength and composition
on particle deposition were all investigated experimentally. In summary,
when nanoparticles transport through porous media with single phase
flow, they can be removed from the dispersion. Nanoparticle filtration
brings either delay in breakthrough or effluent history plateau value less
then unity. During nanoparticle transport in porous media, there exists
a mass exchange between the nanoparticle on the pore surface and in
the fluids by adsorption and desorption from pore surface, and block-
ing at pore throats [7]. The influence of flow velocity, nanoparticle coat-
ing, dispersion concentration and specific surface area of the porous
medium on nanoparticle retention and transport were all investigated
by experiments (Caldelas, 2010; [13]). Murphy [13] conducted some
laboratory experiments to investigate silica nanoparticle transport in a
water-saturated sandpack. The results proved the existence of the revers-
ible adsorption of hydrophilic silica nanoparticles in porous media and
that postflush provides a chance for retained particles to be detached
from solid phase. Most of the effluent concentration histories reach
the nanoparticle injection concentration after about three pore-volume
injection of nanoparticle dispersion. Yu [14] carried out some experi-
ments to determine whether nanoparticles can be treated so that deposi-
tion at oil/water interfaces is maximized while deposition on rock grain
surfaces is minimized. Parvazdavani et al., [15] investigated the effect of
water based nanoparticles addition on hysteresis of oil and water rela-
tive permeability curves. Silica nanoparticles (14nm) and sandstone core
samples (k=10mD) were used. Relative permeability curves with water
flooding and nanofuids flooding were both obtained in their study.
Ju and Fan [7] have used two types of NPs in oil fields to improve oil
recovery and enhance water injection. They introduced experimental inves-
tigations to study both physical properties of the nanoparticles, and pore
characteristics of sandstone. Also, they have presented a 1D mathematical
model to describe the nanoparticles transport carried by two-phase flow

in a porous medium. Sbai and Azarroual developed a mathematical model
that couples the incompressible two phase fluid flow reservoir equations
at a macroscopic level to equations of nanoparticles transport at a smaller,
but still macroscopic, secondary scale. Improvements in the recovered vol-
umes by injecting hydrophobic nanoparticles which enhance or reverse
the initial reservoir wettability favor an increase in the relative permea-
bility of the oil phase and the capillary pressure drop between phase pres-
sures. EL-Amin et al. [16] developed a mathematical model to simulate
nanoparticles transport in porous media based on Ju and Fan’s model. The
model considered capillary forces as well as Brownian diffusion. Variation
of water saturation, nanoparticles concentration and porosity ratio were
investigated as well. The simulation of nanoparticle transport in porous
media from previous researchers are all on laboratory scale.
In this study, a solution procedure coupling the alteration of relative
permeability curves (due to the nanofuids flooding) with the in-house flow
and transport simulation system was proposed, to quantitatively predict
the oil recovery in the reservoir after nanoflooding and analyze the effect
of operational parameters, such as injection time, injection rate and slug
injection, on oil recovery. This would contribute to the migration of nano-
technology on the real oil field.
19.2 Simulation Framework

19.2.1 Background
The in-house simulator was constructed in C code under Linux system
taking advantage of high performance computing (HPC) resources in
Louisiana Optical Network Initiative (http://www.loni.org) [17]. PETSc,
Portable, Extensible Toolkit for Scientific Computation was employed
to solve the assembled transport equations for each species of chemical
components. The simulator input is through reading data files instead of
a graphical manner. The results will be saved as .vtk, so as to be visualized
by external packages such as ParaView (www.paraview.org) [18]. Several
external packages utilized are introduced as follows:
PETSc (Portable, Extensible Toolkit for Scientifc Computation) is used

to solve the partial differential equations in this study. PETSc is a suite of
data structures and routines developed by Argonne National Laboratory
for the scalable (parallel) solution of scientific applications modeled by
partial differential equations. It employs the Message Passing Interface

(MPI) standard for all message-passing communication. PETSc is the
world’s most widely used parallel numerical software library for partial
differential equations and sparse matrix computations (http://www.mcs.
anl.gov/petsc/) [19].
All of the computational efforts were performed on a supercomputer
named ERIC in LONI (Louisiana Optical Network Initiative), a state-of-
the-art, fiber optics network that runs throughout Louisiana, and connects
Louisiana research universities to one another. LONI provides Louisiana
researchers with one of the most advanced optical networks in the country
and the most powerful distributed supercomputer resources available to
any academic community with over 85 teraflops of computational capacity
(http://www.loni.org) [17].
19.2.2 Two Essential Computational Components

19.2.2.1 Flow Model
As the most critical module in the in-house simulator, flow model aims
to predict the water and oil displacement process based on the following
assumptions.
1. Flow is under isothermal condition and the rock and fluids

are supposed to be incompressible.
2. Flow of oil and water in porous media follows Darcy’s law.
3. Both oil and water are Newtonian fluids.
The Darcy’s law and the equations of mass conservation for each phase
govern the two-phase immiscible incompressible flow in a homogeneous
porous medium. The equations are given below.
Kkrw
υw = − ∇Pw (19.1)
µw
Kkro
υo = − ∇Po (19.2)
µo
∂ Sw
φ + ∇ • υw = 0 (19.3)
∂t
∂ So
φ + ∇ • υo = 0 (19.4)
∂t
The saturation of the phases are constrained by,
Sw + So = 1 (19.5)
where the subscripts w and o denote the wetting phase (water) and
non-wetting phase (oil), respectively, P is pressure, v is Darcy velocity vec-
tor, Sw is water saturation, So is oil saturation, kr is the relative permeability
and μ is viscosity.
The sum of the velocities of both the wetting and non-wetting phases is,
Kkrw Kk
υt = υ w + υ o = − ∇Pw − ro ∇Po (19.6)
µw µo
Based on Eqs. (19.1), (19.2) and (19.5), one may get that,
∇ • υt = 0 (19.7)
Or,
 Kk Kk 
∇ •  rw ∇Pw + ro ∇Po  = 0 (19.8)
 µw µo 

The capillary pressure Pc is defned as a pressure difference between the

non-wetting and wetting phase pressures, which is given below,
Pc = Po − Pw (19.9)
In this study, Pc is neglected, so,
Po = Pw (19.10)
Eq. (19.8) can be rewritten as,
 k k  
∇ • K   rw + ro  ∇Pw  = 0 (19.11)
  µw µo  

The nanoparticles retention on the rock surface may cause wettability

alteration and further change the relative permeability curves. When S = S*
the total surfaces per unit bulk volume of the porous media are completely
covered by the nanoparticles adsorbed on the pore surfaces or entrapped
in core throats, and when S < S* the surfaces per unit bulk volume of the
porous media are partially covered by the nanoparticles. Therefore, the
relative permeabilities of the water and oil phases can be expressed as a
linear function of the surface covered by the nanoparticles [16], i.e. 0 < S
< S*. A linear interpolation can be employed to generate the following two
equations.
S
krw ,s = krw + (k ∗ − krw ) (19.12)
S ∗ rw,s
S
kro ,s = kro + (k ∗ − kro ) (19.13)
S ∗ ro,S
where krw,S* and kro,S* are the relative permeabilities of water and oil phases
when the surfaces per unit bulk volume of the porous media are completely
occupied by the nanoparticles, while krw,S and kro,S are the relative permea-
bilities of water and oil phases when the surfaces per unit bulk volume of
the porous media are partially occupied by the nanoparticles.
In this study, two groups of relative permeability data obtained from
experiments were used. One is for S = 0, and the other one is for S = S*. For
all the situations in between (0 < S < S*), a linear relationship was assumed.
The experimental data is based on the study of Parvazdavani et al., in 2012
as shown in Table 19.2 [15].
19.2.2.2 Nanoparticle Transport and Retention Model

There are basically two types of nanoparticles used in oil industry to
improve oil recovery. One is called lipophobic and hydrophilic polysillicon
nanoparticles (LHPN) and exists only in water phase, while the other is
called hydrophobic and lipophilic nanoparticles (HLPN) and exists only in
the oil phase. In this study, to simplify the problem, LHPN was considered.
So nanoparticles can either exist in the water phase or on the surface of the
rock.
The transport of nanoparticles is a coupled process of advection and
dispersion, which is described mathematically as Advection-Dispersion-
Equation (ADE). The relevant assumptions for the transport model include:
1. An intrinsic adsorption capacity exists on the solid surface,

which is independent of the flow conditions.
2. Both the adsorption and desorption rate are uniform in the
porous media and are independent of flow conditions.
3. All nanoparticles in the dispersion are available for adsorp-
tion, and the adsorbed nanoparticles on solid surface can be
desorbed, which refers to as a reversible adsorption process.
Since LHPN was taken into account, injected nanoparticles were

assumed to only exist in water phase, and the ADE for the nanoparticles in
water is written as,
∂ (φ Sw Cw )
+ υw • ∇Cw = ∇ • (φ Sw Dw ∇Cw ) (19.14)
∂t
where is ϕ the porosity of the porous media, Sw is the water saturation, Cw

is the volume concentrations of nanoparticles in the water phase, vw is the
pore velocity, Dw is the dispersion coefficients of nanoparticles in the water
phase.
The retention of nanoparticles due to adsorption and desorption can
be described by a Langmuir-like model as Eqs. (19.15) and (19.16) [20].
Eq. 19.15 represents the amount of nanoparticles change in the fluid,
while Eq. 19.16 represents the amount of nanoparticles change on the
rock surface.
∂ (Sw Cw ) ρb ∂ S
+ =0 (19.15)
∂t φ ∂t
∂ S Swφ  S 
= ka  1 − Cw − kd S (19.16)
∂ t ρb  Smax 
where, S is defined as the mass of nanoparticles adsorbed on unit mass of

solids, Smax is the maximum retention capacity with the same unit as S, ka
and kd are adsorption and desorption rates, respectively.
The local equilibrium can be mathematically described as S ′ = Cw′ = 0,
and the equilibrium adsorption concentration with given dispersion con-
centration can be derived as,
Swφ  S 
ka  1 − Cw − kd S = 0 (19.17)
ρb  Smax 
Smax
S∗ =
ρ k S (19.18)
1 + b d max
φ ka Cw
Which implies that S* is always smaller than Smax when kd is not zero. In
this study, S* is denominated as the equilibrium adsorption at given values
of adsorption and desorption rates, which is also the maximum “possible”
adsorption (less than Smax).
19.3 Coupling of Mathematical Models

Basically, the mathematical model is a coupled system of nonlinear
time-dependent partial differential equations, which consists of the con-
tinuity equations for water and oil phases (Eq. (19.3) and Eq. (19.4)), the
advection dispersion equation (Eq. (19.14)) and several auxiliary equa-
tions. The IMplicit Pressure Explicit Saturation (IMPES) approach was
employed, which solves the pressure equation implicitly and updates the
saturation explicitly. After the pressure was obtained, Darcy’s velocity and
two-phase saturations can be determined.
On the aspect of stability, IMPES is classified as a conditionally stable
method, which means the proper selection of the timestep will result in a
stable result but certainly slow down the simulation speed. That’s why most
of the commercial simulators aim to a fully implicit solution for the uncon-
ditional stability (they consider the stability and speed as top criteria but
not accuracy). However, fully implicit solutions potentially present lower
accuracy because of larger truncation errors in discretions. Verification
case 4.1 is used to prove that the selected timestep can produce stable
outputs.
The solving procedures are: (1) First, the pressure distribution is
obtained by solving Eq. (19.10), and water saturation is calculated by Eq.
(19.11); (2) Then Nanoparticles concentration distribution is obtained by
solving the ADE (Eq. 19.14) and nanoparticles concentration on the rock
surface S can also be determined; (3) Based on the Sw from the first step,
krw,0, krw,S*, kro,0 and kro,S* can be calculated by linear interpolation; (4) Based
on the results from the second step and the third step, krw,S and kro,S can be
calculated by linear interpolation.
So the whole process is a decoupled process. The operator splitting
error is acceptable according to the results of verification cases previously.
The nanoparticles transport and retention model influence flow model by
updating the relative permeability of both water and oil.
The solution for mathematical models for one time step is shown in
Figure 19.1.
The entire process is actually known as an ADRE (Advection-Dispersion-
Reaction-Equation) problem. To simplify the calculation, operator-splitting
was employed and the ADRE was therefore split into an ADE problem (Eq.
19.14) plus Reaction behaviors (Eq. 19.15-19.16), mathematically a PDE
and coupled ODEs, which are solved by ADE solver and Retention model,
respectively (as seen in the yellow box of Figure 19.1). The simultaneous
ADE and Reaction (the reaction here is referred to as retention including
adsorption and desorption of nanoparticles) are considered sequentially in
the presented modeling (ADE first and then retention). This splitting will
introduce numerical error that is why a verification case (4.3) was studied
Start
Solve flow model

Update krw and kro
(Eq. 19.11),
obtain vw and Sw
Sw
Solve ADE
(Eq. 19.14),
obtain Cw
Linear interpolation
table 19.2 and table 19.3
obtain krw, 0 and kro, S* and Solve retention model
krw, 0, kro, S* (w, n) (Eq. 19.15 & 19.16),
obtain S and Cw
Linear interpolation S/S*

figure 19.12
obtain krw, S and kro, S
Figure 19.1 Flowchart of solution for mathematical models for one time step.
to prove that the produced splitting error is negligible. ADE are solved
implicitly and the ODEs can be solved by many methods, such as Rouge-
Kutta or an iterative method as used in this study.
19.4 Verification Cases

Before running nanoflooding simulations, three verification cases were
conducted. First, the performance of the in-house simulator under differ-
ent time steps was analyzed to prove that the selected timestep can pro-
duce stable outputs. Also, Eclipse and MNM1D (Micro and Nanoparticle
Transport Model in porous media in 1D geometry) were used to verify the
predicting capability of the simulator, where Eclipse was used to verify the
accuracy of two-phase displacement, and MNM1D was used to account
for nanoparticles transport and retention.
ECLIPSE is a commercial simulator released by Schlumberger. It can be
used to model any reservoir type (https://www.software.slb.com/products/
eclipse/simulators).
MNM1D is a software developed by Tosco and Sethi [21] in the Matlab
environment using finite difference method. MNM1D is able to simulate
the colloid behavior in porous media in the presence of both constant
and transient hydrochemical parameters, and account for attachment and
detachment phenomena.
19.4.1 Effect of Time Steps on the Performance

of the in House Simulator
The verification was conducted on a homogeneous two-dimensional two-
phase domain. It’s a squared reservoir where there is an injection well on
one corner and a production well on the opposite corner. The parameters
in simulations are shown in Table 19.1.
As a time discretization scheme, the IMPES method is conditionally sta-
ble, and hence it must take a very small time step size [22]. Here two differ-
ent time steps (0.1 and 0.05 day) were used to examine the accuracy for the
decoupling of governing equations (ADE+Retention). Oil recovery factors
calculated using various time steps were compared, and the results present
a satisfactory agreement (Figure 19.2). Based on the presented results, it
can be proved that the selected timestep meets the requirement and can
produce stable outputs. It may somehow slowdown the simulation but can
still complete within a reasonable time, and the speed is not in the scope
of this study. Also, this proves that the simulator is accurate enough for
Table 19.1 Parameters used in simulation.

Phase Perm (md) Poro Dimension
water, oil 200 0.15 51*51*1
Grid Tops (ft)
DX (ft) DY (ft) DZ (ft)
20 20 60 3280
0.9
0.8
0.7
Oil recovery, percent
0.6
0.5
0.4
0.3
0.2 Delta t = 0.1d
0.1 Delta t = 0.05d
0
0 500 1000 1500 2000 2500
Time, d
Figure 19.2 Simulation results under two different time steps.
calculating the oil recovery factor with the time step of 0.1d, which is thus
employed as the time step for the following simulations.
19.4.2 Comparison with Eclipse

The verification was conducted on the homogeneous two-dimensional
two-phase domain mentioned in the last step.
The oil-water relative permeability data is based on Parvazdavani et al.,
[15], who did some core flooding experiments to investigate the effect of
water based nanoparticles addition on hysteresis of oil and water relative
permeability curves. The oil and water relative permeability curves after
water flooding and nanofuids flooding are shown in the paper and by digi-
tizing the data from the curves, the oil and water relative permeability data
are available and are listed in Tables 19.2 and 19.3, respectively. The relative
permeability curves are then re-plotted in Figure 19.3 based on the data in
Table 19.2 and Table 19.3.
Table 19.2 Oil and water relative

permeability data with water flooding.
Sw krw-water kro-water
0.369 0 0.755
0.431 0.01 0.396
0.469 0.015 0.266
0.514 0.023 0.196
0.609 0.033 0.058
0.659 0.045 0.04
0.69 0.068 0.022
0.714 0.1 0.013
0.73 0.111 0
Table 19.3 Oil and water relative permeability data with

nanofuids flooding.
Sw krw-nano kro-nano
0.369 0 0.95
0.465 0.037 0.64
0.527 0.068 0.454
0.597 0.114 0.27
0.723 0.206 0.081
0.808 0.274 0.028
0.836 0.295 0.023
0.852 0.303 0.006
0.857 0.324 0
The residual oil saturation after water flooding is 0.270, and that after
nanofluids flooding is 0.143, proving that nanofuids flooding could
enhance oil recovery significantly.
1.0
0.9
0.8 krw-water
0.7 kro-water
krw-nano
Kro, Krw
0.6
kro-nano
0.5
0.4
0.3
0.2
0.1
0.0
0.0 0.2 0.4 0.6 0.8 1.0
Sw
Figure 19.3 Oil and water relative permeability curves (repotted from Parvazdavani et al.,
[15].
The oil recovery factors (RF) produced from two simulators presented
a great agreement, which proved the consistency of two-phase flow com-
putation by in-house simulator comparing to a well-known commercial
simulator.
19.4.3 Comparison with Software MNM1D

After comparing with Eclipse, results are compared with software MNM1D
(Micro and Nanoparticle Transport Model in porous media in 1D geom-
etry) developed by Tosco and Sethi [21] in the Matlab environment using
finite difference method. MNM1D is able to simulate the colloid behavior in
porous media in the presence of both constant and transient hydrochemical
parameters, and account for attachment and detachment phenomena.
In this case, the 1D nanoparticle transport and retention behavior with
finite particle retention capacity, Smax, is simulated. Other parameters used
are porosity (ϕ) = 0.4, Darcy velocity = 8 × 10–5 m/sec, bulk density (ρb) =
1.5 × 103 kg/m3, distance between inlet and outlet = 0.1m, and hydrody-
namic dispersion (D) = 10–7 m2/sec. Temporal and spatial steps adapted by
Tosco are v t = 2 sec and v x = 10–3 m. A constant nanoparticle concentra-
tion (C0) is injected with the total injection time of 3600s. After that, the
injection continues but the fluid contains no nanoparticles.
The comparison results are presented in Figure 19.5 and Figure 19.6 at
different values for ka and kd, and excellent agreement can be observed.
This could confirm the capability of the in-house simulator in predicting
transport and retention of nanoparticles.
0.9
0.8
Oil recovery, percent 0.7
0.6
0.5
0.4 Nano (eclipse)
0.3 Nano (this study)
Water (eclipse)
0.2
Water (this study)
0.1
0
0 500 1000 1500 2000 2500
Time, d
Figure 19.4 Comparison of oil recovery change with time after water flooding or
nanofuids flooding computed by Eclipse and in-house model.
1.0
Normalized concentration (C/C0)
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1 Tosco et al. (2009)
This study
0.0
0.0 2000.0 4000.0 6000.0
Time (sec)
Figure 19.5 Compared results of normalized effluent concentrations of nanoparticles

after nanoflooding for 3600s followed by water flooding up to 6000s between Tosco model
and in-house model with ka = 0.004sec–1 and kd = 0.008sec–1.
19.5 Results
In this chapter, the effect of nanoparticle injection time, injection rate,
and nanoparticle slug injection were all investigated by simulation. For
all the different injection scenarios, the amount of nanoparticles trapped
in the formation and ultimate oil recovery were calculated for compar-
ison. The distributions of nanoparticles on the rock surface and in the
fluid were plotted.
1.0
Normalized concentration (C/C0)

0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1 Tosco et al. (2009)
This study
0.0
0.0 2000.0 4000.0 6000.0
Time (sec)
Figure 19.6 Compared results of normalized effluent concentrations of nanoparticles

after nanoflooding for 3600s followed by water flooding up to 6000s between Tosco model
and in-house model with ka = 0.008sec–1 and kd = 0.004sec–1.
Table 19.4 Geometric parameters of the model.

Dimension X (ft) 1020
Dimension Y (ft) 1020
Dimension Z (ft) 60
Number of Grid X 51
Number of Grid Y 51
Number of Grid Z 1
Table 19.5 Base data for model analysis.

Water viscosity (cp) 1 Simulation time (days) 2000
Oil viscosity (cp) 2 Delta time (days) 0.1
Water density (lb/cf) 63 Density of nanoparticle (lb/cf) 145
Oil density (lb/cf) 55 Diameter of nanoparticle (nm) 22
Bulk density (lb/cf) 128 Diffusivity coefficient (ft /day)
2
47
Residual oil saturation 0.27 Adsorption rate (ka,1/day) 115.8
Initial water saturation 0.369 Desorption rate (kd,1/day) 259.2
Number of wells 5 Maximum retention capacity 39.6
(Smax, mg/g)
Nanofuids concentration (mg/ml) 1
The model used here is a five-spot flooding model, and the parameters
of the model are shown in Table 19.4. One injection well is in the center of
the squared reservoir and four production wells are at each corner.
The input data used for analysis is based on literature review. Oil and
water relative permeability data is listed in Table 19.2 and Table 19.3
for water flooding and nanoflooding, respectively. Other input data is
based on the experiment #65 conducted by Zhang [20]. The unit of the
data is converted to the unit used in the model and the data is shown in
Table 19.5.
19.5.1 Continuous Injection

Continuous injection means the process of nanofuids injection lasts from
beginning to the end. The effect of nanofuids injection time and injection
rate on oil recovery was investigated with other parameter values set to the
baseline values.
19.5.1.1 Effect of Injection Time on Oil Recovery and Nanoparticle

Adsorption
Nanofuids injection time (the starting time of nanofuids injection) has
some influence on oil recovery. An apparent view for the significance of
this effect is shown in Figure 19.10. The amount of nanoparticles trapped
at 2000 days (3 pore volume) under different nanofuids injection time were
also compared (Figure 19.7).
Based on the graphs, it can be concluded that the oil recovery, as well as
the amount of nanoparticles trapped will be higher with early nanofuids
injection within the same amount of time.
To visualize the evolution of water saturation during nanoflooding in
the reservoir, the water saturation distribution map at injection time of
300d is plotted by Gnuplot. Figure 19.9 shows the result. It can be seen
clearly that oil is pushed to the production well and the reservoir reached a
very high water saturation at the end of the injection.
The distribution of nanoparticles adsorption (on the rock surface) at
different injection time was also plotted by Gnuplot. Comparison result is
shown in Figure 19.10. It indicates that early nanofuids injection, or lon-
ger nanofuids injection, could lead to more residual nanoparticles in the
reservoir.
0.9
0.8
Oil recovery (fraction) 0.7

0.6
0.5
t_inj = 300d
0.4
t_inj = 600d
0.3
t_inj = 900d
0.2 t_inj = 1200d
0.1 t_inj = 1500d
0
0 500 1000 1500 2000 2500
Time (d)
Figure 19.7 Influence of nanofuids injection time on oil recovery.
120
Nanoparticle trapped, ton
100
80
60
40
20
0
300 600 900 1200 1500
Injection time, d
Figure 19.8 Amount of nanoparticles trapped under different injection time.
In conclusion, nanofuids flooding could improve oil recovery as a ter-

tiary recovery method. However, the amount of nanoparticles trapped in
the reservoir is huge, which could lead to an extensive economic loss. Based
on the data, it can be concluded that the amount of nanoparticles trapped
on the rock surface can be determined by nanofuids injection period. In
order to decrease the nanoparticle loss, less time of nanofuids injection
would be necessary.
50 0.84 50 0.86
0.82 0.84
40 0.8 40 0.82
0.78
Water saturation
0.8
30 0.76 30
0.78
0.74
0.76
20 0.72 20
0.74
0.7
10 0.68 10 0.72
T = 300d (0.5 PV) 0.66 T = 1000d (1.5 PV ) 0.7

0 0.64 0 0.68
0 10 20 30 40 50 0 10 20 30 40 50
50 0.86 50 0.858
0.856
0.84
40 40 0.854
0.82 0.852
0.8 0.85
30 30
0.848
0.78
0.846
20 20
0.76 0.844
0.74 0.842
10 10 0.84
T = 1500d (2.25 PV )
0.72 T = 2000d (3 PV)
0.838
0 0.7 0 0.836
0 10 20 30 40 50 0 10 20 30 40 50
Figure 19.9 Water saturation distribution at 0.5 PV, 1.5 PV, 2.25 PV and 3 PV respectively
when the time of nanofuids injection t = 300d (the color bar represents water saturation).
19.5.1.2 Effect of Injection Rate on Oil Recovery and Nanoparticle

Adsorption
Different nanofuids injection rates were also simulated in this study, where
nanofluids were injected from the beginning under different flow rate. The
result shows that injection rate does affect the oil recovery. Higher injection
rate means higher oil recovery within the same amount of time. But when
the time is long enough, they would eventually reach the same oil recovery
(Figure 19.11). This indicates that higher injection rate could improve the
flooding efficiency, but not the ultimate oil recovery for field development.
The amounts of nanoparticles trapped in the reservoir at the end of
injection under different injection rates were also compared. The result is
shown in Figure 19.12. Higher injection rate means more nanoparticles
loss. So for a real field application, the injection rate should not be too
high. For this case, oil recoveries under different flow rate all reached 0.77
50 1 1
0.99
40 0.98 0.95
0.97
30 0.96 0.9
0.95
20 T = 600d 0.94 0.85
T = 900d
0.93
10 0.92 0.8
0.91
0 0.9 0.75
50 1 0.85
0.95 0.84
40 0.9 0.82
0.85
30 0.8
0.8
0.78
0.75
20 0.76
T = 1200d 0.7 T = 1500d
0.65 0.74
10
0.6 0.72
0 0.55 0.7
0 10 20 30 40 50 0 10 20 30 40 50
Figure 19.10 Nanoparticle distribution in the reservoir at the end of the nanofuids
injection (3PV) under different nanofuids injection time (the color bar represents
normalized nanoparticle concentration).
0.9
0.8
0.7
Oil recovery, percent
0.6
0.5 q = 2000 stb/d
0.4 q = 2500 stb/d
0.3 q = 3000 stb/d
q = 5000 stb/d
0.2
q = 10000 stb/d
0.1
0
0 500 1000 1500 2000 2500
Time, d
Figure 19.11 Oil recovery under different nanofuids injection rates.

102
101
Nanoparticle trapped, t
100
99
98
97
96
2000 2500 3000 5000 10000
Injection rate, stb/d
Figure 19.12 The amounts of nanoparticles trapped in the reservoir at the end of
injection under different injection rate.
by 2000d (3PV), but when flow rate is 2000 stb/d, there is the least amount
of nanoparticle trapped (Figure 19.12), which makes it the best flow rate.
Also, the difference between the nanoparticles loss decreased as the
injection rate increased from 2000stb/d to 10000stb/d. This is because of
the existence of nanoparticle equilibrium concentration (S*). No more
nanoparticles would be adsorbed onto the rock surface when the concen-
tration increased to S*.
19.5.2 Slug Injection

Since injecting nanofuids continuously will cause a large amount of
nanoparticles loss, it’s not economic in a real reservoir development. Slug
injection of nanofuids was simulated to find the best time and slug size to
perform nanoflooding after water flooding so that the peak value of oil
recovery could be reached with the least nanoparticles trapped.
19.5.2.1 Effect of Injection Time on Oil Recovery and Nanoparticle

Adsorption
The effect of injection time on slug injection of nanofuids was studied.
Water was first injected, followed by a slug of nanofuids, and then water
injection again. The injection times selected were 300d, 600d and 900d.
The slug size (nanofuids injection time length) was 300d for all the three
processes. The result is shown in Figure 19.13 and Table 19.6.
0.8
0.7
Oil recovery (fraction)
0.6
0.5
0.4
t_inj = 300d
0.3
t_inj = 600d
0.2 t_inj = 900d
0.1
0
0 500 1000 1500 2000 2500
Time (d)
Figure 19.13 Oil recovery with slug injection of nanofuids under different injection time.
Table 19.6 Oil recovery and mass of trapped nanoparticle

under different injection time (slug size 300d).
Injection time, d Oil recovery, fraction NP trapped, t
300 0.76 2.1
600 0.757 4.61
900 0.75 10.08
Based on the result, it can be seen that for slug injection of nanofuids,
early injection would lead to higher oil recovery and less nanoparticle
retention. This is actually related to the desorption property of nanopar-
ticles. When the nanofuids were first injected into the reservoir, a lot of
nanoparticles were adsorbed onto the rock surface. Then after the slug
injection of nanofuids, water is injected into the reservoir, and nanopar-
ticles are desorbed and pushed out by water flooding. So with longer
water flooding, more nanoparticles would be pushed out, resulting in less
nanoparticle loss.
The water cut changes when injection time is 300d is shown in Figure
19.14. It can be seen from the curve that water cut decreased significantly
when nanofuids were in effect, and then increased slowly and stabilized at
100%. This means that oil production increased due to nanofuids injec-
tion. Based on the plot, the water cut reached 98% by 1450d (2.2 PV), so
in field application, the injection should have been stopped for this case
if only considering oil production. For nanofuids flooding, nanoparticle
1.2 1.2
Oil recovery (fraction) 1 1
Water cut (fraction)

0.8 0.8
0.6 0.6
0.4 0.4
t_inj = 300d
0.2 Water cut 0.2
0 0
0 500 1000 1500 2000 2500
Time (d)
Figure 19.14 Oil recovery and Water cut against injection time (nanofuids injection time
and time length were both 300d).
loss is also a problem for economy. Longer water flooding after nanofuids
flooding might be preferred since it can help recover more nanoparticles.
19.5.2.2 Effect of Slug Size on Oil Recovery and Nanoparticle

Adsorption
The effect of slug size on oil recovery and nanoparticle retention was also
investigated. Water was first injected, followed by slug injection of nano-
fuids at time equal to 300d, and then water injection again, which is also
0.9
0.8
0.7
0.6
0.5
0.4 100d
0.3 300d
500d
0.2
0.1
0
0 500 1000 1500 2000 2500
Time (d)
Figure 19.15 Oil recovery with slug injection of nanofuids at different slug size.
Table 19.7 Oil recovery and nanoparticles trapped under

different slug size.
Slug size, d Oil recovery, fraction NP trapped, t
100 0.71 0.41
300 0.76 2.10
500 0.77 4.94
called postflush. Three different slug sizes were selected, 100d, 300d and
500d respectively. The result is shown in Figure 19.15 and Table 19.7.
In conclusion, both injection time and slug size of nanofuids flood-
ing have obvious influence on oil recovery and nanoparticle retention.
Postflush could help to recover more nanoparticles, and long time post-
flush may be needed considering economic and environmental issues.
19.5.3 Water Postflush

In the nanoparticles transport experiments, brine water was injected after
nanofluids flooding into the core sample to push the nanoparticles out.
Researchers discovered that given enough time, nanoparticles effluent
concentration would be less than 20% of the initial nanofuids concentra-
tion, meaning that post flush could help recover nanoparticles [13, 14, 20].
The effect of postflush on nanoparticles recovery at the production well
was studied. Basically, nanofuids flooding is employed as secondary oil
recovery here, where nanofuids was first injected and followed by water
flooding.
19.5.3.1 Effect of Injection Time Length

First, the time duration of nanofuids injection was analyzed. Different
duration of injection time were simulated and the result is shown in Figure
19.16. According to the result, 100d (0.15PV) is not enough to improve the
oil recovery from 55% to 77%, while 300d (0.45 PV) is long enough and
has the least nanoparticles trapped in the reservoir.
19.5.3.2 Effect of Flow Rate Ratio Between Water and Nanofuids

on Oil and Nanoparticle Recovery
Based on the previous study, higher nanofuids injection rate would lead to
a larger amount of nanoparticles trapped. On the other hand, higher post
flush flow rate, e.g. brine water flow rate, might be helpful in nanoparticle
recovery. So the flow rate ratio between water flooding and nanoflood-
ing was looked into here. Three groups of simulation were completed.
Nanofuids injection time length was 300d. The flow rate of nanoflooding
was set at 2000stb/d, and the injection rate of water post flush changed
from 2000stb/d to 4000stb/d. The result is shown in Figure 19.17, Figure
19.18 and Table 19.8.
It can be seen from Figure 19.17 and Table 19.8 that there is not much
difference in the ultimate oil recovery for the post flush processes under
three different injection ratios. But, the difference between the mass of
nanoparticle retention is significant. Higher flow rate ratio results in higher
nanoparticle recovery. However, the water flow rate can’t be too high due to
the pump capacity and water flooding cost.
According to Figure 19.18, the nanofuids concentration was still increas-
ing after the nanoflooding period. This shows that the trapped nanoparti-
cles were washing out by brine water flooding. During the flooding period,
the effluent concentration was always above 0, which means the whole
process, including the postflush, is actually nanoflooding. But the concen-
tration of the fluid is not constant, it will increase first due to nanoparticle
desorption, and then decrease due to nanoparticles being washed out of
the formation. This happens probably because after nanoflooding, there
are a lot of nanoparticles adsorbed onto the rock surface, and the following
brine flooding could help nanoparticles desorb from the substrate and get
back into the water phase due to concentration difference. The desorbed
nanoparticles would be pushed forward and adsorb onto the unsaturated
0.9
0.8
0.7
0.6
0.5
0.4 t_inj = 0d
0.3 t_inj = 100d
t_inj = 300d
0.2
t_inj = 500d
0.1 t_inj = 700d
0
0 500 1000 1500 2000 2500
Time (d)
Figure 19.16 Oil recovery curves under different length of nanofuids injection time
followed by post flush.
0.8
0.7
0.6
0.5
0.4 Qw/Qn = 1
0.3 Qw/Qn = 1.5

Qw/Qn =2
0.2
0.1
0
0 500 1000 1500 2000 2500
Time (d)
Figure 19.17 Oil recovery under different flow rate ratio between water flooding and
nanoflooding.
0.4
0.35
Qw/Qn = 1
Dimensionaless C/C0
0.3
concentration C/C0
Qw/Qn = 1.5
0.25
Qw/Qn = 2
0.2
0.15
0.1
0.05
0
0 500 1000 1500 2000 2500
Time (d)
Figure 19.18 Dimensionless nanoparticle effluent concentration at production well with

post flush.
Table 19.8 Oil recovery and nanoparticle recovery under different

flow rate ratio between water flooding and nanoflooding.
Qw/Qn| Trapped NP,kg Oil recovery Nanoparticle recovery
1 2157.6 75.30% 97.74%
1.5 285.5 75.60% 99.70%
2 36.8 75.60% 99.96%
surface which could contribute to higher oil recovery. The whole process
is like the recycling of nanoparticles. So postflush not only help improve
the nanoparticles recovery, but also extend the nanoflooding period, thus
enhancing the oil recovery.
In conclusion, for this simulation model, the best developing strategy
would be using nanofuids flooding with water postflush as secondary oil
recovery method. Nanofuids would be injected first at 2000stb/d for 300d,
and then water is injected at 4000stb/d for the rest of the time.
19.5.4 3D Model Showcase

In order to analyze the performance of nanoflooding in heterogeneous
formation, a 3D model simulation was performed in this study. The 3D
model employed is a revised subsection of SPE 10B model (http://www.
spe.org/web/csp/datasets/set02.htm). At the fne geological model scale,
the SPE 10B model is described on a regular Cartesian grid. The dimen-
sions are 1200×2200×170 (ft). The top 70 ft (35 layers) represents the
Tarbert formation (a representation of the near shore environment), and
the bottom 100 ft (50 layers) represents Upper Ness, which is fluvial. The
fine scale model size is 60×220×85 cells (1.122×106 cells) and accordingly
the fine scale cell size is 20×10×2 (ft). Figure 19.19 shows the porosity for
the whole model.
In this study, a subdomain of the original fine scale model was also built
up and studied, which is actually a corner of the full model. The size of the
subdomain is 30×110×35 (totally 1.155×105 cells). Two vertical wells were
proposed and completed throughout formation (penetrated the whole 35
layers), where one injector well was located at one corner of the subdomain
with a constant injection rate of 1,500 bbl/day, and one producer well at the
Figure 19.19 Porosity of the SPE 10B model.

opposite corner was maintained under a constant bottom hole pressure of

4,000 psi.
The input data for the 3D simulation is summarized in Table 19.9.
Three different processes were simulated here. The first one is brine
water flooding, the second one is nanofuids flooding, and the last one
is nanofuids injection for 300d followed by water flooding as post flush
to the end. Water saturation distribution under three different processes
were compared in Figure 19.20 on the same color scale. The nanopar-
ticles concentration in water phase was also analyzed, and the result is
shown in Figure 19.21.
Based on the result, it’s obvious that the 3D model has very strong het-
erogeneity. Also, the residual oil saturation after nanoflooding or postflush
is lower than that after water flooding, indicating that nanoflooding could
work as an EOR method. According to Figure 19.20, the water drive per-
formance among the three different scenarios can be investigated. There
is obvious difference in the residual oil saturation between water flooding
and nanofuids flooding. Also, the water front is further in water flooding
compared to postflush scenario. This shows that post flush is better than
water flooding, which is in accordance with the result from 2D simulation.
It can be seen in Figure 19.21 that water postflush could push nanoparticles
forward from the near wellbore area. The concentration of nanoparticles in
the water phase is way much lower after postflush compared to nanofuids
flooding.
Table 19.9 Input data for 3D simulation.

Viscosity Water(cp) 1 Simulation Time(days) 1000
Viscosity Oil(cp) 2 Dt(days) 0.5
Density Water(lb/cf) 63 Density of nanoparticle(lb/cf) 145
Density Oil(lb/cf) 55 Diameter of nanoparticle(nm) 22
Density Bulk(lb/cf) 128 Diffusion Coefficient(ft2/day) 47
Sor 0.27 Adsorption rate(Ka,1/day) 115.8
Swc 0.369 Desorption rate(Kd,1/day) 259.2
Number of Wells 2 Maximum retention 39.6
capacity(Smax,mg/g)
Nanofuids concentration (mg/ml) 1
8.570e-01
0.81
0.72
0.63
0.54
(a)
0.45
3.690e-01
(b)
(c)
Figure 19.20 Comparison results of (a) water saturation under water flooding,
(b) nanofuids flooding and (c) Post flush (300d nanoflooding + 700d water flooding)
under the same color cale.
19.6 Discussions
Mechanism of Postflush Enhancing Oil Recovery with
Less Nanoparticles Loss
The results on postflush may surprise the researchers. Before, postflush
was considered to be a good method to recover nanoparticles trapped in
the reservoir. Based on this study, if it’s used after a short time of nano-
flooding (0.5 PV or more), it could serve as a perfect method to enhance
oil recovery with the least nanoparticles loss (compared to nanoflooding
all the time). In order to find explanations for these results, another group
of simulations were completed. This time, nanofuids were injected into the
Nanoflooding Concentration
1.000e+00
0.75
0.5
(a) 0.25
Nanoflooding + Postflush 0.000e+00
(b)
Figure 19.21 The distribution of Nanoparticles concentration in water phase after

nanoflooding and postflush on the same color scale. (a) The distribution of nanoparticles
concentration in water phase after nanoflooding. (b) The distribution of nanoparticles
concentration in water phase after nanoflooding followed by water postflush.
reservoir for 100 days (0.15 PV) first, then brine water was injected up to
400 days. The nanofuids concentration distribution and nanoparticles con-
centration on the rock were both analyzed. The results are shown in Figure
19.22 and Figure 19.23.
From Figure 19.22, it can be seen that when the nanofuids is injected
for 100d, there is a nano-circle generated around the injection well, and
there is a concentration gradient inside the circle. It’s the same situation
for the nanoparticles concentration on the rock surface (Figure 19.23).
Nanoparticles adsorption onto the rock surface means oil molecules
desorption from the rock. So the 100d nanoflooding actually helps free the
oil around the injection well, inside the nano-circle area.
After nanoflooding, postflush has been conducted for 300 days. It can be
seen that the nano-circle expands and covers a bigger area. While, in the
center area, around the injection well, there are almost no nanoparticles,
indicating that water is pushing nanoparticles forward towards the produc-
tion wells. So the nano-circle becomes a nano-loop. Meanwhile, the same
amount of nanoparticles would desorb from the nano-circle area and adsorb
onto the rocks in the nano-loop area to help extract more oil. So on and so
forth, the nano-loop keeps expanding and the nanoparticles injected in the
20 1 20 0.4
Nano-circle 0.9 Nano-loop
0.35
0.8
15 15 0.3
0.7
0.6 0.25
10 0.5 10 0.2
0.4 0.15
0.3
5 5 0.1
0.2
0.1 0.05
t = 100d t = 200d
0 0 0 0
0 5 10 15 20 0 5 10 15 20
20 0.25 20 0.18
0.16
0.2
15 15 0.14
No-nano-circle 0.12
0.15
0.1
10 10
0.08
0.1
0.06
5 5
0.05 0.04
0.02
t = 300d t = 400d
0 0 0 0
0 5 10 15 20 0 5 10 15 20
Figure 19.22 Nanofuids concentration distribution in the reservoir at different injection

time after 100d nanoflooding chased by 300d water flooding (injection concentration C=1
mg/ml, which is 0.1 wt% if the density of the nanofuids is 1g/ml).
first 100 days keep being pushed forward. At 400d, there are nanoparticles
all around the reservoir except the area around the injection well, which is
called no-nano-circle here. It can be predicted that, if the postflush is contin-
ued for a long time, the no-nano-circle will expand and eventually, most of
the nanoparticles injected at the first 100 days would be washed out.
So in the whole process of this EOR method, the same amount of
nanoparticles would be used again and again to extract oil through the
whole reservoir, thus reducing the amount of nanoparticles required and
enhancing the oil recovery at the same time.
19.7 Conclusions and Future Work

1. Technically, this paper introduces a way to integrate nanopar-
ticle behaviors (regarding EOR) into a simulation procedure.
20 0.03 20 0.01
Nano-loop 0.009
Nano-circle
0.025
0.008
15 15
0.007
0.02
0.006
10 0.015 10 0.005
0.004
0.01
0.003
5 5
0.002
0.005
0.001
t = 100d t = 200d
0 0 0 0
0 5 10 15 20 0 5 10 15 20
20 0.006 20 0.0045
0.004
0.005
15 15 0.0035
0.004 No-nano-circle 0.003
0.0025
10 0.003 10
0.002
0.002 0.0015
5 5
0.001
0.001
0.0005
t = 300d t = 400d
0 0 0 0
0 5 10 15 20 0 5 10 15 20
Figure 19.23 Nanoparticles concentration distribution on the rock surface at different

injection time after 100d nanoflooding chased by 300d water flooding (injection
concentration C=1 mg/ml, which is 0.1 wt% if the density of the nanofluid is 1g/ml).
That is the interpolation between the two relative permeability

Datasets. To present the contribution of nanoparticles in EOR
processes, this study correlates the effect with the adsorption,
S, and performs interpolations based on S to determine kro
and krw. Simply speaking, Kro and Krw is not a function only
regarding Sw, but also related to the adsorption, S.
2. When nanofuids flooding is used after water flooding as ter-
tiary recovery method, early nanofuids injection will lead
to higher oil recovery, but with more nanoparticle loss. The
oil recovery factor decreased by 10% when injection time
increased from 0.45 PV to 2.25 PV with the total injection
period of 3 PV.
3. Higher injection rate of nanofuids could help improve the
flooding efficiency, but not the ultimate oil recovery for field
development. Also, it can cause more nanoparticle loss.
4. Brine water postflush is recommended when doing nano-

flooding. It can significantly improve the recovery of
nanoparticles, and for a homogeneous or heterogeneous
reservoir, oil recovery is better compared to water flooding.
5. The mathematical model used in this study is based on the
derivation from previous researchers [7, 16]. To simplify
the flow model of nanoflooding, they both neglected the
capillary pressure between fluid and oil, and it’s the same
in this study (Eq. 2.9). However, this is not accurate consid-
ering the wettability alteration and ITF reduction caused
by nanoparticles, which is experimentally proved [15]. So
for the next step of this study, a mathematical model con-
sidering the capillary pressure will be derived and the new
simulation results will be compared with this one to find
out the influence of capillary pressure on oil recovery.
References
1. M. Prodanovic, S. Ryoo, A. R. Rahmani, R. Kuranov, C. Kotsmar, T.E.
Milner, K.P. Johnston, S.L. Bryant, and C. Huh, Effects of magnetic field on
the motion of multiphase fluids containing paramagnetic particles in porous
media. SPE 129850 presented at SPE Improved Oil Recovery Symposium,
Tulsa, OK (2010).
2. B. Ju, S. Dai, Z. Luan, T. Zhu, X. Su, and X. Qiu, A study of wettability and
permeability change caused by adsorption of nanometer structured polysili-
con on the surface of porous media. SPE 77938 presented at SPE Asia Pacific
Oil and Gas Conference and Exhibition, Melbourne, Australia (2002).
3. M.K. Chaudhury, Complex fluids: Spread the world about nanofuids. Nature
423(10), 131–132 (2003).
4. D.T. Wasan and A.D. Nikolov, Spreading of nanofuids on solids. Nature 423,
156–159 (2003).
5. P.L.J. Zitha, Smart fluids in the oilfields. Exploration & Production: The Oil &
Gas Review 66–68 (2005).
6. I. N. Evdokimov, N. Y. Eliseev, A. P. Losev, and M. A. Novikov, Emerging
petroleum-oriented nanotechnologies for reservoir engineering. SPE 102060
presented at SPE Russian Oil and Gas Technical Conference and Exhibition,
Moscow, Russia (2006).
7. B. Ju and T. Fan, Experimental study and mathematical model of nanoparti-
cle transport in porous media. Powder Tech. 192, 195–202 (2009).
8. H. Lecoanet, J. Bottero, and M. Wiesner, Laboratory assessment of the mobil-
ity of nanomaterials in porous media. Environ. Sci. Technol. 38, 5164–5169
(2004).
9. S.W. Jeong and S.D. Kim, Aggregation and transport of copper oxide
nanoparticles in porous media. J. Environ. Monit. 11, 1595–1600 (2009).
10. F. He, M. Zhang, T. Qian, and D. Zhao, Transport of carboxymethyl cellu-
lose stabilized iron nanoparticles in porous media: Column experiments and
modeling. J. Colloid Interface Sci. 334, 96–102 (2009).
11. N. Saleh, H. Kim, T. Phenrat, K. Matyjaszewski, R.D. Tilton, and G.V. Lowrey,
Ionic strength and composition affect the mobility of surface-modified Fe0
nanoparticles in water-saturated sand columns. Environ. Sci. Technol. 42,
3349–3355 (2008).
12. A. Tiraferri and R. Sethi, Enhanced transport of Zerovalent Iron nanoparti-
cles in saturated porous media by Guar Gum. J. Nanopart. Res. 11, 635–645
(2009).
13. M.J. Murphy, Experimental analysis of electrostatic and hydrodynamic forces
affecting nanoparticle retention in porous media, MS Thesis, The University
of Texas at Austin (2012).
14. H. Yu, Transport and retention of surface-modified nanoparticles in sedi-
mentary rocks. PhD Dissertation, The University of Texas at Austin (2012).
15. M. Parvazdavani, M. Masihi, and M.H. Ghazanfari, Investigation of the effect
of water based nano-particles addition on hysteresis of oil and water relative
permeability curves. SPE International Oilfeld Nanotechnology Conference,
Noordwijk (2012).
16. M.F. El-Amin, A. Salama, S. Sun, and K. Abdullah, Modeling and simula-
tion of nanoparticles transport in a two-phase flow in Porous Media. SPE
154972, SPE International Oilfield Nanotechnology Conference, Noordwijk,
The Netherlands, June 12–14 (2012).
17. Louisiana Optical Network Initiative. Last updated Tuesday, April 15, 2014,
http://www.loni.org/ (accessed November 17, 2015).
18. ParaView. www.paraview.org (accessed November 17, 2015).
19. PETSc. http://www.mcs.anl.gov/petsc (accessed November 17, 2015).
20. T. Zhang, Modeling of nanoparticle transport in porous media. PhD disser-
tation, The University of Texas at Austin, August (2012).
21. T. Tosco and R. Sethi, MNM1D: A numerical code for colloid transport in
porous media. Implementation and validation. Am. J. Env. Sci. 5, 517–525
(2009).
22. J. Kou and S. Sun, On iterative IMPES formulation for two-phase flow with
capillarity in heterogeneous porous media. Int. J. Numer. Anal. Modeling 1,
20–40 (2010).
23. S.K. Das, S.U.S. Choi, W. Yu, and T. Pradeep, Nanofuids Science and Technology,
John Wiley & Sons, Inc. Publishing, Hoboken, NJ, ISBN 0470074736 (2008).
23. M.A. Sbai and M. Azaroual, Numerical modelling of formation damage by
two-phase particulate transport processes during CO2 injection in deep het-
erogeneous porous media. Adv. Water Resour. 34, 62–82 (2011).
20
3D Seismic-Assisted CO2-EOR
Flow Simulation for the Tensleep
Formation at Teapot Dome, USA
Payam Kavousi Ghahfarokhi1*, Thomas H. Wilson1 and Alan Lee Brown2
Department of Geology and Geography, West Virginia University

1
2
Schlumberger NExT, Director of Geoscience Training for NAM, San Felipe,
Houston, USA
Abstract
The Tensleep Formation at Teapot Dome, Wyoming is a naturally fractured res-
ervoir that has been studied for CO2-EOR. A realization of fractures is generated
using a discrete fracture network (DFN) approach. The model DFN has three frac-
ture sets: a dominant, high intensity N76°W set and two lower-intensity sets that
strike N28°W and N75°E. The fracture intensity in the model DFN is extracted
from a most negative curvature dissimilarity attribute calculated from the post-
stack 3D seismic data. In addition, probable deformation bands are interpreted
from the fracture intensity seismic attribute and incorporated in the reservoir
model as permeability barriers parallel to N40°E striking S1 fault. The gener-
ated model DFN is plugged into a dual porosity compositional reservoir simu-
lator (Eclipse 300) and adjusted to obtain a production history match. A three
parameter Peng-Robinson equation of state (EOS) is calibrated against the CO2
swelling tests laboratory data to obtain accurate phase behavior for CO2-EOR in
the Tensleep reservoir. Two CO2-EOR models are evaluated. CO2 is injected in
the B1 Tensleep Sandstone from the end of the production history (December,
2005) for 3 years at an injection rate of 1 MMSCF/day. The first model includes
three horizontal injectors parallel to the main fracture set and the second model
has three horizontal injection wells perpendicular to the dominant fracture set.
Although immediate CO2 breakthrough is observed in both models, model 1
yields a higher oil recovery and a lower CO2 mole fraction at the surface. The oil
recovery improvement is less than 700 MSTB for both models. Finally, we show
463
that ultimate CO2-EOR recovered oil can vary up to 150 MSTB and depends on
how impermeable the permeability barriers are.
Keywords: Naturally fractured reservoirs, flow simulation, post stack 3D seismic

data, CO2-EOR, DFN, curvature, tensleep, teapot dome
20.1 Presentation Sequence

This paper starts with an introduction to fractured reservoir simula-
tion and CO2-EOR processes. A geological setting section introduces
backgrounds about the Tensleep Formation, the aquifer, and the sealing
formation. Next the static modeling section covers fracture framework
and petrophysical modeling. Subsequently, a PVT (pressure, volume,
and temperature) section covers the steps needed to develop an equa-
tion of state (EOS) for the compositional simulation. Applying a process
termed streamline simulation in this section includes visualization of
streamlines in naturally fractured Tensleep Formation. These processes
lead to creation of two CO2-EOR models which are different in injection
patterns. Finally, a conclusions section summarizes the key findings of
this research.
20.2 Introduction
Naturally fractured reservoirs (NFR) are challenging to model, since
fracture networks are often complex and incorporate fluid transfer from
matrix to the fractures. High permeability fractures typically cause major
fluid flow in fractures rather through the matrix block. However, they can
also act as flow barriers when mineralized or through infilling of fine-
grain materials in formerly open fractures. Wireline image logs, out-
crop studies, core observations, and seismic data analysis provide direct
and indirect measurements of fracture properties at different scales.
Observations made at these different scales must be integrated to pro-
vide a representative model of the reservoir DFN [1–4]. Fracture spacing,
aperture, and length distributions can be incorporated in the reservoir
model using wireline image logs, core data, and field studies [2, 5]. While
uncertainties remain, these models are consistent with observations made
at a variety of scales.
Dual porosity (DP) models have been widely used to simulate the
fluid flow between the fractures and matrix medium [6–9]. Dual porosity
3D Seismic-Assisted CO2-EOR Flow Simulation 465
models incorporate coupled fow continua: one for the matrix and the
other for the fracture network. Fluid flow through the fracture network
along with flow from the matrix into the fracture network can be modeled
with a transfer function referred to as the shape factor [7, 8, 10, 11]. More
recently, discrete fracture models (DFM) have been employed to simulate
the flow in 2D and 3D fractured media [12–21]. Fracture permeability
upscaling is necessary to model the fluid flow in fractured reservoir using
a dual porosity model. Oda [22] algorithm calculates a fracture tensor,
which only considers the geometry of the fractures (aperture, size and
orientation). Oda’s approach is often used as a calculation technique in
commercial software as a fast algorithm to calculate fracture permeability
tensors. However, this method assumes that all fractures are connected
and overestimates effective permeability [17, 18, 23–25]. Oda’s permea-
bility tensor can be corrected using a flow-based numerical method by a
linear transformation.
We use a DFN along with the corrected Oda’s method to simulate pro-
ductions in a fully compositional dual porosity simulator for the Teapot
Dome oilfield, Wyoming (Figure 20.1). The history matched reservoir
model is then used in a compositional simulator to model CO2-EOR.
CO2-EOR can increase ultimate oil recovery by 7–23% of original oil in
place [26]. CO2-EOR efficiency significantly depends on optimization of
the injection process which requires a thorough understanding of reser-
voir heterogeneities and phase behaviors [27]. Injected CO2 usually follows
less resistant flow paths such as fractures or high permeability channels
to production wells. This effect is intensified by the low viscosity of CO2
at reservoir conditions. Although optimum well location relative to reser-
voir heterogeneities increases sweep efficiency, early breakthrough cannot
be avoided in most cases, just postponed [27]. CO2 and oil have multiple
contact miscibility. It is a dynamic fluid-mixing process in which CO2 ini-
tially mixes with the oil and makes it lighter. This process continues by
vaporization of oil components into CO2 rich phase. Exchanges of com-
ponents continue until no interfaces remain between the CO2-enriched
and oil-enriched CO2 phases [26]. Although miscibility is a function of
reservoir temperature and pressure, only pressure controls the miscibility
in isothermal reservoirs [28].
Minimum miscibility pressure (MMP) refers to a pressure at which
miscibility occurs. It can be achieved by injecting CO2 at a higher pres-
sure than MMP or enriching the gas with intermediate-weight hydrocar-
bons [27]. Injected CO2 displaces oil, develops miscibility, swells oil, and
reduces oil viscosity [29]. CO2-EOR sweep efficiency can be greatly influ-
enced by well pattern, well type, injection rate, reservoir heterogeneity,
2000 ft
790000 800000
976000 976000
972000 972000
968000 968000
Depth subsea (ft. )

964000 964000 –100
–200
–300
–400
–500
400 ft
802000 (a) 960000 –600
–700
–800
–900
955600 955600 –1000
956000 –1100
–1200
955200 955200 lt –1300
f au –1400
S1 952000
954800 954800
954400 954400 948000
954000 954000
944000
953000 953000
800000
953200 953200 N
952800 952800
952400 952400
952000 952000
951000 951000
951200 951200
802000
Log of cumulative oil production
1.00 2.00 3.00 4.00 5.00 6.00
(b)
Figure 20.1 (a) The Tensleep horizon interpreted from 3D seismic data shown in subsea
depth (ft.) (b) Contour map of cumulative oil production (bbl) from 1978 to 2004 in
Section 10 overlaid by the reservoir grid for flow simulation. The gridding size is 300 ft by
300 ft and it is approximately oriented parallel to the main fracture set (N76°W). Note that
suffix “-TPX-10” is removed from the well names for visualization.
phase behavior, along with mass transfer of components between gas

and oil phases, density contrast between phases, MMP, and many other
parameters [27, 30–33]. These complexities make CO2-EOR difficult to
accurately simulate.
Wadleigh [34] implemented a CO2-EOR simulation for RMOTC (Rocky
Mountain Oil Testing Center) during the planning of CO2-EOR pilot tests
in the Tensleep reservoir at Teapot Dome. He used a Peng-Robinson [35]
equation of state (EOS) and calibrated it against the swelling test data from
Hycal laboratory data [36]. The model has constant porosity and permea-
bility for each layer.
Friedmann and Stamp [37] used this model and predicted 30–40% oil
recovery improvement by CO2-EOR. Gaviria (2005) developed a model
with variable fracture permeabilities (20000–40000 mD) derived from
fracture apertures obtained from CT scans of cores. He showed that a
pseudo-miscible black oil fluid model could not represent the CO2-EOR
and sequestration project at Teapot Dome and noted that a fully com-
positional simulation is necessary. In addition, fully compositional his-
tory matching was not achieved due to simulation errors. Chiaramonte
[28] implemented a fully compositional simulation using Wadleigh [34]
EOS along with stochastic petrophysical properties for matrix, with the
addition of constant fracture porosity and permeability for each layer.
Fracture permeabilities have lower values perpendicular to the main frac-
ture set which strikes at N76°W. She observed CO2 breakthrough after
16 days when injecting 1 MMSCF/day for 6 weeks in well 44-1-TPX-10
while producing at 56-TPX-10, 43-2-TPX-10, 55-TPX-10, 63-TPX-10,
and 75-TPX-10 in vicinity of 44-1-TPX-10 injector well (refer to Figure
20.1 for well locations). A well control strategy, including shutting in the
wells until the pressure rises, was applied and yielded a 28% oil recovery
improvement within 5 years.
This research is different from previous works mainly in two aspects:
The seismic-assisted fracture intensity attribute introduced by Kavousi
and Wilson [3] is used to generate stochastic discrete facture networks
for dual porosity flow simulation. Additionally, two CO2-EOR scenar-
ios are compared: one with lateral injectors parallel to the dominant
fracture set, and the other with laterals perpendicular to the dominant
fracture set.
20.3 Geological Background

Teapot Dome Field is part of the previously designated Naval Petroleum
Reserve, known as NPR-3. Teapot Dome was owned and operated by
the U.S. Department of Energy and Rocky Mountain Oilfield Testing
Center (RMOTC) from 1977 through 2014. In 2014, the field was
placed on the market and sold in early 2015 to Stranded Oil Resources
Corporation.
The Pennsylvanian Tensleep reservoir in the southern part of Teapot
Dome has been estimated to contain 3.8 million bbl (0.6 million m3) oil
of 32° API gravity and 11 MMscf (0.31 million m3) of natural gas to date.
More than 1.8 million bbl (0.29 million m3) oil and 170 million bbl. (27
million m3) of water have already been produced [37]. A residual oil satu-
ration between 29% and 56% in the Tensleep cores from Teapot Dome is
observed [28].
The aeolian depositonal framework of the Tensleep Formation in Teapot
Dome is composed of shallow marine carbonates, aeolian sandstones, and
extensive beds of low-permeability dolomicrites [29, 38]. The Tensleep
Formation has an average matrix porosity of 8% (1%–19%) and permea-
bility of 80 mD (0–110 mD) [37].
Teapot Dome includes a stratigraphic interval spanning from the
Cretaceous to the Pre-Cambrian. The Pre-Cambrian basement is composed
of granite. The Mississippian age Madison Limestone is a regional aquifer
–500 Alcova limestone

W E Triassic
Chugwater formation
Permian Goose egg formation
Tenseep
2-way time (ms)
Pennsylvanian sandstone
–1000
Alco
va L Amsden formation
Tens s.
leep
Ss. Madison limestone
Mad Mississippian Regional aquifer
ison
Prec LS. and water drive
amb
rian Englewood formation
–1500 gran
ite Devonian
Fremont canyon Ss.
1 km Precambrian Granite
(a) (b)
Figure 20.2 (a) West–east seismic section across the Teapot Dome. The reverse fault
(shown by black line) has a southeast dip. (b) Stratigraphic column Precambrian-Triassic
strata. The Madison Limestone is the regional aquifer that provides the water drive for the
Tensleep reservoir, and the Goose Egg Formation is the seal (taken from Wilson et al. [2]).
that provides the water drive for the Tensleep reservoir. The Chugwater
and Permian Goose Egg formations (a mixture of shale, siltstone with
some sandstone) seal the Pennsylvanian Tensleep reservoir (Figure 20.2).
20.4 Discrete Fracture Network (DFN)

A model discrete fracture network (DFN) is generated using the fracture
aperture and length distributions proposed by Wilson et al. [2] and a frac-
ture intensity attribute introduced by Kavousi and Wilson [3]. Wilson
et al. [2] used wireline image logs and showed that fracture apertures are
log-normally distributed with the mean log electrical aperture of –1.805
(corresponding to 0.01568 mm). Ferguson [39] conducted a computerized
tomography (CT) scan of two cores retrieved from well 48-X-28 at depth of
5565 ft (Core A) and 5566 ft (Core B). They reported open and mineralized
fractures (filled with high density minerals such as crystalline dolomite
in both cores). The measured standard deviation of apertures observed in
Core A and Core B are presented in Table 20.1. Fracture aperture distribu-
tion in cores are not measured under overburden pressure and show larger
values than those measured in wireline image logs (Figure 20.3).
Kavousi and Wilson [3] used a minimum dissimilarity of most negative
curvature attribute extracted from post stack 3D seismic data to generate
a fracture intensity attribute. The designed fracture intensity is upscaled to
the reservoir grid for each zone (Figure 20.4). Fracture intensity average for
all zones varies from 1 fracture per foot to 1 fracture per 10 ft. These two
limits are the maximum and minimum fracture intensities reported for the
Tensleep Formation by Chiaramonte [28]. She assigned a constant fracture
intensity of 1 fracture every 10 ft for the A Sand, 1 fracture every 10 ft for
the B Sand, 1 fracture every 3 ft for the B Dolomite, and one fracture per
feet for the C Dolomite and C Sand. These values were assigned based on
Lorenz and Cooper [40] observations from core data of well 48-X-28.
Table 20.1 The mean and standard deviation of apertures observed in

the CT of Core A and B from well 48-X-28.
Sample Mean [mm] Standard deviation [mm]
Core A 0.222 0.577
Core B 0.537 0.757
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2

% 4 8 12 16 20 24 24 32 36 40 44 48 52
96 %
40
5 10 15 20 25 30 35 40 45 50 55 60 96
10
8 30
88 88
6
20
4
80 10
80
2
0
0 0 0.05 0.1 0.15 0.2 0.25 0.3
72 0 0.07 0.14 0.21 0.28 0.35 0.42 0.49
Wireline_Image_Log_Apertures
72
Core_A_Apertures
64 % 4 8 12 16 20 24 24 32 36 40 44 48 52 64
10
8
56 56
6
48 2 48
0
0 0.35 0.7 1.05 1.4 1.75 2.1 2.45
(%)
Core_B_Apertures
40 40
32 32
24 24
16 16
8 8
0 0
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
Aperture (mm)
Wireline_image_log_Aper tures Core_B_aper tures

Core A aper tures
Figure 20.3 Apertures are log-normally distributed in the wireline image logs, Core A,
and Core B. Note apertures measured in Core A and Core B are under no overburden
pressure.
% 2 4 6 8 10 12 14 16 18 20 22 24 26 % 24 4 6 8 10 12 14 16 18 20 22 24 26 28
9 13
8 A Sand 12
B Dolomit e
11
7 10
6 9
8
5
7
4 6
5
3
4
2 3
2
1
1
0 0
0.4 0.6 0.8 1 0.4 0.6 0.8 1
% 2 4 6 8 10 12 14 16 18 20 22 24 26 28 % 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30
6.5
11
6
10 B Sand Aquifer
9 5
8 4.5
7 4
6 3.5
5 3
2.5
4
2
3
1.5
2 1
1 0.5
0 0
0.3 0.4 0.5 0.6 0.7 0.8 0.9 0.3 0.4 0.5 0.6 0.7 0.8 0.9
Figure 20.4 Distributions of the fracture intensity attribute for each zone in the grid.
Note that the B Sand fracture intensity histogram consists of the B1 Sand and B2 Sands.
The aquifer includes C Dolomite and C sand.
Kavousi and Wilson [3] qualitatively related the fracture intensity attri-
bute to the production history of the field. Wells located on NW seismic
discontinuities are in general more productive than wells located on NE
trending discontinuities. NW discontinuities (Figure 20.5) are interpreted
as possible open fracture zones which mainly include N76°W fracture set.
Wilson et al. [2] showed that the dominant fracture set (N76°W) is roughly
parallel to the present-day maximum horizontal compressive stress (Shmax)
inferred from drilling induced fractures. They also interpreted permeabil-
ity barriers roughly parallel to the S1 fault.
In this study, NE seismic discontinuities (Figure 20.5) are interpreted as
possible low-permeability deformation bands. Chiaramonte [28] showed
that a 3300 psi reservoir pressure increase is necessary to make the S1 fault
SH
ma
x
(a) (b)
0 1000 2000 3000 ft
N Low high
Figure 20.5 (a) Possible deformation bands are shown by red arrows, note that they are
parallel to sub-parallel to the S1 fault. (b) The fracture intensity attribute introduced by
Kavousi and Wilson.2 Wells located on the NW discontinuities (yellow arrows) are more
productive than wells on the NE discontinuities (red arrows).
72-TpX-10
62-TpX-10
52-1-TpX-10
53-TpX-10
73-TpX-10
63-TpX-10
43-TpX-10
43-2-TpX-10
54-TpX-10
74-CMX-10-WD
44-1-TpX-10 75-TpX-10
55-TpX-10
76-TpX-10
46-TpX-10
56-TpX-10
i
67-1-X-10
0 1000 2000 3000 ft
Figure 20.6 The permeability barriers are roughly parallel the S1 fault. The grids j
direction is almost parallel to the dominant fracture set (N76°W). The grid block size is
300 ft by 300 ft. Note Kxx = Ki and Kyy = Kj.
leaky in the Tensleep Formation at Teapot Dome. We assume the same

threshold pressure along the S1 fault and along the interpreted permeabil-
ity barriers roughly parallel to the S1 fault during the flow simulation. This
threshold pressure is included in flow simulation as the maximum bottom
hole pressure for injectors.
The interpreted permeability barriers are incorporated into the static
model of the Tensleep reservoir (Figure 20.6).
20.5 Petrophysical Modeling

Chiaramonte [28] used core data of 8 wells which penetrated Tensleep
Formation at Teapot Dome for reservoir petrophysical modeling: 11-AX-
11, 43-TPX-10, 44-1-TPX-10, 54-TPX-10, 55-TPX-10, 56-TPX-10,
62-TPX-10, and 71-AX-15 (Figure 20.7).
In this study we assign the mean values of porosity and permeability
from the cores for each zone as (Table 20.2).
20.6 PVT Analysis

Oil production from the Tensleep reservoir at Teapot Dome consists of a
sulfurous saturated black oil. The fluid properties are summarized in Table
20.3.
Only one oil sample is available with swelling tests by Hycal Energy
Resource Laboratories which shows oil volume and viscosity variations
with amount of CO2 injected at a reservoir temperature of 190.4 °F [36].
The sample is from well 72-TPX-10 taken in 2004. It has very low amounts
of light components (Methane, Ethane and propane) and abundant heavy
components up to C30+. This could explain the low bubble point pressure
between 40–70 psi and very low GOR of 4 SCF/STB (Table 20.4) [41].
A three parameter Peng-Robinson equation of state (EOS) is necessary
to analytically predict oil phase behavior during CO2 injection process
[35]. Gaviria [41] split C6+ components of Sample A into five pseudo-com-
ponents C6-C12, C13-C19, C20-C27, C28-C29 and C30+. The resultant Peng-
Robinson Equation of State (EOS) has 13 components. The simulation
speed depends on number of components in EOS [41]. We performed a
lumping on Gaviria’s EOS components to fit an 8 component EOS to the
laboratory data. The 8 component EOS is presented in Table 20.5.
X-axis
800000 802000 804000 806000
958000 958000
956000 956000 Subsea depth (ft.)

–50.00
–100.00
–150.00
–200.00
–250.00
954000 954000 –300.00
Y-axis –350.00
–400.00
–450.00
–500.00
–550.00
–600.00
952000 952000 –650.00
950000 950000
800000 802000 804000 806000
Figure 20.7 Locations of 8 Wells with core data used in Chiaramonte (2009) reservoir
modeling study are shown on Tensleep structure. The black polygon over the A Sand
horizon shows areal extent of high oil production zones. Subsea depth and coordinates are
in feet.
The 8 component EOS parameters are tuned to fit the swelling test data
from well 72-TPX-10 oil sample. The resulting EOS properly outputs the
oil sample response to the swelling tests (Figure 20.8). The swelling tests
show that oil responds to CO2 injection with volume increase and viscos-
ity reduction. We modeled the liquid viscosity at each swelling test step
and compared that with laboratory results (Figure 20.9). The 8-component
EOS simulates the liquid viscosity reductions with a reasonable error at
each swelling test step.
Table 20.2 Matrix porosity (ft3/ft3), permeability (mD.), and average cell thickness (ft.) for the
Tensleep static model (modified from Chiaramonte [28]).
Horizontal Vertical Number of Average cell
Zone Porosity% permeability permeability the layers thickness
A Sand 4.74 3.69 1.11 2 23
B Dolomite 3.03 0.21 0.10 1 26.29
B1 Sand 10.54 29.31 6.14 1 19
B2 Sand 6.95 3.45 1.05 2 26.25
Aquifer 20 10 2.53 8 50
Table 20.3 Oil properties of the Tensleep reservoir at Teapot

Dome, Wyoming [41].
Gas-oil ratio 4 SCF/STB
Initial reservoir pressure at -110ft subsea 2300 psi
Oil Gravity 31° API
Bubble point pressure 40–70 psi
Oil viscosity at 60 °F and 42psi 3.5 cp
Minimum Miscibility Pressure (MMP) 1300 psi
Table 20.4 The oil sample composition at the surface

condition [41].
Component Sample mole fraction
CO2 0.08
N2 0.13
C1 0.02
C2 0.12
C3 0.17
i-C4 0.08
n-C4 0.22
i-C5 0.15
n-C5 0.3
C6 1.29
C7+ 97.44
Mole Weight C7+ 303.85
Density C7+ at 60 °F, gr/cm3 0.8972
Temperature(°F) 190.4
Table 20.5 The 8-components Peng-Robinson EOS parameters.
Component Mw PC (atm) Tc (°K) Omega A Omega B Acentric factor Vc (lb./mole)
CO2 44.0 72.8 304.2 0.4572 0.0779 0.225 0.094
C1-C3 22.6 46.1 215.1 0.4572 0.0779 0.048 0.10961
C4-C5 93.9 34.0 458.1 0.4572 0.0779 0.478 0.2868
C6-C12 157.4 25.9 566.1 0.4572 0.0779 0.556 0.4725
C13-C19 215.04 15.6 734.2 0.4572 0.0779 0.558 0.8194
C20-C27 467.12 14.4 809.8 0.4572 0.0779 0.790 1.119636
C28-C29 518.9 9.3 810.2 0.4572 0.0779 0.948 1.306902
C30+ 632.3 5.3 838.9 0.4572 0.0779 1.201 1.888011
CO2
Swelling Calc.-After regression
3000.0 1.50
2800.0 1.40
2600.0 1.30
2400.0 1.20
Saturation pressure (psia)
2200.0 1.10
2000.0 1.00
1800.0 0.90
Swelling factor
1600.0 0.80
1400.0 0.70
1200.0 0.60
1000.0 0.50
800.0 0.40
600.0 0.30
400.0 0.20
200.0 0.10
.0 0.00
0.000 0.100 0.200 0.300 0.400 0.500 0.600 0.700 0.800
Composition (mol fraction)
Legend
Psat S.F. Exp.Psat Exp.S.F.
Figure 20.8 Saturation pressure and swelling factor calculated in the swelling
experiments and model EOS are plotted against mole fraction CO2 added to the oil sample
from well 72-TPX-10 at 190.4°F. The saturation pressure increases as higher mole fraction
of CO2 is added to the fluid sample.
Model viscosity vs. swelling tests at 190.4°F

4.5 80
4 70
3.5
60
3
50
Solvant CO2 (model%
)
Fluid viscosity (cp)
2.5
40
2
30
1.5
20
1
0.5 10
0 0
0 500 1000 1500 2000 2500 3000
Bubblepoint pressure (psia)
Swelling test viscosity (experiment) Viscosity (model) Solvant CO2 (mole %)
Figure 20.9 Fluid viscosity is plotted vs. bubble point pressure at each swelling test step.
The modeled viscosity reduction curve (orange line) is close to the swelling test viscosity
results (blue line).
20.7 Streamline Analysis

We undertook a streamline simulation [42–44] to visualize flow paths in
the history matched model. A streamline is defined as flow path or flow
direction. It is a vector that varies from point-to-point with velocity along
the flow path. The matched model streamlines show flow patterns per-
turbed by permeability barriers. Streamlines do not cross permeability
barriers, which shows, as designed, the reservoir does not communicate
across the permeability barriers (Figure 20.10).
20.8 CO2-EOR
The 8-component EOS is used in a compositional simulator (Eclipse 300)
to model CO2-EOR in the Tensleep reservoir at Teapot Dome, Wyoming.
Two models are designed for CO2-EOR using three horizontal wells within
72
63 73
43 43-2 Water saturation
54 1.0000
0.9500
0.9000
0.8500
44 75 0.8000
55 0.7500
0.7000
76 0.6500
0.6000
0.5500
0.5000
56
67
0 1000 2000 3000 ft
Figure 20.10 Streamline analysis on 03-28-2004 for water saturation around producing
wells. The streamline simulator assigns the streamlines to the center of the grids where
wells are located. The model mainly produced water. Well 56-TPX-10 is the most oil
prolific well in this field. Note how streamlines change around permeability barriers.
the B1 sandstone (Figure 20.11). In model 1 (denoted as m1), the injec-

tors are parallel to the dominant fracture set (N76°W); and in model 2
(m2), injectors are perpendicular to the dominant fracture set (N14°E).
Alternatively, this puts the m1 and m2 laterals perpendicular and parallel,
respectively to the flow barriers.
Models 1 and 2 have wells which inject 1000 MSCF/day (333.33 Mscf/
day per injector) from the end of the production history data (December
01, 2005) to January 01, 2009, a 26 month injection period. This injection
rate is proposed by Chiaramonte [28]. The injectors’ bottom hole pressure
constraint is 3300 psi, equivalent to the threshold pressure for seal integrity
of S1 fault [28]. The bottom hole pressure (BHP) for all the producers is set
to 500 psi [28].
The two models produce significantly different results under CO2
injection. In m2 breakthrough occurs earlier than m1 since injection
is directly into the dominant fracture set along the j direction (Figure
20.12). However, the dominant fracture set accelerates CO2 break-
through in both models. Models 1 and 2 are compared to a case where
no CO2 is injected. Continued production in models 1 and 2 with EOR is
X-axis X-axis
802000 802000
955600 955600 955600 955600
955200 955200 955200 955200
954800 954800 954800 954800
954400 954400 954400 954400
954000 954000 954000 954000
953600 953600 953600 953600

Y-axis
Y-axis
Y-axis
Y-axis
953200 953200 953200 953200
952800 952800 952800 952800
952400 952400 952400 952400
952000 952000 952000 952000
951000 951000 951000 951000
951200 951200 951200 951200
802000 802000
X-axis X-axis
(a) (b)
Figure 20.11 (a) Model 1 has injectors (I4, I5, and I6) parallel to the main fracture set
(N76°W). (b) In model 2, Injectors (I7, I8, and I9) are perpendicular to the dominant
fracture set. Each grid block is 300 ft by 300 ft.
Field
Jan 2006 Jul 2006 Jan 2007 Jul 2007 Jan 2008 Jul 2008 Jan 2009
8000
400 600 800 1000 1200 1400
Injection rate for m1 and m2
6000
Gas flowrate (MSCF/d)
Liquid flowrate (STB/d)

4000
m1 gas production m2 gas production
2000
m1 oil production
200
m2 oil production
0
0
Figure 20.12 A high mole fraction of CO2 is observed at the producers beginning almost
immediately after injection in model 2. Model 1 has a higher oil production and lower
gas production rates during the injection at 1000 Mscf/day. Both models suffer from early
breakthrough because of the fractures which act as easy flow conduits for CO2 to reach the
producers while bypassing the matrix blocks.
640 MSTB and 580 MSTB more, respectively, than the case with no CO2
injection (Figure 20.13).
We undertook a sensitivity analysis on the fault transmissibility mul-
tiplier (MULTFLT) for the CO2-EOR models. We consider two cases: all
permeability barriers are completely sealing (MULTFLT=0), and all per-
meability barriers are completely leaky (MULTFLT=1). Results show that
recovered oil varies up to 150 MSTB and depends on MULTFLT values.
The higher the MULTFLT, the higher the recovered oil (Figure 20.14).
The model studies show that early CO2 breakthrough limits sequestra-
tion effectiveness. Wilson et al. [2] noted that Tensleep fracture lengths
are much less than the spacing between laterals with the majority of frac-
ture lengths between 5 m and 15 m. Their outcrop studies also indicated
considerable variation of fracture intensity within Tensleep sequences
exposed in the Fremont Canyon area. The lower sequences were generally
less intensely fractured than those in the upper Tensleep. The LIDAR mea-
surements of Zahm and Hennings ([38], their Figure 20.8) indicate that
sequence bound fracture intensities are somewhat less in the lower lime-
stone and dolostone dominated sequences exposed in the Alcova anticline.
Sequence bound fracture intensities also appear to form vertical zones of
alternating lower and higher intensity that cut through the Tensleep [38]
Field oil production cumulative

2.4E+06 Jan 2006 Jul 2006 Jan 2007 Jul 2007 Jan 2008 Jul 2008 Jan 2009
2.4E+06
2.2E+06
2.2E+06
Oil production rate (STB/d)
2E+06
2E+06
1.8E+06
1.8E+06
Date
Model 1 Model 2 No_CO2INJ
Figure 20.13 The oil production for Models 1 and 2 are compared to an oil production
case with no CO2 injection. The recovery improvement of 640 MSTB (black curve) is
achieved in Model1 compared to no CO2 injection case (Blue curve).
Field oil production cumulative

2.2E+06 2.4E+06
2.4E+06
Liquid production volume (STB)
2.2E+06
2E+06
2E+06
1.8E+06
1.8E+06
Date
No_CO2INJ Model 1 Model 1 _FLT1 Model 2
Model 2 _FLT1
Figure 20.14 Increasing fault multiplier results in higher oil production for both CO2-
EOR models.
(Figure 20.8). Wilson et al. [2] also observed alternating vertical zones of
high and low fracture intensity in Fremont Canyon Tensleep exposures.
Seismic discontinuities observed in the Teapot Dome 3D seismic data set
are interpreted to represent zones of increased fracture intensity. Estimated
widths of these zones are less than twice the seismic bin spacing or approx-
imately 60 m with lengths that vary from about 200 m to 1000 m. Wilson
et al. [2] suggested that perforation into these zones should be avoided to
reduce breakthrough time. We also suggest that wells land in the lower,
less intensely fractured, Tensleep. Completion in the lower Tensleep away
from seismic discontinuities may reduce breakthrough time and increase
sequestration volume.
20.9 Conclusions
• Streamlined simulation shows that the dominant (N76°W)
fracture set has significant influence on flow orientation.
Two other low intensity sets do not significantly control flow
directions except for the highest producers when they allow
the well to access larger reservoir volume.
• CO2 tends to emerge in the producers with the start of injec-
tion for both models: at a higher rate in model 2 where the
injectors are perpendicular to the dominant fracture set
(N76°W) and at a lower rate in model 1 where the injectors
are oriented parallel to the dominant fracture set.
• Horizontal wells parallel to the main anisotropy axis
resulted in a higher sweep efficiency in model 1 compared
to model 2.
• A threshold pressure of 3300 psi for reservoir sealing was
assumed based on work of Chiaramonte [28], however, at
an injection rate of 1000 MSCF/day, this pressure was never
reached in either model. Thus, the sealing properties of the
S1 fault and other permeability barriers is not compromised
during CO2-EOR.
• We suggest that wells be completed in the lower, less intensely
fractured, Tensleep and that perforations avoid zones of seis-
mic discontinuity to reduce breakthrough time and increase
sequestration volume.
Acknowledgement
I appreciate Rocky Mountain Oil Testing Center for providing me with
3D seismic and log data for Teapot Dome. The use of Eclipse and Petrel
from Schlumberger, and CMG from Computer Modelling Group is greatly
appreciated along with the insights on how to use the software from Dr.
Alan Lee Brown from Schlumberger and adjunct Professor at West Virginia
University; these software resources made the analysis presented in this
paper possible. I also thank Vicki Stamp for providing me with PVT report
of oil samples. Production data were sorted and formatted for Petrel by
Valerie Smith from West Virginia University.
References
1. A. Ouenes, T.C. Anderson, D. Klepacki, A. Bachir, D. Boukhelf, G.C.
Robinson, M. Holmes, B.J. Black, and V.M. Stamp, Integrated characteriza-
tion and simulation of the fractured Tensleep Reservoir at Teapot Dome for
CO2 injection design. SPE Western Regional Meeting (2010).
2. T.H. Wilson, V. Smith, and A. Brown, Developing a model discrete fracture
network, drilling, and enhanced oil recovery strategy in an unconventional
naturally fractured reservoir using integrated field, image log, and three-
dimensional seismic data. AAPG Bulletin. 99, 735–762 (2015).
3. P. Kavousi Ghahfarokhi and T.H. Wilson, Fracture intensity attribute for the
Tensleep reservoir at Teapot Dome, Wyoming, USA. Interpretation. 3, SZ41–
SZ48 (2015).
4. P.K. Ghahfarokhi, Reservoir Characterization and Flow Simulation for CO2-
EOR in the Tensleep Formation Using Discrete Fracture Networks, Teapot
Dome, Wyoming. Diss. West Virginia University (2016).
5. S.P. Cooper, Deformation Within a Basement-Cored Anticline: Teapot Dome,
Wyoming, Citeseer (2000).
6. G. Barenblatt, I.P. ZHELTOV, and I. KOCHINA, Basic concepts in the theory
of seepage of homogeneous liquids in fissured rocks [strata]. J. Appl. Math.
Mech. 24, 1286-1303 (1960).
7. J. Warren and P.J. Root, The behavior of naturally fractured reservoirs. Soc.
Petrol. Eng. J. 3, 245–255 (1963).
8. H. Kazemi, L. Merrill Jr., K. Porterfeld, and P. Zeman, Numerical simula-
tion of water-oil flow in naturally fractured reservoirs. Soc. Petrol. Eng. J. 16,
317–326 (1976).
9. L.K. Thomas, T.N. Dixon, and R.G. Pierson, Fractured reservoir simulation.
Soc. Petrol. Eng. J. 23, 42–54 (1983).
10. G. Penuela, F. Civan, R. Hughes, and M. Wiggins, Time-dependent shape
factors for interporosity flow in naturally fractured gas-condensate reser-
voirs. SPE Gas Technology Symposium (2002).
11. P. Sarma and K. Aziz, New Transfer Functions for Simulation of Naturally
Fractured Reservoirs with Dual Porosity Models. SPE J. 11, 328–340 (2006).
12. M. Karimi-Fard and A. Firoozabadi, Numerical simulation of water injec-
tion in fractured media using the discrete-fracture model and the Galerkin
method. SPE Reserv. Eval. Eng. 6, 117–126 (2003).
13. B. Gong, M. Karimi-Fard, and L.J. Durlofsky, Upscaling discrete fracture

characterizations to dual-porosity, dual-permeability models for efficient
simulation of flow with strong gravitational effects. SPE J. 13, 58–67 (2008).
14. A. Moinfar, W. Narr, M.H. Hui, B.T. Mallison, and S.H. Lee, Comparison of
discrete-fracture and dual-permeability models for multiphase flow in natu-
rally fractured reservoirs. In SPE Reservoir Simulation Symposium (2011).
15. Y. Ding, R. Basquet, and B. Bourbiaux, Upscaling Fracture Networks for
Simulation of Horizontal Wells Using a Dual-Porosity Reservoir Simulator.
SPE Reservoir Simulation Symposium (2005).
16. B. Bourbiaux, M. Cacas, S. Sarda, and J. SABATHIER, A fast and efficient
methodology to convert fractured reservoir images into a dual-porosity
model. SPE Annual Technical Conference and Exhibition (1997).
17. B. Dershowitz, P. Lapointe, T. Eiben, and L. Wei, Integration of Discrete
Feature Network Methods With Conventional Simulator Approaches. SPE
Reserv. Eval. Eng. 3, 165–170 (2000).
18. B. Decroux and O. Gosselin, Computation of effective dynamic properties
of naturally fractured reservoirs: Comparison and validation of methods.
EAGE Annual Conference & Exhibition incorporating SPE Europec (2013).
19. N. Cottereau, M.H. Garcia, O.R. Gosselin, and L. Vigier, Effective fracture
network permeability: Comparative study of calculation methods. SPE
EUROPEC/EAGE Annual Conference and Exhibition (2010).
20. M. Ahmed Elfeel, G. Couples, S. Geiger, and J. Ma, Upscaled multi-phase
flow properties of fracture corridors. SPE Caspian Carbonates Technology
Conference. Society of Petroleum Engineers (2010).
21. M.F. Lough, S.H. Lee, and J. Kamath, A new method to calculate effective
permeability of gridblocks used in the simulation of naturally fractured res-
ervoirs. SPE Reser. Eng. 12, 219–224 (1997).
22. M. Oda, Permeability tensor for discontinuous rock masses. Geotechnique
35, 483–495 (1985).
23. A. Gupta, G. Penuela, and R. Avila, An integrated approach to the deter-
mination of permeability tensors for naturally fractured reservoirs. J. Can.
Petrol. Technol. 40(12), 43–48, (2001).
24. P.K. Ghahfarokhi, The structured gridding implications for upscaling model
Discrete Fracture Networks (DFN) using corrected Oda’s Method. J. Petrol.
Sci. Eng. 153, 708–780 (2017).
25. J. Li, Z. Lei, G. Qin, and B. Gong, Effective Local-Global Upscaling of
Fractured Reservoirs under Discrete Fractured Discretization. Energies. 8,
10178–10197 (2015a).
26. P.M. Jarrell, C. Fox, K. Morgan, M. Stein, and S. Webb, Practical aspects of
CO2 fooding. Richardson, Tex.: Henry L. Doherty Memorial Fund of AIME.
Soc. Petrol. Eng. J. (2002).
27. L. Li, S. Khorsandi, R.T. Johns, and R.M. Dilmore, CO 2 enhanced oil recov-
ery and storage using a gravity-enhanced process. Int. J. Greenhouse Gas
Control. 42, 502–515 (2015b).
28. L. Chiaramonte, Geomechanical Characterization and Reservoir Simulation

of a Carbon Dioxide Sequestration Project in a Mature Oil Field, ProQuest,
Teapot Dome, WY, (2009).
29. L. Chiaramonte, M. Zoback, J. Friedmann, V. Stamp, and C. ZAHM, Fracture
characterization and fluid flow simulation with geomechanical constraints
for a CO2–EOR and sequestration project Teapot Dome Oil Field, Wyoming,
USA. Energy Procedia. 4, 3973–3980 (2011).
30. S. Khorsandi, K. Ahmadi, and R.T. Johns, Analytical solutions for gas dis-
placements with bifurcating phase behavior. SPE J. 19, 943–955 (2014).
31. U.G. Araktingi and F. Orr Jr., Viscous fingering in heterogeneous porous
media. SPE Advanced Technology Series. 1, 71-80 (1983).
32. F.T. Chung, R.A. Jones, and H.T. Nguyen, Measurements and correlations of
the physical properties of CO2-heavy crude oil mixtures. SPE Reservoir Eng.
3, 822–828 (1988).
33. A. Shahkarami, S. Mohaghegh, V. Gholami, A. Haghighat, and D. Moreno,
Modeling pressure and saturation distribution in a CO2 storage project
using a Surrogate Reservoir Model (SRM). Greenhouse Gases: Science and
Technology. 4(3), 289–315 (2014).
34. E.F. Wadleigh, RMOTC in-house reservoir simulation work product (2005).
35. D.-Y. Peng and D.B. Robinson, A new two-constant equation of state. Ind.
Eng. Chem. Fund. 15, 59–64 (1976).
36. Hycal Energy Reasearch Laboratories Ltd., DOE-RMOTC, Teapot Dome
Miscibility study, final report, 29 (2004).
37. S.J. Friedmann and V.M. Stamp, Teapot Dome: Characterization of a CO2-
enhanced oil recovery and storage site in Eastern Wyoming. Environmental
Geosciences. 13, 181–199 (2006).
38. C.K. Zahm and P.H. Hennings, Complex fracture development related to
stratigraphic architecture: Challenges for structural deformation prediction,
Tensleep Sandstone at the Alcova anticline, Wyoming. AAPG Bulletin. 93,
1427–1446 (2009).
39. D. Ferguson, Investigation of Efficiency Improvements during CO2 Injection
in Hydraulically and Naturally Fractured Reservoirs (2003).
40. J. Lorenz and S. Cooper, Analysis of fracture characteristics and distribution
of 48-X-28 well. Rocky Mountain Oilfield Testing Center Report (2004).
41. R. Gaviria Garcia, Reservoir Simulation of CO2 Sequestration and Enhanced oil
Recovery in Tensleep Formation, Teapot Dome Field, Texas A&M University,
Texas (2006).
42. M. Muskat and R. Wyckoff, A theoretical analysis of water-flooding net-
works. Transactions of the AIME. 107, 62–76 (1934).
43. R. Batycky, M.J. Blunt, and M.R. Thiele, A 3D field-scale streamline-based
reservoir simulator. SPE Reservoir Eng. 12, 246-254 (1997).
44. M.J. King and A. Datta-Gupta, Streamline simulation: A current perspective.
In Situ. 22, 91–140 (1998).
Part 7
NEW ADVANCES IN RESERVOIR
CHARACTERIZATION-MACHINE
LEARNING APPLICATIONS
21
Application of Machine Learning
in Reservoir Characterization
Fred Aminzadeh *
FACT Inc., Santa Barbara, CA, USA
Abstract
This chapter gives a brief overview of reservoir characterization, with a focus
on how the emerging artificial intelligence, machine learning and data ana-
lytics techniques can improve the ability to do reservoir characterization (RC).
Specifically, the ability to use AI to integrate different data types with different
Scale, Uncertainty, Resolution, and Environment, that we refer to it as the SURE
challenge is discussed. We show how AI offers a natural toolbox for reservoir prop-
erty estimation, and their uncertainties. Machine/Deep Learning-based methods
perform much like a human brain. They can receive variety of data from many
different sources with drastically different characteristics, and undertake neces-
sary evaluations and perception-based measures, and eventually make the right
decisions and/or solve complicated problems. Human intelligence (engineers,
geoscentists) will always have a superior performance with qualitative data than
computers that are better in dealing with quantitative data. Several examples on
how effective human-machine interfaces to create hybrid solutions to address dif-
ferent reservoir characterization problems are provided.
Keywords: Reservoir characterization, big data artificial intelligence (AI),
support vector machines, data analytics, neural networks, fuzzy logic
21.1 Brief Introduction to Reservoir Characterization

Characterizing a reservoir is the process of assessing reservoir conditions
using the data from various interdisciplinary sources: core samples, log data,
seismic survey, production data. The purpose of characterization is to assist
with delineating and describing a reservoir and its dynamics during produc-
tion and stimulation. Reservoir characterization and modeling are important
489
for optimizing field development and operation, and for reservoir valuation.
Dynamic reservoir characterization allows us to understand and predict the
changes in reservoir properties and to monitor its performance as hydrocar-
bons are produced from the reservoir. Characterization is also critical when
stimulating the reservoir for enhancing production. This is accomplished by
the analysis of data from a combination of different sources, to extract addi-
tional information about the in situ conditions of the reservoir, including
the formation temperature, pressure, and the properties of the oil, gas, and
brine. Other reservoir properties that can affect measured data are density,
hydrocarbon viscosity, stresses, and fractures.
As shown in Figure 21.1, an integrated reservoir characterization starts
with collecting data from geological, petrophysical, seismic and engineer-
ing data. A multi-disciplinary data analysis process creates a model which
includes characterizing the architecture, lithologies, facies, fluid composi-
tions, rock-fluid interactions, the geometry of the flow units and physical
rock properties such as porosities and permeabilities of flow layers. Three
properties are related to the pore space: porosity—the fraction of the entire
volume part occupied by pores, cracks and fractures; internal surface: the
magnitude of the surface of pores as related to the rock mass pore volume
and controls interface-effects at the boundary grain; pore fluid, permeabil-
ity: the ability to flow fluid through rock pores. Porosity and specific inter-
nal surface are scalar properties, permeability is a tensor. Given different
Geologic Petrophysical Seismic Engineering

Pressure
w1
w2
Time, days
Integrated Reservoir Characterization/Modeling
Uncertainty Analysis
Multiple Disciplinary
Analysis
Rate
Production profile Time

Model framework, Facies, Ø, k, Sw...
Figure 21.1 Different components of Reservoir Characterization [32].

Machine Learning in Reservoir Characterization 491
levels of uncertainty discussed under the SURE challenge in Part 1, our

estimates of reservoir properties should also be accompanied with their
respective levels of uncertainty, derived from calibration and extent of val-
idation with well data. Reservoir description is an iterative process and is
based on different components, as shown in Figure 21.1, from the input
data to the process (e.g., well data, seismic data, and production data).

Learning Review
In Part 1 of the book and other references cited in Part 1, including
Aminzadeh and Dasgupta [4] and Aminzadeh [6], we introduced the
concept of “SURE” challenge that is associated with the complications of
integration of different data types with vastly different Scale, Uncertainty,
Resolution and Environment characteristics. We maintained that advances
in various AI technology and their correct application to the E&P prob-
lems may provide the necessary toolbox to address the SURE challenge.
Before we go into various AI applications in reservoir characterization that
may pave the way to solve the SURE challenge, we will provide a general
overview of different aspects of AI and related fields. Some of the material
in this section is adapted from Aminzadeh et al. [3].
Computational Intelligence (CI) is the study of designing and develop-
ing computational paradigm inspired from biological and linguistical pro-
cesses. Traditionally the three main pillars of CI have been Neural Networks
(NNs), Fuzzy Systems and Evolutionary Computation. In fact, some of the
most successful AI systems are based on CI. Over the last few decades there
has been a renewed interest in NNs, largely driven by the pioneering work
by Rumelhart et al. [45] on back propagation. Nowadays, deep learning (the
commonly accepted term for NNs with multiple hidden layers) has become
the core method for artificial intelligence (AI). We recognize the term AI
implies a sense of reasoning and “intelligence” in algorithms. A lot of work
is being done to include causality and concept whitening focused on over-
coming the “black-box” nature of deep learning algorithms. These efforts
are being led by various leaders. However, for the purpose of this book, AI
is used as an umbrella term for deep learning, machine learning, computer
vision and other advanced data analysis techniques.
In this section, we provide a high-level introduction to typical machine
learning and deep learning models that have been used by authors of the
reviewed literature. Given how rapidly the field is growing, it is fair to say
that the material covered in this section is non-exhaustive.
21.2.1 Support Vector Machines

This technique has its roots in statistical learning theory and has shown
promising results in many practical applications, for both regression and
classification problems. SVM also works very well for high dimensional
data and avoids the curse of the dimensionality problem. SVM learning
can be formulated as a convex optimization problem, in which efficient
algorithms are available to find the global minimum of the objective func-
tion. Other classification methods, such as neural networks and rule-based
classifiers, employ a greedy-based strategy to search the hypothesis space;
such methods find only locally optimum solutions.
For the application in regression problems, Vapnik and Cortes [12]
extended the concept of maximum margin in SVM by developing a new
term, ε-insensitive loss function technique. In contrast with the Ordinary
Least Square, the objective function of SVR is to minimize the l2-norm of
the coefficient vector, not the squared error. The objective function and
constraints are as follows:
1
min || w ||2 +C ∑ni =1 | ξ i |
2 (21.1)
| yi − wi xi | ≤ ε + | ξ i |
In Eq. (21.1) the respective parameters are: y – target, w – coefficient,
x – predictor (feature), ξi -slack variable, ε – maximum error or acceptable
margin of error.
In case of a nonlinear problem, the learning task is performed on
the transformed vector Ø(x) instead of the original coordinate space x.
However, the mapping of x -> Ø(x), is a not a trivial task. To do this, Kernel
functions (KF) are used. KF calculate the relationship between observa-
tions as if they are in higher dimensions; KF don’t actually transform
them to higher dimensions. This is done to save the computation time.
Polynomial KF and Radial kernels, are two popular examples of KFs.
21.2.2 Clustering (Unsupervised Classification)

The aim of clustering is to find groups of observations based only on the
observations (or data) and their relationships. The objective is to group in
a manner that within-group relationship (or similarity or homogeneity)
is highest and group-group dissimilarity is also highest. Various forms of
clusters are illustrated in Figure 21.2.
A variety of factors need to be considered when selecting a clustering
technique. One of the key factors is the type of cluster for the intended
(a) Well-separated clusters. Each point is (b) Center-based clusters. Each point is
closer to all of the points in its cluster than closer to the center of its cluster than to
to any point in another cluster. the center of any other cluster.
(c) Contiguity-based clusters. Each point is (d) Density-based clusters. Clusters are
closer to at least one point in its cluster than regions of high density separated by
to any point in another cluster. regions of low density.
(e) Conceptual clusters. Points in a cluster share some general

property that derives from the entire set of points. (Points in the
intersection of the circles belong to both.)
Figure 21.2 Different types of clusters [52].
application. This section aims to summarize different type of clusters and

provides some examples of clustering methods for each type.
K-means clustering is an example of a prototype-based (center-based in
case of K-means) clustering technique that attempts to identify k (a user
defined number) of clusters. K-means defines a prototype in terms of a cen-
troid, which is usually the mean of a group of observations. Seldom does
the centroid correspond to a real observation in the data; when required
K-Mediod is used. A mediod is the most representative point of the group.
For assigning the points to a cluster, the most common similarity met-
ric is the Euclidean distance. However, when appropriate, other distance
functions (also known as proximity functions) measures such as cosine
similarity, and Manhattan, among others are used. The K-Means algorithm
attempts to minimize the inertia, as described in the equation below
inertia = ∑ni =0 min ( | xi − µi | ) (21.2)

In some engineering tasks, observations are not in a well-separated

group. Therefore, portioning observations into clean groups is not feasible.
In such cases use of fuzzy version of clustering makes most practical sense.
Fuzzy clustering techniques are based on fuzzy set theory. In fuzzy cluster-
ing, observations can have membership to multiple clusters. A good exam-
ple of fuzzy clustering is fuzzy c-means, which is very similar to k-means in
terms of the use case, with the following equations describing the centroid
(Cj) and inertia (sum of squared error)
SSE(C1, C 2.., Ck ) = ∑ kj =1 ∑mi =1 wijpdist ( xi , c j )2
Σ mi =1wijp xi (21.3)
cj =
Σ mi =1wijp
In fuzzy clustering (for example, fuzzy k-means) the assignment of any
point to more than one cluster is made through “membership function” or
degree of the belonging of the point to a particular cluster. As an example,
Figure 21.3 shows fuzzy clustering for segmentation of different fractures
represented by their corresponding microseismic event, in a geothermal
field in California. In this case cluster number 2 is divided to 4 overlapping
sub-clusters. The degree of belonging of all fracture to anyone of cluster
numbers 1, 2-1,2-2, 2-3, 2-4 and 3 is given by their respective fuzzy mem-
bership numbers. Fuzzy cluster centers in red and yellow (calculated in
different time windows) in Figure 21.3-b show the direction of hot water
movement in the geothermal reservoir.
Another prototype cluster-based approach that has gained a lot of atten-
tion in the geosciences is Self-Organizing Maps (SOM). The distinguishing
feature of SOM is its ability to retain topological organization of the data.
It does this by applying a non-linear projection of observation space onto a
user defined grid in a lower dimensional setting. Even though SOM is sim-
ilar to K-Means (they are both unsupervised clustering techniques), the
fundamental difference exists in that SOM is a neural network approach,
wherein there is one neuron for each centroid. There are many SOM neu-
ral network (NN) architectures; however, the most commonly used are
rectangular and hexagonal organization of the centroids [52]. The funda-
mentals of the NN-based models, along with other examples, have been
discussed in Section 21.2.4. As expected, the SOM clusters are sensitive
to the selection of the grid selected. It is possible for natural clusters to get
projected on different neurons, thereby implying dissimilarity.
Hierarchical clustering technique are a relatively old clustering technique,
but they still enjoy widespread application. Some of these techniques are an
example of graph-based clustering, while others have an interpretation as a
38.84
1
38.83 2-1
0
1
38.82
2-2 2-4
2-3 -1
38.81
Depth
(1m)
Latitude [deg]
-2
38.8
3 -3
38.79
-4
38.78
-122.84 -122.83 -122.82 -122.81 -122.8 -122.79 -122.78 -122.77
Longitude [deg] -5
Figure 21.3 Fuzzy clustering to monitor fluid movement in a geothermal reservoir [5].
prototype-based approach, depending on the selection of the proximity func-

tion. Following are some examples of proximity methods and their classifica-
tion in terms of the classes described in Figure 21.4, from Pedregosa et al. [42].
a. Single link, complete link and group average – graph-based
cluster proximity
b. Ward’s method – prototype-based cluster proximity (centroid).
Figure 21.4 is a comparison of the clustering algorithms generated under
Python environment, which consists of k-means, mini batch k-means,
GMM, mean shift, DBSCAN, OPTICS, and Birch clustering. As shown in
the figure, there is no objectively “correct” or “best” clustering algorithm.
Single Linkage Average Linkage Complete Linkage Word Linkage
.01s .05s .04s .04s
.01s .07s .06s .06s
.02s .09s .07s .07s
.02s .09s .07s .07s
.02s .09s .07s .06s
.02s .08s .08s .07s
Figure 21.4 Hierarchical clustering with different proximity functions [42].

The most appropriate algorithm for a particular problem often needs to be

chosen experimentally, unless there is a mathematical reason to prefer one
cluster model over another. Thus the choice of the clustering algorithm
depends highly on the data spreads. As an example, k-means cannot find non-
convex clusters.
DBSCAN is an example of density-based clustering algorithm that
locates regions of high density that are separated by points of low density,
which are classified as noise and thus omitted.
21.2.3 Ensemble Methods

The main idea of ensemble methods is to construct a set of base classifiers (or
regressors) from training data and perform analysis by taking a vote on the
predictions from each of the base classifiers. By building several, independent
models it is possible to reduce the variance and bias of model. The rationale
for an ensemble method is explained best using an example [52]. If the error
rate of a base classifier is e=0.35, and say there are 25 such base classifiers, then
the ensemble makes a wrong prediction only if more than half of the classifi-
ers predict incorrectly. The error is given by equation (21.5)
eensemble = ∑13 (i )e (1 − e )25−i = 0.06

25 25 i
(21.5)
Broadly, as highlighted in Figure 21.5, the following are steps to con-

struct an ensemble model:
Original
D Training data
Step 1:
Create Multiple D1 D2 Dt-1 Dt
Data Sets
Step 2:
Build Multiple C1 C2 Ct-1 Ct
Classifiers
Step 3:
Combine C*
Classifiers
Figure 21.5 Overview of the ensemble learning method [52].

1. By manipulating the training set – In this approach, multi-

ple training sets are created by resampling the original data.
Bagging and boosting are two examples of methods that
manipulate the training dataset.
2. By manipulating the input features – A subset of input fea-
tures is chosen to form each training set. Random forest is
an example of an ensemble that does this.
3. By manipulating class labels – This method is used when the
number of classes is large. Error-correcting output coding is
an example of this.
4. By manipulating the learning algorithm.
21.2.4 Artificial Neural Networks (ANN)-Based Methods

This section is focused on introducing the fundamentals of ANN and the
various architectures that have been covered in the reviewed literature.
Fundamentally, all NN-based approaches are inspired by the biological
neural system. The simplest model of NN, called a multi-layer perceptron
(MLP), is shown in Figure 21.6. Sometimes MLP is used loosely to refer to
any feedforward neural network; however, the difference lies in the acti-
vation function used (in MLP the activation function is a step function).
An MLP is quite similar to a modern neural network. By including the
following key elements, an MLP system becomes a full-fledged deep learn-
ing system:
a. Activation functions and hyper-parameters – a full NN

system uses a variety of activation functions, as discussed
later, that output real values, not just Booleans like in the
case of MLP. Through inclusion of hyper-parameters such
as learning rate, epochs and iterations, a full NN system is
more flexible.
b. Backpropagation – a full NN system uses backpropagation to
find optimal weights. This is done calculating the partial deriv-
atives of the error function (that quantifies difference between
true value and predicted value) and updating the weights at
each neuron in different layers of the network. Weights can be
updated after every sample in the training; however, in prac-
tice, weights are updated for a batch of samples.
Each neuron in a layer has a weight and has directed connection to the
neuron(s) in the next layer. The inputs are multiplied by the weight and fed
into the activation function. The output of the activation function cascades
to the next neuron; eventually the output layer generates a prediction. In
the backpropagation, the weights are updated as per the original connec-
tions between the neurons, but in the opposite direction, back from the
output layer.
To give the network an ability to learn non-trivial solutions, non-linear
activation functions are important. There are many activation functions –
both linear and non-linear; following are a few examples of the non-linear
functions:
Weights
Constant 1
W0
X1 Weights Sum
W1
Out
Inputs Wn-1
Xn-1 Step Function
Wn
Xn
(a)
Input Hidden Layer Output

Layer Layer
Input #1
Input #2
Output
Input #3
Input #4
(b)
Figure 21.6 (a) Perceptron learning process [11]. (b) Multi layer perceptron architecture
[22].
a. Sigmoid function. For very high or low values of the input,

the derivative of sigmoid is really small and that can slow
convergence to a prediction. This is called the vanishing gra-
dient problem.
b. Tanh function. This works better than a sigmoid function.
However, hyperbolic tangent also has the vanishing gradient
problem.
c. Rectified Linear Unity (ReLU) which is defined as f(x) =
max(0,x). Computationally efficient but when the input
approach zero, or are negative, the gradient becomes zero.
This means the network cannot perform backpropagation.
This is called the dying ReLU problem.
d. Leaky ReLU and parametric ReLU which is defined as f(x) =
max(ax,x), where a is a hyperparameter. Provides a small
positive slope and fixes the dying ReLU problem.
e. Tanh and Sigmoid. Two historically used functions but are
now known to have vanishing gradient problems.
f. Softmax. Typically used only for the output layer, as it can
handle multiple classes.
g. Swish which is defined as f(x) = x × sigmoid(x)
There are several architectures of the ANN which are popular and widely
deployed to solve challenges that are difficult to solve with traditional NN
structure. In this chapter, we have restricted ourselves to Convolutional
Neural Network (CNN), Recurrent Neural Networks (RNN), Auto Encoders
(AE) and Generative Adversarial Networks (GAN). Although the reviewed
literature makes use of variants of the architectures discussed earlier.
CNNs are very effective at tasks involving pattern recognition in images
(multi-band images). The primary benefit of a CNN is that they success-
fully capture spatial and temporal dependencies in an image (or sequence).
Practically CNNs are the driving force behind most computer vision tasks –
semantic segmentation, instance segmentation, panoptic segmentation.
As the name suggests, the CNN use convolution between a kernel and the
Convolution Pooling Convolution Pooling Fully Fully Output
+ReLU +ReLU Connected Connected predictions
dog (0.01)
Cat (0.01)
Boat (0.94)
Bird (0.94)
Figure 21.7 An illustrative example of CNN architecture [11].

image, as part of the analysis. The objective of the convolution operation is

to extract the high-level features such as edges, color, and gradient orienta-
tion among others, from the input image. Traditionally, such features in an
image were identified using filters such as edge detection filter (Canny filter,
Sobel filter, etc.). The beauty of CNN lies in its ability to learn such filters.
As shown in Figure 21.7 CNN first convolves a filter with the image, ana-
lyzing a small part of it each time. Through convolution, it creates a feature
map with probabilities that each feature belongs to the class. A pooling oper-
ation is done on the map, to reduce the dimensionality of each feature, with-
out losing the features or information. A pooling operation can be done in
multiple ways; however, max pooling or average pooling is most popular.
RNN is an architecture that is good at dealing with data that is sequen-
tial in nature. An RNN network (Figure 21.8) accepts a series of inputs over
y<1> y<2> y<t> y<t+1>
a<t-1> a<t> a<t+1>

a<0>
x<1> x<2> x<t> x<t+1>
Figure 21.8 Illustration of the RNN architecture [49].

‹
X1 X1
‹
X2 X2
‹
X3 X3
hw,b(x)
‹
X4 X4
‹
X5 X5
+1
‹
X6 X6
Layer L2 Layer L3
+1
Layer L1
Figure 21.9 Illustration of the Auto encoder architecture [59].

Discriminator
loss
Real images Sample
Discriminator
Random input
Denerator
loss
Generator Sample
Figure 21.10 Illustration of GANs architecture [41].
time and performs backpropagation through time (BPTT) to update. As

expected, BPTT starts with the latest time stamp (or the last value in case of
non-temporal sequence), cascading up the chain to learn optimal weights.
So far, we have discussed different architectures of ANN that are used
in a supervised setting. However, ANN architectures also exist that can be
used with unlabeled training examples. Self-Organizing maps were men-
tioned in Section 21.2.2, and are an example of this. Auto Encoders (AE)
is another architecture that can be used when no labeled data is available.
In AE, as shown in Figure 21.9, the target variables are equal to the inputs.
The encoder attempts to learn an identity function in the compressed form
(the neurons in the hidden layer < output layer). This architecture is very
helpful when denoising a dataset. They are also an example of generative
model (they can be used to generate random outcomes).
GANs or Generative Adversarial Networks are another example of gen-
erative model. GANs initiates a game between a generator that attempts
to learn the mapping from latent space to true data distribution of interest
and a discriminator that is tasked to distinguish “generated” candidates
from the true data distribution. GANs have been applied for detecting pat-
terns in images, 3D reconstruction of objects from images, improving res-
olution of images, and for some malicious applications such as Deepfakes.
Figure 21.10 illustrates this process.

Learning Applications to Reservoir
Characterization
The application of AI has gained a lot of interest because of its ability
to increase exploration success, improve reservoir characterization and
management and ultimately, optimize the hydrocarbon production. The

ability of deep learning to identify and learn features and patterns in the
data has pushed the envelope even further. It is not unfair to say that
such advancements in the data analysis have an impact that is akin to
that of the hydraulic fracturing and the subsequent increase in shale oil
and gas production. Starting from the prospecting phase of resources via
geophysical surveys, high resolution well logs, to well production data,
the petroleum system has various direct and indirect data sources that
collectively enable the industry to visualize a reservoir. Ojukwu et al.
[40] commented that in order to thoroughly characterize the reservoir
and make reservoir management decisions, it is important to integrate all
available data – static and dynamic. It also highlights the critical nature
of data management and data pipelines. We know the data size (volume
and velocity) and complexity (ex. high dimensions, multicollinearity) is
inversely correlated to both cost of computation and the model’s fidelity.
This is a typical example of challenges faced during reservoir character-
ization. In such situations, traditional methods struggle but improve-
ments can be gained by using hybrid models or even AI based models
[8, 38, 39, 47].
21.3.1 3D Structural Model Development

A 3D structural model, as shown in Figure 21.11, consists of largely two
elements – the bounding surfaces and the faults.
Computer vision for seismic interpretation has seen a lot of successful
research and application. The key goal is to identify the location of geo-
logical features that can indicate the presence of oil and gas reservoirs.
For this task, the evaluation of structures is an important aspect since a
large number of hydrocarbon reservoirs are contained in some kind of
structural trap. Guillen-Rondon et al. [20] presented a case study where
Convolutional Neural Networks (CNN) was used to infer and classify anti-
clines structures. Saiful Islam [37] used semantic segmentation algorithms –
U-Net and Se-ResNet to identify salt bodies in seismic images. Chopra
and Marfurt [10] used unsupervised techniques—K-means, principal
component analysis, self-organizing mapping (SOM) and generative topo-
graphic mapping—and found that SOM and the generative topographic
mapping provide additional information of interpretation interest. Wang
and Nealon [16] used CNN for denoising and enhancing the signals in a
3D seismic image. They noted, geologic structures of different scales, from
high-resolution faults and diffractors to deep subsalt sediments, were all
well-preserved by the deep neural network.
Well data Seismic data
• Tops • Two-way-time map

Integrate
• Well time • Root-mean-square velocity
• Well velocity • Model-based stacking velocity
• Fault polygon
Figure 21.11 Elements of a 3D structural model [43].
The conventional fault interpretation techniques are highly demand-

ing. They require a lot of experience, pre-processing and computational
time. Xiong et al. [61] developed a method that uses CNN to automatically
detect and map fault zones using 3D seismic images in a similar fashion to
the way done by interpreters. This was the first literature where no precom-
puted attributes were required to predict the faults. The system relied only
on labeling the images in binary classes – faults and no faults. This work
is a good example showing how well the NNs are able to abstract data and
find patterns, with limited intervention outside of data preparation steps.
Hussein et al. [24] showed the application of SOM and PCA to cluster and
combine the data from mathematically independent geometric seismic
attributes by using to identify small faults. The small faults can provide an
idea of the hydrocarbon migration pathways and aid in locating bypassed
oil accumulation within a reservoir.
Prediction of structural complexity for geohazard and subsurface

resource applications requires constraining and interpreting data that
are often ambiguous or lack key information [27]. Moreover, structural
complexity is a subjective term, requiring context for quantification.
Recognizing this, a new knowledge-data framework and a geospatial fuzzy
logic method was developed to represent and predict structural complexity
in the subsurface. A conceptual model for known structural complexity
serves as a basis for associating geospatial representations with types of
damage zones. A second conceptual model for zones of structural com-
plexity facilitates its prediction, notably in areas with limited explicit struc-
tural data. For each conceptual model, a fuzzy logic inference model is
developed to incorporate geospatial data and estimate structural complex-
ity potential. As is shown in Figure 21.12 and Figure 21.13 from Justman
et al. [27] demonstrated this approach using several public geospatial data-
sets within the state of Oklahoma. Explicit fault and earthquake location
data were integrated using a fuzzy model of known structural complexity
to train topographic, lithologic, and geophysical proxy datasets, applied to
a fuzzy model to predict structural complexity, and evaluated with Receiver
Operating Characteristic analyses and error classification. Together, these
results demonstrated an approach for predicting structurally complex
areas.
Surface slope 17.0
0.1
Topographic
Topographic 16.5
Wetness Index 7.2
(TWI)
Surface
Change in 69.2
sedimentary 0
thickness
Lithologic Sedimentary subsurface
Basement surface 81.0
curvature 0
Crystalline basement
Horizontal derivative of
total magnetic intensity
Geophysical Depth representation of proxy data

Constrained
Horizontal derivative of .282 Unconstrained
isostatic residual gravity .002
Figure 21.12 Proxy dataset maps aligned to data types (right) and concept layers (left)
from Justman et al. [27].
Identify and modify explicit

structural data to represent SC
• Fault
density
Maximum known SC potential
• Fault density weighted by Apply fuzzy
Faults
feature length inference model
• Fault intersection points High
Earthquakes
weighted by feature length
• Earthquake density Low
• Earthquakes weighted by
energy released
Training dataset Test dataset
Evaluation
Identify and modify proxy Training Process

structural data to represent SC Frequency ratios Fuzzy membership Fuzzy maps
• Surface slope Predict SC for all zones
Surface DEM
Basement • Topographic Wetness Index
DEM (TWI) Apply fuzzy
Isostatic • Chnage in sedimentary inference model High
residual thickness
gravity (IRG) • Basement surface curvature Low
Total
magnetic
• Horizontal derivative of
intensity isostatic residual gravity
(TMI) • Horizontal derivative of total
magnetic intensity
Figure 21.13 Illustrative workflow from Justman et al. [27].
As a proof of concept, the knowledge-data framework and fuzzy logic

methods were applied to geospatial data from the state of Oklahoma to
define and predict structurally complex areas. Fuzzy logic modeling
was performed using the open-source geospatial tool SIMPA (Spatially
Integrated Multi-variate Probabilistic Assessment [62]. Subsequent model-
ing and analyses were performed using ArcGIS Desktop 10.6 (ESRI, 2018),
IHS Petra®, and Microsoft Excel. Figure 21.4 describes the workflow to
encode geologic background, data and model design.
21.3.2 Sedimentary Modeling

Lithology identification, obtained through the analysis of several geophys-
ical properties, is critical for adding missing details (reservoir modeling,
volumetric assessment, production allocation) in a 3D structural model.
This step is key to make a complete shared-earth model. Lithology iden-
tification can be made by direct (from a physical sample) and/or indirect
methods (for example, Petrophysical logs). These methods have some lim-
itations, the main ones amongst them are cost and imprecision. There is
a great scope for using AI and uncertainty quantification techniques to
address some of these challenges.
Tokpanov et al. [58] presented a novel automated stratigraphic correlator
to create well top and zonation interpretations using CNN and Recurrent
NNs architecture using measurements from logs such as gamma-ray,
resistivity, neutron density, etc. They note that both BiLSTM and Inception
convolutional autoencoder reached human-level performance in the blind
test, with accuracy of nearly 90%.
The traditional procedure used for estimating permeability and poros-
ity, such as the experimental method and statistical methods used to
determine these properties of the reservoir, is very expensive and imprac-
tical due to the complexity of realistic conditions. Anifowose et al. [17],
Priezzhev et al. [25], Otchere et al. [14], and Saikia and Baruah [46]
studied and demonstrated the use of AI for predicting reservoir prop-
erties using machine learning techniques. Saikia and Baruah [46] used
an ensemble method, named StackNet to predict well log data as output.
Otchere et al. [14] performed a comparative study on the logging data as
the target variable but used different supervised learning models—ANN,
Support Vector Machines (SVM) and Relevant Vector Machine—and
determined SVM worked better than ANN in data scarce environment.
They also commented on achieving improved performance by hybridiz-
ing multiple algorithms and the clear benefit that had on enhanced reser-
voir characterization.
He et al. [23] developed and tested a framework to predict sedimentary
facies of uncored areas with high accuracy using Multi-Layer Perceptron
(MLP) trained on data characterizing cored sampled petrophysical logs.
Imamverdiyev and Sukhostat [26] proposed a new CNN architecture with
one dimensional kernels to identify lithological facies from petrophysical
logs. Alizadeh et al. [2] used hierarchical methods to cluster Total Organic
Column (TOC) according to their geochemical composition and used
NNs to relate Logging While Drilling data to source rock. This allowed
the authors to draw burial thermal history diagrams and identify potential
zones for hydrocarbon production.
In shale plays, fracture network and fracture identification is a critical
part of the reservoir characterization process. Tian and Daigle [57] used
single shot object detection method on SEM images to identification and
characterization microfractures in shales (carbonate-rich shale and a sili-
ceous shale). With this approach they were able to identify abundance,
obtained statistics of length and areal porosities of the fractures as well.
Suhag et al. [50] evaluated the performance of NNs for forecasting shale
oil production using data such as petrophysical logs, past-production, and
well interventions – stimulation. This analysis focused on the Bakken,
North Dakota, formation. Through a selection of physical properties from
different sources, they built an NN model that fits with the production
data in wells that have a diverse production history. Their work has shown
the importance of including physical parameters into a process that was

heretofore seen as a time series regression problem. This work clearly illus-
trated the importance of reservoir characterization from the point of view
of estimating recoverable production and stimulation strategies for an
unconventional well.
Given the low signal-to-noise ratio nature of the subsurface data, it
is very important to consider and account for error and uncertainty. A
classical stochastic approach to uncertainty management is by drawing
multiple samples at random and running statistical tests to see the confi-
dence intervals. Similarly, in the subsurface data multiple realizations are
developed to account for uncertainty in measurements. Kang and Choe
[29] illustrate a good example by using Generative Adversarial Networks
(GAN) architecture, specifically DCGAN, to generate realization and
quantify uncertainty.
21.3.3 3D Petrophysical Modeling

Seismic petro-facies characterization in low net-to-gross reservoirs is
challenging. Moreover, when the elastic properties of the desired petro-fa-
cies significantly overlap, the characterization becomes further compli-
cated. Yenwongfai et al. [63] used impedance and facies as input and used
supervised feed forward NN to discriminate porosity profile. The pro-
posed workflow improved the seismic reservoir delineation of the Snadd
Formation (low net-to-gross) in the Goliat Field.
Predicting the petrophysical properties of rock samples using
micro-CT images is a very involved process with high computation
costs. However, this is a critical piece of information for properly char-
acterizing a reservoir since the CT scan images provide a lot of valuable
information – porosity and permeability included. Tembely et al. [55]
developed a supervised learning framework based on physics-informed
CNN. They applied this method to predict the petrophysical properties
of porous rock from 3D micro-CT images and achieved over 90% accu-
racy, while achieving 3x gains in computational performance.
Non-stochastic simulation models often can be highly complex and
time-consuming. In their place, the meta-models provide an approxima-
tion that is adapted to in order to capture local minima or maxima. The
Kriging model is perhaps the most popular example of a meta-model. The
petrophysical models that have been discussed previously have relied on
direct or indirect methods of evaluation. However, there is still a need
for performing logical extrapolation of these properties to make realistic,
spatially continuous predictions for properties in unsampled locations.

Some common methods include:
• Geostatistics: Ordinary Kriging (OK) [30]

• Geostatistics: Geographically weighted regression (GWR)
[7]
• ML/DL: Support vector machines for regression (SVR)
[13, 28, 48]
• Fuzzy Kriging similar to what was discussed under fuzzy
clustering. For on oil and gas application see Zhao et al.
[64]
• Hybrid: Geographically weighted regression kriging
(GWRK) [21, 31]
• Hybrid: Artificial neural networks kriging (ANNK) [15, 53].
Chen et al. [9] studied the techniques mentioned above to map organic
carbon content in topsoil. They noted accuracy gains in using hybrid and
machine learning models – ANNK, SVR, respectively, with ANNK per-
forming the best.
Present subsurface predictions often rely upon disparate and limited
a priori information. Even regions with concentrated subsurface explo-
ration still face uncertainties that can obstruct safe and efficient explora-
tion of the subsurface. Uncertainty may be reduced, even for areas with
little or no subsurface measurements, using methodical, science-driven
geologic knowledge and data. Rose et al. [44] identified this and devel-
oped a hybrid spatiotemporal statistical-geologic approach, subsurface
trend analysis (STA), that provides improved understanding of subsur-
face systems. The STA method assumes that the present-day subsurface is
not random, but is a product of its history, which is a sum of its system-
atic processes. With even limited data and geologic knowledge, the STA
method can be used to methodically improve prediction of subsurface
properties. To demonstrate and validate the improved prediction poten-
tial of the STA method, it was applied in an analysis of the northern Gulf
of Mexico. This evaluation was prepared using only existing, publicly
available well data and geologic literature. Using the STA method, this
information was used to predict subsurface trends for in situ pressure,
in situ temperature, porosity, and permeability. The results of this STA-
based analysis were validated against new reservoir data. STA-driven
results were also contrasted with previous studies. Both indicated that
STA predictions were an improvement over other methods. Overall, their
Alabama
Mississippi
Florida
Louisiana
Texas
N
0 25 50 100 150 200
Miles
Texas Texas
0 5 10 20 30 40 0 5 10 20 30 40
Miles Miles
(a) (c)
0
R2 = 0.9284 R2 = 0.9999 R2 = 0.9318
Depth (TVD ft)
5000
10000
0 5000 10,000 0 5000 10,000 0 5000 10,000
Initial pressure (psi) Initial pressure (psi) Initial pressure (psi)
(b)
Figure 21.14 Example of domain boundary revision and iteration using an example from
central GoM. Rose et al. [44].
approach could provide critical information to evaluate and reduce risks,

identify and improve areas of scarce or discontinuous data, and provide
inputs for multi-scale modeling efforts, from reservoir scale to basin
scale. This method offered a framework for guiding future science-based
machine learning and natural language processing to optimize subsur-
face analyses and predictions.
A combination of statistical analyses and geologic system knowledge
was selected to provide constraint on subsurface attributes. Geologic con-
text of the GoM basin was established using the a priori geologic systems
publications mentioned in the previous section. The three primary pro-
cesses influencing subsurface properties are lithology, structure, and sec-
ondary alteration. An initial iteration of STA domains was then formulated
by combining structural and depositional domains. Upon review of petro-
leum systems literature and initial statistical results of the BOEM Sands
dataset and petroleum systems in the GOM, it was determined that the
most useful depositional domains to target for STA domain formulation

would be those representing the Miocene, Pliocene, and Pleistocene, as
these chronozones host the premier hydrocarbon production fields in the
GOM and the majority of BOEM Sands data points. These domains were
then combined in ArcGIS, creating 21 discrete domains.
The hybrid deductive-probabilistic approach demonstrated with the
STA method integrates contextual geologic information with quantita-
tive analytical data and tools to improve prediction of subsurface prop-
erties and reduce uncertainty. As demonstrated in Figure 21.15, the STA
approach leverages information from areas with data to offer insights into
probable subsurface properties for areas with sparse or nonexistent data.
The method is flexible to allow for inclusion and use of different data types
and sources. The STA method’s systematic framework ensures integration
of deductive geologic process models with more probabilistic approaches
to constraining subsurface properties.
Many subsurface properties (e.g., pressure, temperature, porosity,
and formation thicknesses) vary in relation to themselves as a result of
the geologic processes that drive their distribution. This information is
Florida
Louisiana
Texas
Validation error
0.4
0.2
N 0 50 100 150 0.0

Sands validation points
Kilometers Best model fit -0.2
Co-Kriging with STA influence
Kriging - no STA influence -0.4
Kriging Co-kriging
STA domains
Figure 21.15 Subsurface Trend Analysis (STA) results. Data-based prediction vs.
knowledge-data-driven predictions [44].
critical for reservoir characterization. Geologic expertise is leveraged to

identify domains thought to have experienced a common geologic history.
Mark-Moser et al. [33] conducted a detailed study of the north-central off-
shore GOM using bottom hole data from more than 13,000 data points
and information from numerous studies that define geologic domains of
the GOM that were likely to have experienced a common geologic his-
tory. Linear regressions for the full data set were contrasted against linear
regression results for each domain. Trends within each domain showed an
improved fit of the distribution relative to the entire GOM data set. New
data points acquired since 2012 were used for further validation. In each
domain, new data points were plotted against the original pre-2012 data
set. The new data fell along the predicted trend, resulting in the same or
improved R2 for all domains, demonstrating that predictions for locations
within domains were robust.
21.3.4 Dynamic Modeling and Simulations

A commonly deployed dynamic technique is time‐lapse seismic reser-
voir monitoring (also known as 4‐D seismic), which involves comparing
3‐D seismic survey data taken over different time intervals. Analyses of
4‐D seismic data volumes can help improve the quality of storage reser-
voir characterization, track the movement of injected plume, and identify
potential carbon dioxide leakage from the storage reservoir.
In this research, they apply a physics‐based deep learning method to
facilitate the solution of both the forward and inverse problems in seismic
inversion while honoring physical constraints. A cycle generative adversar-
ial neural network (CycleGAN) model is trained to learn the bidirectional
functional mappings between the reservoir dynamic property changes and
seismic attribute changes, such that both forward and inverse solutions can
be obtained efficiently from the trained model. Zhong et al. [65] show that
our CycleGAN‐based approach not only improves the reliability of 4‐D
seismic inversion but also expedites the quantitative interpretation. Our
deep learning‐based workflow is generic and can be readily used for reser-
voir characterization and reservoir model updates involving the use of 4‐D
seismic data.
Mudunuru et al. [36] presented a physics-informed machine learning
(PIML) workflow for real-time unconventional reservoir management.
Through this application workflow, Machine Learning, field scale data,
and high-fidelity physics, model simulation models are combined and cre-
ated for real-time forecasting which are a cumulative production (e.g., gas,
water), a production rate, a stage-specific production, and spatial evolution
of quantities of interest (e.g., residual gas, reservoir pressure). They pro-

posed three main keys (site behavior libraries based on fast and accurate
physics; ML-based inverse models to refine key site parameters; a fast for-
ward model that combines physical models and ML to forecast production
and reservoir conditions) for the PIML workflow. First, they used synthetic
production data to develop the site behavior library. After that, ML-based
reverse models were created to enable the forecasting of production behav-
ior. This ML-based application is significantly faster and a low-cost alter-
native to accurately simulate gas flow and transport in fractured shale-gas
fields.
21.4 Machine (Deep) Learning and Enhanced Oil

Recovery (EOR)
Data collected from the well logs, cores, completion designs, real-time
production data, maintenance records and financial restrictions plays
an important role to locate resources, find the volume of reservoirs, and
extract them in a cost-effective manner while being conscious about envi-
ronmental safety. Many systems exist which collect, analyze, visualize and
contextualize this huge amount of the data. Even though the current simu-
lations and software can provide critical analytics for engineers and work-
ers, machine learning techniques provide a pathway to integrate them in a
sensible manner. An integrated approach can aid in development of a tool
that can explore millions of what-if scenarios. This is critical for predicting
production responses, mechanical equipment failure, changing fieldwide
forecasts conditions, and drilling critical decisions.
Many EOR projects have been applied in many regions around the
world. It is very important to understand how to process the EOR in the
petroleum industry and predict the critical decisions on the applicability
of certain techniques. EOR techniques can take days if engineers clarify a
small number of data manually; however, machine learning can run the
same amount of data in minutes with less man-hours work and computa-
tional time, and less mistakes than human analysis. On the other hand, it
takes seconds to assimilate new data and build a suite of statistical models
that investigate the relationships between many different measurements
[34]. Using machine learning techniques in enhanced oil recovery, three
steps can apply to perform ML in EOR projects. Selecting a suitable EOR
process depends on the reservoir characteristics, preparing correct input
data sets to design the selected EOR process, and making optimization
studies on key parameters selected by the user [18].
21.4.1 ANNs for EOR Performance and Economics

Thanh et al. [60] studied CO2-EOR and storage in Residual Oil Zones
(ROZ) using Artificial Neural Networks. They used ANN for predicting
field oil recovery, CO2 stored, and CO2 retained in the real field ROZs in
the Permian basin. They began by performing reservoir simulations of
residual oil zones to generate the training dataset. The ANN model com-
prised 7 hidden layers and 20 neurons of each hidden layers, minimum
Mean Square Error, the maximum correlation factor (R2) of testing data
set, details of the workflow in Figure 21.8. The ANN models can predict the
CO2-EOR and storage performance with high accuracy of in ROZs and can
also be used to determine other aspect of EOR and CO2 sequestration, such
as prediction trapping index, CO2 leakage from the cap-rock or CO2 plume
migration area.
Using the architecture described in Figure 21.16 and Particle Swarm
Optimization (PSO), the authors were able to obtain the best oil produc-
tion, CO2 storage, and economic parameters such as the Net-Present Value
(NPV). From the perspective of project economics, they demonstrated the
application of this approach for a range $(30–60) per barrel oil prices. Their
developed ANN models provide a fast and robust tool to estimate project
economic benefits.
This study shows that a similar analysis can be carried for other EOR
methods – thermal, chemical, gas to determine project economics, optimal
Uncertainty parameters Design reservoir model Start
CMOST-AI Thickness
Permeability Oil production
Residual oil
Latin Hypercube Design saturation to CO2 stored
water flood
Residual oil
Training & Testing
saturation to
gas flood DDM
N samples Injection rate
Producer BHP
CO2 retained
Porosity
Blind validation DDM
Training Simulation using
CMG-GEM
Input layer Hidden layer Output layer
Prediction real fields
Artificial Neural Network ROZs using DDM
Hand-shaking
Objective functions Reservoir simulation and ANN
(Cumulative CO2 storage,
retained and Oil Output collection End
Production)
a
Figure 21.16 ANN-based workflow from Thanh et al. [60].

Table 21.1 Chemical EOR input and output data from [31].
Parameter Unit Type Min Max Average Standard deviation
Surfactant slug size PV Input 0.097 0.259 0.177 0.072
Surfactant concentration Vol. fraction Input 0.005 0.03 0.017 0.011
Polymer concentration in wt.% Input 0.1 0.25 0.177 0.067
surfactant slug
Polymer drive size PV Input 0.324 0.648 0.482 0.144
Polymer concentration in wt.% Input 0.1 0.2 0.148 0.044
polymer drive
Kv/Kh ratio - Input 0.01 0.25 0.129 0.107
Salinity ofpdymer drive Meq/ml Input 0.3 0.4 0.349 0.045
Recovery factor (RF) % Output 14.82 56.99 39.67 9.24
Net present value (NPV) $ MM Output 1.781 7.229 4.45 1.53
injection, optimal recovery. Moreover, an extension of this work also has

application to study the feasibility of CCUS.
Ahmadi and Pournik [31] studied and developed an SVM model to
predict performance and efficiency of a chemical flooding simulation. The
data utilized was generated using UTCHEM simulator, based on Benoist
sand reservoir. Details of the simulator input and output are in Table 21.1.
They concluded a strong agreement (R2 99.3% for both RF and NPV)
between SVM predicted and real data.
21.4.2 ANNs for EOR Screening

The selection process for the right EOR techniques is a complex and
very important task which involves a combination of a set of rock and
fluid parameters controlling technical and economic performance of a
reservoir. Screening the EOR is highly risk for financial for specialized
lab investigation, analysis, simulations, pilot study and field operations.
Fluid properties, remaining volume, lithology, depth, temperature and
permeability parameters are the core technical parameters for EOR
screening [54].
As shown by [35] in Figure 21.17, there are four stages in EOR screening
workflow. The first two stages were designed to read, audit, and interpret
First Second
Data Data Stage Stage
Validation Analysis Screening Screening
Figure 21.17 EOR screening process workflow [35].

relevant EOR data. Data analysis and validations are the other stages of the
screening flow.
Giro et al. [19] introduce a methodology that uses Naïve Bayes Classifier
and selects EOR materials for specific reservoir conditions. They used
this model with physical and chemical representations of injection fluids,
including EOR materials, with reservoir-specific information on lithology,
porosity, permeability, as well as oil, water and salt conditions to provide
recommendation for EOR “cocktail” for injection fluids.
Suhag et al. [51] and Temizel et al. [56] worked on data-driven mod-
els with application that focused on determining sensitivity of controlling
parameters and controlling agents. This work provides a better under-
standing of the influence of surfactant adsorption and thus, a number of
chemicals to be used in an efficient manner. Optimum values for control-
lable decisions were determined by coupling a commercial optimization
software with the reservoir simulator. This work aids in screening and
deciding operation parameters.
21.5 Conclusion
Reservoir characterization is a critical step for understanding the sub-
surface. It is crucial for making optimal decisions to recover resources
from hydrocarbon reservoirs, geothermal assets, and ground water from
aquifers. Given the characterization process involves integration of multi-
disciplinary processes, computational intelligence tools are ideal for this.
Keeping this in mind, this chapter introduces various machine learning
techniques – supervised and unsupervised methods, deep learning archi-
tectures, and many other soft computing techniques and their applications
to reservoir characterization and enhanced oil recovery.
This chapter begins with a detailed overview of some popular machine
learning and deep learning techniques that have been used for various
subtasks of reservoir characterization. Broadly, the chapter breaks down
different subtasks of reservoir characterization—structural model devel-
opment, sedimentary modeling, petrophysical modeling, and dynamic
modeling. Similarly for the enhanced oil recovery section, the chapter
focuses on EOR performance and economics, and EOR screening. The
goal is to introduce the rationale behind a method and list some relevant
literature that might assist readers with their own research problem. With
regards to the deep learning methods, the goal has been to introduce the
fundamental architecture, different activation functions and the process
of selecting them for a specific problem statement. With deep learning
growing rapidly, this section should serve as a bedrock to build a deeper

understanding of the advanced variants.
Acknowledgement
The author acknowledges Anuj Suhag’s contributions in reviewing and
adding some of the refences to a few case histories.
References
1. Ahmadi, M. A. and Pournik, M. (2016), A predictive model of chemical-
flooding for enhanced oil recovery purposes: Application of least square sup-
port vector machine, Petroleum, Volume 2, Issue 2, 2016, pp. 177-182, ISSN
2405-6561, https://doi.org/10.1016/j.petlm.2015.10.002.
2. Alizadeh, B., Najjari, S., & Kadkhodaie-Ilkhchi, A. (2012). Artificial neural
network modeling and cluster analysis for organic facies and burial history
estimation using well log data: A case study of the South Pars Gas Field,
Persian Gulf, Iran. Computers & Geosciences, 45, 261-269.
3. Aminzadeh, F., Temizel, C. and Y. Hajizadeh., 2022, Intelligent Data Analysis
for Exploration and Production, Wiley & Sons.
4. Aminzadeh, F., 2021, Reservoir Characterization: Combining Machine
Intelligence with Human Intelligence, E&P Plus, April 2021, Vol. 96 Issue 4,
E&P Plus, Hart Energy.
5. Aminzadeh, F., 2015, Characterizing Fractures in Geysers Geothermal
Field by Micro-seismic Data, Using Soft Computing, Fractals, and Shear
Wave Anisotropy, Topic Area 23: Fracture Characterization Technology
DE-FOA-0000075-23, Final Report.
6. Aminzadeh, F. and Dasgupta, S., 2013. Geophysics for Petroleum Engineers,
Elsevier.
7. Brunsdon, C., Fotheringham, S., & Charlton, M. (1998). Geographically
weighted regression. Journal of the Royal Statistical Society: Series D (The
Statistician), 47(3), 431-443.
8. Bruyelle, J., & Guérillot, D. (2019, October). Proxy Model Based on Artificial
Intelligence Technique for History Matching-Application to Brugge Field. In
SPE Gas & Oil Technology Showcase and Conference. Society of Petroleum
Engineers.
9. Chen L, Ren C, Li L, Wang Y, Zhang B, & Wang Z, A Comparative Assessment
of Geostatistical, Machine Learning, and Hybrid Approaches for Mapping
Topsoil Organic Carbon Content. ISPRS International Journal of Geo-
Information. 2019; 8(4):174.
10. Chopra, S., & Marfurt, K. J. (2018). Seismic facies classification using
some unsupervised machine-learning methods. Society of Exploration
Geophysicists.
11. Complete Guide to Artificial Neural Network Concepts & Models. (n.d.).
MissingLink.Ai. Retrieved February 16, 2021, from https://missinglink.ai/
guides/neural-network-concepts/complete-guide-artificial-neural-networks/.
12. Cortes, C., and Vapnik, V. Support-vector networks. Mach Learn 20, 273–
297 (1995). https://doi.org/10.1007/BF00994018.
13. Da Silva, L. M., Avansi, G. D., & Schiozer, D. J. (2020). Support Vector
Regression for Petroleum Reservoir Production Forecast Considering
Geostatistical Realizations. SPE Reservoir Evaluation & Engineering.
14. Daniel Asante Otchere, Tarek Omar Arbi Ganat, Raoof Gholami, & Syahrir
Ridha, Application of supervised machine learning paradigms in the predic-
tion of petroleum reservoir properties: Comparative analysis of ANN and
SVM models, Journal of Petroleum Science and Engineering.
15. Demyanov, V., Kanevsky, M., Chernov, S., Savelieva, E., & Timonin, V.
(1998). Neural network residual kriging application for climatic data. Journal
of Geographic Information and Decision Analysis, 2(2), 215-232.
16. Enning Wang, and Jeff Nealon; Applying machine learning to 3D seismic
image denoising and enhancement. Interpretation; 7 (3): SE131–SE139. doi:
https://doi.org/10.1190/INT-2018-0224.1.
17. Fatai Anifowose, Suli Adeniye, Abdulazeez Abdulraheem, & Abdullatif
Al-Shuhail, Integrating seismic and log data for improved petroleum res-
ervoir properties estimation using non-linear feature-selection based
hybrid computational intelligence models, Journal of Petroleum Science and
Engineering, Volume 145, 2016, pp. 230-237,ISSN 0920 4105,https://doi.
org/10.1016/j.petrol.2016.05.019.(http://www.sciencedirect.com/science/
article/pii/S0920410516301814).
18. Gharbi, R. B. (2000). An expert system for selecting and designing EOR pro-
cesses. Journal of Petroleum Science and Engineering, 27(1-2), 33-47.
19. Giro, Ronaldo, Lima Filho, Silas Pereira, Neumann Barros Ferreira, Rodrigo,
Engel, Michael, and Mathias Bernhard Steiner. “Artificial Intelligence-
Based Screening of Enhanced Oil Recovery Materials for Reservoir-Specific
Applications.” Paper presented at the Offshore Technology Conference Brasil,
Rio de Janeiro, Brazil, October 2019. doi: https://doi.org/10.4043/29754-MS.
20. Guillen-Rondon, P., Cobos, C., Larrazabal, G., & Diz, A. (2019, August 28).
Machine learning: A deep learning approach for seismic structural evalua-
tion. Society of Exploration Geophysicists.
21. Harris, P., Fotheringham, A. S., Crespo, R., & Charlton, M. (2010). The use
of geographically weighted regression for spatial prediction: an evaluation of
models using simulated data sets. Mathematical Geosciences, 42(6), 657-680.
22. Hassan, Hassan & Negm, Abdelazim & Zahran, Mohamed & Saavedra,
Oliver. (2015). Assessment of artificial neural network for bathymetry
estimation using high resolution satellite imagery in shallow lakes: Case

study El Burullus lake. International Water Technology Journal. 5.
23. He, J., La Croix, A. D., Wang, J., Ding, W., & Underschultz, J. R. (2019). Using
neural networks and the Markov Chain approach for facies analysis and pre-
diction from well logs in the Precipice Sandstone and Evergreen Formation,
Surat Basin, Australia. Marine and Petroleum Geology, 101, 410-427.
24. Hussein, M., Stewart, R., & Wu, J. (2020, December). Unsupervised Machine
Learning Techniques for Subtle Fault Detection. In 82nd EAGE Annual
Conference & Exhibition (Vol. 2020, No. 1, pp. 1-5). European Association of
Geoscientists & Engineers.
25. Priezzhev, I., Veeken, P. C. H. Egorov, S. V. and Strecker, U. (2019), “Direct
prediction of petrophysical and petroelastic reservoir properties from seis-
mic and well-log data using nonlinear machine learning algorithms,” The
Leading Edge 38: 949–958.
26. Imamverdiyev, Y., & Sukhostat, L. (2019). Lithological facies classification
using deep convolutional neural network. Journal of Petroleum Science and
Engineering, 174, 216-228.
27. Justman, D., Creason, C. G., Rose, K., & Bauer, J. (2020). A knowledge-data
framework and geospatial fuzzy logic-based approach to model and predict
structural complexity. Journal of Structural Geology, 141, 104153.
28. Kanevski M., Pozdnukhov A., Canu S., Maignan M., Wong P.M., Shibli S.A.R.
(2002) Support Vector Machines for Classification and Mapping of Reservoir
Data. In: Wong P., Aminzadeh F., Nikravesh M. (eds.) Soft Computing for Reservoir
Characterization and Modeling. Studies in Fuzziness and Soft Computing, vol.
80. Physica, Heidelberg. https://doi.org/10.1007/978-3-7908-1807-9_21.
29. Kang, B., & Choe, J. (2020). Uncertainty quantification of channel reservoirs
assisted by cluster analysis and deep convolutional generative adversarial
networks. Journal of Petroleum Science and Engineering, 187, 106742.
30. Krige DG (1951) A statistical approach to some basic mine valuation prob-
lems on the witwatersrand. Journal of the Chemical Metallurgical and Mining
Society of South Africa, 52:119-139.
31. Kumar, S., Lal, R., & Liu, D. (2012). A geographically weighted regression
kriging approach for mapping soil organic carbon stock. Geoderma, 189,
627-634.
32. Ma, Zee. (2011). Reservoir Characterization: A look back and ways forward.
AAPG Bulletin.
33. Mark-Moser, M., Miller, R., Rose, K., Bauer, J., & Disenhof, C. (2018). Detailed
analysis of geospatial trends of hydrocarbon accumulations, offshore Gulf of
Mexico (No. 07e31174-2af8-410b-8fb8-653540baf1b7). National Energy
Technology Laboratory (NETL), Pittsburgh, PA, Morgantown, WV, and
Albany, OR (United States). Energy Data eXchange; National Energy
Technology Laboratory (NETL), Pittsburgh, PA, Morgantown, WV (United
States).
34. Martin, E., Wills, P., Hohl, D., & Lopez, J. L. (2017, February). Using machine
learning to predict production at a Peace River thermal EOR site. In SPE
Reservoir Simulation Conference. Society of Petroleum Engineers.
35. Moreno, J. E., Gurpinar, O. M., Liu, Y., Al-Kinani, A., & Cakir, N. (2014,
December). EOR advisor system: a comprehensive approach to EOR selec-
tion. In International Petroleum Technology Conference. International
Petroleum Technology Conference.
36. Mudunuru, M.K., O’Malley, D., Srinivasan, S., Hyman, J., Sweeney, M.,
Frash, L., Carey, B., Gross, M., Welch, N., Karra, S., Vesselinov, V., Kang, Q.,
Xu, H., Pawar, R., Carr, T., Li, L., Guthrie, G.D., & Viswanathan, H. (2020).
Physics-Informed Machine Learning for Real-time Reservoir Management.
AAAI Spring Symposium: MLPS.
37. Muhammad Saiful Islam, Using deep learning based methods to classify salt
bodies in seismic images, Journal of Applied Geophysics, Volume 178, 2020,
104054, ISSN 0926-9851, https://doi.org/10.1016/j.jappgeo.2020.104054.
38. Müller, M., Macklin, M., Chentanez, N., Jeschke, S., & Kim, T. Y. (2020,
December). Detailed rigid body simulation with extended position based
dynamics. In Computer Graphics Forum (Vol. 39, No. 8, pp. 101-112).
39. Navrátil Jiří, King Alan, Rios Jesus, Kollias Georgios, Torrado Ruben, Codas
Andrés, Accelerating Physics-Based Simulations Using End-to-End Neural
Network Proxies: An Application in Oil Reservoir Modeling, Frontiers in
Big Data Vol. 2, 2019. URL=https://www.frontiersin.org/article/10.3389/
fdata.2019.00033, DOI=10.3389/fdata.2019.00033/
40. Ojukwu, C., Smith, K., Kadkhodayan, N., Leung, M., & Baldwin, K. (2020,
August 11). Reservoir Characterization, Machine Learning and Big Data –
An Offshore California Case Study. Society of Petroleum Engineers.
doi:10.2118/203642-MS/
41. Overview of GAN Structure | Generative Adversarial Networks. (n.d.). Google
Developers. Retrieved February 16, 2021, from https://developers.google.
com/machine-learning/gan/gan_structure/
42. Pedregosa, F., Varoquaux, G., Gramfort, A, Michel, V., Thirion, B., Grisel, O.,
Blondel, M., Prettenhofer, P., Weiss, R., Dubourg, V., Vanderplas, J., Passos,
A., Cournapeau, D., Brucher, M., Perrot M., and Duchesnay, E., 2011, Scikit-
learn: Machine Learning in Python., Journal of Machine Learning Research
12 (2011) 2825-2830.
43. PetroWiki. (2016, January 18). Basic elements of a reservoir characterization study.
https://petrowiki.spe.org/Basic_elements_of_a_reservoir_characterization_
study.
44. Rose, K. K., Bauer, J. R., & Mark-Moser, M. (2020). A systematic, science-driven
approach for predicting subsurface properties. Interpretation, 8(1),
T167-T181.
45. Rumelhart, D., Hinton, G. & Williams, R. Learning representations
by back-propagating errors. Nature 323, 533–536 (1986). https://doi.
org/10.1038/323533a0.
46. Saikia, P., & Baruah, R. (2019). Investigating Stacked Ensemble Model for Oil
Reservoir Characterisation. 2019 IEEE International Conference on Systems,
Man and Cybernetics (SMC), 13-20.
47. SHAHKARAMI, A. MOHAGHEGH, S. Applications of smart proxies for
subsurface modeling, Petroleum Exploration and Development, Volume 47,
Issue 2, 2020, pp. 400–412, ISSN 1876-3804.
48. Sitharam, T. G., Samui, P., & Anbazhagan, P. (2008). Spatial variability of rock
depth in Bangalore using geostatistical, neural network and support vector
machine models. Geotechnical and Geological Engineering, 26(5), 503-517.
49. Stanford University & MIT University. (n.d.). CS 230 - Recurrent Neural
Networks Cheatsheet. Stanford.Edu. Retrieved February 16, 2021, from
https://stanford.edu/%7Eshervine/teaching/cs-230/cheatsheet-recurrent-
neural-networks
50. Suhag, A, Ranjith, R, and Aminzadeh, F. Comparison of Shale Oil Production
Forecasting using Empirical Methods and Artificial Neural Networks. SPE
Annual Technical Conference and Exhibition, San Antonio, Texas, USA,
October 2017.
51. Suhag, Anuj, Ranjith, Rahul, Balaji, Karthik, Peksaglam, Zumra, Malik,
Vidhi, Zhang, Ming, Biopharm, Frontida, Putra, Dike, Energy, Rafflesia,
Wijaya, Zein, Dhannoon, Diyar, Temizel, Cenk, and Fred Aminzadeh.
“Optimization of Steamflooding Heavy Oil Reservoirs.” Paper presented
at the SPE Western Regional Meeting, Bakersfield, California, April 2017.
doi: https://doi.org/10.2118/185653-MS.
52. Tan PN, Steinbach M, Kumar V (2005) Introduction to data mining.
Addison-Wesley Longman Publishing Co., Inc, Boston.
53. Tapoglou, E., Karatzas, G. P., Trichakis, I. C., & Varouchakis, E. A. (2014).
A spatio-temporal hybrid neural network-Kriging model for groundwater
level simulation. Journal of Hydrology, 519, 3193-3203.
54. Tarrahi, M., Afra, S., & Surovets, I. (2015, October). A Novel Automated
and Probabilistic EOR Screening Method to Integrate Theoretical Screening
Criteria and Real Field EOR Practices Using Machine Learning Algorithms.
In SPE Russian Petroleum Technology Conference. Society of Petroleum
Engineers.
55. Tembely, M., AlSumaiti, A., & Alameri, W. (2020). A deep learning perspec-
tive on predicting permeability in porous media from network modeling to
direct simulation. Computational Geosciences, 24, 1541-1556.
56. Temizel, C., Tiwari, A., Kirmaci, H., Aktas, S., Ranjith, R., Zhu, Y., Tahir, S.,
Aminzadeh, F., Yegin, C. (2016-January). “Improved optimization through
procedures as pseudo objective functions in nonlinear optimization of oil
recovery with next-generation reservoir simulators,” Proc. - SPE Annu. Tech.
Conf. Exhib., vol. 2016.
57. Tian, X., & Daigle, H. (2018). Machine-learning-based object detection in
images for reservoir characterization: A case study of fracture detection in
shales. Geophysics, 37, 435-442.
58. Tokpanov, Y., Smith, J., Ma, Z., Deng, L., Benhallam, W., Salehi, A., … Castineira,
D. (2020, October 21). Deep-Learning-Based Automated Stratigraphic
Correlation. Society of Petroleum Engineers. doi:10.2118/201459-MS.
59. Unsupervised Feature Learning and Deep Learning Tutorial. (n.d.). Stanford.
Edu. Retrieved February 16, 2021, from http://ufldl.stanford.edu/tutorial/
unsupervised/Autoencoders/.
60. Vo Thanh, H., Sugai, Y. & Sasaki, K. Application of artificial neural network
for predicting the performance of CO2 enhanced oil recovery and storage
in residual oil zones. Sci Rep 10, 18204 (2020). https://doi.org/10.1038/
s41598-020-73931-2.
61. Wei Xiong, Xu Ji, Yue Ma, Yuxiang Wang, Nasher M. AlBinHassan, Mustafa
N. Ali, and Yi Luo, (2018), “Seismic fault detection with convolutional neural
network,” Geophysics 83: O97-O103.
62. Wingo, P., Justman, D., Creason, G., Jones, K., Bauer, J., Rose, K., SIMPA,
2019-03-29. https://edx.netl.doe.gov/dataset/simpa-tool.
63. Yenwongfai, H.D., Mondol, N., Lecomte, I., Faleide, J.I., & Leutscher, J. (2019).
Integrating facies‐based Bayesian inversion and supervised machine learn-
ing for petro‐facies characterization in the Snadd Formation of the Goliat
Field, south‐western Barents Sea. Geophysical Prospecting, 67, 1020–1039.
64. Zhao, X, Popa, A.S., Ershaghi, I, Aminzadeh, F., Li, Y., and Cassidy, S., 2019,
Reservoir Geostatistical Estimation of Imprecise Information Using Fuzzy
Kriging Approach, Accepted for SPE-190051- Reservoir Evaluation &
Engineering.
65. Zhong, Z., Sun, A. Y., & Wu, X. (2020). Inversion of Time‐Lapse Seismic
Reservoir Monitoring Data Using CycleGAN: A Deep Learning‐Based
Approach for Estimating Dynamic Reservoir Property Changes. Journal of
Geophysical Research: Solid Earth, 125(3), e2019JB018408.
Index
2XDCR07D.FOR software, 116–117 Additive committee machines (CM),

2XDCR07E.FOR software, 128 203–206
2XDCR07F.FOR software, 128–129 Additive permeability model, 198–200
2XDCR*FOR software series, 114 Advanced geochemical technologies
3D Model Showcase, 455–456 (AGTs), 70
3D petrophysical modeling, 508–512 Advection-Dispersion-Equation
3D seismic-assisted CO2-EOR flow (ADE), 435–437
simulation, 464–483 Aggregated AD classifier, optimization,
CO2-EOR, 479–483 58–61
discrete fracture network (DFN), Ahmadi, K., 516
469, 471, 473 Akrad, O. M., 407, 413
geological background, 468–469 Alcova anticline, 481
introduction, 464–465, 467 Alizadeh, B., 507
petrophysical modeling, 473 Alqahtani, A. A., 410
presentation sequence, 464 al-Tahini, A. M., 401
PVT analysis, 473–474 EL-Amin, M. F., 432
streamline analysis, 479 Aminzadeh, F., 491
3D seismic interpretation, with well Amplitudes, 129–132
measurements, 17 high/low amplitude, 86, 87–88,
4D seismic data, for DRC, 12, 13–14 99–100
large-amplitude mud pump noise,
Abousleiman, Y. N., 401, 408, 413 112
Acoustic(s), small amplitude faults, 226, 227f
impedance (AI), seismic attribute, time lapse seismic, 14
11, 14 variations in, 11
inefficiencies, 107 wave, 87–88, 91
multiwave acoustic velocity log, 219 Anadarko, 302, 308
MWD mud pulse telemetry, 82, 83, Analysis methodology,
84, 87, 89 non-Newtonian fluid flow tests,
nodes and antinodes, 102 177–180
parameters, nonlinear, 228 dimensionless wellbore storage, 178
properties, gross, 17 finding n value, 177–178
signal measurement, 97 match point, 179–180
wind tunnel, 99f type curves, use, 178–179
AD. see Anomaly detection (AD) uncertainty, 180
525
526 Index
Analytical model for predicting fluid p-values for hypothesis tests, 63t,
profile. see Wellbore temperature 64
profile during drilling quality characteristics of, 52
gas-hydrates reservoirs sparsity, 51, 55, 56–58, 59–60, 61,
Analytical thermal-model for 63f, 64
optimization of gas-drilling, trainSetRecords, 59
340–351 universal classifiers, 58–61
introduction, 340–341 identification rule, 50
mathematical model, 341–345 methodology, 49–50
model applications, 349–351 overview, 48–49
model comparison, 346–348 performance
sensitivity analysis, 348 one-class anomaly detection
Anifowose, F., 507 technique, 53
Anisotropy, 37, 406–407, 413 posterior quality characteristics,
Anomalous cluster sets, prior FD rates 52–55
for, 142–143 prior quality characteristics,
Anomalous records, 52–55
absence of, 142 ROC-curve analysis, 52
efficiency and reliability of posterior analysis of efficiency of,
identification, 145 144–146
prior and posterior TD and FD rates ROC curve analysis, 55–58
for, 142 training and test sets, 49, 55, 59, 60
in randomized test sets, 144f Anomaly index(es),
Anomaly cutoff, 53 cluster, 140–141, 144–146
calculation of, 52 gas and brine datasets, 146–148
defined, 50 individual records, 141, 144–146
functions of, 52 regular records, 145
Anomaly detection (AD), Anomaly indicator,
bootstrap bootstrap based tests of anomaly
based tests, 61–64 type hypothesis, 61, 62–63
for statistical analysis of, 52 percent of values of, 63t
classifiers for three types of classifiers, 63f,
adaptive aggregated, 52, 63f, 64 64
aggregated, optimization of, Apache, 297, 308
58–61 ArcGIS Desktop 10.6, 506, 511
Anomaly records, 59, 60 Area under ROC curve (AUC)
AUC for, 55–58 analysis, 159, 160
construction of, 49, 50 posterior, 55–58, 61t
distance, 50–51, 55, 56–58, 61, values, ranking parameters
63f, 64 according to, 161–162
divergence, 51, 53–55, 56–58, 61 Arnold, F. C., 364
histograms of bootstrap generated Arps hyperbolic model, 270
values, 63f, 64 Artificial intelligence (AI), 491
Index 527
Artificial intelligence and machine modeling. see Forecast and

(deep) learning applications uncertainty analysis of
to reservoir characterization, production trends
502–513 for statistical analysis of AD, 52
3D petrophysical modeling, training and test sets, 49, 55
508–512 Borehole environment, physical and
3D structural model development, mechanical coalbed properties in,
503–506 225–228
dynamic modeling and simulations, Bottom hole assembly (BHA), 340–341
512–513 Bottom hole pressure (BHP), 480
sedimentary modeling, 506–508 Bouroullec, R., 259, 263
Artificial intelligence and machine BP, 294, 296, 298, 305
(deep) learning review, 491–494, Bradley, W. B., 413
496–502 Bregman divergence, 51
artificial neural networks (ANN)- Brine-filled sands,
based methods, 498–502 combination of records from, 60
clustering (unsupervised dataset as data with regular records,
classification), 492–494, 496–497 146–148
ensemble methods, 497–498 parameter ranking and efficiency,
support vector machines, 492 155–159
Artificial neural networks (ANNs), records from, 136, 137
for EOR performance and Brine-filled shaly formations,
economics, 514, 516 classification of. see Monte-Carlo
for EOR screening, 516–517 simulation of seismic velocities in
Auto encoders (AE), 502 shaly sands
Azarroual, 432 Brine saturated sands,
Poisson’s ratio, 62
Back pressure regulator (BPR), 33 shear wave velocity of, 29
Backpropagation through time Bulk modulus (K),
(BPTT), 502 pore fluid composition, 29
Bakken, North Dakota, 407, 507 rock sample, 30, 31, 39
Bandyopadhyay, K., 406 Bureau of Ocean Energy Management
Barnett shale, 404, 414 (BOEM), 256, 259, 261
Baruah, R., 507
Basis function expansion, 200–201 Calcite-rich shale rock, 413
Bedding plane orientation, 408–410 Canadian National, 297
Bentley, L. R., 401 Capital equipment, 300
Bentsen, R. G., 347, 364 Carbonate, 413, 420, 468
BG Group, 320, 323 Carbonate-derived oil,
BHP Billiton (Australia), 297 biomarker characteristics, 72, 77t
BiLSTM, 507 GC-fingerprint, 72, 76f
Bitumen, 72, 75f triterpane fingerprints, 78
Bootstrap(s), Carbonate source-derived
based tests, 61–64 hydrocarbons,
528 Index
overview, 70–71 trainSetRecords, 59

results and discussions, 72–79 universal classifiers, 58–61
bulk compositions of oils and Clay content, 407
bitumen, 72, 75f Clay-rich rocks, 407
GC-fingerprint, 72, 76f, 77f Closmann, J., 413
kerogen conversion and maturity, Clustering assemblies,
72, 75f anomaly indexes of individual
TOC and Rock-Eval pyrolysis, records and, 141
72, 73t construction of, 138–139
samples and analysis, 71–72 defined, 141
Carbon dioxide, injecting, 28, 33–34, 35 for detection of geologic anomalies.
Cascade fields, 263, 499 see Geologic anomalies, detection
CBM. see Coalbeds methane (CBM) instability of process, 136–138
C29/C30 hopane, 78 posterior analysis of efficiency of
Chen, L., 509 anomaly identification, 144–146
Cheng, Y., 270 prior FD rates for anomalous cluster
Chesapeake, 308 sets, clusters, and individual
Chevron, 294, 296, 298, 301, 305 records, 142–143
Chiaramonte, L., 467, 469, 471, 473, Cluster set, irregularity index of
475f, 480 individual clusters in, 139–141
China National Petroleum CMs. see Committee machines (CMs)
Corporation (CNPC), 82, 97, 102, CNOOC, 320, 323
296 Coalbeds methane (CBM),
Chinook fields, 263 developing technologies for, 221
Choe, J., 508 development of, 218
C30-hopane (hopane), 78 resources, 217–218, 221, 226, 228
C29-hopane (norhopane), 78 section, high permeability zones in,
Chopra, S., 503 225
Chugwater and Permian Goose Egg sorption capacity of, 220
formations, 469 Coal reservoir, physical properties
Classifiers, AD, and external load conditions on,
adaptive aggregated, 52, 63f, 64 219–225
aggregated, optimization of, 58–61 Coates, R., 402
Anomaly Records, 59, 60 CO2-EOR. see 3D seismic-assisted
AUC for, 55–58 CO2-EOR flow simulation
construction of, 49, 50 Committee machines (CMs),
distance, 50–51, 55, 56–58, 61, 63f, 64 additive, multiplicative, and
divergence, 51, 53–55, 56–58, 61 exponential, 203–206
histograms of bootstrap generated first level, permeability prediction
values, 63f, 64 with
p-values for hypothesis tests, 63t, 64 accuracy of outliers replacement
quality characteristics of, 52 by, 212–214
sparsity, 51, 55, 56–58, 59–60, 61, carbonate reservoirs, 210–212
63f, 64 sandstone dataset, 206–210
Index 529
Monte Carlo Cooper, S. P., 469

construction of, 236–237 Core data, 3, 8, 9–10
rock permeability forecasts Cortes, C., 492
using, 233–236; see also Rock Costs,
permeability forecasts extraction, 299
overview of, 249–251 finding, 299–300
second level, accuracy of outliers lifting, 299
replacement by, 212–214 production, 299
support vector machines (SVM), 246 Cretaceous oils, GC-fingerprint of, 72,
two-level, 200 77f
Commodity prices, 298–299 Cross validation, MC,
Company value, factors that impact, extended, performance of, 236–237
295–303 parameters of distribution of
assets, 300 number of individual forecasts,
capital structure, 300–301 237–238
commodity prices, 298–299 rock permeability forecasts using,
degree of integration, 297–298 233–236
finding cost, 299–300 Cross validation, ROC curve analysis
geographic diversification, 301–302 with, 159–161
geologic diversification, 301 Crystalline dolomite, 469
ownership, 295–297 Cycle generative adversarial neural
production cost, 299 network (CycleGAN) model, 512
product mix, 298
unobservable factors, 302–303 Daigle, H., 507
Completely connected perceptron Darwis, S., 270
(CCP), 251 Dasgupta, F., 491
Compressional wave modulus, Data, for reservoir characterization,
of brine saturated rock, 31 geological, 5, 6, 10
calculation, 30–31 integration challenges, 7–10
of rock’s minerals, 31 measurements, 7
Compressional wave velocity, microearthquake, 12, 14, 15f
cross plot of, 36f, 37 rate(s)
Greenberg-Castagna formula, 34f,35 high, MWD mud pulse telemetry.
rate of variability of experimental/ see Mud pulse telemetry,
estimated velocities, 38, 39f measurement-while-drilling
rock saturated with brine, 30 (MWD)
shear wave velocity, plot of, 37–38 siren, 84
shear wave velocity using, 29 requirements, 5–6
Computational intelligence (CI), 491 seismic. see Seismic data
COMSOL, 175, 176 types, 7–10
ConocoPhillips, 294, 296 DBSCAN, 497
Consolidated subsidiaries, 304n7 Decline curves. see Forecast
Continuous injection, 445–447, 449 and uncertainty analysis of
Convolutional neural networks production trends
(CNN), 500–501, 503 De-hua Han, 385
530 Index
Denoo, A., 402 Eagle Ford, 410

Desurger, 82, 84, 86, 111, 112, 114 Earthquake engineering, 86
Devils Tower fields, 263 Echo cancellation,
Devon, 297, 308 data for, 98
Dewhurst, D. N., 401, 413 multiple-transducer methods, 103
Dibenzothiophene/phenanthrene MWD, design, 112
ratio, 78, 79f ECLIPSE, 439
Diegel, F. A., 257, 262 EDM (elastic deformation modules),
Dipole Shear Sonic Imager, 27 of coalbeds, 219, 221, 222f
Directional wave filtering, 111–132 Effective pressure, influence on shale’s
background remarks, 111–112 rock-mechanical properties,
calculations, 116–132 401–402
theory, 112–116 Elastic coefficients,
Discrete fracture models (DFM), calculation, 27–28
465 of rock samples, 39
Discrete fracture network (DFN), 469, Elastic deformation modules (EDM),
471, 473 of coalbeds, 219, 221, 222f
Dissimilarity analysis of petrophysical Elastic modulus of rock sample,
parameters. see Petrophysical Greenberg-Castagna model, 28–29,
parameters, dissimilarity analysis 31, 35, 37, 38–39, 40
Distance classifier, AD, 50–51, 55, laboratory set up and
56–58, 61, 63f, 64 measurements, 32–34
Divergence classifier, AD, 51, 53–55, methodology, 28–31
56–58, 61 overview, 26–28
Dolomite, 469 results and discussion, 35–41
Downhole source and signal Elastic properties of shale, 404, 407,
optimization, 89–92 412, 413, 419
Drill cuttings, 341, 343, 348, 351, 364, Elastic waves,
374, 375 empirical regression formulas for, 27
Drilling optimization, 340 velocity and coefficients, 27–28
Dual porosity (DP) models, 464–465 Empirical permeability models, linear
Duhamel’s principle, application of, regression permeability forecast
170–171 with, 238–242
Duong model, 271, 275f Energy sustainability, high data rates
Dvorkin, J. P., 402 and, 82–83
Dynamic (non-destructive) method, Enhanced oil recovery (EOR), 5, 12
for elastic coefficients, 28 Enhanced oil recovery (EOR) using
Dynamic reservoir characterization hydrophilic nanofluids, 430–461
(DRC), 4, 12–15 coupling of mathematical models,
4D seismic for, 12, 13–14 437–439
microseismic data for, 12, 14–15 discussions, 457–459
Dynamic reservoir model, 18 introduction, 430–432
Index 531
results, 443, 445–447, 449–453, ROC curve analysis, 55, 159–161

455–456 true and, 52, 53, 55
simulation framework, 432–437 Fan, T., 431, 432
verification cases, 439–443 Fayetteville shale gas reserves, 340
Eni, Italy, 296–297 FDR. see False discovery rates (FDR)
EOG Resources, 308 Ferguson, D., 469
Eppelbaum, L. V., 341, 365 Feynman, R., 430
Equation of state (EOS), 467 Finance-Yahoo, 305
Erosion analysis, MWD mud pulse Financial Times Global 500, 305
telemetry and, 93–96 Finding cost, 299–300
Erosion computer modeling, 93–96 Finite element use, 176–177
Espinosa, 430 First level CMs, permeability
Exhaustive search, prediction,
of high accuracy linear regression accuracy of outlier replacement,
models, 198 212–214
for optimum subset of predictors, carbonate reservoirs, 210–212
200–201 sandstone dataset, 206–210
Expected false discovery rate (expFD), Flow-based numerical method, 465
50, 52, 53, 54, 55, 60, 61 Fluid content dissimilarity of values,
Expected-posterior ROC curve, 55 petrophysical parameters, 154
Exploration phase, reservoir Ford, R., 340
characterization in, 10–11 Forecast and uncertainty analysis of
Exponential committee machines production trends, 270–287
(CM), 203–206 block bootstrap forecast and
Exponential permeability model, analysis of forecast uncertainty,
198–200 284–285
External load conditions, on coal comparative analysis of results
reservoir, 219–225 of Monte Carlo and bootstrap
Extraction cost, 299 simulations, 285–287
Exxon, 294, 296, 298, 301, 305 grid search followed by iterative
minimization with Levenberg-
False discovery rates (FDR), Marquardt algorithm, 276–277
for anomalous cluster sets, clusters, introduction, 270–271
and individual records, 142–143 iterative minimization of
of anomalous records, 147, 148 least squares with multiple
anomaly detection with, 149 approximating models, 275–276
defined, 160 Monte Carlo forecast and analysis of
expected, 50, 52, 53, 54, 55, 60, 61 forecast uncertainty, 280–282
identification of gas-sands with, 159 new method of grid search for
with KNN and LDA classifiers, 163 approximation and forecast of
posterior, 53, 54–55, 62, 142 decline curves, 273–275
prior and posterior, 142 nonlinear least squares for decline
probabilities of, 155 curve approximation, 273
532 Index
simulated decline curves, 271–272 temperature profile during

two methods for aggregated drilling gas-hydrates reservoirs
forecast and analysis of forecast Gas predictors, multidimensional
uncertainty, 277–278 parameters as, 163–164
uncertainty quantile ranges Gas-prone organic facies, 71
obtained using Monte Carlo and Gas sands,
bootstrap methods, 279–280 anomaly, detection of, 146–148
Form 10-K, 293 identification, petrophysical
Fort Worth and Permian Basin in parameters for, 152–154
Texas, 414 Gas-saturated sands,
Fourier Transform Infrared (FTIR) Poisson’s ratio, 62
technique, 407 records from, 136
Fractures, 219, 220, 223, 225, 228, 465 Gassman relations, 27, 28, 29, 35, 38,
Fracture zone identifier (FZI) 40
attributes, 14 Gas temperature profiles, 341
“Frequency-shift-keying” (FSK), 87, Gaviria, R., 467, 473
88, 100, 118 GazProm, 297
Friedmann, J., 467 General linear model (GLM), AUC
FSK (“frequency-shift-keying”), 87, 88, values, 161, 162
100, 118 Generative adversarial networks
Fuzzy clustering, 494 (GANs), 502, 508
Fuzzy logic method, 505–506 Geo-cellular model, 17–18
Fuzzy set theory, 494 Geological and geophysical criteria,
overview, 217–218
Galloway, W. E., 257 physical and mechanical coalbed
Gas chromatograms (GC-fingerprints), properties in borehole
of carbonate and shale-derived environment, 225–228
oils, 72, 76f physical properties and external
Gas-drilling. see Analytical thermal- load conditions on coal reservoir,
model for optimization of 219–225
gas-drilling Geological data,
Gas-filled sands, geophysical and, 5, 6
anomaly, 62 modeling petrophysical properties,
combination of records from, 60, 10
61 Geological reservoir characterization,
parameter ranking and efficiency, 18
155–159 Geologic anomalies, detection,
Gas-filled shaly formations, anomaly indexes of individual
classification of. see Monte-Carlo records, 141
simulation of seismic velocities in brine sand dataset as data with
shaly sands regular records, 146–148
Gas hydrates deposits, temperature inhomogeneity of training and test
profiles in drilling. see Wellbore sets and instability, 136–138
Index 533
irregularity index of individual basin development and geologic

clusters in cluster set, 139–141 overview, 257–259
multiple randomized test sets, hydrocarbons, 261–262
formation, 138–139 introduction, 256–257
notations, 149 petroleum system, 259
overview, 135–136 reservoir geology, 259–261
posterior analysis of efficiency of salt and structure, 262–263
anomaly identification, 144–146 Gupta, N., 406
prior and posterior TD and FD
rates, 142 Hager, R., 413
prior FD rates for anomalous cluster Handin, J, 413
sets, clusters, and individual Han’s equation, for shale sand rocks, 27
records, 142–143 Hasan, A. R., 341, 347, 365
records in gas sand dataset as He, J., 507
anomalous, 146–148 Heat capacity of gas, 342, 347
Geologic diversification, 301 Heidari, Z., 411–412
Geomechanical parameters, Hendraningrat, L., 430
estimating, 31–32 Hennings, P. H., 481
Geophysical data, Hierarchical clustering technique, 494,
borehole, 9 496–497
with geologic data, 5, 6 High data rate MWD mud pulse
integration of, 20 telemetry, 82–83; see also Mud
monitoring, 19f pulse telemetry, measurement-
and petrophysical properties, 17 while-drilling (MWD)
tools, 6 High gas permeability of coalbeds,
vertical and spatial resolution, 9f zones of, 218; see also
George, B., 340 Geological and geophysical
Geostatic pressure in wells, 223f criteria
Giro, R., 517 Holmes, C. W., 364
Glowka, D. A., 350 Holt, C., 340
Gnuplot, 445 Hood, K. C., 261–262
Goal Seek tool, 345 Hudec, R., 262
Government sponsored enterprises Hussein, M., 504
(GSEs), 296–297, 300, 320, Hybrid permeability model, 198–200
321t–322t, 323 Hydrogen index (HI), 72
Grayson, E. S., 340 Hydrophobic and lipophilic
Great White field, 263 nanoparticles (HLPN), 435
Greenberg-Castagna model, 28–29, 31,
35, 37, 38–39, 40 IHS Herold, 294
Green River, 410 IHS PetraR, 506
Guillen-Rondon, P., 503 Imamverdiyev, Y., 507
Gulf of Mexico (GOM) petroleum IMplicit Pressure Explicit Saturation
system, (IMPES) approach, 437
534 Index
Inclined direct shear testing device Iraq National Oil Company (INOC), 320
(IDSTD), 413 Irregularity index, of individual
Incompressibility, of reservoir rock, clusters in cluster set, 139–141
14 Islam, M. A., 402
Independent oil and gas companies,
307t, 308–309, 327t–335t Jeong, S. W., 431
Individual clusters, irregularity index Jochen, V., 270
of, 139–141 Joukowski formula, 87
Individual forecasts, Joule-Thomson cooling effect, 341–
enhancement of stability of MC 342, 348, 350, 365, 372, 375–377
CMs forecast via increase of Ju, B., 431, 432
number, 246–247 Justman, D., 505
instability of, analysis, 244–246
by linear regression and machine Kabir, C. S., 341, 347, 364–365
methods, 239–240 Kang, Q., 508
with ML methods, 242–244 Karstifed Limestone (Papua New
parameters of distribution of Guinea), 340
number of, 237–238 Kavousi, P., 467, 469, 471
Individual records, KazMunaiGas (Kazakhstan), 296
anomaly indexes of, 141 Keller, H. H., 346, 364
prior FD rates for anomalous cluster Kernel functions (KF), 492
sets, clusters and, 142–143 Kerogen, 71, 407, 413
Inhomogeneity of training and test conversion and maturity, 72, 75f
sets, 136–138 type, 72, 74f
Instability, Kim, T. Y., 431
of clustering process, 136–138 K-means clustering, 493
of individual forecasts, analysis, K-nearest neighbor (KNN), 391–392
244–246 AUC values, 161, 162
Instability index, individual forecasts with, 242–244
of individual forecasts, 246 true and false discovery rates with,
for MC CM forecast, 235 163
Integrated oil company reserves, 306t Kumar, M., 402, 407, 413
Integration, Kutasov, I. M., 341, 365
conventional seismic, 15f Kuwait Oil Company (KOC), 320
data Kuybyshev, Along-Volga Region,
challenges, 7–10 Russia, 249
geophysical, 6, 20 Kuzbass (Russia), CBM resources,
seismic and geologic data, 7–10 217–218, 221, 226, 228
3D seismic interpretation, 17
information from seismic data, 18 Lai, B. T., 404
reservoir structure, 5f Langmuir-like model, 436
Intel i5 class chipsets, 112 Laplace space solution, 172
International oil companies, 295–296, Laplace transform, 170, 189–190
300, 305, 310–311 Laramide Orogeny, 257
Index 535
Lashkaripour, G. R., 404 to reservoir characterization,

Lays, R., 340 502–513
Lecoanet, H., 430–431 artificial intelligence and machine
Leco carbon analyzer, 72 (deep) learning review, 491–494,
Leung, P. K., 404 496–502
Levenberg-Marquardt algorithm, 275, brief introduction to reservoir
276–277 characterization, 489–491
Levenberg-Marquardt type machine (deep) learning and
minimization, 276–278, 284 enhanced oil recovery (EOR),
Li, J., 341, 347–348, 409 513–514, 516–517
LIDAR, 481 Machine learning (ML),
Lifting costs, 299 permeability prediction using; see
Lin. Z-Sh., 270, 404 also Permeability prediction
Linear discriminant analysis (LDA), committee machines (CM). see
160, 394 Committee machines (CMs)
AUC values, 161, 162 nonlinear ML method, 198, 199
true and false discovery rates with, rock permeability forecasts using;
163 see also Rock permeability
Linear regression permeability forecasts
forecast, 238–242 accuracy of forecasts with,
Lipophobic and hydrophilic 242–244
polysilicon nanoparticles Macondo oil spill, 310
(LHPN), 435–436 Macroporous coal permeability, 219
Liquid-drilling, 341 Mancos shales, 410
Lithologic dissimilarity of values, Maranuk, C., 340
petrophysical parameters, 154 Marathon Oil, 308
Lithology identification, 506 Marcellus shale, 340
Liu, C. L., 406 Marfurt, K. J., 503
Log data, 3, 8, 9, 10 Mark-Moser, M., 512
LONI (Louisiana Optical Network Marouby, R., 411–412
Initiative), 433 Mars, 263
Lorenz, J., 469 Marshall, D. W., 347, 364
Lost circulation material (LCM), 106 MC. see Monte Carlo (MC)
Louann Salt, 257–259 Mean absolute bias (MAB), 206–210,
Lower Tertiary Wilcox, 260 234, 242
Mean absolute error, defined, 234
Machine (deep) learning and Mean instability of individual forecasts
enhanced oil recovery (EOR), (MINST), 206–210
513–517 Mean square error, 514
Machine learning in reservoir Message Passing Interface (MPI)
characterization, 489–517 standard, 433
artificial intelligence and machine Method 4-3, directional wave filtering,
(deep) learning applications 112, 114, 115, 116–125, 128–129
536 Index
Method 4-4, directional wave filtering, random models for seismic

114, 115, 116, 126f, 127f, 128f, velocities, 385–386
129–132 reliability analysis of identifying gas-
Microearthquake (MEQ) data, 12, 14, filled formations, 389–396
15f separability of (Vp, Vs) clusters
Microporous coal permeability, 219 for Gasand Brine-saturated
Microseismic data, for DRC, 12, formations, 388–389
14–15 variability of seismic velocities
Mineralogy, 411–413 predicted by random models,
Minimum miscibility pressure (MMP), 387–388
465, 467 Mud pulse telemetry, measurement-
Miocene reservoirs, 261 while-drilling (MWD),
ML-based inverse models, 513 directional wave filtering, 111–132
MNM1D (Micro and Nanoparticle background remarks, 111–112
Transport Model in porous media calculations, 116–132
in 1D geometry), 439, 442 theory, 112–116
Mobil Oil, 85 downhole source and signal
Modular neural networks, 249–251 optimization, 89–92
Mohr’s theory, 402 high data rates and energy
Monomineralic rocks, 35 sustainability, 82–83
Monte Carlo (MC) clustering new telemetry approach, 87–88
assemblies, geologic anomalies overview, 83–85
with. see Geologic anomalies, pressure, torque and erosion
detection computer modeling, 93–96
Monte Carlo (MC) rock permeability prototype single-siren tool, 88f, 89f
forecasts using; see also Rock surface signal processing and noise
permeability forecasts, removal, 92–93
committee machines, 203, 233–236 telemetry basics, 85–87
cross validation, 233–236 test results, example, 108–111
enhancement of stability of MC wind tunnel analysis, 96–107
CMs forecast, 246–247 Mud sirens, high data rates, 82, 84
parameters of distribution of Mudunuru, M. K., 512
number of individual forecasts, Multidimensional parameters as gas
237–238 predictors, 163–164
performance of extended MC cross Multi-layer perceptron (MLP), 250,
validation and construction of 251, 498, 507
MC CMs, 236–237 Multiple randomized test sets,
Monte Carlo modeling. see Forecast formation, 138–139
and uncertainty analysis of Multiplicative committee machines
production trends (CM), 203–206
Monte Carlo simulation, 279 Multiplicative permeability model,
Monte-Carlo simulation of seismic 198–200
velocities in shaly sands, Multiwave VSP technology, in CBM
384–396 section, 219, 226, 227f
introduction, 384–385 Murphy, M. J., 431
Index 537
Naive Bayes Classifier, 517 finite element, use, 176–177

Nano-indentation technique, 402 objective, 173
Nanotechnology in oil industry. see overview, 170–173
Enhanced oil recovery (EOR) problem analysis, 173–176
using hydrophilic nanofluids model assumptions, 174
National Gas Hydrate Expedition solution by mathematical
Programme (NGHP-01), India, inspection, 175–176
373–374 solution verification, 176
National Iranian Oil Company wellbore storage and skin effects,
(NIOC), 320 175
National oil companies, 294, 296, 320 wellbore storage distortion,
Natural gas permeability of coalbeds, solution without, 175
218 test data examples, 180–188
Naturally fractured reservoirs. see 3D analysis recommendations, 185,
seismic-assisted CO2-EOR flow 186
simulation match point, 182, 183–184, 185,
NbClust R clustering package, 136 186
Nealon, J., 503 real field data, 181–182
Neural networks (NNs), 242–244, 491, simulated data, 182–185
494, 504 Non-stochastic simulation models,
New Power-Exponential model, 271 508
Newton-Raphson iteration, 345 North Sea Shale, 413
Nguyen, D., 365 NPR-3, 468
Nigeria National Petroleum Company, Nuclear test detection, 86
296 Nur, A., 385
Noble energy, 308 Nyquist-Shannon Theorem, 112
Noise removal, Occidental, 297
downhole signal enhancement and, Oda, M., 465
95 Oil and gas company production,
surface signal processing and, reserves, and valuation,
92–93, 97 290–323
Non-linear boundary condition, factors that impact company value,
189–190 295–303
Nonlinear least squares (NLS), 273 government sponsored enterprises
Nonlinear permeability models, 199 and other international
Non-Newtonian fluid flow tests, companies, 320, 323
analyzing, international oil companies,
analysis methodology, 177–180 310–311
dimensionless wellbore storage, introduction, 290–291
178 market capitalization, 309–310
finding n value, 177–178 National Oil Companies of OPEC,
match point, 179–180 320
type curves, use, 178–179 private companies, 318
uncertainty, 180 production, 294–295
538 Index
reserves, 292–294 additive, multiplicative, exponential,

summary statistics, 303–305, and hybrid permeability models,
308–309 198–200
U.S. Independents, 312–318 basis function expansion, 200–201
Oil-base muds, 377 with first level committee machines
Oil recovery factors, 439 carbonate reservoirs, 210–212
Oil water contact (OWC), 13f sandstone dataset, 206–210
Ojukwu, C., 503 high gas permeability of coalbeds,
Oletu, J., 259 225; see also Geological and
One-class anomaly detection geophysical criteria
technique, 53 of coals, 220
ONGC, 297 geological and geophysical gas,
Optimum subset of predictors, 218
exhaustive search for, 200–201 macroporous, 219
Orenburg Field, Russia, 249 microporous, 219
Otchere, D. A., 507 natural gas, 218
Outliers, optimum subset of predictors,
identification of, 201–203 exhaustive search for, 200–201
replacement outliers, identification of, 201–203
algorithms for, 204–206 outliers replacement
by first and second level CMs, algorithms for, 204–206
accuracy, 212–214 by first and second level CMs,
accuracy, 212–214
ParaView, 432 overview, 197–198
Particle swarm optimization (PSO), rock, forecasts using ML. see Rock
514 permeability forecasts
Partitioning algorithm, 392–393 workflow for, 214f
Parvazdavani, M., 431, 435, 440 Petrobras, Brazil, 296–297, 320, 323
Passaris, E. K., 404 PetroChina, China, 296–297, 320, 323
PCA, 504 Petroleum Resources Management
Peng-Robinson equation of state, 467, System (PRMS), 293
477f Petronas (Malaysia), 296
Perdido Fold Belt, 263 Petrophysical modeling, 473
Performance, anomaly detection, Petrophysical parameters, dissimilarity
one-class anomaly detection analysis,
technique, 53 for gas-sand identification, 152–154
posterior quality characteristics, lithologic and fluid content
52–55 dissimilarities of values, 154
prior quality characteristics, 52–55 multidimensional parameters as gas
ROC-curve analysis, 52 predictors, 163–164
Permeability dataset, 142–143 overview, 152
Permeability prediction, ranking and efficiency of
additive, multiplicative, and identification of gas-sands,
exponential CMs, 203–206 155–159
Index 539
ranking parameters according to Pressure analysis, MWD mud pulse

AUC values, 161–162 telemetry and, 93–96
ROC curve analysis with cross Pressure transient analysis problems,
validation, 159–161 180
Pham, Q. T., 404 Priezzhev, I., 507
Phananthrene/dibenzothiophene, 78, Principal component regression
79f (PCR), individual forecasts with,
Phase-shift-keying (PSK), 86, 87 242–244
Physics-informed machine learning Prior and posterior TD and FD rates,
(PIML) workflow, 512–513 142
Phytane/pristane ratio, 72, 78, 79f Pristane/phytane ratio, 72, 78, 79f
Pioneer, 308 Private companies in the oil and gas
Pleistocene reservoirs, 261 industry, 308, 319t
Pletcher, J. P., 340 Probability of true discovery, 384,
Poisson’s ratio, 389–395
for calculation of AD classifiers, 55 Production cost, 299
porosity and, 402, 404 Production phase, reservoir
as single anomaly characterization characterization in,
function, 62 primary, 11
Polynomial Kernel functions (KF), 492 secondary/tertiary, 11–12
Pore fluid composition, effective bulk Pubs.usgs.gov website, 136, 198,
modulus of, 29–30 233
Porosity, PVT (pressure, volume, and
influence on shale’s rock-mechanical temperature) analysis, 473–474
properties, 402 P-wave anisotropy, 413
physical geology information, 26 P-wave velocities, 27
pore space, properties, 4 at different effective pressures, 34f
rocks property, 16, 17, 18 vs. laboratory measurements, 36f, 37
shear wave velocity to, 28
static reservoir model, 18 Quartz, 413
wave propagation velocity to, 27
Porous media, fluid flow through, 173, Radial kernels, 492
174, 179 Ramey, H. J., 364
Postflush, 452–453, 455, 456, 457–459 Random forest (RF), AUC values, 161,
Pournik, M., 516 162
Power law indices, type curve charts Ranking parameters according to AUC
for, 191–194 values, 161–162
Predicting fluid profile, analytical Rate of penetration (ROP), 349–351,
model for. see Wellbore 375, 377
temperature profile during Raymond, L. R., 364
drilling gas-hydrates reservoirs Rectified Linear Unity (ReLU), 500
Predictors, Recurrent neural networks (RNN),
optimum subset of, 200–201 501–502
quantitative, 239–240 Recursive partitioning, 392–394
540 Index
Regression coefficients, Greenberg- SURE challenge, 7–10

Castagna formula, 29 Reservoir modeling for reservoir
Regression tree (RT), individual simulation, 15–20
forecasts with, 242–244 components, 18, 19f
Regression with partial least squares integrated, 18, 19f
(PLSR), 242–244 process of, 17–20
Regular records, rock physics, 15, 16–17
brine sand dataset as data with, Reservoir monitoring,
146–148 drainage patterns, 12
prior and posterior TD and FD rates EOR operation, 12
for, 142 geophysical, 18–20
Relevant Vector Machine, 507 hydraulic fracturing, 14
Reliability analysis of identifying gas- Reservoir pressure condition,
filled formations, 389–396 laboratory measurements at,
classification with K-nearest 32–34
neighbor, 391–392 Reservoir simulation, reservoir
classification with linear modeling for, 15–20
discriminant analysis, 394 process, 17–20
classification with recursive rock physics, 15, 16–17
partitioning, 392–394 Residual oil zones (ROZ), 514
comparison of the three R function lm(), 233
classification techniques, 395–396 RMOTC (Rocky Mountain Oil Testing
Reserves, 292–294 Center), 467
contractual differences, 294 ROC curve analysis,
probable and possible reserves, 293 AD performance, 52
production, 294–295 with cross validation, 159–161
proved reserves categories, results of AD, 55–58
292–293 AD classifiers, 55–58
reserves reporting, 293 AUC, 55–58
Reservoir characterization, expected-posterior, 55
components of, 4, 5f posterior, 55
data requirements for, 5–6 types, 55
DRC, 12–15 Rock-Eval pyrolysis, results of, 72, 73t
4D seismic for, 12, 13–14 Rock permeability forecasts,
microseismic data for, 12, 14–15 accuracy of forecasts with ML
exploration, development and methods, 242–244
production phases, 10–12 analysis of instability of forecast,
modeling for reservoir simulation, 244–246
15–20 bias of forecast with permeability,
components, 18, 19f 241f
integrated, 18, 19f enhancement of stability of MC
process of, 17–20 CMs forecast, 246–247
rock physics, 15, 16–17 errors, bias, instability, and
overview, 3–5 correlations of forecasts, 240t
Index 541
individual forecasts, increase of dataset, permeability forecast

number of, 246–247 first level CMs, 206–210
linear regression permeability outlier replacement by first and
forecast with empirical second level CMs, 212–214
permeability models, 238–242 sample, 32, 35
MC cross validation and MC CMs, Sapropelic organic matter, 70–71
233–236 Sapropelic (Type II) kerogen, 72, 74f
MC CMs, construction of, SAS14D.FOR software, 129–132
236–237 Saudi Aramco (Saudi Arabia), 296, 320
parameters of distribution of Sbai, M. A., 432
number of individual forecasts, Scale, Uncertainty, Resolution, and
237–238 Environment (SURE) challenge,
performance of extended MC 7–10, 491
cross validation, 236–237 Scottrade, 305
modular networks or CMs, Sedimentary modeling, 506–508
249–251 Seismic data, 3, 7, 8–10
ordered values of permeability and 3D, 6
their forecasts, 241f 4D, 4–5, 6
overview, 232–233 for DRC, 12, 13–14
permeability models from different microseismic, 4–5
fields, description, 248–249 for DRC, 12, 14–15
Rock physics, 15, 16–17 time-lapse, 13–14
analysis, on logs and cores, 28 Seismic velocities; see also Monte-
bulk modulus, 29–30 Carlo simulation of seismic
compressional wave velocity, 29, 30 velocities in shaly sands
shear wave velocity, 28–31 random models for, 385–386
velocity and porosity, 28 variability of, 387
ROCR R software package, 160 variability predicted by random
Root mean square error (RMSE), models, 387
forecast by committee machine, Self-organizing maps (SOM), 494, 502,
234–235 503
of individual forecasts, 206–210 SEPD model, 271, 274, 275f, 281
Rose, K., 509 Sethi, R., 442
Rosneft, 297 Shale-derived oil,
R package ‘leaps,’ 198, 201 biomarker characteristics, 72, 77t
R/P ratio, 304 GC-fingerprint, 72, 76f
R rminer package, 233 triterpane fingerprints, 78
Rumelhart, D., 491 Shale gas, 400
Russia (Lukoil), 296 Shale mechanical properties influence
factors, 400–418
Saiful Islam, 503 experimental investigation of water
Saikia, P., 507 saturation effects on Shale’s
Salt solutions, 405–406 mechanical properties, 414–418
Sandstone, influence factors, 400–413
542 Index
Shales, SIMPA (Spatially Integrated Multi-

Barnett, 404, 414 variate Probabilistic Assessment),
calcite-rich shale rock, 413 506
combination of records from, 60 Sinopec, China, 296–297
elastic properties of, 404, 407, 412, Sirens, mud,
413, 419 concepts and shapes, evaluation, 99
humic-type, 70–71 frequency, MWD collar for, 90f
Mancos, 410 high data rates, 82, 84
Marcellus, 340 pair of ganged or tandem, 100f
parameter ranking and efficiency, wind tunnel analysis, 97–103,
155–159 105–107
Shear modulus, changes in, 40 Skin distortion, problem, 175; see also
Shear wave velocity, estimated, Non-Newtonian fluid flow tests,
common methods for, 26–27 analyzing
compressional and, 34f Slug injection, 449–452
compressional wave velocity, plot of, Softmax, 500
37–38 Solvent extraction, bitumen from
cross plot of, 36f, 37 source rocks, 72
geomechanical parameters, 31–32 Sone, H., 407
Greenberg-Castagna model, 28–29, Sonic logs, 27
31, 35, 37, 38–39, 40 Sparsity classifier, AD, 51, 55, 56–58,
laboratory measurement, 27 59–60, 61, 63f, 64
laboratory set up and Specific internal surface, 4
measurements, 32–34 Spikes, K. T., 402
laboratory vs. estimated shear StackNet, 507
modulus, 40f Stamp, V., 467
laboratory vs. estimated Young’s Static method, for elastic coefficients,
modulus, 41f 27–28
methodology, 28–31 Static reservoir model, 15, 18
overview, 26–28 Statoil, Norway, 296–297
P-wave velocity, 34f Steady heat transfer solution, 355–361
rate of variability of experimental/ Steal drill pipe, 370
estimated velocities, 38, 39f Steiger, R. P., 404
results and discussion, 35–41 Stone, C., 350
Shell, 296, 298, 305 Streamline analysis, 479
Short wind tunnel analysis, 93–94, Structural complexity, 505
97 Subsurface trend analysis (STA),
Sigmoid function, 500 509–511
Signal optimization, downhole source Suhag, A., 507, 517
and, 89–92 Sukhostat, L., 507
Signal-to-noise ratio, 508 Sump, G. D., 364
SIGPROC-1.FOR software, 129 Support vector machines (SVM),
Sigsbee salt, 263 242–244, 246, 507
Index 543
SURE (Scale, Uncertainty, Resolution, Total, 296

and Environment) challenge, Total-organic-carbon (TOC), 71, 72,
7–10, 491 73t, 406–407, 507
Surface signal processing and noise Towler, B. F., 270
removal, 92–93 Training sets,
S-wave anisotropy, 413 inhomogeneity of, 136–138
S-wave velocities, in MC cross validation, 233–236
at different effective pressures, 35, Triterpane fingerprints,
36f carbonate-derived oil, 78
vs. laboratory measurements, 36f, 37 shale-derived oil, 78
Swift, S. C., 364 True discovery rates (TDR),
Swish, 500 anomalous records with anomaly
index, 145
Talisman energy, 308 anomaly detection with, 149
Tan, C. P., 409 for clustering assemblies, 148
Tanh function, 500 evaluation of, 152
Tarbert formation, 455 false and, 52, 53, 55
TDR. see True discovery rates (TDR) identification of gas-sands with, 159
Teapot Dome Field. see 3D seismic- with KNN and LDA classifiers, 163
assisted CO2-EOR flow mean values of, 55
simulation posterior, 54, 55, 60, 61, 65, 142
Tembely, M., 508 prior and posterior, 142
Temizel, C., 517 ROC curve analysis, 159–161
Tensleep Formation at Teapot Dome, Turbidite deposits, 259–260
USA. see 3D seismic-assisted Turbidite sands, 260–261
CO2-EOR flow simulation Type curve,
Tensleep static model, 475f charts for various power law indices,
Tertiary oils, GC-fingerprint of, 72, 77f 191–194
Test sets, for non-Newtonian fluid flow tests.
inhomogeneity of, 136–138 see Non-Newtonian fluid flow
in MC cross validation, 233–236 tests, analyzing
Thanh, V., 514 for non-Newtonian fluids, 172–173,
Thermal conductivities of steal drill 176, 177, 178–179, 181–188
pipe, 370 pressure and pressure derivative,
Thunder Horse, 263 191–194
Tian, X., 507
Time-lapse seismic data, 13–14 Uncertainty analysis. see Forecast
Tmax values, 72 and uncertainty analysis of
Tokpanov, Y., 506–507 production trends
Torque, Underbalanced drilling, 340
air-to-mud, 97 Uniaxial compressive strength (UCS),
MWD mud pulse telemetry and, 415
93–96 Uniaxial compressive tests, 401, 414
Tosco, T., 442 Uniaxial tests, 404
544 Index
Universal classifiers, for AD, 58–61 solution without, 175

Upper Ness, 455 Wellbore temperature profile during
Ursa, 263 drilling gas-hydrates reservoirs,
U.S. Independents, 312–316, 318 364–377
consolidated small-caps, 314 case study, 373–374
conventional vs. unconventional, introduction, 364–365
315–316 mathematical model, 365–372
large vs. small cap, oil vs. gas, sensitivity analysis, 374–377
312–313 Well data, 4, 5, 7, 8f
multinational vs. domestic, 314–315 Wilhide, S., 340
private companies, 318 Williams, B. B., 364
production and reserves, 316 Willson, S. M., 408
regression models, 316, 318 Wilson, T. H., 467, 469, 471, 481, 482
UTCHEM simulator, 516 Wind tunnels, in modeling downhole
mud flow, 96–107
Van Eeckhout, E. M., 404 distant multiple transducer array
Vanishing gradient problem, 500 setup, 106f
Vapnik, V., 492 flow meter, 103f
Vernik, L., 406 flow straighteners for upstream and
Versatile seismic imager (VSI), 226 downstream use, 102f
Vertical seismic profiling (VSP), ganged/tandem mud sirens, 100f
multiwave, 219, 226, 227f hub convergence effects on signal
Vuktylskiy Gas-Condensate Field, strength and torque, 101f
Russia, 249 long wind tunnel, 105
piezoelectric transducer closest to
Wadleigh, E. F., 467 siren, 106f
Wang, C., 341, 503 real-time data acquisition and
Water-base drilling mud, 370 control system, 104f
Water content effects on shale rocks. short “hydraulic,” 97, 98f
see Shale mechanical properties short wind tunnel, 104f
influence factors sirens tested in, 100
Water postflush, 452–453, 455 siren test section with differential
Wave velocity estimate. see Shear wave transducers, 103f
velocity torque, position and rpm counter, 104f
Weimer, P., 259, 262–263 very long “acoustic,” 98, 99f
Wellbore storage, distortion; see also Woodford shale, 413
Non-Newtonian fluid flow tests, Woodside (Australia), 297, 320, 323
analyzing, Wooley, G. R., 347, 364
dimensionless, 178 Wu, B., 365, 409
mathematical inspection, solution Wuerker, R. G., 402
by, 175–176
skin effects and, 175 Xiong, W., 504
solution verification, 176 XTO energy, 301
Index 545
Yenwongfai, H. D., 508 Zhang, M., 341, 365, 401, 402,

Young’s modulus, 41, 402 405–406, 430, 445
YPF SA, Argentina, Zhong, Z., 512
302n6 Zoback, M. D., 407
Yu, H., 431 Zones of high gas permeability of
coalbeds, 218; see also Geological
Zahm, C. K., 481 and geophysical criteria
Zargari, S., 407 Zreik, A., 340
Also of Interest
Check out these other related titles from Scrivener

Publishing
Other Books in the Series, “Sustainable Energy Engineeering”
Hydraulic Fracturing and Well Stimulation, edited by Fred Aminzadeh,

ISBN 9781119555698. The first volume in the series, Sustainable Energy
Engineering, written by some of the foremost authorities in the world on
well stimulation, this groundbreaking new volume presents the advan-
tages, drawbacks, and methods of one of the hottest topics in the energy
industry: hydraulic fracturing (“fracking”). NOW AVAILABLE!
Other Related Titles from Scrivener Publishing
Hybrid Renewable Energy Systems, edited by Umakanta Sahoo, ISBN

9781119555575. Edited and written by some of the world’s top experts in
renewable energy, this is the most comprehensive and in-depth volume on
hybrid renewable energy systems available, a must-have for any engineer,
scientist, or student. NOW AVAILABLE!
Energy Storage 2nd Edition, by Ralph Zito and Haleh Ardibili, ISBN
9781119083597. A revision of the groundbreaking study of methods for
storing energy on a massive scale to be used in wind, solar, and other
renewable energy systems. NOW AVAILABLE!
Nuclear Power: Policies, Practices, and the Future, by Darryl Siemer, ISBN
9781119657781. Written from an engineer’s perspective, this is a treatise
on the state of nuclear power today, its benefits, and its future, focusing on
both policy and technological issues. NOW AVAILABLE!
Progress in Solar Energy Technology and Applications, edited by Umakanta

Sahoo, ISBN 9781119555605. This first volume in the new groundbreaking
series, Advances in Renewable Energy, covers the latest concepts, trends,
techniques, processes, and materials in solar energy, focusing on the state-
of-the-art for the field and written by a group of world-renowned experts.
NOW AVAILABLE!
A Polygeneration Process Concept for Hybrid Solar and Biomass Power

Plants: Simulation, Modeling, and Optimization, by Umakanta Sahoo, ISBN
9781119536093. This is the most comprehensive and in-depth study of the
theory and practical applications of a new and groundbreaking method for
the energy industry to “go green” with renewable and alternative energy
sources. NOW AVAILABLE!
Zero-Waste Engineering 2nd Edition: A New Era of Sustainable Technology

Development, by M. M. Kahn and M. R. Islam, ISBN 9781119184898. This
book outlines how to develop zero-waste engineering following natural
pathways that are truly sustainable using methods that have been devel-
oped for sustainability, such as solar air conditioning, natural desalination,
green building, chemical-free biofuel, fuel cells, scientifically renewable
energy, and new mathematical and economic models. NOW AVAILABLE!
Sustainable Energy Pricing, by Gary Zatzman, ISBN 9780470901632. In

this controversial new volume, the author explores a new science of energy
pricing and how it can be done in a way that is sustainable for the world’s
economy and environment. NOW AVAILABLE!
Sustainable Resource Development, by Gary Zatzman, ISBN 9781118290392.

Taking a new, fresh look at how the energy industry and we, as a planet,
are developing our energy resources, this book looks at what is right and
wrong about energy resource development. This book aids engineers and
scientists in achieving a true sustainability in this field, both from an eco-
nomic and environmental perspective. NOW AVAILABLE!
The Greening of Petroleum Operations, by M. R. Islam et al., ISBN

9780470625903. The state of the art in petroleum operations, from a
“green” perspective. NOW AVAILABLE!
Emergency Response Management for Offshore Oil Spills, by Nicholas P.

Cheremisinoff, PhD, and Anton Davletshin, ISBN 9780470927120. The
first book to examine the Deepwater Horizon disaster and offer processes
for safety and environmental protection. NOW AVAILABLE!
Biogas Production, Edited by Ackmez Mudhoo, ISBN 9781118062852. This
volume covers the most cutting-edge pretreatment processes being used
and studied today for the production of biogas during anaerobic digestion
processes using different feedstocks, in the most efficient and economical
methods possible. NOW AVAILABLE!
Bioremediation and Sustainability: Research and Applications, Edited

by Romeela Mohee and Ackmez Mudhoo, ISBN 9781118062845.
Bioremediation and Sustainability is an up-to-date and comprehensive
treatment of research and applications for some of the most important
low-cost, “green,” emerging technologies in chemical and environmental
engineering. NOW AVAILABLE!
Green Chemistry and Environmental Remediation, Edited by Rashmi Sanghi

and Vandana Singh, ISBN 9780470943083. Presents high quality research
papers as well as in depth review articles on the new emerging green face of
multidimensional environmental chemistry. NOW AVAILABLE!
Bioremediation of Petroleum and Petroleum Products, by James Speight

and Karuna Arjoon, ISBN 9780470938492. With petroleum-related spills,
explosions, and health issues in the headlines almost every day, the issue
of remediation of petroleum and petroleum products is taking on increas-
ing importance, for the survival of our environment, our planet, and our
future. This book is the first of its kind to explore this difficult issue from an
engineering and scientific point of view and offer solutions and reasonable
courses of action. NOW AVAILABLE!
WILEY END USER LICENSE AGREEMENT
Go to www.wiley.com/go/eula to access Wiley’s ebook EULA.

Reservoir Characterization - F&a

Uploaded by

Copyright:

Available Formats

Reservoir Characterization - F&a

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Reservoir Characterization - F&a

Uploaded by

Copyright:

Available Formats

Sustainable Energy Engineering

Series Editor: Fred Aminzadeh

Fundamentals and Applications

Wiley Global Headquarters

Limit of Liability/Disclaimer of Warranty

Library of Congress Cataloging-in-Publication Data

Cover image: Geo/Rock Wall, 31647625 © Pzaxe | Dreamstime.com

Printed in the USA

Part 2: General Reservoir Characterization

4.4 Summary and Conclusions 79

6.7 Estimates of Prior False Discovery Rates

8.4 Use of Finite Element 176

Part 3: Reservoir Permeability Detection 195

9.8 Analysis of Accuracy of Outlier Replacement

Appendix 1- Description of Permeability Models from

Part 4: Reserves Evaluation/Decision Making 253

13.11 Comparative Analysis of Results of Monte Carlo

14.8 U.S. Independents 312

Part 5: Unconventional Reservoirs 337

16.4 Sensitivity Analysis 374

18.3.1 Experiment Description 414

Part 6: Enhance Oil Recovery 427

19.5.4 3D Model Showcase 455

Part 7: New Advances in Reservoir

21.4 Machine (Deep) Learning and Enhanced Oil

petroleum engineering, from laboratory measurements to field applica-

Dr. Ali Tura

An important step in exploration development, monitoring, and man-

This book will provide a comprehensive body of technical material on

application of machine learning in reservoir characterization. It also dis-

FACT Inc., Santa Barbara, CA, USA

Keywords: Reservoir characterization, data integration challenges, 3D/4D

1.1 Introduction to Reservoir Characterization?

Reservoir valuation and producing from a field demands a realistic descrip-

Reservoir grid Upscaling

Facies proportions Geological model: Flow

in geo-model updating monitoring production and the EOR/reservoir

1.2 Data Requirements for Reservoir

Integration of geophysical data with geologic data, and engineering

------sand bodies---river channels---oil/gas traps------

1mm 1m 10m 100m 1km 10km

1.3 SURE Challenge

Uncertainty/Coverage Pyramid SURE Challenge

Information Knowledge &

AI-DA Offers solutions to

with different levels of uncertainties. For example, the direct measure-

At the Well Increasing Areal Coverage Across the Asset

We refer to these four key challenges: Scale, Uncertainty, Resolution

1.4 Reservoir Characterization in the Exploration,

1.4.1 Exploration Stage/Development Stage

by correlating lithology, porosity, net pay thickness and other properties at

1.4.2 Primary Production Stage

1.4.3 Secondary/Tertiary Production Stage

develop an optimum secondary/tertiary recovery mechanism and plan for

1.5 Dynamic Reservoir Characterization (DRC)

1.5.1 4D Seismic for DRC

1.5.2 Microseismic Data for DRC

1.6 More on Reservoir Characterization and