Data Science and

Machine Learning Applications in

Subsurface Engineering
Editor
Daniel Asante Otchere
Pennsylvania State University
University Park, PA, USA

A SCIENCE PUBLISHERS BOOK

First edition published 2024
by CRC Press
2385 NW Executive Center Drive, Suite 320, Boca Raton FL 33431
and by CRC Press
4 Park Square, Milton Park, Abingdon, Oxon, OX14 4RN

© 2024 Daniel Asante Otchere

CRC Press is an imprint of Taylor & Francis Group, LLC

Reasonable efforts have been made to publish reliable data and information, but the author and
publisher cannot assume responsibility for the validity of all materials or the consequences
of their use. The authors and publishers have attempted to trace the copyright holders of
all material reproduced in this publication and apologize to copyright holders if permission
to publish in this form has not been obtained. If any copyright material has not been
acknowledged please write and let us know so we may rectify in any future reprint.
Except as permitted under U.S. Copyright Law, no part of this book may be reprinted,
reproduced, transmitted, or utilized in any form by any electronic, mechanical, or other means,
now known or hereafter invented, including photocopying, microfilming, and recording, or in
any information storage or retrieval system, without written permission from the publishers.
For permission to photocopy or use material electronically from this work, access
www.copyright.com or contact the Copyright Clearance Center, Inc. (CCC), 222 Rosewood
Drive, Danvers, MA 01923, 978-750-8400. For works that are not available on CCC please
contact [email protected]
Trademark notice: Product or corporate names may be trademarks or registered trademarks and
are used only for identification and explanation without intent to infringe.

Library of Congress Cataloging‑in‑Publication Data (applied for)

ISBN: 978-1-032-43364-6 (hbk)

ISBN: 978-1-032-43365-3 (pbk)
ISBN: 978-1-003-36698-0 (ebk)
DOI: 10.1201/9781003366980

Typeset in Times New Roman

by Radiant Productions
 Dedication

To God, who has been my rock and guide throughout my life’s journey,
I dedicate this book with all my heart.
To my dad, Retired Assistant Commissioner of Ghana Police, Daniel
Asante Otchere, thank you for instilling in me a strong work ethic and
dedication to excellence. You have always been a source of inspiration, and
I am grateful for your unwavering support.
To my mum, Jane Duah Otchere, your love and encouragement have been
a constant source of strength. Thank you for being my pillar of support and for
believing in me even when I doubted myself.
To my beloved, Annabelle, thank you for being my partner, my best friend,
and my biggest cheerleader. Your unwavering love and support have been
instrumental in my success, and I am grateful for your presence in my life.
To my siblings, Nicholas, Elliot, Yvonne, and Thelma, thank you for your
love, support, and unwavering belief in me. You have been a constant source
of motivation, and I am grateful for the joy and laughter you bring into my life.
This book is a testament to the love, support, and guidance of these
amazing individuals in my life. Thank you all for being part of my life’s
journey and for being my inspiration.
 Foreword

It is with great pleasure that I write the foreword for this book on Data Science
and Machine Learning Applications in Subsurface Engineering.
The field of subsurface engineering is a critical aspect of the global energy
industry, and it has undergone significant transformations in recent years with
the advent of data science and machine learning. This has led to increased
efficiency, improved decision-making, and reduced costs in the exploration,
production, and development of subsurface resources. Applying data science
and machine learning techniques to subsurface engineering is transforming
how we approach and solve complex problems in this domain. The
combination of vast amounts of data generated by sensors and other sources
and the development of powerful algorithms and computing capabilities has
enabled us to extract valuable insights and make informed decisions.
The edited book on Data Science and Machine Learning Applications
in Subsurface Engineering is an important contribution to this rapidly
evolving field. This book brings together a collection of chapters highlighting
innovative and impactful applications of these technologies in subsurface
engineering. From reservoir characterisation to production optimisation
and drilling engineering, the contributors provide various perspectives,
demonstrating potentials and challenges. The expert and knowledgeable
contributors come from diverse backgrounds, including academia, industry,
and government. The diversity of their perspectives enriches the discussion
and highlights the need for cross-disciplinary approaches to solving the
complex challenges of subsurface engineering.
The chapters in this book cover a wide range of application areas, including
data-driven workflows for subsurface characterisation at the well scale and the
reservoir scale, machine learning techniques for well performance analysis,
and smart completions. Diverse data types are analyzed including well flow
data, wireline log data and seismic images. A variety of machine learning
algorithms are described, from traditional multivariate statistical methods
and tree-based methods to artificial neural networks and deep convolutional
neural networks. This book offers valuable insights into the ongoing research
and development in the field of subsurface engineering.
 Foreword v

Looking forward, we can expect advancements in artificial intelligence

and generative machine learning models to improve our ability to analyse
subsurface data, extract meaningful insights, and make decisions. However,
as with any rapidly evolving field, there will be challenges along the way. It
will be essential to continue to develop and refine algorithms to ensure that
the models are accurate and reliable and to address the ethical implications
of using these technologies in subsurface engineering. I hope this book will
serve as a valuable resource for researchers, engineers, and decision-makers
in the energy industry and inspire continued innovation and progress in the
dynamic field of subsurface engineering.
Tapan Mukerji
Professor (Research)
Energy Science & Engineering
Stanford University
 Preface

Data science and machine learning have revolutionised various industries, and
the subsurface engineering industry is no exception. Subsurface engineering
involves the exploration, development, and production of natural resources
like oil and gas, and is critical to meet the world’s energy demands. The
application of data science and machine learning techniques in subsurface
engineering has the potential to significantly improve the efficiency, accuracy,
and safety of operations in this industry.
This research book on Data Science and Machine Learning Applications
in Subsurface Engineering provides a comprehensive overview of this field’s
latest research and developments. The book is intended for professionals,
researchers, and students interested in understanding the potential of data
science and machine learning in subsurface engineering. I hope that readers
value each chapter in this book because subsurface engineering plays a vital
role in meeting the world’s energy demands, and the application of data
science and machine learning techniques has the potential to significantly
improve efficiency and accuracy, and safety of operations in this industry.
This book brings together a collection of chapters that showcase some
of the most innovative and impactful applications of these technologies in
subsurface engineering. The contributors to this book represent a diverse
range of backgrounds and expertise, from academic researchers to industry
professionals. Each chapter offers a unique perspective on using data science
and machine learning in subsurface engineering. The success of this book
is a testament to our contributors’ hard work and dedication. I extend my
deepest gratitude to Ramez, Daniel, Eric Thompson, Ayoub, Halim, and
Nikita for their contributions. The contributions of our authors have made this
book possible, and their research provides valuable insights into the potential
of data science and machine learning in subsurface engineering. I want to
sincerely thank each of our contributors for their hard work, dedication, and
commitment to advancing knowledge in this field.
Throughout the book, several key themes emerge, including the importance
of interdisciplinary collaboration, the need for robust and transparent data
management practices, and the challenges of implementing data science and
 Preface vii

machine learning techniques in real-world environments. These themes reflect

the complex and evolving nature of subsurface engineering and highlight the
need for continued innovation and collaboration in this field.
One of the most significant challenges in this field is the integration of data
science and machine learning techniques with existing engineering practices.
The contributors to this book demonstrate that successful integration requires
a deep understanding of both the technical and practical aspects of subsurface
engineering. This research book contributes to the advancement of the field
by providing a platform for collaboration and sharing knowledge, expertise,
and experiences in applying data science and machine learning in subsurface
engineering.
Another key theme that emerges from this book is the importance of
transparency and ethical considerations in the use of data science and machine
learning in subsurface engineering. The contributors to this book emphasise
the need for transparency in data collection, analysis, and interpretation,
as well as the importance of considering the potential impacts of these
techniques on the environment and society. It is known that many articles and
texts prevent the reader from reproducing the results either because the data
were not freely available or because the software was inaccessible, or only
available for purchase. Therefore, it was my goal to be as hands-on as possible
in stating the methods used, enabling the readers to reproduce the results and
extend the methodology to their own data. Furthermore, the authors opted
to use the Python language, an open-access software, for all stages of the
machine learning process.
Most of the chapters contain links to the data sets, Python notebooks,
and software used here to reproduce the analyses in each chapter. I selected
Python as the computational engine of this book for several reasons. First,
Python is freely available for multiple operating systems. I encourage you
to take the time to compare each of your solutions with the results in this
book. Performing this comparison may help you become more familiar with a
technique you could have used to solve a specific problem more efficiently. In
some cases, it could also reveal that you have discovered a better or simpler
way to solve the problem than the chapters had.
This research book is an important contribution to the field. The
application of data science and machine learning techniques in subsurface
engineering has the potential to significantly improve the efficiency, accuracy,
and safety of operations in this industry. The research presented in this book
showcases novel techniques and applications that can address some of the
most pressing challenges in subsurface engineering, including seismic
interpretation, reservoir characterisation, production optimisation, and
data-driven decision-making. The research offers a unique perspective on the
viii Data Science and Machine Learning Applications in Subsurface Engineering

current state-of-the-art and future directions in this field, making it a valuable

resource for researchers, practitioners, and students in this domain.
This edited research book on Data Science and Machine Learning
Applications in Subsurface Engineering features a collection of research
papers highlighting the latest field advances. The book covers a wide range
of topics related to the application of data science and machine learning
techniques in subsurface engineering. The book’s main contributions can be
summed as novel techniques for seismic interpretation, improved reservoir
characterisation, data-driven decision-making, and applications beyond
oil and gas. Overall, the contributions of this edited research book advance
our understanding of the potential of data science and machine learning in
subsurface engineering. The book provides a valuable resource for researchers,
practitioners, and students interested in this field and highlights the latest
developments that can be applied to solve complex problems in subsurface
engineering.
The future of subsurface engineering is exciting and filled with
opportunities, but it also presents significant challenges. The rapid pace of
technological innovation and the increasing demand for sustainable energy
solutions require the industry’s continuous and dynamic response. The
contributors to this book provide valuable insights into the potential of data
science and machine learning techniques to address these challenges, as well
as the need for continued collaboration and innovation in this field.
In conclusion, this book offers a comprehensive collection of research
papers covering various aspects of applying data science and machine
learning in subsurface engineering. The book highlights the potential of
these technologies to transform how we approach subsurface engineering
challenges and highlights the importance of collaboration, transparency, and
ethical considerations in their use. As the editor, I encourage readers to delve
into the chapters and explore the exciting research, as they can expect to gain
insight into the latest techniques and developments in this exciting field.
 Contents

Dedication iii
Foreword iv
Preface vi
1. Introduction 1
2. Enhancing Drilling Fluid Lost-circulation Prediction: Using 6
Model Agnostic and Supervised Machine Learning
1. Introduction 6
2. Background of Machine Learning Regression Models 9
3. Data Collection and Description 11
4. Methodology 11
4.1 Data Analysis and Visualisation 11
4.2 Machine Learning Model Application 13
4.3 Explainable AI 15
4.3.1 Permutation Feature Importance 15
4.3.2 Shapley Values 16
5. Results and Discussion 17
5.1 Evaluation of Model Performance 17
5.2 Model Agnostic Results 20
5.3 Analysis of Features Using Model Agnostic Metrics 22
5.4 Analysis of Features Using Shapley Values Model 23
Agnostic Metrics
5.5 Evaluation of Top Features 26
5.6 Model Optimisation 26
5.7 Sensitivity Analysis 27
6. Conclusions 28
Acknowledgement 29
Data Availability 29
References 30
x Data Science and Machine Learning Applications in Subsurface Engineering

3. Application of a Novel Stacked Ensemble Model in Predicting 33

Total Porosity and Free Fluid Index via Wireline and NMR Logs
1. Introduction 33
2. Nuclear Magnetic Resonance 36
2.1 Concept and Application 36
2.2 Works Related to the Use of Machine Learning in NMR 38
for Reservoir Characterisation
3. Methodology 40
3.1 Data Collection and Description 40
3.2 Data Analysis and Feature Engineering 42
3.3 Machine Learning Model Application 43
3.3.1 Building Deep Learning Models 43
3.3.2 Building a Hybrid Stacked Ensemble Model 44
3.4 Criteria for Model Evaluation 46
4. Results and Discussion 47
4.1 Evaluation of Models’ Performances 47
5. Conclusions 54
Acknowledgement 55
References 55
4. Compressional and Shear Sonic Log Determination: Using 57
Data-Driven Machine Learning Techniques
1. Introduction 57
2. Literature Review 58
3. Background of Machine Learning Regression Models 64
3.1 Decision Tree Conceptual Overview 64
3.1.1 Attribute Selection Measures 65
3.2 Random Forest Conceptual Overview 68
3.3 Extremely Randomised Trees Conceptual Overview 69
4. Data Collection and Description 70
5. Methodology 73
5.1 Data Analysis and Visualisation 73
5.2 Machine Learning Model Application 73
6. Results and Discussion 75
6.1 Evaluation of Model Performance 75
6.2 Model Optimisation 79
6.3 Model Deployment 82
7. Conclusions 82
Acknowledgement 83
Data Availability 83
References 83
 Contents xi

5. Data-Driven Virtual Flow Metering Systems 87

1. Introduction 87
2. VFM Key Characteristics 88
3. Data Driven VFM Main Application Areas 90
3.1 Virtual Sensing in ESP Wells 90
3.2 Virtual Sensing for SRP Wells 91
3.2.1 Virtual Flow Meter on Rod Pumping Systems 93
3.2.2 Virtual Sensing of the Dynamometer Card 93
3.3 Virtual Sensing for Gas Lifted Wells 94
3.4 Virtual Sensing for Gas Wells and Plunger Lifted Wells 95
3.5 Miscellaneous Applications for Identifying Flow Regimes 96
4. Methodology of Building Data-driven VFMs 97
4.1 Data Collection and Preprocessing 97
4.2 Model Development 98
5. Field Experience with a Data-driven VFM System 99
References 100
6. Data-driven and Machine Learning Approach in Estimating 104
Multi-zonal ICV Water Injection Rates in a Smart
Well Completion
1. Introduction 104
2. Brief Overview of Intelligent Well Completion 107
2.1 ICV Setting and Determination 107
2.2 Literature Review of ICV Innovations and Machine 108
Learning Applications
3. Methodology 110
3.1 A Brief Overview of Models Used in This Study 111
3.2 Criteria for Model Evaluation 112
4. Results and Discussion 114
4.1 Explainable AI 114
4.2 Model Evaluation 116
4.3 Sensitivity Analysis 117
4.4 Model Deployment 120
5. Conclusions 122
Code Availability 122
Acknowledgement 123
References 123
7. Carbon Dioxide Low Salinity Water Alternating Gas 125
(CO2 LSWAG) Oil Recovery Factor Prediction in Carbonate
Reservoir: Using Supervised Machine Learning Models
1. Introduction 125
2. Methodology 132
xii Data Science and Machine Learning Applications in Subsurface Engineering

2.1 Modeling of CO2-LSWAG 132

2.2 Geochemical Reactions of CO2-LSWAG 133
2.3 Machine Learning Methods 133
2.3.1 Multivariate Adaptive Regression Splines (MARS) 134
2.3.2 Group Method of Data Handling (GMDH) 136
2.3.3 Performance Metrics 137
2.3.4 Dataset Standardisation 138
3. Results and Discussion 138
3.1 Numerical Model Description 138
3.2 Input and Target Dataset 141
3.3 MARS Modeling 143
3.4 GMDH Modeling 148
3.5 Numerical Simulator and Machine Learning 153
Computational Time
4. Conclusion 153
Acknowledgment 154
References 154
8. Improving Seismic Salt Mapping through Transfer Learning 159
Using A Pre-trained Deep Convolutional Neural Network: A Case
Study on Groningen Field
1. Introduction 159
2. Method 164
2.1 Collection and Description of Data 164
2.2 Deep Convolutional Neural Network in Salt Mapping 164
and Post-processing
2.2.1 Simplified Architecture of Residual U-net 168
2.3 Transfer Learning Application 169
2.4 Criteria for Model Evaluation 169
3. Results and Discussion 172
3.1 Calculated Salt Body Volume 172
3.2 Semantic Segmentation – Transfer Learning Application 174
3.3 Sensitivity Analysis of Model and Expert Interpretations 177
4. Conclusions 178
Data and Software Availability 179
Acknowledgement 179
References 179
9. Super-Vertical-Resolution Reconstruction of Seismic Volume 181
Using A Pre-trained Deep Convolutional Neural Network:
A Case Study on Opunake Field
1. Introduction 181
2. Brief Overview 183
 Contents xiii

3. Methodology 187
3.1 Regional Geological Overview of the Opunake Field 187
3.2 Local Geological Overview of the Opunake Field 188
3.3 Deep Convolutional Neural Network in Seismic 189
Image Resolution
3.3.1 Simplified Architecture of Residual U-net 190
3.4 Training and Testing Process 192
3.5 Criteria for Model Evaluation 193
4. Results and Discussion 194
4.1 Conditioned Seismic Volume 194
4.2 Model Evaluation 197
5. Conclusions 202
Data and Software Availability 203
Acknowledgement 203
References 203
10. Petroleum Reservoir Characterisation: A Review from 207
Empirical to Computer-Based Applications
1. Introduction 207
2. Empirical Models for Petrophysical Property Prediction 209
2.1 Porosity and Permeability Prediction Models 209
2.2 Saturation Prediction Models 210
3. Fractal Analysis in Reservoir Characterisation 213
4. Application of Artificial Intelligence in Petrophysical 214
Property Prediction
4.1 Artificial Neural Networks (ANNs) 215
4.1.1 ANN Application in Petrophysical 218
Reservoir Prediction
4.2 Support Vector Machine (SVM) 228
4.2.1 Machine Learning (ML) Application in 229
Petrophysical Reservoir Prediction
5. Lithology and Facies Analysis 234
5.1 AI Applications in Lithology and Facies Analysis 234
6. Seismic Guided Petrophysical Property Prediction 239
7. Hybrid Models of AI for Petrophysical Property Prediction 244
8. Summary 247
9. Challenges and Perspectives 248
9.1 AI Perspective 248
9.2 Rock Physics Perspective 250
10. Conclusions 251
References 253
xiv Data Science and Machine Learning Applications in Subsurface Engineering

11. Artificial Lift Design for Future Inflow and Outflow 261
Performance for Jubilee Oilfield: Using Historical Production
Data and Artificial Neural Network Models
1. Introduction 261
2. Methodology 263
2.1 Artificial Lift Screening Techniques 263
2.2 Inflow Performance Relationship Production Forecast 263
2.3 Outflow Performance Relationship Production Forecast 264
2.4 PROSPER Procedure for Well Model Set-Up 264
2.4.1 Deviation Survey Data Input 266
2.4.2 Surface Equipment Data Input 266
2.4.3 Downhole Equipment Data Input 266
2.4.4 Average Heat Capacities Data Input 267
2.5 Artificial Neural Networks 267
2.5.1 Back Propagation Neural Network 268
2.5.2 Radial Basis Function Neural Network 268
2.5.3 ANN Procedure 268
3. Results and Discussion 269
3.1 Production and Well Data of the Study Area 269
3.1.1 Base Case Flow Rates 271
3.2 Artificial Lift Screening 272
3.3 PROSPER Simulation Results 273
3.3.1 IPR Curves 273
3.3.2 Vertical Lift Performance Correlations 275
3.3.3 Desired Flow Rates 276
3.4 Gas Lift Results 277
3.4.1 Optimum Production Rates 278
3.5 ANN Results 279
3.5.1 ANN Architecture 279
3.5.2 Model Visualization 280
3.6 Discussion 282
4. Conclusions 283
Acknowledgment 284
References 284
12. Modelling Two-phase Flow Parameters Utilizing 286
Machine-learning Methodology
1. Introduction 286
2. Data Sources and Existing Correlations 288
3. Methodology 289
 Contents xv

4. Results and Discussions 291

4.1 Data Pre-processing 291
4.2 Model Development and Evaluation 293
5. Comparison between ML Algorithms and Existing Correlations 295
6. Conclusions and Recommendations 298
Nomenclature 298
References 299
Index 303
 Chapter 1
Introduction

“Data is the new oil”—this buzz phrase has become a ubiquitous adage
in today’s digital age. Data science and machine learning have become
indispensable tools for extracting insights and value from vast amounts of
data. The field of subsurface engineering is no exception to this trend, and
this book brings together a collection of chapters from experts in this field,
exploring various data science and machine learning applications in subsurface
engineering.
In this book, we delve into the potential of these technologies for subsurface
engineering, including topics such as data-driven reservoir characterisation,
machine learning in drilling operations, and computer vision application in
seismic image processing and interpretation. The book is divided into several
chapters, each offering a unique perspective and case studies on the different
applications of data science and machine learning in subsurface engineering.
Chapter 2 focuses on predicting drilling fluid loss in the Marun oil
field using machine learning models. The research begins by assessing
the importance of input features through model agnostic metrics. Once a
suitable dataset is established, several machine learning models will be used,
with the best-performing one optimised using the Bayesian Optimisation
algorithm. This research aims to generate new insights into the generalisation
of individual features to explain the target. The study’s main contributions
are twofold. Firstly, it provides a global explanation of variables in mud-loss
prediction using explainable artificial intelligence (AI). Secondly, it develops
a machine learning workflow that utilises explainable AI to enhance drilling
fluid lost circulation prediction. Overall, this research offers valuable insights
into applying machine learning to subsurface engineering, and the potential
benefits of explainable AI in improving drilling fluid loss prediction.
Chapter 3 describes a study aimed at developing an AI-based model that
can predict porosity and producible pore volume fraction using wireline logs
and NMR-measured total porosity and free fluid index. Wireline logs are
2 Data Science and Machine Learning Applications in Subsurface Engineering

more economical and time-saving than running nuclear magnetic resonance

(NMR) logs in all wells, making it a desirable field-scale technique for
quantitative porosity and pore volume fraction measurement. The study aims
to provide a more efficient and accurate method for reservoir characterisation
and management, especially given the current demand for inexpensive and
reliable techniques in the face of high oil prices. By developing an accurate
prediction model, the study hopes to assist in improving reservoir management
and decision-making processes.
Chapter 4 delves into the use of data-driven machine learning techniques
to determine compressional and shear sonic logs, which are crucial in
subsurface engineering. The authors discuss the importance of using relevant
input variables to train supervised machine learning models, as the model’s
performance can degrade if irrelevant data or essential features are omitted. The
authors identify and select input features based on their demonstrated impact
on Vs and Vp measurements to address this issue. The study is distinguished
from earlier research by its multi-output prediction of Vp and Vs, with Vp not
being used as an input for Vs prediction. The authors use data from three wells
in the Volve oil field to evaluate the relevance of input features, and different
machine learning models are employed to predict Vp and Vs. The authors
then use statistical metrics to determine the best-performing model, which is
optimised using the Bayesian Optimisation (BO) algorithm.
Chapter 5 explores the concept of virtual flow metering (VFM) in the
oil and gas industry and focuses on data-driven solutions. VFM technology
relies on analytical or data-driven models for real-time calculations of
phase production. The chapter first introduces the classification of VFM
based on modelling paradigms, including the first principle and data-driven
VFMs. The applications of data-driven modelling in VFM systems are then
discussed, emphasising the models’ features, predicted variables, input
data, and respective papers. The chapter also describes the components
and methodology used to develop data-driven virtual flow meters using the
cross-industry standard process for a data mining framework. An
implementation of this methodology is presented to estimate flow rate and
water cut prediction from the electrical submersible pump’s sensor data.
Finally, the chapter discusses the real operational experience reported in
the literature using data-driven models. Overall, this chapter provides a
comprehensive overview of the concept of VFM and highlights the potential
applications of data-driven solutions in VFM systems in the oil and gas
industry.
Chapter 6 presents a novel approach to optimising inflow control valve
(ICV) settings to increase oil production using machine learning techniques.
The chapter proposes a data-driven approach to estimate the injected water
 Introduction 3

volumes into each reservoir when ICV settings formula is discordant rather
than recalibrating empirical correlation, which leads to interruptions in field
production. The study employs eight machine learning models to estimate each
reservoir unit’s volume of production fluids based on operational parameters,
providing a simple and accurate approach to reservoir management plans. The
method can also provide real-time estimates of produced volumes, allowing
daily production operational changes to meet critical targets and develop
domestic oil resources responsibly. This initiative has tremendous research
possibilities, and the proposed approach could be applied to oil production
wells with various input parameters upon acceptable results.
Chapter 7 of this book explores using Carbon Dioxide Low Salinity
Water Alternating Gas (CO2 LSWAG) flooding as an Enhanced Oil Recovery
(EOR) technique in carbonate reservoirs. The chapter highlights the benefits
of this technique, including a high recovery factor and improved displacement
efficiencies. The authors use a compositional simulator with geochemical
models to develop proxy models for predicting the oil recovery factor.
Multivariate Adaptive Regression Splines (MARS) and Group Method of Data
Handling (GMDH) machine learning methods are used in the study to develop
these proxy models. The authors advocate for using machine learning proxy
models as prediction tools to enhance the efficient full-field implementation
of this technique and reduce the computational time associated with numerical
simulations in carbonate reservoirs.
Chapter 8 showcases the application of transfer learning to a convolutional
neural network pre-trained with synthetic labels, generating salt probability
models for use in seismic imaging and velocity modelling phases. The use of
deep learning techniques in object and edge detection has succeeded in various
fields, making them a promising approach for seismic salt mapping. The
study’s main contributions are improved accuracy in salt segmentation, time
and cost savings, enhanced data analysis, and potential for future research.
By using transfer learning to automate the process of salt segmentation in
seismic images, we can save time and reduce costs while improving accuracy,
ultimately improving our understanding of the subsurface geology and
advancing the energy transition journey. The findings of this study will provide
new insights into the use of transfer learning for salt segmentation in seismic
images, highlighting the potential for developing new energy resources, and
mitigating climate change.
Chapter 9 explores advanced AI techniques to improve seismic image
resolution and quality, which is crucial for accurate subsurface exploration
and analysis. This study focuses on using a pre-trained deep convolutional
neural network (DCNN) to enhance the signal-to-noise ratio (SNR) and
vertical resolution in seismic images of the Opunake field. The study compares
4 Data Science and Machine Learning Applications in Subsurface Engineering

two-image resolution DCNN techniques to identify the most suitable one

for vertical resolution improvements. The success of using synthetic data to
pre-train the DCNN and deploying it on a real field makes AI a promising
candidate for enhancing vertical and lateral seismic image resolution. This
research highlights the potential of AI techniques in improving seismic imaging
and interpretation, paving the way for future developments in this area.
Finally, Chapter 10 provides a structured overview of the application of
AI in reservoir characterisation. The paper focuses on the empirical approach,
AI application, and its enhancement with other computational models,
as well as recent AI advancements for reservoir property prediction. The
paper is divided into 10 sections, with Section 2 providing an overview of
the empirical correlation used for petrophysical reservoir characterisation.
The subsequent sections delve into the application of fractals analysis, AI,
and machine learning techniques such as ANN and SVM and the review of
lithology and facies analysis. The contribution of seismic data in reservoir
property prediction is also examined, with Section 7 exploring the application
of hybrid AI in reservoir characterisation. Section 8 highlights the summary
of individual AI algorithms used in reservoir characterisation, while
Section 9 addresses the challenges and the way forward for reservoir
characterisation. Finally, Section 10 concludes the study and spells out the
significant keys drawn from the present review. Overall, this structured
overview provides valuable insights into the application of AI in reservoir
characterisation, highlighting its potential to enhance the accuracy and
efficiency of characterisation processes.
Together, these chapters provide a comprehensive overview of how
data science and machine learning can be applied to subsurface engineering,
demonstrating the potential of these technologies for revolutionising the way
we explore, produce, and manage subsurface resources.
Subsurface engineering is a complex and challenging field that requires
a deep understanding of the physical properties of the subsurface, including
geology, fluid mechanics, and rock mechanics. Traditionally, subsurface
engineering relied on deterministic models and manual interpretation of data,
which were often time-consuming and prone to errors. The advent of data
science and machine learning has transformed this field, offering new tools
and approaches for handling and analysing large volumes of subsurface data.
These technologies have revolutionised subsurface engineering by enabling
more accurate predictions, faster decision-making, and improved resource
management. This book provides a timely and comprehensive overview of
the different applications of data science and machine learning in subsurface
engineering, highlighting the potential of these technologies for unlocking
new insights and value from subsurface data. By exploring the different
 Introduction 5

topics covered in this book, readers will gain a deeper understanding of the
transformative power of data science and machine learning in subsurface
engineering and the challenges and opportunities of integrating these
technologies into subsurface workflows.
In conclusion, this book offers a comprehensive overview of the different
applications of data science and machine learning in subsurface engineering,
demonstrating the potential of these technologies for unlocking new
insights and value from subsurface data. The book’s central themes include
the importance of interdisciplinary collaboration, the need for robust and
transparent data management practices, and the challenges of implementing
data science and machine learning techniques in real-world environments.
In addition, the book highlights the importance of transparency and ethical
considerations in using data science and machine learning in subsurface
engineering, emphasising the need for responsible and ethical practices in this
rapidly evolving field.
By bringing together experts from various fields, this book offers a
unique and interdisciplinary perspective on applying data science and
machine learning in subsurface engineering, highlighting the potential
of these technologies for transforming the way we explore, produce, and
manage subsurface resources. The book’s contributions to the field include
case studies, best practices, and critical analyses of the opportunities and
challenges of implementing data science and machine learning techniques in
subsurface engineering. Overall, this book serves as a valuable resource for
researchers, practitioners, and students in the field of subsurface engineering,
offering insights and perspectives that are critical for staying up-to-date with
the latest developments in this rapidly evolving field.
 Chapter 2
Enhancing Drilling Fluid
Lost-circulation Prediction
Using Model Agnostic and Supervised
Machine Learning
Daniel Asante Otchere,1,2,*
Mohammed Ayoub Abdalla Mohammed,3 Hamoud Al-Hadrami 4
and Thomas Boahen Boakye5

1. Introduction
Drilling is a complex and high-risk element of oil and gas field development
and production (Zarrouk and McLean, 2019). Drilling fluid serves several
purposes in rotary drilling, where the drilling mud is cycled through the
drill string to remove cuttings and improve drill bit performance (Alkinani
et al., 2020). The size, shape, and density of the drilling fluid’s cuttings, as
well as its annular velocity, all affect the drilling fluid’s ability to remove
accumulated cuttings. One of the most prevalent drilling industry difficulties
is drilling fluid lost returns. This problem occurs whenever the volume of mud
injected during drilling partially or totally filtrates into the formation rather
than flowing back up to the surface (Toreifi et al., 2014). Due to the volume
1
Centre of Research for Subsurface Seismic Imaging, Universiti Teknologi PETRONAS, 32610, Seri
Iskandar, Perak Darul Ridzuan, Malaysia.
2
Institute for Computational and Data Sciences, Pennsylvania State University, University Park, PA,
USA.
3
Chemical and Petroleum Engineering, UAE University, Sheik Khalifa Street at Tawam R/A, Maqam
District, Al Ain, United Arab Emirates.
4
Department of Petroleum and Chemical Engineering, Sultan Qaboos University, Muscat, Oman.
5
Subsurface Department, Tullow Ghana Limited, North Dzorwuku, Off George Bush Highway,
PMB, Accra, Ghana.
* Corresponding author: [email protected]
 Enhancing Drilling Fluid Lost-circulation Prediction 7

of mud lost during drilling, Pilehvari and Nyshadham (2002) proposed three
classifications due to the severity or otherwise. The three classifications are
defined as seepage (ranging from 0.5 to 10 bbls/hr), partial (ranging from
10 to 500 bbls/hr), and complete (above 500 bbls/hr) losses of all the fluid. The
losses can be attributed to the formation type and structural and petrophysical
properties of the formation (Moazzeni et al., 2012). The loss of these drilling
fluids leads to increased costs in drilling operations, differential sticking, a
blowout, damage to reservoir intervals, and, most seriously leading to a loss
of the well (Alkinani et al., 2019a; Feng and Gray, 2017; Sabah et al., 2019;
Toreifi et al., 2014).
Minimising the loss of drilling fluids is of the utmost importance
considering the severe consequences. This fluid loss has led to the development
of different preventive and remedial treatments employed to prevent fluid loss.
Before deciding on the appropriate lost circulation solution to utilise to reduce
mud loss, the degree of the mud loss should be determined. It should, however,
be noted that finding a single solution to reduce lost circulation is challenging.
Consequently, several lost circulation remedial alternatives are available,
including but not limited to fibrous, high viscosity pills, granular, brittle
materials, and cement sludges. Each treatment method or material is chosen
depending on how much of the mud is lost, the time and expense involved, the
drilling phase, the fluid type, and the theft zone (Alkinani et al., 2020). The
main reason why these treatments are used is to bridge existing fractures and
prevent the development of new fractures (Alkinani et al., 2019b).
Several factors lead to this drilling problem, the most severe being
formation, drilling operation, and time-dependent. Since these factors are
highly dimensional and have complex relationships, traditional mathematical
techniques have proven futile in predicting drilling fluid loss (Sabah et al.,
2019; Toreifi et al., 2014). As a result, a significant amount of effort and
money is spent harnessing technical breakthroughs in data acquisition to
potentially limit its likelihood. One method for accomplishing this is accurately
forecasting lost circulation using drilling parameters. Artificial Intelligence
(AI) techniques are now widely employed in the oil and gas sector, resulting
in substantial advancements in their use to increase accuracy (Otchere
et al., 2021a). Several studies have also employed machine learning models to
forecast drilling fluid loss circulation, demonstrating their efficacy in creating
patterns among complicated drilling parametric interactions (Alkinani et al.,
2020; Sabah et al., 2019; Toreifi et al., 2014).
Sabah et al. (2021) introduced a hybrid machine learning approach that
enhances the prediction of lost circulation using the Multilayer Perceptron
Neural Network (MLP-NN), and the Least Squares Support Vector Machine
(LSSVM) models. To enhance the accuracy of the predictions, feature
selection was employed, which helped to eliminate input parameters that
8 Data Science and Machine Learning Applications in Subsurface Engineering

were not beneficial to the machine learning model. The researchers applied
a Savitzky–Golay (SG) filter to reduce noise recorded in the data. From their
results, the wrapper method, the most efficient feature selection technique, was
used in conjunction with several evolutionary algorithms to improve model
performances. Their study concluded that the LSSVM-Cuckoo Optimisation
Algorithm recorded the highest prediction accuracy of 0.94 compared to all
the hybrid models developed. Alkinani et al. (2020) also employed an artificial
neural network to estimate lost circulation in induced and natural fractures.
Their work entailed separating the datasets into three categories: 60% for
training, 20% for testing, and 20% for validating. Databases, consisting of
nine input features for each fracture type, were created for both natural and
induced fractures to give room for data normalisation. The application of
feature selection justified the selection of the Levenberg–Marquardt function
to train the dataset as it resulted in the highest accuracy. Their results achieved
an accuracy of 0.96 and 0.93, respectively, for both natural and induced
fracture networks.
Abbas et al. (2019b) assessed the potential of machine learning in mining
and analysing drilling data to predict lost circulation while drilling. For the
selection of input parameters, the work made use of the feature ranking method
to reduce the dimensionality of the dataset. Eighteen out of the 23 studied
parameters were selected as input variables to predict the lost–circulation.
The feature ranking method enhanced the efficiency of the dataset, thereby
improving computational processing time. The results showed that the
Gaussian kernel support vector machine (SVM) recorded the highest accuracy
of 0.92. Their work led to more research, such as using two different machine
learning models that consider geological and operational parameters in
making lost-circulation predictions (Abbas et al., 2019a). Toreifi et al. (2014)
also proposed a new technique using the modular neural network (MNN) and
particle swarm optimisation (PSO) capable of predicting drilling fluid loss.
Data normalisation was used to transform the data into a range between 0 and
+1 to develop a more efficient model, for which 60% was used for training,
20% for testing, and 20% for validating. Their results concluded that the
PSO achieved a more accurate output by optimising the parameter variation
process. Other influential research and review work in the field of integrating
AI with drilling activities include Aalizad and Rashidinejad (2012), Abbas
et al. (2019c), Ahmadi (2016), Al-Baiyat and Heinze (2012), Barbosa et al.
(2019), Brankovic et al. (2021).
Considering different machine learning models and techniques used
in predicting drilling fluid loss, the main difference was the type of model
being used and the input variables. Supervised machine learning models are
only as efficient as the information they are trained with. Therefore, when
irrelevant data is included as input, the model’s performance suffers (Otchere
 Enhancing Drilling Fluid Lost-circulation Prediction 9

et al., 2021c). As such, there is the need to develop a workflow capable of

determining the most appropriate features, which involves engineering new
features from existing input variables (Otchere et al., 2021b). These new
features should be capable of improving model performance compared to
when they are inputted singularly. Again, the choice of machine learning
models should not be a hindrance if a more robust model appending all robust
models into one super machine learning model is used. This approach tends
to mitigate the inherent problems of each model while complementing each
other’s strengths.
This study uses data from the Marun oil field used by Sabah et al. (2021)
and Toreifi et al. (2014). The input features and their importance will be
assessed using model agnostic metrics when predicting drilling fluid loss.
Once a suitable dataset of relevant features is established, several machine
learning models will be used to predict drilling fluid loss. The model that
gives the best results will then be optimised using the Bayesian Optimisation
(BO) algorithm. The main objective of this research is to analyse the effect
of individual features to generate new insight into their generalisation of
explaining the target. The main contributions of this research are the global
explanation of variables in mud-loss prediction using an explainable AI and
the development of a machine learning workflow to enhance the prediction of
drilling fluid lost circulation using an explainable AI.

2. Background of Machine Learning Regression Models

This research stems from earlier studies where several workflows and
techniques were used to predict drilling fluid loss using available input
variables. As academics increasingly move away from empirical correlations,
the use of machine learning has become entrenched. Individual drilling
parameters offer critical information concerning drilling fluid lost circulation.
However, when some subsets are used, they can predict drilling fluid loss more
accurately. In some instances, researchers employ the Pearson correlation to
identify appropriate variables, while others use wrapper approaches, intrinsic
techniques, and metaheuristics algorithms to identify important features. The
use of several techniques indicates that no definitive collection of features
exists for making this prediction. However, a robust workflow is needed to
utilise model agnostic techniques to solve this prediction problem. This study
aims to understand the relationship among key input variables relevant for
drilling fluid lost-circulation prediction. The relevant features selected by
the model agnostic techniques will then be used as input features. Several
machine learning models will be used to predict fluid loss based on their
learning theory and ability to work with high-dimensional and complex data.
The reviewed models perform differently based on learning theory,
dimensionality, small or large data, and different data distribution. Identifying
10 Data Science and Machine Learning Applications in Subsurface Engineering

a suitable model that can handle all these problems has become necessary since
drilling data are commonly highly dimensional, either small or large, with
different data distributions. Based on these assertions, this study establishes
relevant features that can lead to lower prediction errors. Improving the
accuracy in estimating mud loss has a massive impact on drilling operations
and the integrity of a well. Hence, the application of machine learning models,
although dependent on data type, can be applied to drilling operations with
similar input features. The most common issue with machine learning models
and their varying performance is centred on data. Having the capability of
explaining the input variables and determining causation has been the pinnacle
of this study as new approaches are created to solve this issue. The slightest
gain in accuracy is critical in improving decision-making in the petroleum
business, making this field of research critical. Table 2.1 presents a summary
analysis of the models reviewed in this research.

Table 2.1. Summary of algorithms used in this study and corresponding authors (Otchere
et al., 2022b).

Model Description Developed or

Implemented by
Ridge Regression Robust technique against multicollinearity (Hoerl and
minimises standard errors by applying some bias Kennard, 1970)
to the model estimates, resulting in a more reliable
prediction. Ridge regression model works by
calculating the difference between the means of the
standardised dependent and independent variables
and dividing by their standard deviations.
Least Absolute Robust against multicollinearity, perform better with (Tibshirani, 1996)
Shrinkage and Selection lower-dimensional data, prevents overfitting, and
Operator (Lasso) reduces standard errors.
Support vector machine Utilises the Structural Risk Minimisation induction (Vapnik and
(SVM) principle, hence, produces improved generalised Lerner, 1963)
global, sparse, and unique results with a simple
geometric interpretation.
Decision Tree (DT) Uses induction and pruning techniques to build (Gordon et al.,
hierarchical decision boundaries and remove 1984)
unnecessary structures from the decision tree to
battle overfitting.
Extreme Gradient An implementation of gradient boosted decision (Chen and
Boosting (XGBoost) trees of the first- and second-order to maximise the Guestrin, 2016)
loss function, adding an extra regularisation term to
avoid overfitting by adjusting the final weights.
Extremely Randomised Potentially achieves better performance than the (Geurts et al.,
Trees (Extra Tree or random forest, a simpler algorithm is employed 2006)
ET) to construct the decision trees used as ensemble
members.
Random Forest (RF) An advanced decision tree that is robust against (Breiman, 2001)
overfitting and offers easy interpretability.
 Enhancing Drilling Fluid Lost-circulation Prediction 11

3. Data Collection and Description

This research aims to improve the prediction of drilling fluid loss using data
from the onshore Marun oilfield during drilling operations. Fracturing in
the reservoir interval is a defining feature of this field in Iran’s Khuzestan
province. The data consists of 2,820 data points with 19 features compiled
from resources such as final well reports, daily drilling reports (DDRs),
literature, and daily mud reports (Sabah et al., 2021; Toreifi et al., 2014).
However, after removing the missing data, 2783 points remained. Table 2.2
shows the descriptive statistics of some selected features. The main input
features for this research are;
1. Drilling operation features: Geographic coordinates, depth where the
loss occurred, weight on bit (WOB), hole size, drilling time, penetration
rate (RPM), mud-pump pressure, the flow rate of the mud pump, Marsh
funnel viscosity (MFVIS), and bit rotational speed.
2. Characteristics of drilling mud: Viscosity, shear stresses at shear rates of
600/300 rpm, Gel strength (10-s and 10-min), retort solid per cent, and
mud weight (MW).
3. Formation characteristics: Formation type, pore pressure, and fracture
pressure.
One of the main advantages of applying data analytics and machine
learning to data is to find patterns and hidden information in high-dimensional
data. As such, some input variables will be deemed irrelevant in predicting
drilling fluid lost circulation. For this study, the lithologies were encoded into
numeric variables ranging from 1–15 (Table 2.3). The coordinates for each
well were also removed.

4. Methodology
4.1 Data Analysis and Visualisation
Data analysis and visualisation were used to aid in understanding how the
input features related to the output. The degree of correlation among each
input variable pair and output was quantified using the Spearman rho
covariance matrix. From the heatmap, the pump pressure was the only feature
to show a moderate correlation to the target. The pair plot also depicted the
nonlinear distribution between some of the input and the target. Based on
the nonlinear distribution depicted in Fig. 2.1, it is evident that none of the
features showed a linear correlation to the target. Hence, nonlinear models
were deemed appropriate for this research.
 Table 2.2. Summary statistics of some input and target variables.

12 Data Science and Machine Learning Applications in Subsurface Engineering

Hole WOB Flow MW MFVIS Retort Pore Fracture Fan Gel10min/ Pump RPM Mud Loss
Size (1000lb) Rate (pcf) Solid Pressure Pressure 600/Fan Gel10s Pressure (bbl/hr)
(gpm) (%) (psi) (psi) 300 (psi)
count 2783 2783 2783 2783 2783 2783 2783 2783 2783 2783 2783 2783 2783
mean 12.3 20.9 545.6 95.3 47 23 5421.4 7929.8 1.6 1.5 1934.4 137.9 56.6
std 4.9 9.4 275.1 34.6 11.9 16.6 2508.2 2760.4 0.2 0.2 721.7 34.6 97.8
min 4.1 1 80 30.5 27 0 1254.8 1970.6 1.2 1.1 297 42.8 0
25% 8.4 15 280 66 38 8 3470 5468.9 1.5 1.3 1423.4 111.6 0
50% 12.3 20 520 79.3 44 18 4473.4 8736.7 1.6 1.5 2228.5 146.7 17.7
75% 17.5 25 830 140 56 42 7554.4 9687.2 1.8 1.6 2472.3 166.7 68.4
max 26 70 1000 160.5 100 61 10507.9 13610.4 2 2.9 2954.2 198.1 696.9
 Enhancing Drilling Fluid Lost-circulation Prediction 13

Table 2.3. Assigned numbers for the Marun Field Formations (Sabah et al., 2021).

Formation Type Numeric Code

Aghajery 1
Mishan 2
Gachsaran 7 3
Gachsaran 6 4
Gachsaran 5 5
Gachsaran 4 6
Gachsaran 3 7
Gachsaran 2 8
Gachsaran 1 9
Asmary 10
Pabdeh 11
Gurpi 12
Ilam 13
Sarvak 14
Fars 15

4.2 Machine Learning Model Application

Cross-validation involves re-sampling input data from multiple subsets and
assessing the model’s predictive performance in the overall dataset using a
secondary subsets not utilised in the training process (Otchere et al., 2021a;
Tarafder et al., 2021). This method may be used to minimise prediction
errors on unknown data. The holdout cross-validation approach was used to
randomly partition the data into 85:15 training and testing groups, representing
418 data points for testing. This approach guarantees that separate data points
not observed by the trained model are used to validate the model leading to
the avoidance of overfitting and selection bias. Several models were used in
this research to effectively know the model that predicts mud loss with lower
errors. The same out-of-sample data was used throughout this research to
ensure that similar data points were used in all models.
Hyperparameter tuning of the final model was performed to further
improve the model’s performance. The BO algorithm, which is classed as
a sequential model-based optimisation procedure, was used to improve the
model (Tarafder et al., 2022). BO is a flexible and efficient approach that uses
the Bayes Theorem to provide a rational mechanism for directing a global
optimisation issue’s investigation to the extrema of objective functions. The
algorithm works by iteratively training a Bayesian approximation probabilistic
model of the objective function based on prior outcomes approximation.
 14 Data Science and Machine Learning Applications in Subsurface Engineering

Fig. 2.1. Pair plot distribution of all variables colour-coded against the formation type.
 Enhancing Drilling Fluid Lost-circulation Prediction 15

The probabilistic model, a response surface that maps hyperparameters to

the objective function probability score, is successfully evaluated with an
acquisition function before selecting suitable samples to assess the actual
objective function. Hyperparameter tuning is the act of updating the model’s
parameters to regulate the training procedure’s performance and optimise the
model’s performance by achieving a reasonably lower cost function (Otchere
et al., 2022c).

4.3 Explainable AI
In most cases, the application of feature selection techniques is unsupervised;
hence, there is no right or wrong answer, making each technique different
for each data. For this research, model agnostic methods will be used to
analyse the importance of input features to separate its explanations from the
machine learning model. The desirable characteristics sought after are model
representation and explanation flexibility, which are not limited to a specific
type and make sense in the context of the model being explained. For this
purpose, the Permutation Feature Importance (PFI) and Shapley values will
be used to explain the results generated by the models. The inputs to achieve
this are the model, the feature vectors, the target, and the error metrics. After
applying these techniques, features that do not explain the target will be
removed from the input feature vector.

4.3.1 Permutation Feature Importance

Permutation feature importance (PFI) is a model agnostic approach for
computing the significance of features (Otchere et al., 2022a). This approach
works by permuting the relationship between a feature and the model output
by assigning the feature a nearly random value. The basic working principle
to achieve this is to compute the importance of a feature by measuring the
increase in estimation error after fifty feature permutations. As a result, a
feature is considered essential if omitting its value increases the model error.
Otherwise, it is unimportant if the error remains constant (Breiman, 2001).
The main advantages of applying PFI are its straightforward interpretation,
highly compressed global insight into a model’s behaviour, and not require
retraining. However, one disadvantage is that it is unclear whether the training
or test data should be used to measure PFI. The in-built function, permutation_
importance, in Scikit-learn (Pedregosa et al., 2011) was used to run the PFI
inspection technique. Given a regressor model, a function is created where the
feature importance of all the input features is computed. An importance score
criteria are computed whereby features achieving high values represent better
predictive power.
16 Data Science and Machine Learning Applications in Subsurface Engineering

4.3.2 Shapley Values

The Shapley value, introduced by Lloyd Shapley (Shapley, 1953) from
Cooperative Game Theory, is a critical concept in measuring feature importance
and the contribution of each feature to the models’ performances. Shapley
values work by allocating a unique distribution among the participants of the
entire surplus value created by the coalition of all players in each cooperative
game. The theory defines four properties that must be satisfied: efficiency,
symmetry, dummy, and additivity. The players represent the input features
in machine learning and interpretability, and the Shapley value determines
how much each feature contributes to the output. Knowing how the features
contribute to the model output helps in understanding the essential factors
in generating the model’s results. The Shapley value is not specific to any
particular model type; hence it can be used regardless of the model type and
architecture. Perhaps, unlike any other method of interpreting model results,
Shapley’s values are based on a solid conceptual framework. While making
intuitive sense is essential for interpretability, other methods lack the same
rigorous conceptual framework. Shapley values have one key advantage
of being evenly distributed among an instance’s feature values. It has been
suggested that Shapley may be the only way to offer a complete explanation
in cases when the law mandates it, such as the right to explanations. Shapley,
like any other approach, has inherent drawbacks. The most crucial is that it
is computationally expensive, implying that only an approximate solution
can be calculated in a high percentage of real-world scenarios. It is also easy
to misunderstand. After removing the feature from the model training, the
Shapley value is not the difference between the expected and actual values.
It is how much an input value contributes to the gap between the actual and
average prediction.
The open-source Shap library is a robust tool for working with Shap values
and other metrics. Shap establishes a link between efficient credit allocation
and local explanations, making it model agnostic. It ties together optimum
credit allocation and local explanations. Using game theory’s traditional
Shapley values and associated expansions have been the focus of numerous
recent works. Recent academics have expanded on this notion by developing
approaches for Shap values in deep learning models and gradient explainers.
These approaches estimate Shap values for any model using specifically
weighted local linear regression. It also provides various charts to help view
the data and comprehend the model.
This study does not include the mathematical foundations of the models
utilised because there is a vast collection of published material detailing them.
An overview of the methodology workflow used in this study is demonstrated
in Fig. 2.2.
 Enhancing Drilling Fluid Lost-circulation Prediction 17

Fig. 2.2. Workflow summarising the methodology used in this study.

5. Results and Discussion

5.1 Evaluation of Model Performance
All of the models used in this study were evaluated using Akaike Information
Criterion (AIC) to find the best fit for the data set. The assessment of the
result shows that lower AIC values have a higher probability of fitting data
compared to a larger AIC. The following factors were considered when
comparing and selecting the best model. A difference in AIC findings across
models is considered insignificant if it is less than 10, moderate between
10 and 50, significant between 50–100, and very strong if above 100. From
the results, it can be asserted that the extra tree model ended up being the best
for drilling fluid loss predictions. There was a difference of ~ 260 AIC when
the extra tree model was compared to the random forest model representing
the two models with the lowest AIC. Hence, the extra tree model showed a
strong disparity and improved likelihood of fit compared to the other models.
The tendency of the model to over- or underfit the data still exists, even though
AIC can measure the degree to which the model converges to the data.
The models’ performances were assessed based on the out-of-sample data
results. Since the models were trained with the training data, any attempt to
reproduce the values will likely result in high accuracy. On the other hand, a
high training accuracy is not always desirable since it may overfit the data,
collecting intrinsic noise and yielding a non-generalised model. Test accuracy
based on the out-of-sample data is the proportion of correct predictions made
by the models on data that has not yet been observed.
18 Data Science and Machine Learning Applications in Subsurface Engineering

Fig. 2.3. AIC results comparing all the models on the test data.

Since the model aims to improve unseen data, it is particularly crucial

that models have high test accuracy. In this work, correlation coefficient
(R2) values greater than 0.85 for both training and test accuracy scores were
deemed a good trade-off over bias and variance. Table 2.4 shows that the
Extra Tree model had th e highest R2 on the test data for predicting mud loss.
Figure 2.4 below depicts a comparison of all of the model’s predictions to the
actual data.
Any meaningful gain in accuracy, and hence a reduction in
inaccuracy, achieved by machine learning models significantly influence
decision-making. Each model’s statistical and theoretical foundations resulted
in widely variable performances. Figure 2.5 depicts the models’ performances
as a cross-validation evaluation to analyse their exactness, efficiency, and

Table 2.4. Train and test correlation coefficient score of all models used in this study.

Models Train R2 Score Test R2 Score

Extra Tree Regression 1 0.959819
Random Forest 0.984328 0.92394
XGBoost 0.99873 0.921476
Decision Tree 1 0.916176
Gradient Boosting Regression 0.868513 0.801631
Ridge Regression 0.372394 0.381738
Lasso Regression 0.367822 0.37421
Support Vector Machine 0.994892 –0.027344
 Enhancing Drilling Fluid Lost-circulation Prediction 19

Fig. 2.4. Cross-plot of predicted vs actual mud loss for all models based on test data.

reliability. The Root Mean Squared Error (RMSE) findings for all the models
indicate that the Extra Tree model evaluation is the most reliable compared
to the actual mud loss values. The Mean Absolute Error (MAE) also suggests
that the Extra Tree model is the most exact model for predicting mud loss. In
selecting the best model, the ranking feature used in this study is as follows;
MAE, RMSE, AIC, and R2.
20 Data Science and Machine Learning Applications in Subsurface Engineering

Fig. 2.5. Comparison of test data prediction errors of all models based on RMSE and MAE.

The performance of the extra tree model is attributed to the

bias-variance concept used to build this model. The explicit randomisation
of the cut-point and features combined with ensemble averaging can reduce
variance more strongly than the weaker randomisation techniques employed
by other algorithms. Aside from having similar advantages as Decision Trees
based on consistency for universal generalisation and approximation, Extra
Trees additionally give resilience regarding gross model errors because
outliers affect its predictions slightly and locally. When the input features are
significantly more than the random splits, Extra Trees potentially outperform
Decision Trees relating to computational efficiency while resisting irrelevant
input features.

5.2 Model Agnostic Results

Permutation importance based on the XGBoost model was computed to
evaluate the importance of and analyse the effect of multicollinearity on
the input variables. The results in Fig. 2.6 show that the pump and pore
pressures highly correlate to the prediction of mud loss. The results indicate
that permuting an input feature reduces the model’s accuracy by, at most,
0.3, suggesting that some of the inputs are relevant. However, drilling time,
hole size, meterage, WOB, and MFVIS do not offer any improvements to
the model prediction. This result is compared with the feature importance
output from the XGBoost model shown in Fig. 2.7. The discrepancy in the
result indicates the importance of model agnostic that can give an inference
into causality. This inference is necessary because correlation does not mean
causation; hence, feature selection techniques based on correlation may not
entirely improve model performance.
 Enhancing Drilling Fluid Lost-circulation Prediction 21

Fig. 2.6. Permutation importance plot indicating the importance of all the input variables.

Fig. 2.7. XGboost Feature Importance of the input variables to the target.
22 Data Science and Machine Learning Applications in Subsurface Engineering

5.3 Analysis of Features Using Model Agnostic Metrics

To maintain consistency in the final model prediction, a similar train/test split
was employed, and the PFI results of the input features’ performance on the
entire data are presented in Fig. 2.8. The RF model yielded an initial testing
R2 of 0.92. The importance score is computed so that higher values represent
a better predictive power. Generally, the importance score is calculated such
that a higher value indicates a greater predictive power. A sizeable portion
of the 0.92 accuracy score can be attributed to the importance of the values
for the most relevant features. The results showed that some input variables
were given a relatively low importance score. The results point to a small set
of input variables that account for most of the predictability. Inferences can
be drawn from this finding about the relative importance of the features. This
result is in line with the high-test accuracy that was calculated.
The initial permutation importance calculated on the training data shows
the model dependency on each feature during training. Nevertheless, it is
essential to state that this analysis was done for both the training and testing
dataset to help account for features that may help with the generalisation power
of the model. Overfitting is more likely to occur when a feature is considered
important during training but is dismissed as important during testing.
Sensitivity analysis was performed to select the relevant features for
further prediction. The RF model was used to predict mud loss using the top

Fig. 2.8. Permutation feature importance of the input variables.

 Enhancing Drilling Fluid Lost-circulation Prediction 23

Table 2.5. Selection of input variables based on their permuting accuracy.

Features Features Dropped Accuracy

Selected
16 None 0.9221
15 Drilling time 0.9238
14 Drilling time and Hole size 0.9226
13 Drilling time, Hole size, and Meterage 0.9348
12 Drilling time, Hole size, Meterage, and WOB 0.9325
11 Drilling time, Hole size, Meterage, WOB, and MFVIS 0.9304
10 Drilling time, Hole size, Meterage, WOB, MFVIS, and Retort solids 0.9339

15–10 features, and the result is illustrated in Table 2.5. The results indicate
that the optimal number of features to predict mud loss is thirteen.

5.4 Analysis of Features Using Shapley Values Model Agnostic Metrics

The prediction of mud loss was performed using all the features to train the Extra
Trees model. The same training and testing datasets used in the PFI analysis
were used. Since machine learning models are supposed to be interpretable
to help understand how a model made a particular prediction, model agnostic
metrics are helpful. Shapley’s values originated from the game theory, and
it is necessary to clarify its application to supervised machine learning
interpretability. The game represents the prediction assignment for a specific
event in the dataset. The gain is the difference between the actual forecast for
a specific instance and the average forecast for all instances. The instance’s
feature values collaborate to achieve the gain in the mud loss forecast.
Figure 2.9 visualises the shapely values as absolute values in a feature
importance plot. The Shap feature importance works with a simple principle
where the traits with high absolute values are the features of most importance.
The absolute mean of the Shapley values for each attribute across the data is
computed to evaluate the global importance. The pump pressure and flow rate
feature variables were highly important, with pump pressure being the most
important and changing the target on average by 56.35 bbl/hr on the x-axis. The
plot also realised that the correlation of meterage, drilling time, and WOB to the
target was low, with the inclusion of WOB disagreeing with PFI results. Based
on this result, further investigation was necessary to ascertain which traits were
irrelevant to the target. The importance of Shapley values is an alternative
to the importance of permutation features. Both the importance measures
have a significant difference. The decline in model performance determines
the permutation attribute’s relevance, whereas Shapley values are primarily
based on the magnitude of feature provenances. The feature importance plot is
24 Data Science and Machine Learning Applications in Subsurface Engineering

Fig. 2.9. SHAP feature importance measured as the mean absolute Shapley values.

informative but provides no other information outside the significance. The bee
swarm technique, a more informative plot, is used for analysis.
Figures 2.10 and 2.11 visualise the Shapely values as absolute values in a
bee swarm summary plot for both train and test data, respectively. The y-axis
is determined by the feature, and the x-axis by the Shapley value. The feature
importance with characteristic effects is combined into global explanations
whereby a feature matrix of Shapley values is achieved for every instance. In
the summary plot, the first indications of the positive and negative relationship
between the value of a feature and the impact on the target is identified. The
data points, made up of all the training data points, overlap in a scattered
manner on the y-axis. This type of visualisation depicts the distribution of
Shapley values per feature. The features are arranged in descending order of
significance. From the plots, low feature values are represented by blue, while
high feature values are denoted by red. In analysing the influence of pump
pressure, it is observed that low values predict high mud-loss values, whereas
high values predict low mud-loss values. The widespread of the pump pressure
data points also indicates a global explanation that explains the entire model
behaviour. This analysis confirms that this correlation on its own cannot
be termed causation. This conclusion enforces the importance and need for
model agnostic techniques to understand the influence the other features have
in mud-loss prediction. From the bee swarm plot, it can be observed that about
12 features can globally explain how the predictions were made.
The results of the bee swarm plot confirmed that features that have high
importance in both the train and test results and exhibit their importance in the
 Enhancing Drilling Fluid Lost-circulation Prediction 25

Fig. 2.10. A bee swarm summary plot of Shap value impact on model target for the train data.

Fig. 2.11. A bee swarm summary plot of Shap value impact on model target for the test data.

global explanation of the target. From the demonstrated results, the following
analysis was derived;
1. Feature importance: The features are ranked in descending order, and
from the train and test data plot, the pump pressure is ranked first. The
meterage feature has close to zero importance because it does not have
any causal effect in predicting mud loss.
2. Impact: The horizontal location of the data points shows that the pump
pressure has a negative correlation and a high prediction effect in general.
26 Data Science and Machine Learning Applications in Subsurface Engineering

Shapley values are based on a solid theoretical foundation in game theory

and compute feature values reasonably distributed to the target. This method
also contributes to the unification of interpretable machine learning. The fast
computation allows for the computation of multiple Shapley values required
for the global model interpretations. Since the Shapley values represent the
smallest element of the global interpretations, it makes global interpretations
compatible with local explanations.

5.5 Evaluation of Top Features

Further analysis was performed using the top 5 performing models for the
top 13 features based on PFI and Shapley values to ascertain if they will
outperform the initial model predictions. PFI resulted in the removal of
Drilling time, Hole size, and meterage. Meanwhile, Shapley values resulted
in the removal of Drilling time, WOB, and Meterage. This result is shown
in Table 2.6. All the top-performing models improved when the features
based on PFI and Shapley values were used. However, it was noticeable that
Shapley values selected the most relevant features, predominantly observed in
the gargantuan improvement in model accuracy.

Table 2.6. Computed accuracy on test data using top 5 models.

Models Accuracy Based Accuracy Based on Accuracy Based on Shapley

on All Features PFI Top 13 Features Values Top 13 Features
Gradient Boosting 0.8016 0.8073 0.9999
Decision Tree 0.9162 0.9298 0.9990
Random Forest 0.9239 0.9290 0.997
Extra Tree 0.9598 0.9710 0.999
XGBoost 0.9215 0.9266 0.996

5.6 Model Optimisation

The BO algorithm was used to tune the model parameters to improve model
performance. The results in Fig. 2.12 show the model performance of the
extra tree model using the top 13 features identified by Shapley values
and the BO–ET model. The results were compared to the results of Sabah
et al. (2021). This comparison is necessary because the author used the same
data as this study but employed different techniques and models in predicting
fluid lost circulation. It is noteworthy to mention that although machine
learning models depend on the data used, most of the data used in training the
model in this study may not be entirely similar to the other study. However,
it is expected that the majority of the data will be the same. As such, this
comparison is mainly based on the use of the input features and how they
 Enhancing Drilling Fluid Lost-circulation Prediction 27

Fig. 2.12. Comparing the estimation errors of top-performing models and Sabah et al. (2021) models
based on RMSE and MAE.

explain the target and not downgrade the results of the other researchers. From
the error analysis performed, it was observed that the features selected based
on the Shapley values significantly reduced the model’s error. The initial
extra tree model recorded an MAE of 12.6 bbls/hr, while the MAE from the
Shapley selected features was 0.3 bbls/hr. This result represents about a 97%
reduction in MAE.
Similarly, RMSE for the Extra Trees model using all 16 features was
24.0 bbls/hr. The RMSE from the Shapley selected features was recorded as
1.4 bbls/hr. Again, this represents about a 94% reduction in this error metric.
Using hyperparameter tuning, BO–ET was able to reduce the extra tree’s
mean absolute error (MAE), and root mean squared error (RMSE) to 0.2 and
1.2 bbls/hr, respectively. The superiority of the Extra Trees model is mainly
attributed to the bias-variance concept used to build this model, which makes it
resilient to outliers. Based on all the evaluation criteria, it was determined that
the Shapley selected features are highly relevant, offer a global generalisation
of the target, and improve model efficiency.

5.7 Sensitivity Analysis

The kernel density estimate (KDE) of the estimated and actual mud-loss data
is shown in Fig. 2.13. The estimated values are in blue, while the actual values
are in green. According to the test data, the BO–ET is much closer to the
actual data. This investigation indicates that the BO–ET can capture a broader
range of values than the other models. This observation makes it suitable for
use in evaluating drilling fluid lost circulation in other wells using the relevant
drilling parameters. As a result, the sensitivity analysis verifies the assessment
metrics outlined above.
28 Data Science and Machine Learning Applications in Subsurface Engineering

Fig. 2.13. Kernel density estimation for BO–ET model mud-loss prediction demonstrating the
closeness of projected values to actual values.

The methodologies evaluated in this research demonstrated that explainable

AI through the use of model agnostic metrics could give an insight into the
input features. Since correlation does not mean causation, simple statistical
techniques may not be able to indicate the causal effects of each input to the
target. The findings show that gradually increasing the input variables above
13 causes a degradation in the model’s performance. According to the Shapley
values, drilling duration, meterage, and WOB are among the variables that have
no direct influence on drilling fluid lost circulation. Hence, it should not be used
to train the model. Model agnostic has provided trustworthiness and valuable
solutions for predicting fluid loss using relevant causal features. The ability to
adequately predict lost circulation prior to drilling gives vital information about
fracture aperture and the optimal LCM particle-size distribution. As a result,
LCM may be formulated to fill existing fractures or not to percolate deep into
high permeable formations to minimise excessive drilling fluid loss. As such,
predicting drilling fluid lost circulation can give valuable insight into the type
of interventions required to be implemented.

6. Conclusions
This study demonstrated a methodology for identifying significant variables
for drilling fluid lost-circulation estimations. The effectiveness of this approach
 Enhancing Drilling Fluid Lost-circulation Prediction 29

has been assessed to surpass previously reported techniques based on the same
dataset. The evaluation criteria discussed presented the probability of fitting
held-out data of the models. The exactness, reliability, and dependability
of the models were also assessed based on the R2, MAE, and RMSE on the
out-of-sample data. The models and how the target is predicted, based on
the PFI and Shapley values, have made the input features explainable as the
following traits are displayed;
1. Fairness: The results have ensured that the predictions are neutral and
are not biased against any input variable, either implicitly or overtly.
An easy-to-interpret model can explain why a particular prediction was
made, making it more straightforward to determine which input variable
had a more significant influence.
2. Reliability: The results help capture the differences in input variables that
result in significant changes in the target and vice versa.
3. Causality: Causal relationships were identified, and the pump pressure
feature for training and test data is ranked high amongst all input variables.
4. Trust: The results have yielded some level of trust in input features used
for the final prediction. This analysis is based on how the employed
techniques, PFI and Shapley values, explain their judgements on the input
features to the target.
The application of Explainable AI enhanced model prediction accuracy
by 97% and 94% in terms of MAE and RMSE, respectively. The comparison
with other published results based on the same data resulted in a 98% and
95% reduction in MAE and RMSE, respectively. This analysis is proof that
correlation does not mean causation. Without the PFI and Shapley values, the
XGBoost feature importance would not have indicated features that will cause
over- or underfitting and not capable of global interpretation of features.

Acknowledgement
The authors express their sincere appreciation to Universiti Teknologi
Petronas, the Centre of Research in Enhanced Oil recovery, and the Centre for
Subsurface Seismic Imaging for supporting this work.

Data Availability
The drilling data was obtained from Sabah et al. (2021).
30 Data Science and Machine Learning Applications in Subsurface Engineering

References
Aalizad, S.A. and Rashidinejad, F. 2012. Prediction of penetration rate of rotary-percussive
drilling using artificial neural networks: A case study/Prognozowanie postępu wiercenia
przy użyciu wiertła udarowo-obrotowego przy wykorzystaniu sztucznych sieci
neuronowych – studium przypadku. Archives of Mining Sciences 57: 715–728. https://doi.
org/10.2478/v10267-012-0046-x.
Abbas, A.K., Al-haideri, N.A. and Bashikh, A.A. 2019a. Implementing artificial neural networks
and support vector machines to predict lost circulation. Egyptian Journal of Petroleum
28: 339–347. https://doi.org/10.1016/J.EJPE.2019.06.006.
Abbas, A.K., Bashikh, A.A., Abbas, H. and Mohammed, H.Q. 2019b. Intelligent decisions
to stop or mitigate lost circulation based on machine learning. Energy 183: 1104–1113.
https://doi.org/10.1016/J.ENERGY.2019.07.020.
Abbas, A.K., Flori, R., Almubarak, H., Dawood, J., Abbas, H. and Alsaedi, A. 2019c. Intelligent
prediction of stuck pipe remediation using machine learning algorithms. In: SPE Annual
Technical Conference and Exhibition. SPE, Calgary. https://doi.org/10.2118/196229-MS.
Ahmadi, M.A. 2016. Toward reliable model for prediction Drilling Fluid Density at wellbore
conditions: A LSSVM model. Neurocomputing 211: 143–149. https://doi.org/10.1016/j.
neucom.2016.01.106.
Al-Baiyat, I. and Heinze, L. 2012. Implementing artificial neural networks and support vector
machines in stuck pipe prediction. In: Kuwait International Petroleum Conference and
Exhibition. SPE. https://doi.org/10.2118/163370-MS.
Alkinani, H.H., Al-Hameedi, A.T.T., Dunn-Norman, S., Flori, R.E., Alsaba, M.T., Amer, A.S.
and Hilgedick, S.A. 2019a. Using data mining to stop or mitigate lost circulation. J. Pet.
Sci. Eng. 173: 1097–1108. https://doi.org/10.1016/J.PETROL.2018.10.078.
Alkinani, H.H., Al-Hameedi, A.T.T., Dunn-Norman, S., Flori, R.E., Hilgedick, S.A., Al-Maliki,
M.A., Alshawi, Y.Q., Alsaba, M.T. and Amer, A.S. 2019b. Examination of the relationship
between rate of penetration and mud weight based on unconfined compressive strength of
the rock. J. King Saud. Univ. Sci. 31: 966–972. https://doi.org/10.1016/j.jksus.2018.07.020.
Alkinani, H.H., Al-Hameedi, A.T.T. and Dunn-Norman, S. 2020. Artificial neural network
models to predict lost circulation in natural and induced fractures. SN Appl. Sci. 2: 1980.
https://doi.org/10.1007/s42452-020-03827-3.
Barbosa, L.F.F.M., Nascimento, A., Mathias, M.H. and de Carvalho, J.A. 2019. Machine learning
methods applied to drilling rate of penetration prediction and optimization: A review.
J. Pet. Sci. Eng. 183: 106332. https://doi.org/10.1016/j.petrol.2019.106332.
Brankovic, A., Matteucci, M., Restelli, M., Ferrarini, L., Piroddi, L., Spelta, A. and Zausa, F.
2021. Data-driven indicators for the detection and prediction of stuck-pipe events in oil
& amp; gas drilling operations. Upstream Oil and Gas Technology 7: 100043. https://doi.
org/10.1016/j.upstre.2021.100043.
Breiman, L. 2001. Random forests. Mach. Learn. 45: 5–32. https://doi.
org/10.1023/A:1010933404324.
Chen, T. and Guestrin, C. 2016. XGBoost: A scalable tree boosting system. pp. 785–794.
In: Proceedings of the ACM SIGKDD International Conference on Knowledge Discovery
and Data Mining. ACM, New York, NY, USA. https://doi.org/10.1145/2939672.2939785.
Feng, Y. and Gray, K.E. 2017. Review of fundamental studies on lost circulation and
wellbore strengthening. J. Pet. Sci. Eng. 152: 511–522. https://doi.org/10.1016/J.
PETROL.2017.01.052.
Geurts, P., Ernst, D. and Wehenkel, L. 2006. Extremely randomized trees. Mach. Learn.
63: 3–42. https://doi.org/10.1007/s10994-006-6226-1.
 Enhancing Drilling Fluid Lost-circulation Prediction 31

Gordon, A.D., Breiman, L., Friedman, J.H., Olshen, R.A. and Stone, C.J. 1984. Classification
and regression trees. Biometrics 40: 874. https://doi.org/10.2307/2530946.
Hoerl, A.E. and Kennard, R.W. 1970. Ridge regression: biased estimation for nonorthogonal
problems. Technometrics 12: 55–67. https://doi.org/10.1080/00401706.1970.10488634.
Moazzeni, A., Nabaei, M. and Jegarluei, S.G. 2012. Decision making for reduction of
nonproductive time through an integrated lost circulation prediction. Petroleum Science
and Technology 30(20): 2097–2107. http://dx.doi.org/10.1080/10916466.2010.495961 30,
2097–2107. https://doi.org/10.1080/10916466.2010.495961.
Otchere, D.A., Arbi Ganat, T.O., Gholami, R. and Ridha, S. 2021a. Application of supervised
machine learning paradigms in the prediction of petroleum reservoir properties:
Comparative analysis of ANN and SVM models. J. Pet. Sci. Eng. 200: 108–182.
https://doi.org/10.1016/j.petrol.2020.108182.
Otchere, D.A., Ganat, T.O.A., Gholami, R., Lawal, M., Arbi Ganat, T.O., Gholami, R. and
Lawal, M. 2021b. A novel custom ensemble learning model for an improved reservoir
permeability and water saturation prediction. J. Nat. Gas Sci. Eng. 91: 103962. https://doi.
org/10.1016/j.jngse.2021.103962.
Otchere, D.A., Ganat, T.O.A., Ojero, J.O., Taki, M.Y. and Tackie-Otoo, B.N. 2021c. Application
of gradient boosting regression model for the evaluation of feature selection techniques
in improving reservoir characterisation predictions. J. Pet. Sci. Eng. 109244. https://doi.
org/10.1016/J.PETROL.2021.109244.
Otchere, D.A., Abdalla Ayoub Mohammed, M., Ganat, T.O.A., Gholami, R. and Aljunid
Merican, Z.M. 2022a. A novel empirical and deep ensemble super learning approach
in predicting reservoir wettability via well logs. Applied Sciences 12: 2942. https://doi.
org/10.3390/app12062942.
Otchere, D.A., Aboagye, M., Mohammed, M.A.A. and Boakye, T.B. 2022b. Enhancing drilling
fluid lost-circulation prediction using model agnostic and supervised machine learning.
SSRN Electronic Journal. https://doi.org/10.2139/ssrn.4085366.
Otchere, D.A., Ganat, T.O.A., Nta, V., Brantson, E.T. and Sharma, T. 2022c. Data analytics
and Bayesian Optimised Extreme Gradient Boosting approach to estimate cut-offs from
wireline logs for net reservoir and pay classification. Appl. Soft Comput. 120: 108680.
https://doi.org/10.1016/j.asoc.2022.108680.
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, É. 2011. Scikit-learn: Machine learning in Python.
Journal of Machine Learning Research 12: 2825–2830.
Pilehvari, A.A. and Nyshadham, V.R. 2002. Effect of material type and size distribution on
performance of loss/seepage control material. Paper presented at the International
Symposium and Exhibition on Formation Damage Control, Lafayette, Louisiana.
pp. 863–875. In: All Days. SPE. https://doi.org/10.2118/73791-MS.
Sabah, M., Talebkeikhah, M., Agin, F., Talebkeikhah, F. and Hasheminasab, E. 2019.
Application of decision tree, artificial neural networks, and adaptive neuro-fuzzy inference
system on predicting lost circulation: A case study from Marun oil field. J. Pet. Sci. Eng.
177: 236–249. https://doi.org/10.1016/J.PETROL.2019.02.045.
Sabah, M., Mehrad, M., Ashrafi, S.B., Wood, D.A. and Fathi, S. 2021. Hybrid machine learning
algorithms to enhance lost-circulation prediction and management in the Marun oil field.
J. Pet. Sci. Eng. 198: 108125. https://doi.org/10.1016/j.petrol.2020.108125.
Shapley, L.S. 1953. A value for n-person games. pp. 307–318. In: Contributions to the
Theory of Games (AM-28), Volume II. Princeton University Press. https://doi.
org/10.1515/9781400881970-018.
32 Data Science and Machine Learning Applications in Subsurface Engineering

Tarafder, S., Badruddin, N., Yahya, N. and Egambaram, A. 2021. EEG-based drowsiness
detection from ocular indices using ensemble classification. pp. 21–24. In: 2021 IEEE 3rd
Eurasia Conference on Biomedical Engineering, Healthcare and Sustainability (ECBIOS).
IEEE. https://doi.org/10.1109/ECBIOS51820.2021.9510848.
Tarafder, S., Badruddin, N., Yahya, N. and Nasution, A.H. 2022. Drowsiness detection using
ocular indices from EEG signal. Sensors 22: 4764. https://doi.org/10.3390/s22134764.
Tibshirani, R. 1996. Regression shrinkage and selection via the lasso. Journal of the
Royal Statistical Society: Series B (Methodological) 58: 267–288. https://doi.
org/10.1111/j.2517-6161.1996.tb02080.x.
Toreifi, H., Rostami, H. and Manshad, A.K. 2014. New method for prediction and solving
the problem of drilling fluid loss using modular neural network and particle swarm
optimization algorithm. J. Pet. Explor. Prod. Technol. 4: 371–379. https://doi.org/10.1007/
s13202-014-0102-5.
Vapnik, V. and Lerner, A. 1963. Pattern recognition using generalised portrait method.
Automation and Remote Control 24: 774–780.
Zarrouk, S.J. and McLean, K. 2019. Geothermal wells. Geothermal Well Test Analysis 39–61.
https://doi.org/10.1016/B978-0-12-814946-1.00003-7.
 Chapter 3
Application of a Novel Stacked
Ensemble Model in Predicting Total
Porosity and Free Fluid Index via
Wireline and NMR Logs
Daniel Asante Otchere1,2

1. Introduction
With the growth in energy needs and advances in drilling and hydraulic
fracturing techniques, the extraction of hydrocarbons from reservoirs
innovatively has become a focus for exploration and exploitation in recent
years (Li et al., 2022). Therefore, a complete understanding of bound and
moveable fluids is critical for accurately evaluating net pay reservoirs and
optimising completion, production, and enhanced oil recovery activities
(Otchere et al., 2022b). In reservoir terms, pore fluids refer to the fluid
distribution within the pore structure of the reservoir rock. There are two
main types of pore volume fractions: bound fluid, the fluid that is physically
trapped within the pore structure and cannot be easily displaced; and free fluid
volume, the fluid that can easily move within the pore structure. Knowing
the specific distinction between the proportion of bound and free fluid in a
reservoir is essential for predicting the ease of fluid flow and production (Cai
and Hu, 2019). For example, a reservoir with a higher proportion of bound
fluid may require more aggressive enhanced recovery methods to increase

1
Centre of Research for Subsurface Seismic Imaging, Universiti Teknologi PETRONAS, 32610, Seri
Iskandar, Perak Daril Ridzuan, Malaysia.
2
Institute for Computational and Data Sciences, Pennsylvania State University, University Park, PA,
USA.
Email: [email protected]
34 Data Science and Machine Learning Applications in Subsurface Engineering

production. Understanding pore fluid type and volume fractions is crucial for
effective reservoir management and optimisation (Cai and Hu, 2019; Otchere
et al., 2021c).
Numerous approaches have been formulated to ascertain the fluid
distribution in a reservoir. The most appropriate and accurate method to
measure pore types in reservoirs depends on the specific characteristics of
the reservoir being studied. One standard method is the mercury injection
capillary pressure (MICP) analysis, which measures the amount of mercury
that can be injected into the pore structure at a given pressure (Peng et al.,
2017). The amount of mercury that can be injected into the pore structure is
directly related to the proportion of free fluid in the reservoir. MICP is a widely
used method for measuring fluid distribution, and it is considered one of the
most accurate and reliable methods for measuring free fluid volume (Dugan,
2015; Mitchell et al., 2008). Another method is nuclear magnetic resonance
(NMR), which can provide a detailed description of the pore structure and fluid
distribution within the reservoir (Xu et al., 2015). NMR is particularly useful
for complex pore structures and can provide information on the distribution
of multiple fluid phases. Additionally, NMR is a non-destructive method that
uses magnetic fields and radio waves to measure the magnetic susceptibility
of the pore fluids (Branco and Gil, 2017; Otchere et al., 2022a). The magnetic
susceptibility of the pore fluids is used to distinguish between the bound and
free fluid in the reservoir, making it a valuable tool for measuring pore fluid
types and volume fractions on a large scale (Heaton et al., 2000).
Both methods, however, have their inherent set of challenges. A
fundamental challenge in measuring or determining pore type is the
complexity of the pore structure. Reservoirs can have a variety of pore sizes
and shapes, and different methods may be more effective for different types of
pore structures. One challenge associated with using MICP is that it requires
a core sample from the reservoir, which can be costly and time-consuming.
Additionally, the availability of core samples may be limited, making it
difficult to obtain enough samples to provide a representative analysis of the
reservoir (Wu et al., 2021). Also, MICP is a destructive method, where the
core sample is destroyed after analysis, which could make it difficult to repeat
the analysis for the same sample if needed. Another challenge is the sensitivity
of MICP to the saturation of the fluid, requiring the fluid to be in a state of
equilibrium, which may not be the case in some reservoirs. This challenge
can affect the accuracy of the results obtained from MICP analysis (Wu
et al., 2020). On the other hand, NMR is sensitive to the reservoir’s porosity
and cannot provide quantitative measurements of pore types. Furthermore,
NMR is responsive to the existence of clay minerals, which can impact the
precision of the outcomes (Elsayed et al., 2020). Table 3.1 summarises the
unique advantages and limitations of these two techniques.
 Application of a Novel Stacked Ensemble Model in Predicting Total Porosity 35

Table 3.1. Summary of MICP and NMR techniques identifying key advantages, limitations, and
measurement principles.

Technique Measurement Developed Advantages Limitations Observations

Principle by
Mercury Injects mercury Purcell Accurate Requires a MICP is
Injection into the pore (1949) measurement core sample considered
Capillary structure at a of free fluid one of the
Pressure given pressure, volume most accurate
(MICP) and the amount of (FFV) and reliable
mercury that can methods for
be injected into measuring
the pore structure FFV, but it is
is directly related a destructive
to the proportion method.
of free fluid in the
reservoir.
Nuclear Uses magnetic (Bloch, Provides Not able NMR imaging
Magnetic fields and radio 1946) detailed to provide is particularly
Resonance waves to obtain images of quantitative useful for
(NMR) detailed images the pore measurements complex pore
of the fluid structure of pore types structures and
distribution within and fluid and sensitive can provide
the reservoir. distribution to the information on
within the presence of the distribution
reservoir. clay minerals. of multiple
fluid phases.

Overall, understanding pore volume fraction is crucial for effective

reservoir management and optimisation. However, accurately measuring and
determining porosity and its associated producible volume can be challenging
due to the complexity of the pore structure and the limited availability of
core samples.
The measurement of pore volume fraction is complicated and challenging.
Many researchers in this discipline are always experimenting with new ideas
and methodologies. Many researchers have been drawn to estimate pore
volume fractions and total porosity using NMR data over the last two decades.
This preference for NMR is owing to its non-invasive nature (Otchere et al.,
2022a). Artificial intelligence (AI) can potentially revolutionise how pore
volumes are predicted in reservoirs. One potential application is using AI to
predict pore volumes measured by NMR, using wireline logs as input.
Conventional wireline logs are ubiquitous in all wells drilled and provide
essential information on the formation properties of the reservoir. These logs
exhibit an intricate and nonlinear correlation with the distribution of pore
volume fractions. On the other hand, NMR measurements provide information
on the fluid distribution within the pore structure. Nonetheless, this approach
36 Data Science and Machine Learning Applications in Subsurface Engineering

incurs significant expenses and may not appear practical to execute in every
well during periods of low oil prices (Otchere et al., 2022a). By combining
these two types of data, AI can be used to predict porosity and volume fractions
with a high degree of accuracy. An ideal way to accomplish this is through the
use of machine learning and data analytical techniques, which can be trained
on a large dataset of wireline logs and NMR porosity and volume fraction
measurements. These techniques exhibit resilience to noise and demonstrate a
high efficacy in identifying intricate or nonlinear patterns or features indicative
of specific input data and pore volume fraction (Otchere et al., 2021a). Once
trained, the AI model can then be used to predict total porosity and pore volume
fractions for new reservoirs based on the wireline logs.
Utilising wireline logs to assist in estimating this indispensable property
is comparatively more economical and time-saving and provides real-time
information than running NMR logs in all wells. As such, considering current
oil prices, the demand for an inexpensive and reliable field-scale technique
for quantitative porosity and pore volume fraction measurement has been
of interest. Hence, this study endeavours to develop an AI-based model that
can accurately predict the porosity and producible pore volume fraction in
a carbonate gas reservoir using wireline logs as input and NMR measured
total porosity and FFV as the target output. This study aims to provide a more
efficient and accurate method for reservoir characterisation and management.

2. Nuclear Magnetic Resonance

2.1 Concept and Application
Atomic nuclei with an odd number of protons and neutrons exhibit NMR
when subjected to magnetic fields at a specific resonance frequency (Coates
et al., 1999). The proton in hydrogen has a strong magnetic moment due to
its abundance in both water and oil, which is the outcome of the nucleus’
angular momentum. At that point, the proton acts like a spinning bar magnet.
In an external magnetic field, it undergoes polarisation, which manifests as
a quantifiable quantum feature known as the induced frequency (Larmor
frequency) (Amani et al., 2017). Figure 3.1 shows a simplified representation
of the reaction of a nucleus to an external magnetic field (Otchere et al., 2022a).
Combinable Magnetic Resonance (CMR) tool is a recently developed
technique that is used for porosity partitioning to determine the bound fluid
volume (BFV) and FFV in reservoirs (Gubelin and Boyd, 1997). CMR can
be used in conjunction with NMR to provide a more accurate and complete
picture of the fluid distribution within the reservoir. CMR measurements
are particularly useful in reservoirs with high water saturation, where other
techniques may not be effective. Additionally, CMR can provide information
on both bound and FFVs, which is not possible with other techniques.
 Application of a Novel Stacked Ensemble Model in Predicting Total Porosity 37

Fig. 3.1. The excitation and relaxation of polarised hydrogen atomic nuclei in response to an external
magnetic field (Otchere et al., 2022a).

The concept of using NMR to predict pore fluid types is to use the magnetic
properties of certain nuclei to provide information about a reservoir’s fluid
composition and rock properties. This prediction is achieved by measuring
the proton density and relaxation times of the fluids in the reservoir, which
are unique to different fluid types, such as water, oil, and gas. When CMR is
used for porosity partitioning, the total porosity is calculated by measuring the
magnetic susceptibility of the pore fluids and using the relationship between
magnetic susceptibility and fluid volume fraction. This method is based on
the assumption that the magnetic susceptibility of the bound fluid is different
from the free fluid, which is a key concept in CMR. The following equations
are used to calculate total porosity, BFV, and FFV in reservoirs:
(ϕT) = ϕB + ϕF (3.1)
where the total porosity is represented as ϕT, which is a dimensionless value
between 0 and 1, representing the fraction of the rock sample that is made up
of pores, ϕB is the bound fluid volume fraction, and ϕF is the FFV fraction.
BFV = ϕB × Vp (3.2)
where Vp is the pore volume.
FFV = ϕF × Vp (3.3)
38 Data Science and Machine Learning Applications in Subsurface Engineering

Vp is typically calculated by multiplying the bulk volume of the rock

sample (Vb) by the total porosity (ϕT).
Vp = ϕT × Vb (3.4)
where Vb is the bulk volume of the rock sample, which can be measured
by using various techniques such as helium pycnometry, mercury intrusion
porosimetry, or gas adsorption. It is worth mentioning that the pore volume
can also be calculated by measuring the rock sample’s bulk volume and
subtracting the solid grains’ volume from it. This method is known as the
grain volume method and is another way to calculate the pore volume. These
equations are based on the assumption that the magnetic susceptibility of the
bound fluid is different from the free fluid. This means that the bound fluid can
be distinguished from the free fluid by measuring the magnetic susceptibility
of the pore fluids.

2.2 Works Related to the Use of Machine Learning in NMR for

Reservoir Characterisation
Machine learning has been extensively employed to complement conventional
methods and, in some instances, provide certain advantages. The deployment
of machine learning techniques and AI algorithms to overcome intricate issues
has gained traction in numerous sectors, including the petroleum industry
(Otchere et al., 2021a). The application of machine learning in reservoir
characterisation has seen massive growth underlining its importance in this
field. Machine learning models have become quite useful in situations where
not much data is collected. These models have been used to predict various
logs, which, due to time and cost, were not run in wells. These models have
illustrated high accuracy compared to actual logs, saving time and money
while providing a real-time estimate of reservoir properties during drilling.
A review of articles utilising machine learning techniques and NMR data to
provide accurate alternatives is presented below.
Rezaee (2022) expounds upon a study that employed an extensive
repository of CMR to engender outputs from an NMR logging tool for clastic
rocks. The study intended to evaluate the effectiveness of diverse machine
learning (ML) methodologies employing customary well-logs as input data.
The study generated outputs for CMR data, including irreducible pore fluid
(BVI), clay bound water (CBW), producible fluid (FFI), irreducible water
saturation (Swirr), logarithmic mean of T2 relaxation time (T2LM), and
permeability. To this end, well logs from 14 Western Australian wells were
gathered, of which 80% were used for model training and validation and the
remaining 20% for testing purposes. Upon comparing the outcomes, it was
revealed that the Adaptive Boosting (AdaBoost) model demonstrated the most
notable results, with R2 values exceeding 0.9 for the blind set. These findings
 Application of a Novel Stacked Ensemble Model in Predicting Total Porosity 39

indicate that the ML methodology can be employed for the generation of

NMR logs with a high degree of precision.
Masroor et al. (2022) describe an ML procedure for estimating NMR
permeability from conventional well logs. The methodology employs three
supervised (ML) algorithms: group method of data handling, Random Forest
(RF), and one-dimensional convolutional neural network (CNN). Additionally,
a modified version of the two-dimensional CNN (Residual 2D-CNN) is
devised, receiving artificial 2D feature maps derived from conventional logs.
The ML and deep learning (DL) models’ hyperparameters are optimised
using genetic algorithms to enhance their efficacy. The results indicate that
nonlinear machine and DL techniques help estimate NMR permeability, with
the Res 2D-CNN model providing the most efficient results (accuracy of
0.97). The study underscores the significance of employing produced feature
maps to train the Res 2D-CNN model and the favourable impact of integrating
residual and deeper bottleneck architectures in enhancing prediction accuracy
and minimising training duration.
Tamoto et al. (2023) also describe a study aimed at developing ML
models for the estimation of NMR porosity from well logs. NMR logs are
considered valuable for porosity quantification but are limited by high costs
and adverse subsurface acquisition conditions. The adjusted RMSE and R2
and metrics compared and evaluated four supervised ML models. The best
results were achieved with the CatBoost regressor, producing an RMSE of less
than 0.01 and an adjusted R2 of 0.87. The ML models substantially improved
total porosity prediction compared to traditional empirical computations. The
dissimilarities between actual NMR logs and the ML predictions were mostly
below 5%. Additionally, a porosity model based on well logs was developed
using tree boosting, and the impacts of input variables on model estimates were
investigated. The study also scrutinised the behaviour of the model’s linear and
nonlinear characteristics to comprehend their relationships with the dataset.
Li et al. (2020) present a study on the use of two neural-network-based
ML models, long short-term memory (LSTM) network and a variational
autoencoder with a convolutional layer (VAEc) network, for the generation of
synthetic NMR T2 distributions from formation mineral and fluid saturation
logs. The models are trained and tested on a limited sample of data from the
Bakken shale formation. The results show that both models perform well,
synthesising fluid-filled pore size distributions even when the input data
is corrupted by noise. The proposed method has the potential to improve
reservoir characterisation under data constraints. NMR logs provide valuable
information on fluid mobility, fluid-filled pore size distribution, fluid mobility,
porosity, and permeability in the near-wellbore reservoir volume. However,
their acquisition is limited due to financial and operational challenges. Hence,
the use of ML can make NMR logs available in all wells.
40 Data Science and Machine Learning Applications in Subsurface Engineering

Gu et al. (2021) present a ML approach to estimate fluid saturation in

reservoirs by evaluating fluid components in the T2-D spectrum. The method
integrates a GABP neural network and Gaussian Mixture Model (GMM) to
determine a new saturation formula based on the physical significance between
the T2-D spectrum and fluid saturation. The approach is evaluated through
numerical simulations, and results demonstrate improved accuracy and
stability compared to traditional BP neural networks. The GMM determines
different Gaussian probability distribution functions to characterise different
fluids in the T2-D spectrum, leading to more accurate saturation estimates. The
proposed method has the ability to quantitatively evaluate fluid components
in the T2-D spectrum.
These studies demonstrate the potential of using ML algorithms to analyse
wireline log and NMR data and predict various reservoir properties such as
pore-scale properties, permeability, fluid saturations, and fluid types. These
methods have been shown to be accurate and have the potential to improve the
efficiency and accuracy of reservoir characterisation.

3. Methodology
3.1 Data Collection and Description
Conventional wireline logs from a vertical exploration gas well were used in
this study. The CMR tool was logged in the well. The objective of the CMR
logging was to evaluate the formation’s total porosity, bound and free fluid,
and permeability. CMR was logged over a carbonate section capped by shale.
Over the zone of interest, we do not observe borehole effects on the CMR signal
due to a relatively smooth borehole combined with moderate mud salinity
(approximately 58 ppk equivalent NaCl using chart GEN-9 with Rmf = 0.0694
ohmm at 21 deg C). No anomalies were observed on the raw echoes, T2
distribution, total porosity, and bin porosities. Since the well is drilled with
water-based mud, the T2 distribution reflects the pore size distribution. The data
used in this study underwent quality control and reprocessing. The parameters
selection for the CMR reprocessing used in this study is captured in Table 3.2.
Figure 3.2 shows the incremental increase in porosity and multiple pore size
distribution with favourable connectivity between pores, identified through the
multi-exponential decay time analysis of NMR T2 distribution in 100% brine
saturated samples. The partitioning of the porosity was done as follows:
1. Small Pore Porosity (T2 min to 3 ms) – this can be considered as
water-filled porosity associated with clay (water resistivity of Rwb in the
Dual Water Saturation model).
2. Capillary Bound Fluid (3 ms to T2 cutoff) – this can be considered part
of effective porosity. In clastic rocks, this would be expected to be water
 Application of a Novel Stacked Ensemble Model in Predicting Total Porosity 41

Table 3.2. Processing parameters selection for CMR reprocessing.

Processing Parameters
Porosity Algorithm
Starting Echo 2nd
Number of averaging levels 3
T2 minimum 0.5 msec
T2 maximum 6000 msec
Number of Spectral Components 30
T1/T2 Ratio Minimum 1
T1/T2 Ratio Maximum 3
Polarization Correction Threshold 0.015 v/v
Producibility Parameters
Free Fluid Cut-off 0.02 v/v
Porosity Parameters
T2 Cut-off 100 msec (default for Carbonate)
Taper Cut-off Start 8 msec
T1/T2 Ratio 2
Bin Porosity T2 Cutoffs (ms) 1, 3, 10, 33, 100, 300, 1000, 3000
Small Pore Porosity T2 Cutoffs (ms) 1 to 3
Capillary Bound Fluid T2 Cutoffs (ms) 3 to 100

Fig. 3.2. NMR T2 multimodal pore size distribution of different core samples at 100% water saturation
(Otchere et al., 2022a).
42 Data Science and Machine Learning Applications in Subsurface Engineering

filled (with water resistivity equal to Rw in the Dual Water Saturation

model); however, in carbonate rocks, this could be filled with water,
hydrocarbon or both.

3.2 Data Analysis and Feature Engineering

The relationship between output and input features was established by
computing a covariance heatmap and pair plots through Jupyter notebook, a
Python-based computing platform. Tables 3.3 and 3.4 present the statistical
summary of the data, aiding in identifying the data distribution and
expected trends.
These logs were chosen due to their established relationship to porosity
using the studies by Otchere et al. (2021b). In the next stage, several ensemble
and DL models were used to predict the porosity and FFV using the selected
wireline logs.
Table 3.3. Descriptive statistics of the input data used in this study.

S-sonic Gamma SP R(S) Rt Rxo Neutron RHOB SXGAS

HLLS
mean 162.64 41.78 –388.62 102.37 335.15 4.33 0.14 2.33 0.51
std 39.73 19.45 27.70 259.54 1628.80 3.84 0.06 0.20 0.35
min 99.98 24.57 –431.31 1.45 1.32 0.86 0.02 1.76 0.00
25% 138.79 31.47 –411.00 12.17 12.44 2.09 0.10 2.17 0.22
50% 149.13 34.83 –397.50 35.94 46.82 3.21 0.13 2.37 0.56
75% 172.14 41.96 –369.56 111.24 180.35 5.19 0.16 2.47 0.82
max 309.04 114.09 –314.13 4986.61 28651.10 51.44 0.37 2.66 1.00

Table 3.4. Descriptive statistics of the targets used in this study.

TCMR CMFF
mean 0.11 0.07
std 0.03 0.04
min 0.01 0.00
25% 0.09 0.04
50% 0.11 0.07
75% 0.13 0.09
max 0.31 0.16
 Application of a Novel Stacked Ensemble Model in Predicting Total Porosity 43

3.3 Machine Learning Model Application

It is widely accepted that DL algorithms perform better than their traditional,
more simplistic equivalents in the field of ML. The ability of DL models to
manage substantial amounts of data has resulted in their widespread popularity
for tasks involving predictive modelling and pattern recognition (Otchere
et al., 2021a). Despite all of the advancements that have been made thanks
to DL in the field of AI, it still has significant flaws. Hyperparameter tuning,
slow convergence on short datasets, infinite structures, and complexities are
among the most often encountered problems. While traditional ML models
can mitigate these issues on smaller datasets, they are limited in their ability to
produce higher model performance overall. It is difficult to know exactly which
model to use because of the hazy definition of small datasets. Based on these
assertions, a new custom stacked ensemble model is developed in this research.
This custom ensemble model integrates multiple varied base learners in its
architecture to enhance model effectiveness. However, compared to the custom
stacked ensemble model, the performances of the individual models are low.

3.3.1 Building Deep Learning Models

A CNN model is a type of DL model that is particularly well-suited for
image and signal processing tasks. However, its performance was evaluated
on a regression task in this study. It comprises multiple layers, including
convolutional, pooling, and fully connected layers. To build a CNN model for
this study, the following steps were taken:
1. The input data, wireline logs with 10 dimensions, were reshaped to match
the expected input shape for a CNN model.
2. The model was composed of a series of convolutional layers designed to
extract features from the input data. These layers use filters to scan over
the input data and extract features at different scales.
3. The output of the convolutional layers was then passed through one or
more pooling layers, which minimised the dimensionality of the feature
maps and offered some form of translation invariance.
4. The final output of the pooling layers was then passed through one or
more fully connected layers to make the final predictions.
5. The model was then trained using the training dataset and evaluated using
the holdout dataset.
6. The model was fine-tuned by adjusting the number of layers, the number
of filters in each layer and the size of the filters to achieve the desired
level of performance.
44 Data Science and Machine Learning Applications in Subsurface Engineering

7. The model was trained using the Adam optimiser with MAE as the loss
function.
8. The model was trained using early stopping to prevent overfitting.
Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU)
models are types of Recurrent Neural Network (RNN) models that are
particularly well-suited for sequential data such as time series, speech, and
text. A GRU model is similar to an LSTM model but has fewer parameters,
which can be useful when working with limited data or computational
resources. The key feature of GRU and LSTM is the ability to remember
information over a more extended period of time by using gates to control
the flow of information in and out of the cell. This model type can be useful
when working with sequential data where the order of the data is essential.
However, this model type can be computationally intensive and requires a
large amount of data for training. To build the GRU and LSTM model for this
study, the following steps were taken:
1. The input data was transformed into a suitable format for GRU/LSTM,
usually in the form of a 3D array where the first dimension represents the
number of samples, the second dimension represents the time steps, and
the third dimension represents the number of features.
2. The model was composed of one or more GRU/LSTM layers, which are
responsible for processing the input data and extracting features.
3. The output of the GRU/LSTM layers was then passed through one or
more fully connected layers to make the final predictions.
4. The model was then trained using the training dataset and evaluated using
the holdout dataset.
5. The model was fine-tuned by adjusting the number of layers, the number
of neurons in each layer, and the dropout rate to achieve the desired level
of performance.
6. The model was trained with regularisation techniques such as dropout
and early stopping to prevent overfitting.

3.3.2 Building a Hybrid Stacked Ensemble Model

A hybrid ensemble model is a combination of multiple models working
together to improve the overall performance of the model. A hybrid ensemble
model of Extra Trees, Random Forest, and XGBoost was built using stacking
in this case. Stacking is a technique where the predictions of the base models
(Extra Trees, Random Forest, and XGBoost) are used as input to train a
meta-model, which makes the final prediction. The base models were trained
on the same dataset, and their predictions were combined into a single array,
 Application of a Novel Stacked Ensemble Model in Predicting Total Porosity 45

Table 3.5. Model architecture stacking ensemble models as base learners for the Custom Ensemble.
Adapted from Otchere et al. (2022a).

Algorithm 1: Stacking Base Learners

1: def ensemble_predict(models, X_train, y_train, X_test):
2: # create an empty array to store predictions
3: predictions = list()
4:
5: # loop through the models
6: for model in models:
7: model.fit(X_train, y_train)
8: y_pred = model.predict(X_test)
9: predictions.append(y_pred)
10:
11: # return the predictions
12: return predictions
13:
14: def train_stacking_ensemble(X_train, y_train, X_test, y_test):
15: models = get_models()
16: ensemble_predictions = ensemble_predict(models, X_train, y_train, X_test)
17: meta_model = LinearRegression()
18: meta_model.fit(ensemble_predictions, y_test)
19:
20: # predict on the test set
21: y_pred = meta_model.predict(ensemble_predictions)
22:
23: # evaluate performance
24: r2 = r2_score(y_test, y_pred)
25: mae = mean_absolute_error(y_test, y_pred)
26: rmse = sqrt(mean_squared_error(y_test, y_pred))
27: fit_likelihood = OLS(y_test, add_constant(y_pred)).fit()
28: print(“R2: {:.4f}”.format(r2))
29: print(“MAE: {:.4f}”.format(mae))
30: print(“RMSE: {:.4f}”.format(rmse))
31: print(“AIC: {:.4f}”.format(fit_likelihood.aic))

which was used as input to train a meta-model using linear regression. The
meta-model was trained on the combined predictions and the target variables
of the training dataset, and it was then used to make predictions on the test
set. This approach can improve the model’s performance by combining the
strengths of different models and reducing the prediction variance (Tarafder
et al., 2022). Table 3.5 summarises the function used to develop the stacked
ensemble using the Extra Trees, Random Forest, and XGBoost models, whiles
Fig. 3.3 summarises the entire workflow.
There are several different combinations of models and different
parameters for each model, as well as different meta-models to find the best
hybrid ensemble for a specific task, dataset, and desired performance.
46 Data Science and Machine Learning Applications in Subsurface Engineering

Fig. 3.3. Simplified workflow of the methodology used in this study.

3.4 Criteria for Model Evaluation

The discrepancy between estimated and actual data shows the models’
prediction errors when assessing their performance. The performance of the
models in this study was analysed via these evaluation criteria;
1. Mean Absolute Error (MAE): This assessment criterion, also known as L1
loss, is vulnerable to relative errors and resistant to global scaling of the
expected outcome. MAE is mathematically written as:

∑
n
yi − ŷi
MAE = i =1
(3.5)
n
2. Root Mean Squared Error (RMSE): This assessment criterion is the
standard deviation of models’ prediction errors and indicates how close
predicted data is to actual data. This is written as:

∑
n
( yi − yˆi ) 2
RMSE = i =1 (3.6)
n
3. Akaike Information Criterion (AIC): This assessment metric is based on
a frequentist probability approach, which scores a model according to
its maximum probability estimation. This technique is used to determine
the precision and excellence of models, indicating a more robust match
between the model and new data. AIC is statistically written as:
AIC = 2K – 2(log – likelihood) (3.7)
 Application of a Novel Stacked Ensemble Model in Predicting Total Porosity 47

4. Coefficient of determination (R2): R2 denotes the closeness of the

dependent value to the best-fit regression line. R2 is mathematically
expressed as:

R2 = 1 −
∑ ( yi − ŷi )2 (3.8)
∑ ( yi − yi )2
Besides R2, low errors are indicative of good model performance.

4. Results and Discussion

4.1 Evaluation of Models’ Performances
Given the expense associated with operating an NMR log and the efficacy
of machine learning models in reservoir characterisation, a variety of models
were employed to estimate the total porosity and free fluid volume of a
carbonate reservoir from wireline logs. The findings in Fig. 3.4 indicate that
the hybrid ensemble exhibited an improved accuracy of holdout data in the
prediction of total porosity and moveable fluid compared to the other models.
Although some models, Extra Trees and XGBoost, achieved the highest
training accuracy but comparatively lower test accuracy compared to the
hybrid model, the models likely overfitted to the training data. Overfitting
occurs when a model is too complex and can fit the training data very well
but cannot generalise well to the test data. This issue is usually caused by
having too many parameters or too much capacity in the model, which allows
it to fit the noise in the training data. In this case, the Extra Trees model
has a high training accuracy of 1, which is an indication that it is able to fit
the training data very well. However, the test accuracy of 0.82 is lower than

Fig. 3.4. Comparison of train/test correlation coefficient results.

48 Data Science and Machine Learning Applications in Subsurface Engineering

the hybrid model, which indicates that the Extra Trees model is not able to
generalise well to new data. On the other hand, the hybrid model combines the
predictions from multiple models, which reduces the variance and increases
the robustness of the final prediction. Therefore, it is less likely to overfit the
training data and more likely to generalise well to new data. Based on this
alone, it is vital to note that the hybrid model performed more robustly than
the other models. Additional analysis was conducted to assess the prediction
errors of all the models.
Error measurements were used to assess the consistency and precision
of the models and are presented in Fig. 3.5. The MAE and results indicate
the hybrid model’s superiority over other models, indicating its consistency
and precision in multioutput prediction. The Extra Trees Regression model’s
performance was comparative, with an RMSE of 0.0143 and an MAE of
0.0102. Its AIC score is also higher than the hybrid ensemble model indicating
that the hybrid model has a higher probability of fitting new data. The
Random Forest and XGBoost models performed relatively well, with RMSE
of 0.0.0152 and 0.0.0155, respectively. Their MAE was higher than the hybrid
ensemble model.
On the other hand, the LSTM and GRU models performed poorly
compared to the other base models, with RMSE of 0.0231 and 0.0202 and
MAE of 0.0176 and 0.0151, respectively. The AIC scores are also higher than
the other models, indicating that they are less likely to generalise well to new
data. Comparing the two RNN models indicate that the GRU model recorded
lower errors than the LSTM model. The worst performing model, the CNN
model, achieved an MAE of 0.0273 and RMSE of 0.034, indicating that it is
not able to fit the data well. The relatively high AIC score also indicates that
it is not the best model for this task.
The results suggest that the hybrid ensemble model is the best performing
model for this study, as it has the highest train and test scores, the best AIC
score, and the lowest MAE and RMSE. The Extra Trees Regression, Random
Forest, and XGBoost models performed well, but some overfitting issues were
observed. The ensemble models should be further investigated and optimised
to improve their performance. Moreover, it would be interesting to explore
other ensemble methods, such as bagging and boosting, to improve the model
performance. Furthermore, a larger dataset and more computational resources
may be necessary to enhance the performance of the LSTM and GRU models
(Fig. 3.5).
Figures 3.6 and 3.7 below illustrate the kernel density estimation (KDE)
of the Hybrid, GRU, and ET predicted and actual total porosity and FFV. From
the observation, the hybrid ensemble predicted outputs are significantly closer
to the total porosity and free fluid volume actual data. This result suggests
that the hybrid model is better equipped to capture a wide range of values
 Application of a Novel Stacked Ensemble Model in Predicting Total Porosity 49

Fig. 3.5. Comparison of the models’ estimation errors based on RMSE, MAE, and AIC on the
holdout data.

compared to the other models and is suitable for deployment in predicting

total porosity and FFV. A joint plot in Fig. 3.8 was used for visualisation to
further evaluate the hybrid model’s performance. The shaded area on the plot
shows a 95% confidence interval, which indicates high confidence in the
predicted total porosity and FFV values when compared to the actual. The
low confidence area is mainly in the prediction of low values as a result of
the low porosity values in the training data. The low values are due to the
data restriction in the reservoir interval, which will mostly have high porosity
values. Thus, the previously mentioned evaluation metrics results are verified
by conducting this sensitivity analysis.
The excellent performance of the hybrid model is mainly attributed to
its learning theory, model architecture, and mathematical formulation. From
a learning theory perspective, the hybrid model combines the strengths of
multiple base models, such as Extra Trees Regression, Random Forest, and
XGBoost, all ensemble methods known to enhance a single model estimation
performance by reducing overfitting and increasing generalisation. This is
achieved by combining predictions from multiple models, which reduces the
variance and increases the robustness of the final prediction. From a model
architecture perspective, the hybrid model uses a stacking ensemble technique
to combine the base models. This technique involves training multiple base
models on the input data and then using their predictions as input to a higher-
level meta-model that makes the final predictions. This allows the hybrid
model to capture both the low- and high-level features of the data, which leads
to improved performance. From a mathematical formulation perspective,
stacking ensemble models are based on the concept of a linear combination
of predictions from multiple models. The predictions from each base model
are combined using a linear function, and the weights of this function are
50 Data Science and Machine Learning Applications in Subsurface Engineering

Fig. 3.6. KDE indicating the closeness of estimated to actual total porosity where (a) is hybrid model
prediction, (b) is GRU model prediction, and (c) is ET model prediction.
 Application of a Novel Stacked Ensemble Model in Predicting Total Porosity 51

Fig. 3.7. KDE showing the closeness of estimated to actual FFV where (a) is hybrid model prediction,
(b) is GRU model prediction, and (c) is ET model prediction.
52 Data Science and Machine Learning Applications in Subsurface Engineering

Fig. 3.8. Joint plot of the hybrid model predicted vs actual values where total porosity is on the left
and FFV is on the right.

learned by training a meta-model on the predictions from the base models.

This combination of predictions from multiple models results in a more
robust model that generalises well to new data, as exhibited by the results.
This establishes the foundation of its superior prediction performance over the
independent base learners.
It is important to note that the LSTM, CNN, and GRU models performed
poorly, which may be due to the limited amount of data or the complexity
of the problem. Training a DL model on small datasets has the propensity
to initiate data memorisation, leading to poor performance on the holdout
data. The ineffectiveness of DL involving small datasets and its incapability
to interpret results confirms that it cannot be excessively adjudged the aptest
solution in all cases
Training a DL model on small datasets causes data memorisation, often
leading to overfitting, resulting in weak performance on the holdout data.
The limited effectiveness of DL on small data and its incapability to interpret
results suggest that it cannot be solely considered the optimal solution in all
scenarios. Overall, the excellent performance of the hybrid model can be
attributed to its ability to combine the strengths of multiple base models and its
use of a stacking ensemble technique and mathematical formulation that leads
to a more robust model that generalises well to new data. The hybrid model
was used to predict the total porosity and FFV for the entire well, as shown in
Fig. 3.9. The predicted model includes the training and the testing data.
Fig. 3.9. Wireline plot of carbonate well indicating the actual and predicted total porosity and FFV.
Application of a Novel Stacked Ensemble Model in Predicting Total Porosity 53
54 Data Science and Machine Learning Applications in Subsurface Engineering

5. Conclusions
To obtain a detailed understanding of the reservoir, it is crucial to quantify the
surface wettability at an early stage. This information is essential for enhanced
oil recovery (EOR), particularly chemical EOR, to maximise production.
This study aimed to develop a suitable model for predicting total porosity
and the free fluid volume using wireline logs as input. The study evaluated
several models, including a Hybrid ensemble model, Extra Trees Regression,
Random Forest, XGBoost, LSTM, GRU, and CNN. The performance of
the models was evaluated using several metrics, including train score, test
score, AIC, MAE, and RMSE. The incorporation of these computational
methods provides a more accurate depiction of surface wettability when core
wettability data and NMR logs are unavailable.
In the light of the results obtained in this study, the following inferences
can be made;
1. The hybrid ensemble model achieved a high train and test accuracy score,
indicating that it is able to fit the training data well and generalise well
to new data. This is further supported by the recorded lowest MAE, and
RMSE in its predictions.
2. The likelihood of the model fitting new data is also high, as indicated by
the AIC score, which is the lowest among all the models. AIC is a measure
of the probability of the model’s fit on new data. A low AIC score indicates
that the model is more likely to generalise well to new data.
3. The DL models, LSTM, CNN, and GRU, performed poorly due to their
high complexity and volume of data used.
4. The hybrid ensemble model performed better than the others because it
combined the strengths of multiple base models, Extra Trees Regression,
Random Forest, and XGBoost, and used a stacking ensemble technique.
This technique involves training multiple base models on the input data
and then using their predictions as input to a higher-level meta-model that
makes the final predictions. This allows the hybrid model to capture both
the low-level and high-level features of the data, which leads to improved
performance.
The results of this study suggest that the hybrid ensemble model is a
robust fit for this problem and is suitable for predicting total porosity and FFV.
It is able to achieve high accuracy, consistency, and precision, and generalise
well to new data. This makes it a valuable tool for the petroleum industry in
estimating the porosity of reservoirs, determining the pore volume fraction of
a reservoir that can be produced, and evaluating the potential of a prospect.
This can help improve hydrocarbon exploration and production efficiency,
ultimately leading to increased productivity and reduced costs.
 Application of a Novel Stacked Ensemble Model in Predicting Total Porosity 55

Acknowledgement
The authors express their sincere appreciation to Universiti Teknologi Petronas
and the Centre of Research for Subsurface Seismic Imaging for supporting
this work.

References
Amani, M., Al-Jubouri, M.B., Khadr, S. and Sayed, A.M. 2017. A comprehensive review on
the use of NMR technology in formation evaluation. https://www.semanticscholar.org/
paper/A-Comprehensive-Review-on-the-Use-of-NMR-Technology-Amani-Al-Jubouri/
fe15a754ea33ae2a3c1f2fa689b81f067b398fef.
Bloch, F. 1946. Nuclear induction. Physical Review 70: 460. https://doi.org/10.1103/
PhysRev.70.460.
Branco, F.R. and Gil, N.A. 2017. NMR study of carbonates wettability. J. Pet. Sci. Eng.
157: 288–294. https://doi.org/10.1016/j.petrol.2017.06.023.
Cai, J. and Hu, X. 2019. Petrophysical Characterization and Fluids Transport in Unconventional
Reservoirs. pp. 1–332. https://doi.org/10.1016/C2018-0-00934-2.
Coates, G.R., Xiao, L. and Prammer, M.G. 1999. NMR Logging Principles and Applications.
Houston: Halliburton Energy Services Publication.
Dugan, B. 2015. Data report: Porosity and pore size characteristics of sediments from Site
C0002 of the Nankai Trough determined by mercury injection. https://doi.org/10.2204/
IODP.PROC.338.202.2015.
Elsayed, M., Glatz, G., El-Husseiny, A., Alqubalee, A., Adebayo, A., Al-Garadi, K. and Mahmoud,
M. 2020. The effect of clay content on the spin–spin NMR relaxation time measured in
porous media. ACS Omega 5: 6545–6555. https://doi.org/10.1021/acsomega.9b04228.
Gu, M., Xie, R. and Jin, G. 2021. A machine-learning based quantitative evaluation of the fluid
components on T2-D spectrum. Mar. Pet. Geol. 134: 105353. https://doi.org/10.1016/J.
MARPETGEO.2021.105353.
Gubelin, G. and Boyd, A. 1997. Total porosity and bound-fluid measurements from an NMR
tool. Journal of Petroleum Technology 49: 718–718. https://doi.org/10.2118/39096-JPT.
Heaton, N., Minh, C.C., Freedman, R. and Flaum, C. 2000. High-resolution bound-fluid, free-
fluid and total porosity with fast NMR logging. Paper presented at the SPWLA 41st Annual
Logging Symposium, Dallas Texas (June). https://www.slideshare net>armhaggag>high
resolute…
Li, C., Liu, X., You, F., Wang, P., Feng, X. and Hu, Z. 2022. Pore size distribution characterization
by joint interpretation of MICP and NMR: A case study of chang 7 tight sandstone in the
ordos basin. Processes 10: 1941. https://doi.org/10.3390/pr10101941.
Li, H., Misra, S. and He, J. 2020. Neural network modeling of in situ fluid-filled pore size
distributions in subsurface shale reservoirs under data constraints. Neural Comput. Appl.
32: 3873–3885. https://doi.org/10.1007/S00521-019-04124-W.
Masroor, M., Emami Niri, M., Rajabi-Ghozloo, A.H., Sharifinasab, M.H. and Sajjadi, M.
2022. Application of machine and deep learning techniques to estimate NMR-derived
permeability from conventional well logs and artificial 2D feature maps. Journal of
Petroleum Exploration and Production Technology 12(3): 2937–2953. https://doi.
org/10.1007/S13202-022-01492-3.
Mitchell, P., Al-Hosani, I., Mehairi, Y. Al and Kalam, M.Z. 2008. Importance of mercury
injectiion capillary pressure (MICP) measurements at pseudo reservoir conditions.
56 Data Science and Machine Learning Applications in Subsurface Engineering

Society of Petroleum Engineers –13th Abu Dhabi International Petroleum Exhibition and
Conference. ADIPEC 2008 2: 952–962. https://doi.org/10.2118/117945-MS.
Otchere, D.A., Arbi Ganat, T.O., Gholami, R. and Ridha, S. 2021a. Application of supervised
machine learning paradigms in the prediction of petroleum reservoir properties:
Comparative analysis of ANN and SVM models. J. Pet. Sci. Eng. 200: 108–182.
https://doi.org/10.1016/j.petrol.2020.108182.
Otchere, D.A., Ganat, T.O.A., Ojero, J.O., Taki, M.Y. and Tackie-Otoo, B.N. 2021b. Application
of gradient boosting regression model for the evaluation of feature selection techniques
in improving reservoir characterisation predictions. J. Pet. Sci. Eng. 109244. https://doi.
org/10.1016/J.Petrol.2021.109244.
Otchere, D.A., Hodgetts, D., Ganat, T.A.O., Ullah, N. and Rashid, A. 2021c. Static reservoir
modeling comparing inverse distance weighting to kriging interpolation algorithm in
volumetric estimation. case study: gullfaks field. In: Offshore Technology Conference.
OnePetro, Virtual and Houston, Texas. https://doi.org/10.4043/30919-MS.
Otchere, D.A., Abdalla Ayoub Mohammed, M., Ganat, T.O.A., Gholami, R. and Aljunid
Merican, Z.M. 2022a. A novel empirical and deep ensemble super learning approach
in predicting reservoir wettability via well logs. Applied Sciences 12: 2942. https://doi.
org/10.3390/app12062942.
Otchere, D.A., Ganat, T.O.A., Nta, V., Brantson, E.T. and Sharma, T. 2022b. Data analytics
and Bayesian Optimised Extreme Gradient Boosting approach to estimate cut-offs from
wireline logs for net reservoir and pay classification. Appl. Soft. Comput. 120: 108680.
https://doi.org/10.1016/j.asoc.2022.108680.
Peng, S., Zhang, T., Loucks, R.G. and Shultz, J. 2017. Application of mercury injection capillary
pressure to mudrocks: Conformance and compression corrections. Mar. Pet. Geol.
88: 30–40. https://doi.org/10.1016/J.MARPETGEO.2017.08.006.
Purcell, W.R. 1949. Capillary pressures –their measurement using mercury and the calculation
of permeability therefrom. Journal of Petroleum Technology 1: 39–48. https://doi.
org/10.2118/949039-G.
Rezaee, R. 2022. Synthesizing Nuclear Magnetic Resonance (NMR) outputs for clastic rocks
using machine learning methods, examples from north west shelf and perth basin, Western
Australia. Energies (Basel) 15: 518. https://doi.org/10.3390/en15020518.
Tamoto, H., Gioria, R. dos S. and Carneiro, de C. 2023. Prediction of nuclear magnetic resonance
porosity well-logs in a carbonate reservoir using supervised machine learning models.
J. Pet. Sci. Eng. 220: 111169. https://doi.org/10.1016/j.petrol.2022.111169.
Tarafder, S., Badruddin, N., Yahya, N. and Nasution, A.H. 2022. Drowsiness detection using
ocular indices from EEG signal. Sensors 22: 4764. https://doi.org/10.3390/s22134764.
Wu, B., Xie, R., Wang, X., Wang, T. and Yue, W. 2020. Characterization of pore structure of tight
sandstone reservoirs based on fractal analysis of NMR echo data. J. Nat. Gas Sci. Eng.
81: 103483. https://doi.org/10.1016/j.jngse.2020.103483.
Wu, B., Xie, R., Xu, C., Wei, H., Wang, S. and Liu, J. 2021. A new method for predicting capillary
pressure curves based on NMR echo data: Sandstone as an example. J. Pet. Sci. Eng.
202: 108581. https://doi.org/10.1016/j.petrol.2021.108581.
Xu, H., Tang, D., Zhao, J. and Li, S. 2015. A precise measurement method for shale porosity with
low-field nuclear magnetic resonance: A case study of the Carboniferous–Permian strata in
the Linxing area, eastern Ordos Basin, China. Fuel 143: 47–54. https://doi.org/10.1016/j.
fuel.2014.11.034.
 Chapter 4
Compressional and Shear Sonic
Log Determination
Using Data-Driven Machine
Learning Techniques
Daniel Asante Otchere,1,2,* Raoof Gholami,3 Vanessa Nta 4
and Tarek Omar Arbi Ganat 5

1. Introduction
Shear (Vs) and Compressional (Vp) sonic waves are significant parameters
in subsurface engineering. They provide valuable information for reservoir
exploration, development, recovery and fluid sequestration (Azizia et al.,
2017; Zoveidavianpoor et al., 2013). These parameters or its ratio are useful
in reflection seismology, lithologic identification, and formation evaluation
(Castagna et al., 1985), pore fluid and pore pressure information (Duffaut and
Landrø, 2007; Rojas, 2008), reservoir characterisation (Eberli et al., 2003;
Pickett, 1963), geophysics (Phadke et al., 2000; Waluyo et al., 1995), and
geomechanical properties (Asoodeh and Bagheripour, 2014; Rasouli et al.,
2011). There are several ways of measuring these parameters, as summarised
in Table 4.1.
1
Centre of Research for Subsurface Seismic Imaging, Universiti Teknologi PETRONAS, 32610, Seri
Iskandar, Perak Daril Ridzuan, Malaysia.
2
Institute for Computational and Data Sciences, Pennsylvania State University, University Park, PA,
USA.
3
Department of Energy Resources, University of Stavanger, Kjell Arholms gate 41, Stavanger, 4021,
Norway.
4
Shell Oil Company, 150 N Dairy Ashford Rd, Houston, TX, 77079, United States of America.
5
Department of Petroleum and Chemical Engineering, Sultan Qaboos University, Muscat, Oman.
* Corresponding author: [email protected]
58 Data Science and Machine Learning Applications in Subsurface Engineering

Table 4.1. Common methods of estimating Vs sonic velocity.

Methods Types of Measurement

Laboratory Core measurements
Direct Dipole Sonic Log
Empirical Correlations Castagna, Gassman, Brocher, Carroll and Greenberg–Castagna
correlations
Machine Learning Examples of models

The acquisition of continuous Vs and Vp subsurface data is reliably done

via wireline logs. The Dipole Shear Sonic Log, invented in the 1980s, is the
most commonly used logging tool used to record the waveform values and
process the data to compute Vs and Vp sonic velocity and became ubiquitous
in the 1990s (Close et al., 2009). This measuring tool indicates how fast sound
propagates through subsurface interfaces. Although the dipole sonic log is
available, its expensiveness has rendered the direct measurement of Vs not
worth the cost, as such, Vs is not ubiquitous in all wells. This limitation has
necessitated further research in estimating Vs from other logs, mainly from
the Vp log (Brocher, 2005; Carroll, 1969; Castagna et al., 1985; Dvorkin and
Mavko, 2014; Gassmann, 1951; Greenberg and Castagna, 1992). With the
advancement of fluid sequestration, geothermal energy, and unconventional
resources, Vs and Vp have become more prevalent. However, older wells never
acquired Vs and Vp data before the dipole sonic log. Acquiring the data now
is not feasible. As such, extensive research is required to provide substantial
study in estimating these two logs from other conventional wireline logs.
In this study, estimating Vs and Vp from wireline logs using a multioutput
supervised machine learning approach is investigated for the Volve field
siliciclastic oil-bearing reservoirs. This study is of the utmost significance,
especially in more mature fields, as shear and compressional sonic logs were
never obtained in old wells. This method would result in better accuracy in
predicting compressional and shear sonic logs for hundreds of wells useful for
geomechanics, performing well-to-seismic inversion, and other operations.

2. Literature Review
Prior to the influx of data, empirical and laboratory techniques have been useful
in estimating Vs from Vp. One of the pioneering articles that were published
by (Pickett, 1963) established the concept of using the compressional to shear
wave velocities ratio as a lithology indicator. Figure 4.1, which has been
modified from Pickett’s study, demonstrates the clear distinction in Vp/V that
exists between dolomites, limestones, and clean sandstones. According to
(Castagna et al., 1985), the Vp/Vs ratio tends to change fairly linearly between
the velocity ratios of the end members with increasing composition in binary
 Compressional and Shear Sonic Log Determination 59

Fig. 4.1. Vs and Vp sonic velocities from in-situ sonic and field seismic measurements for mudrocks.
Modified after Castagna et al. (1985).

combinations of quartz and carbonates. The results of this study served as the
foundation of several empirical approaches. This study will discuss two of the
most commonly used techniques for estimating Vs.
Research conducted by (Castagna et al., 1985) concluded that Vs sonic
velocity is directly related to Vp sonic velocity for both water-saturated and
dry siliciclastic sedimentary rocks, as shown in Fig. 4.1. Shales tend to have a
relatively greater Vp/Vs than clean sandstones, given a similar Vp. From their
results, Vp/Vs is almost homogeneous for dry sandstones. Wet sandstones
and mudstones exhibited an indirect monotonic relationship between Vp
and Vp/Vs. The Vs sonic velocities measured in water-saturated sandstone
agreed with those predicted by Gassmann’s equations. This empirical linear
relationship between the Vp and Vs sonic velocity in brine-saturated clastic
silicate rocks is known as the mudrock line and is expressed from in-situ sonic
and field seismic measurements as:
Vp = 1.16Vs + 1.36 (4.1)
This linear equation is partly explained by the location of the clay point
near a line joining the quartz point with the water velocity. The equation was
formulated assuming that Vp and Vs velocities decrease linearly as increasing
clay porosity approaches water. Similarly, when pure clay is mixed with
quartz, velocities increase monotonically as the quartz point is approached.
These constraints correspond with those deduced from Tosaya and Nur (1982)
empirical relations, except for high porosity behaviour.
Greenberg and Castagna (1992) proposed a correlation, the
Greenberg-Castagna formulation, for pure unit completely brine-saturated
60 Data Science and Machine Learning Applications in Subsurface Engineering

mineral rocks based on linear relations of Vp to estimate the Vs sonic velocity

from other petrophysical parameters. To predict Vs sonic velocity of a 100%
saturated rock, it is required to perform fluid replacement calculations by
utilising Gassman relations. In order to mitigate the major limitation of
the Greenberg-Castagna formula, fluid substitution is done when there are
hydrocarbon-bearing rocks. It requires several sets of operations known as
Gassmann fluid substitutions (Gassmann, 1951). They utilised the regression
weights for different lithologies shown in Table 4.2 to satisfy the expression:
2
Vs = ai2 V p + ai1 Vp + ai0 (4.2)
The lines of fit for the lithological regression weights are based on Simm
and Bacon (2014). The method’s fundamental process entails completing
multiple iterations beginning with Vp assumptions for brine until it reaches
a considerable convergence, assuring the validity of the Vs prediction. The
regression coefficients of Vs are presented in Table 4.2.
Techniques based on artificial intelligence (AI) are currently being utilised
on a massive scale in the renewable and oil and gas industry. These techniques
have led to significant developments in their application to improve the
accuracy of their predictions (Otchere et al., 2021b). Numerous research has
used machine learning techniques to predict Vs, highlighting their usefulness
in identifying patterns amongst highly dimensional and complex wireline
data (Anemangely et al., 2019; Bagheripour et al., 2015; Onalo et al., 2018).
Table 4.3 summarises the various machine learning research conducted
to predict Vs and, in some cases, Vp from conventional wireline logs. The
summary will indicate the differences in this current study.
The fundamental differences between the various studies and techniques
used in predicting Vs or Vp were the model type, formation, reservoir fluid
and input variables. Whereas supervised machine learning models are only as
efficient as the information they are trained with when irrelevant data is used
as input or important features is omitted, the model’s performance degrades
(Otchere et al., 2021c). As a result, identifying and selecting features based on
a demonstrated record of their effect on Vs and Vp measurements is required,
invalidating the choice of machine learning models used as relevant. The
main distinction between this study and earlier studies is the multi-output
prediction of Vp and Vs, with Vp not being used as an input for Vs prediction.

Table 4.2. Regression coefficients of Vs presented by Castagna et al. (1993).

Lithology ai2 ai1 ai0 R2

Sandstone 0 0.80416 –0.85588 0.98352
Limestone –0.05508 1.01677 –1.03049 0.99096
Dolomite 0 0.58321 –0.07775 0.87444
Shale 0 0.76969 –0.86735 0.97939
 Table 4.3. Summary of previous research applying machine learning in predicting Vs and Vp.

Author Model Used Lithology Input Variables Output Variable(s) Study Area
Bagheripour et Support Vector Regression Carbonate • Compressional sonic Shear sonic Iran Gas
al. (2015) • Density fields
• Neutron
• True resistivity
• Shallow resistivity
• Photoelectric factor
• Gamma ray
Bukar et al. Exponential Gaussian Process Sandstone • Compressional sonic Shear sonic Australia Gas
(2019) Regression Model • Caliper fields
• Bulk density correction

Compressional and Shear Sonic Log Determination 61

• True resistivity
• Neutron
• Porosity
• Density
• Water saturation
• Gamma ray
Anemangely et Least Square Support Vector Machine-Cuckoo Carbonate • Compressional sonic Shear sonic Iran’s
al. (2019) Optimisation Algorithm • True resistivity Ahvaz and
• Neutron Ab-Teymour
• Density oilfields
• Gamma ray
Nourafkan and Hybrid ant colony–fuzzy inference system Carbonate • Compressional sonic Shear wave velocity Iran
Kadkhodaie- (ACOFIS) • Neutron Cheshmeh–
Ilkhchi (2015) • Density Khosh
• Gamma ray oilfield
• Photoelectric factor
Elkatatny et al. • Adaptive neuro-fuzzy inference system (ANFIS) Not stated • Bulk density • Shear wave Not stated
(2018) • Support vector machine (SVM) • Neutron porosity • Compressional wave
• Artificial Neural Network (ANN) • Gamma ray
Table 4.3 contd. ...
...Table 4.3 contd.

62 Data Science and Machine Learning Applications in Subsurface Engineering

Author Model Used Lithology Input Variables Output Variable(s) Study Area
Gamal et al. • Decision Tree (DT) • Limestone surface drilling parameters • Shear slowness Not stated
(2022) • Random Forest (RF) • Sandstone • Compressional
• Shale slowness
LeCompte et al. Artificial neural network (ANN) Sandstone • Drilling diagnostics • Compressional sonic Gulf of
(2021) (EDR) data • Density Mexico
• Resistivity log
• Gamma ray
• Adaptive neuro-fuzzy inference system (ANFIS) Not stated • Weight on bit (WOB) • Shear wave slowness Not stated
Suleymanov et • Support vector machine (SVM) • Rate of penetration • Compressional wave
al. (2021) (ROP) slowness
• Standpipe pressure
(SPP)
• Torque (T)
• Drill pipe rotation
(RPM)
• Mud flow rate (GPM)
Ali et al. (2021) Deep Neural Network (DNN) • Shale • Sonic log Shear sonic log Lower Goru
• Sandstone • Deep resistivity Formation in
• Neutron the Middle
• Depth Indus Basin,
• Gamma ray Pakistan
Asoodeh and • Artificial neural network Carbonate • Thermal neutron • Compressional wave Asmari
Bagheripour • Fuzzy logic algorithm porosity (NPHI) velocity formation
(2012) • Neuro-fuzzy algorithm • Bulk density (RHOB) • Shear wave velocity
• Electrical resistivity (Rt) • Stoneley wave velocity
• Shale volume (Vsh)
Safaei-Farouji et • Random Forest (RF) Carbonate • Compressional sonic Shear sonic wave velocity Sarvak
al. (2022) • Extra Tree (ET) wave velocity oilfield, Iran
• Gaussian Process Regression (GPR) • Neutron
• Integration of Adaptive Neuro Fuzzy Inference • Density
System (ANFIS) with differential evolution (DE) • Gamma ray
and imperialist competitive algorithm (ICA)
optimisers
Hatampour and Fuzzy Logic Carbonate • Neutron • Compressional wave Iran
Ghiasi-Freez • Density velocity
(2013) • Electrical resistivity • Shear wave velocity
(Rt) • Stoneley wave velocity
Chaikine and Hybrid convolutional-recurrent neural network • Siltstone • Well deviation survey Shear slowness Montney

Compressional and Shear Sonic Log Determination 63

Gates (2020) (cRNN) • Shale • Density Formation
• Compressional in Alberta,
slowness Canada
Zoveidavianpoor Adaptive neuro-fuzzy inference system Carbonate • Neutron Compressional wave Middle
et al. (2013) • Density Eastern
• Gamma ray Oilfield
64 Data Science and Machine Learning Applications in Subsurface Engineering

Data from three wells in the Volve oil field is used in this study. The input
features and their relevance will be evaluated using the laboratory-verified
justification of their influence on Vs and Vp prediction. After establishing
a suitable dataset of relevant input variables, different machine learning
models will be employed to predict Vp and Vs. The best-performing model
based on several statistical metrics will then be optimised using the Bayesian
Optimisation (BO) algorithm.

3. Background of Machine Learning Regression Models

This study builds on previous research in which multiple procedures and
strategies were utilised to forecast Vs or Vp utilising available input information.
Machine learning is becoming increasingly popular as academics move away
from empirical correlations. Individual well log parameters provide crucial
information about Vp and Vs. However, when particular subsets are employed,
the outcomes may be estimated more precisely. In some instances, researchers
use the filter method to select critical features, while others use wrapper
methods, intrinsic techniques, and metaheuristics algorithms to select relevant
input features. This shows that there is no standard set of features for this
prediction. These strategies, however, are based on statistical measurements
rather than domain knowledge. As such, understanding Vp and Vs and how
subsurface parameters influence their measurements is utilised to solve this
prediction problem. The appropriate features selected will subsequently be
used as input features. Several machine learning models based on Decision
Tree learning will be used to predict Vp and Vs.

3.1 Decision Tree Conceptual Overview

The Decision Tree algorithm has been around for decades, but it is currently
known as Classification and Regression Trees (CART). CART, in contrast
to other supervised learning models, can be used to address regression and
classification problems (Tarafder et al., 2022). The aim of utilising a Decision
Tree is to build a training model to estimate the value or class of an output
variable by learning basic decision rules using prior data (Breiman, 2001). The
CART model is the basis for major machine learning models like the Random
Forests, bagged decision trees, and boosted decision trees. The CART model
is a binary tree similar to algorithms and data structures, with no, single or
two child nodes for each node. Assuming the parameter is numeric, a node
defines a specific input parameter (X) and a split point on that parameter. The
tree’s terminal (leaf) nodes include a target parameter (y) utilised to make a
decision. Upon building the tree, the tree can be traversed by inserting a new
row of data after each branch with splits until the final prediction. Making
binary decision trees is a technique for splitting up the input data. A greedy
 Compressional and Shear Sonic Log Determination 65

technique known as recursive binary splitting is utilised for the data partition.
Recursive binary splitting is a statistical approach in which all values are
aligned, and several strategic split points are explored and assessed using an
objective function. The split that results in the lowest cost is thence chosen.
Based on the objective function, all input parameters and feasible splits are
assessed and selected in a greedy approach.
The Disjunctive Normal Form concept, also known as the Sum of
Products (SOP), is the foundation of the CART model. Various branches that
end in the same class combine to form a sum for each branch that extends
from the tree’s root to a leaf node of that class. In implementing the decision
tree, determining which qualities should be regarded as the root node at each
level is the primary issue. Addressing this is referred to as attribute selection.
Different attribute selection techniques determine the attribute designated as
the root note at each level.
The preference to do judicious splits has a major impact on tree’s
reliability. There are specific selection principles for CART models. To
determine whether a node should be split into multiple sub-nodes, different
algorithms are used by CART models. The formation of sub-nodes improves
the uniformity of newly emerging sub-nodes (Gordon et al., 1984). The
decision to do strategic splits heavily affects a tree’s accuracy, meaning that
the integrity of the node grows with the target variable. The CART model
divides the nodes based on all available parameters and then chooses the split
that produces the most homogenous subnodes. The algorithm selection is also
chosen depending on the type of target data.

3.1.1 Attribute Selection Measures

Considering a dataset has N parameters, selecting which ones to put at the
root or various levels of the tree as internal nodes is a complicated process.
Choosing any node at random to be the root will not address the problem.
Choosing a random approach may result in bad outcomes with low precision.
Researchers worked on this attribute selection issue and proposed several
approaches utilising parameters such as (Breiman, 2001):
1. Entropy (E): The measure of how random information is being processed
is called entropy. High entropy makes it difficult for inferences to be made
from the data. A branch with zero entropy is a leaf node, while a branch
with a zero entropy needs to be split further. Entropy for a parameter is
expressed mathematically as:
n
E( S=
) ∑ − p log
i=1
i 2 pi (4.3)
66 Data Science and Machine Learning Applications in Subsurface Engineering

where S denotes the present state, and Pi is the probability of an event

i occurring in a node of state S. Entropy for multiple parameters is
expressed mathematically as:
E( S , Y ) = ∑ P(n) E (n) (4.4)
n∈Y

where Y represents the selected parameter.

2. Information Gain (IG): IG is a numerical attribute that quantifies how
efficiently a specific attribute distinguishes training samples according
to their target. Building a decision tree is to identify a characteristic with
the most considerable information gain and the lowest entropy. Entropy
decreases as a result of information gain. Based on the specific attribute
values provided, it calculates the difference between the data’s entropy
before splitting and the average entropy after splitting. IG is denoted
mathematically as:
IG(S,Y) = E(S) – E(S,Y) (4.5)
3. Gini Index (GI): The Gini index may be considered an objective function
to assess splits in a dataset. It is computed by subtracting the total of
each class’s squared probability from one. GI prefers large divisions
that are simple to implement, whereas information gain prefers fewer
divisions with unique values. The Gini Index only performs binary
splits by employing the binary target variable ‘Success’ or ‘Failure’. A
high GI value indicates greater inequality and heterogeneity. GI can be
mathematically denoted as:
n
GI =1 − ∑ − ( pi )2 (4.6)
i=1

4. Gain Ratio: IG is biased towards choosing attributes with many values as

root nodes. It means it prefers the attribute with a large number of distinct
values. Gain ratio, a variation of IG that minimises bias, is typically the
best option compared to IG. The gain ratio solves the issues related to IG
by considering the number of branches that might occur before splitting.
It adjusts IG by considering the inherent information of a split. The gain
ratio is expressed as:
E (S ) − E (S , Y )
GI = K (4.7)
∑ w log w
j =0 j 2 j

where K refers to the number of subsets produced by the split, and j,

refers to a subset.
 Compressional and Shear Sonic Log Determination 67

5. Reduction in Variance (RIV): RIV is a technique used for regression

problems. This method uses the traditional variance formula to select the
optimum split. The split achieving the lowest variance is chosen as the
criterion for splitting the data:

Variance =
∑(X − X ) 2
(4.8)
n
–
The mean of the values is shown as X. X is the actual value, and n is the
number of values.
6. Chi-Squared Automatic Interaction Detector (CHAID): CHAID is a
tree classification approach that determines the statistically significant
differences between sub-nodes and the parent node. The sum of squares
of the standardised discrepancies between the observed and predicted
frequencies of the target is used to calculate it. CHAID employs the binary
target variable ‘Success’ or ‘Failure’ by performing two or more splits.
The greater the Chi-Square value, the greater the statistical significance
of discrepancies between the sub-node and Parent node. Chi-squared is
denoted mathematically as:
(O − E) 2
x2 = ∑ (4.9)
E
where x2 represents the obtained Chi-Square, O is the observed score, and
E denotes the expected score.
These criteria will be used to calculate each attribute selection value.
Each value is ranked, while the characteristics are orderly put in the tree. The
attribute with the highest value regarding knowledge gain is at the root. The
categorical output is assumed when utilising IG as a criterion, whereas GI
assumes continuous attributes. CART models are simple to understand since
they result in precepts, but one main disadvantage is the practical possibility of
overfitting. Overfitting generally happens when CART builds many branches
due to outliers and irregularities in data, but it can be minimised using the
pre- and post-pruning approaches. Pre-pruning terminates tree construction
early, and it is preferable not to split a node if its measure of purity falls
below a certain threshold. However, deciding on an acceptable stopping point
is challenging. Post-pruning begins by going deeper into the tree to construct
an entire tree. If the tree exhibits overfitting, pruning is performed as a
post-pruning phase. Cross-validation is employed to determine if extending
a node would improve or not to evaluate the efficacy of pruning. If there is
an improvement, extending that node can be done. However, if it indicates
a degraded performance, the node should not be extended and changed to a
terminal node (Breiman, 2001; Gordon et al., 1984).
68 Data Science and Machine Learning Applications in Subsurface Engineering

3.2 Random Forest Conceptual Overview

Another well-known robust bagging algorithm is the Random Forest (RF)
developed by Breiman (Breiman, 2001). RF ensembles numerous Decision
Tree base learners randomly using averaging (Dinov, 2018). Each base
learner is trained on a random vector independently chosen with similar
vector distribution to the other forest learners. The wisdom of crowds is a
simple yet powerful notion at the core of RF. This concept means that many
relatively uncorrelated learners working together surpass any individual base
learners. The crucial point is the poor correlation amongst models, which
allows the excellent performance of RF. The trees shield each other from their
prediction errors, so RF generates excellent outcomes. While some trees may
be inaccurate, many others will be accurate, allowing the trees to grow in the
appropriate direction as a forest. Decision trees are susceptible to the dataset
on which they are trained such that minor changes to the training data typically
result in a radically altered tree structure. This disadvantage is solved by the
random forest using the bagging principle, which allows each unique tree to be
randomly selected from the dataset with replacement, resulting in diverse trees.
Any given tree’s prediction becomes, in a sense, a characteristic of the
overall model prediction. Due to its robustness against overfitting and simple
interpretability, RF is frequently employed to provide good model results.
This merit results from its assessment of the target’s critical input factors
and the high priority given to correlated aspects. However, RF has a greater
bias against high cardinal characteristics. The averaging of the base learners’
predictions is mathematically represented as:
1 n
ŷ = ∑ hk ( X )
n k =1
(4.10)

where ŷ is the output, h(X) is the number of trees, X is the input vector, and n
and k are the overall numbers of trees grown where n is greater than k and k
is greater than 1.
The out-of-bag (OOB) score is another essential aspect of the RF algorithm.
Specific samples will be excluded from the subsamples used to train the base
learners when bootstrapping. These out-of-sample examples may be used to
assess the learner and generate an OOB score, acting as a pseudo-validation
subset for the random forest model. Set oob score = True when initialising
the random forest object to acquire the OOB score. Another critical aspect of
the RF algorithm is its feature engineering capability entailing selecting the
most relevant features from the input variables in the training dataset. Feature
selection is an essential part of the machine learning workflow.
The primary disadvantage of random forest is that many trees may slow
down and render the model inefficient for predictions. In general, the model
 Compressional and Shear Sonic Log Determination 69

tends to learn rapidly but makes slow predictions. A more precise prediction
necessitates more trees, resulting in a computationally exhaustive model.

3.3 Extremely Randomised Trees Conceptual Overview

Extremely Randomised Trees (Extra-Trees) is an ensemble learning technique
proposed by Geurts et al. (Geurts et al., 2006). Tree-based regressors are
composed of a hierarchy set of rules that may predict the numerical values of
the output (Gordon et al., 1984). The Extra-Trees technique uses the traditional
top-down strategy to construct an ensemble of unpruned decision trees. The
two key distinctions from previous tree-based ensemble algorithms are that it
separates nodes by selecting cut-points entirely at random. It grows the trees
using the entire training sample instead of replicating a bootstrap replica.
Extra-Trees appropriately average the randomised predictions of Decision Trees
to enhance precision while minimising computational complexity significantly.
The Extra-Trees approach is built on the bias-variance concept. The
explicit randomisation of the cut-point and features combined with ensemble
averaging can reduce variance more strongly than the weaker randomisation
techniques employed by other algorithms. The entire original training data
should be used to reduce bias rather than bootstrap replicas. The computational
cost of the tree growing method, assuming balanced trees, is on the order of
N log N regarding the learning sample size, as is the case with most other
tree-growing procedures. Moreover, given the ease of the node splitting
technique, it is anticipated that the constant factor will be considerably lower
than in previous ensemble-based methods that optimise cut points locally
(Geurts et al., 2006).
Constructing the nodes and branches that make up a tree depends on
dividing the input vector into independently exclusive areas based on a
predetermined splitting criterion, gradually reducing the size of the regions.
When the number of samples in a tree section goes below a predefined
threshold, the splitting terminates to form a leaf. When inputting a new sample
into the tree, a specific path is taken based on the splitting criteria provided
in the tree-building process. The predicted target is generated by aggregating
the values stored in the leaf. The requirements for splitting, termination into
leaves, tree quantity developed, and assigning weight to each leaf are the
critical aspects that distinguish Extra-Trees from other tree-based techniques.
Aside from having similar advantages as single trees based on
consistency for universal generalisation and approximation, Extra-Trees also
give resilience in gross model errors because outliers affect its predictions
slightly and locally. Extra-Trees potentially outperform single trees in relation
to computational efficiency while resisting irrelevant input features when
the input features are significantly more than the random splits. Further
comparison with single trees and dependent on the issue at hand indicate that
70 Data Science and Machine Learning Applications in Subsurface Engineering

Extra-Trees exhibit low variance, and their prediction accuracy improves with
an increasing number of trees (Ernst et al., 2006).

4. Data Collection and Description

The Volve field, located in the Norwegian part of the North Sea, is a sandstone
oil-bearing reservoir interval. The Volve field formation predominantly
consists of near-shore, shallow, to marginal marine deposits. As a result
of the extensive core sampling, it was discovered that six distinct facies
associations represent the Heather Formation; coastal plain, shoreface,
fluvio-tidal channel-fill, bay-fill, mouth bar, and offshore open marine (Folkestad
and Satur, 2008; Otchere et al., 2022b, 2021a; Sneider et al., 1995). The data
used has 12 dimensions compiled from three wells, Well 15_9-F-1_A (Well
A), 15_9-F-1_B (Well B) and 15_9-F-15_D (Well C). However, well A and B
were combined, making up 4,000 data points (2,125 for Well A and 1,875 for
Well B), to train and test the model. A wireline plot of Wells A and B showing
the shear (DTS) and compressional (DT) logs is illustrated in Figs. 4.2 and 4.3.
The model was then deployed to estimate DTS and DT for Well C, which had a
total data point of 11,935 without Vp and Vs logs. One of the main advantages
of applying data analytics to data is to find patterns and hidden information in
high dimensional data. As such, the primary input features for this research are
summarised and justified in Table 4.4. Table 4.5 shows the descriptive statistics
of the features for the combined Well A and Well B.

Table 4.4. Justification for log selection as input variables.

Log Purpose Reason

Gamma ray (GR) Lithology determinant Lithogies exhibit different Vs and Vp wave
velocities (Otchere et al., 2021c).
Neutron (NPHI) Lithology and porosity Lithogies and porosity influences Vs and
determinant Vp differently (Otchere et al., 2021c).
Density (RHOB) Lithology and porosity Lithogies and porosity influences Vs and
determinant Vp differently (Otchere et al., 2021c).
Photoelectric factor Determination of minerals The response of Vs and Vp varies with
(PEF) for lithology interpretation different mineral types or compositions
(Castagna et al., 1985).
Porosity (PHIF) Lithology determinant Vs velocity is more sensitive than Vp to
porosity (Hamada and Joseph, 2020).
Water saturation (SW) Fluid determinant Different formation fluids affect Vp and Vs
responses differently (Castagna et al., 1985).
Resistivity [true (RT), Lithology and hydrocarbon Different fluids have different effects on Vs
shallow (RACELM), detection logs and Vp (Otchere et al., 2021c).
medium (RACEHM),
and deep (RPCELM)]
Caliper (CALI) Detection of permeable Different permeabilities affect Vs and Vp
zones hence relates to differently (Hamada and Joseph, 2020).
formation types
Fig. 4.2. Wireline plot of Well A showing the input logs and the shear (DTS) and compressional (DT) logs.
Compressional and Shear Sonic Log Determination 71
 72 Data Science and Machine Learning Applications in Subsurface Engineering

Fig. 4.3. Wireline plot of Well B showing the input logs and the shear (DTS) and compressional (DT) logs.
 Compressional and Shear Sonic Log Determination 73

Table 4.5. Descriptive statistics of the outputs and input variables for the combined Wells A and B.

CALI DT DTS GR NPHI PEF RACEHM RACELM RHOB RPCELM RT PHIF

count 4000 4000 4000 4000 4000 4000 4000 4000 4000 4000 4000 4000
mean 8.66 75.98 129.96 48.75 0.16 6.07 4.25 3.40 2.47 3.02 3.96 0.11
std 0.04 6.80 13.56 20.18 0.05 0.75 87.41 50.11 0.13 7.85 11.99 0.06
min 8.47 58.63 96.90 8.00 0.05 4.30 0.20 0.23 2.15 0.14 0.19 0.02
25% 8.63 71.68 122.19 36.88 0.13 5.62 1.04 1.03 2.40 1.06 1.08 0.06
50% 8.67 75.14 130.42 46.23 0.16 6.05 1.44 1.38 2.48 1.50 1.58 0.10
75% 8.70 80.93 136.00 56.55 0.19 6.37 2.53 2.42 2.56 2.64 3.03 0.15
max 8.87 96.22 186.09 127.06 0.41 10.99 5464.37 2189.60 3.05 96.43 134.70 0.27

5. Methodology
5.1 Data Analysis and Visualisation
The data for this study were analysed and visualised to help understand
how the input variables correlate to both targets. The Kendall correlation
covariance heatmap was used to calculate the degree of correlation between
the inputs and the targets. The NPHI, RHOB, and PHIF were the parameters
that strongly correlated to both targets. The nonlinear distribution between the
input and the targets is also illustrated in the pair plot in Fig. 4.4. Based on
the nonlinear distribution, it is clear that none of the input parameters had a
linear relationship to both targets, hence advising on the selection of nonlinear
models for this study.

5.2 Machine Learning Model Application

As part of the machine learning process, the models are trained on data
which were randomly sampled from the total dataset using the holdout
cross-validation technique. This technique is used to assess the performance
of models and prevents selection bias and overfitting by predicting the targets
using unseen data (Otchere et al., 2022a; Tarafder et al., 2021). The holdout
approach divided the data into a 90:10 split, making up 3,600 training and
400 testing data points. The selection of the split is to ensure that the majority
of the data are used to train the model due to the heterogeneity of the subsurface
but to ensure a suitable number unknown to the models is kept for testing.
Different multi-output models were utilised in this study to determine which
model accurately predicts the Vs and Vp accurately. This study utilised the
same out-of-sample data to guarantee that all the models were predicted on
equivalent data points.
The performance of the models was assessed, in order of ranking, using
the mean absolute error (MAE), root mean squared error (RMSE), mean
 74 Data Science and Machine Learning Applications in Subsurface Engineering

Fig. 4.4. Pair plot distribution of all input features colour-coded using sand flag.
 Compressional and Shear Sonic Log Determination 75

absolute per cent error (MAPE) and R2. The models were also compared
using the Akaike Information Criterion (AIC), a frequentist probability
framework that scores models based on the maximum likelihood of fitting
unknown data. Upon selecting the best model based on these criteria (BO),
hyperparameter tuning, a sequential model-based optimisation procedure,
was used to enhance the model’s prediction accuracy. The BO technique is
a convenient and dynamic paradigm that uses the Bayes Theorem to give a
reasonable mechanism for guiding the search of a global optimisation problem
to the extrema of objective functions (Otchere et al., 2022b). The technique
operates by training a Bayesian approximation probabilistic model of the
objective function repeatedly based on previous result estimation. Before
choosing acceptable samples for evaluation on the real objective function, the
probabilistic model is evaluated using an acquisition function. The process
of adjusting the model’s parameters to improve the learning algorithm’s
efficiency and optimise model performance by reaching a sufficiently reduced
cost function is known as hyperparameter tuning (Otchere et al., 2022b).
The optimised model is then deployed on a new well from the same field to
predict Vs and Vp from the wireline logs. The predicted Vs and Vp log will
be compared to that obtained from empirical correlations. Figure 4.5 shows
the research workflow.

Fig. 4.5. Schematic workflow of the methods used in this study.

6. Results and Discussion

6.1 Evaluation of Model Performance
The models employed in this study were assessed using AIC to identify
the model with the highest likelihood of fitting out-of-sample data. The
76 Data Science and Machine Learning Applications in Subsurface Engineering

procedure in using this metric to evaluate the models suggests that the lower
the AIC values, the greater the possibility of fitting new data. The following
considerations were taken for this study when comparing and selecting the best
model. There are four levels of significance when it comes to differences in
AIC results between models: negligible (< 20), moderate (21–50), substantial
(51–100) and extremely strong (> 100). The Extra-Trees model resulted in the
highest probability of fit for Vp and Vs predictions by recording the lowest
AIC value. When the Extra-Trees model was compared to the Random Forest
model, representing the two models with the lowest AIC, there was a difference
of 180 AIC. The result is illustrated in Fig. 4.6. Based on this outcome, the
Extra-Trees model showed extreme differences and an enhanced chance of
fit. AIC determines the model’s convergence to actual data. However, over- or
underfitting may persist. As such, further model assessment is required using
other error metrics.
Out-of-sample data results were used to evaluate the models’
performance. Since the models were trained using the training dataset, any
attempt to duplicate the data will be highly accurate. As a result, the models
are expected to have a high training accuracy score. However, a high level of
training accuracy might lead to an unrepresentative model due to overfitting
the data by matching inherent noise in the data. The proportion of accurate
predictions generated by the models on the test data is used to assess the model
performance on data that has not yet been observed. A high test accuracy is
highly desired in the prediction of Vp and Vs to confirm model accuracy and
robustness on new data since inaccurate measurements will have long-term
defects in many operations. In this study, R2 greater than 0.9 was regarded
as an acceptable compromise between bias and variance. As can be seen in
Fig. 4.7, the Extra-Trees model performed the best in terms of estimating Vp
and Vs during testing.

Fig. 4.6. AIC results comparing all the models on the test data.
 Compressional and Shear Sonic Log Determination 77

Fig. 4.7. Train and test correlation coefficient score of all models used in this study.

Fig. 4.8. Comparison of test data prediction errors of all models based on MAE, RMSE, and MAPE.

Any noteworthy improvement in accuracy attained by models significantly

impacts decision-making. Although the models are from the same family,
their theoretical and statistical foundations will give a wide range of model
performances. Figure 4.8 displays the performance of the models on the test
data to assess their precision, consistency, and robustness. The Extra-Trees
model showed the most consistency, accuracy, and robustness when assessed
using MAE, MAPE, and RMSE. The Extra-Trees thus surpassed the other
models based on all the evaluation metrics.
Figure 4.9 shows the joint plot and regression combination used to
visualise the predicted Vp and Vs test data and actual data and its error margins.
The joint plot displays a univariate histogram of KDE curves and a bivariate
cross-plot. Analysis of the regression plot and 95% confidence interval
78 Data Science and Machine Learning Applications in Subsurface Engineering

Fig. 4.9. A joint plot of Predicted vs Actual Vs (left) and Vp (right) on test data showing cross plot,
data distribution, and confidence interval. (A) is Decision Tree, (B) is Random Forest and (C) is
Extra-Trees.
 Compressional and Shear Sonic Log Determination 79

illustrated in these figures indicates that the Extra-Trees model outperformed

the other models. The Extra-Trees model showed a higher confidence interval
and matched the range of values based on the histogram and KDE plot.
Although all the models showed higher confidence in predicting Vp, the
confidence level tends to deteriorate as high Vs values are predicted for the
Decision Tree (above 140 μs/ft) and Random Forest (above 160 μs/ft) models.
The two key reasons that make Extra-Trees superior to Random Forests
and Decision Trees are the combination of the separation of nodes by random
selection of cut-points and the use of the entire training sample to reduce bias.
Extra-Trees appropriately average the randomised predictions of Decision
Trees to enhance precision while minimising computational complexity
significantly. The Extra-Trees approach is built on the bias-variance concept.
In addition to sharing similar advantages with the other models based on
consistency for universal generalisation and approximation, Extra-Trees
provide robustness in gross model errors since outliers have a small and local
impact on its predictions. Extra-Trees may surpass Decision Trees in terms
of computational efficiency by remaining impervious to irrelevant input
features and when the input features exceed the random splits. The Decision
Trees model exhibited a robust overfitting with 100% training accuracy but
a relatively low out-of-sample accuracy compared to the Extra-Trees model.
The Extra-Trees is then selected as the model with the most robust fit to the
test data based on all the metrics.

6.2 Model Optimisation

The BO technique was employed to adjust its parameters to increase the
Extra-Trees model’s performance. Although the increase in accuracy was
minimal, in subsurface engineering, any minimal increase in accuracy
significantly impacts decision-making and can have a long-term effect on
field development. Table 4.6 shows a comparison of all models against the
BO-Extra-Trees (BO–ET) model. The predicted Vp and Vs were further
evaluated based on formation type.
The results were further plotted in Fig. 4.10 and classified into rock
types to evaluate the model’s performance. The figure shows that the
BO–ET model exhibited higher confidence in predicting Vs in both rock
types. However, it was observed that most of the error is from the prediction
of Vp and Vs in the sandstone interval. It is known that the sandstone interval
is the hydrocarbon-bearing zone. As such, the effect of different fluids will
influence the measurement of sonic velocities. Figure 4.11 shows a plot of the
BO–ET predicted Vs and Vp compared to the empirical formulas Castagna
and Greenberg-Castagna. The two empirical correlations only predict Vs,
while the BO–ET predicts both Vs and Vp. The data were sorted based on
80 Data Science and Machine Learning Applications in Subsurface Engineering

Table 4.6. Performance of all models compared to the Bayesian Optimised Extra-Trees.

R2 Decision Tree ||||||||

Random Forest |||||||||
Extra-Trees |||||||||
BO-Extra-Trees |||||||||
MAE Decision Tree ||||||||||
Random Forest ||||||||
Extra-Trees ||||||
BO-Extra-Trees ||||||
RMSE Decision Tree ||||||||
Random Forest |||||
Extra-Trees ||||
BO-Extra-Trees |||
MAPE Decision Tree ||||
Random Forest |||
Extra-Trees ||
BO-Extra-Trees |

Fig. 4.10. Cross-plot of test and predicted data indicating confidence interval for BO–ET model based
on formation type.

increasing saturation. By comparison, the BO–ET exhibited a better match

done with the empirical correlations due to its data-driven approach. It was
observed that at the zone of higher water saturation (mostly shales), the
predicted Vp and Vs matched the actual better than the low water saturation
zone (primarily sandstones). This observation confirms that different types
 Compressional and Shear Sonic Log Determination 81

Fig. 4.11. Wireline log comparison of empirical and BO–ET predicted values against actual.

of fluids influence the measurements on sonic velocities, contributing to

the model’s error. The influence of the fluids can be adequately modelled to
improve model confidence by retraining the model with additional data from
the hydrocarbon-bearing interval.
The methodologies evaluated in this research demonstrated that a
data-driven model is essential for Vs and Vp prediction. The shear and
compressive sonic velocities are affected by various subsurface conditions.
Wireline logs carry vital information about the subsurface. Hence, it is suitable
to be used to measure Vs and Vp. The reason for most of the errors occurring in
the sandstone interval is attributed to the effect of water saturation. There are
significant variations in Vs and Vp as water saturation increases. When there
is a high water saturation, grains tend to shear against one another. Elastic
features can be seen in subsurface formations where the pore spaces are not
completely saturated, allowing them to remain compressible. However, when
more fluid percolates into a porous medium, the rock increases in rigidity,
whiles elasticity reduces as the entire rock is occupied. The type of fluid or
fluids coexisting in the pore spaces also influences rigidity and elasticity. The
superior performance of the BO–ET model over the empirical correlations
for Vs prediction is because the model is data-driven and considers both dry
and saturated conditions and even the type of fluid saturation. There is an
apparent interaction between fluids and the rock texture that influences Vs
and Vp measurements.
82 Data Science and Machine Learning Applications in Subsurface Engineering

Fig. 4.12. Wireline log illustration of the BO–ET predicted Vp and Vs for the deployment Well C.

6.3 Model Deployment

After the successful evaluation of the BO–ET model, the model was retrained
with all data from Well A and Well B. The retraining was necessary to
ensure that the model learns from all the data sets. The BO–ET model was
successfully deployed on Well C to predict Vs and Vp from the wireline logs.
Figure 4.12 shows the wireline plot of the predicted Vs and Vp logs where
the top reservoir is indicated at a depth of 3,570.6 m. The reservoir has three
different formation water salinity in the three different formations. Based on
observations made in the hydrocarbon-bearing zone, the predicted log within
that interval was of high interest. The sensitivity towards fluid presence was
adequately predicted for all three formations confirming the robustness of
the model. This result demonstrates the appropriateness of the model for
predicting Vs and Vp in similar formation types.

7. Conclusions
This study demonstrated the use of data-driven domain knowledge and an
optimised Extra-Trees model for predicting shear and compressional sonic
logs in wells that they are absent. This approach has been found to outperform
empirical correlations that rely on the compressional sonic log to predict the
 Compressional and Shear Sonic Log Determination 83

shear sonic log in terms of accuracy. The assessment criterion indicated the
models’ likelihood of fitting held-out data. The models’ accuracy, precision,
and dependability were further evaluated based on the MAE, MAPE, RMSE,
and R2 on the held-out data. The findings are outlined as follows:
1. The input data were clearly justified to understand their influence on
sonic velocity.
2. The Extra-Trees model outperformed the Random Forest and Decision
Tree models in terms of precision, robustness, and probability of fit.
3. The BO algorithm made a marginal but meaningful improvement in the
accuracy of the Extra-Trees model.
4. In the multi-output regression prediction, the predicted Vs in the sandstone
interval exhibited the highest error on the test data.
5. The error in the sandstone interval is due to the influence of different
formation fluids in the reservoir zone. This error was minimised by
retraining the model with all the data from Wells A and B.
6. After deploying the model on Well C, the predicted Vs and Vp matched
the influence of different fluid saturations in the three different
formation types.
The proposed methodology offers the geothermal, fluid sequestration,
and oil and gas industry a cheaper and time-efficient alternative to running the
sonic logging tool in every drilled well. It also provides a way to measure Vs
and Vp in older wells that were not previously measured.

Acknowledgement
The authors express their sincere appreciation to University Teknologi Petronas
and the Centre for Subsurface Seismic Imaging for supporting this work.

Data Availability
The wireline log data used was obtained from Equinor Volve Field Datasets at
https://www.equinor.com/en/how-and-why/digitalisation-in-our-dna/volve-
field-data-village-download.html ([Dataset], 2018).

References
Ali, M., Jiang, R., Ma, H., Pan, H., Abbas, K., Ashraf, U. and Ullah, J. 2021. Machine learning: A
novel approach of well logs similarity based on synchronization measures to predict shear
sonic logs. J. Pet. Sci. Eng. 203: 108602. https://doi.org/10.1016/j.petrol.2021.108602.
Anemangely, M., Ramezanzadeh, A., Amiri, H. and Hoseinpour, S.A. 2019. Machine learning
technique for the prediction of shear wave velocity using petrophysical logs. J. Pet. Sci.
Eng. 174: 306–327. https://doi.org/10.1016/j.petrol.2018.11.032.
84 Data Science and Machine Learning Applications in Subsurface Engineering

Asoodeh, M. and Bagheripour, P. 2012. Prediction of compressional, shear, and stoneley wave
velocities from conventional well log data using a committee machine with intelligent
systems. Rock Mech. Rock Eng. 45: 45–63. https://doi.org/10.1007/s00603-011-0181-2.
Asoodeh, M. and Bagheripour, P. 2014. ACE stimulated neural network for shear wave velocity
determination from well logs. J. Appl. Geophy. 107: 102–107. https://doi.org/10.1016/j.
jappgeo.2014.05.014.
Azizia, H., Siahkoohi, H.R., Evans, B., Farajkhah, N.K. and Kazemzadeh, E. 2017. A comparison
between estimated shear wave velocity and elastic modulus by empirical equations and that
of laboratory measurements at reservoir pressure condition. Journal of Sustainable Energy
Engineering 5: 29–46. https://doi.org/10.7569/JSEE.2017.629502.
Bagheripour, P., Gholami, A., Asoodeh, M. and Vaezzadeh-Asadi, M. 2015. Support vector
regression based determination of shear wave velocity. J. Pet. Sci. Eng. 125: 95–99.
https://doi.org/10.1016/j.petrol.2014.11.025.
Breiman, L. 2001. Random forests. Mach. Learn. 45: 5–32. https://doi.
org/10.1023/A:1010933404324.
Brocher, T.M. 2005. Empirical relations between elastic wavespeeds and density in the earth’s
crust. Bulletin of the Seismological Society of America 95: 2081–2092. https://doi.
org/10.1785/0120050077.
Bukar, I., Adamu, M.B. and Hassan, U. 2019. A machine learning approach to shear sonic
log prediction. In: Society of Petroleum Engineers–SPE Nigeria Annual International
Conference and Exhibition 2019, NAIC 2019. Society of Petroleum Engineers, Lagos.
https://doi.org/10.2118/198764-MS.
Carroll, R.D. 1969. The determination of the acoustic parameters of volcanic rocks from
compressional velocity measurements. International Journal of Rock Mechanics and
Mining Sciences & Geomechanics Abstracts 6: 557–579. https://doi.org/10.1016/0148-
9062(69)90022-9.
Castagna, J.P., Batzle, M.L. and Eastwood, R.L. 1985. Relationships between Compressional‐
wave and Shear‐wave Velocities in Clastic Silicate Rocks. pdf Vol. 50: 571–581.
Castagna, J.P., Batzle, M.L., Kan, T.K. and Backus, M.M. 1993. Rock physics—The link
between rock properties and AVO response. Offset-dependent reflectivity—Theory and
practice of AVO analysis. SEG 8: 135–171.
Chaikine, I.A. and Gates, I.D. 2020. A new machine learning procedure to generate highly
accurate synthetic shear sonic logs in unconventional reservoirs. In: SPE Annual Technical
Conference and Exhibition. SPE. https://doi.org/10.2118/201453-MS.
Close, D., Cho, D., Horn, F. and Edmundson, H. 2009. The sound of sonic: A historical
perspective and introduction to acoustic logging. Canadian Society of Exploration
Geophysicists Recorder 34: 34–43.
[Dataset]. 2018. Volve field data village download - data 2008-2016 - equinor.com [WWW
Document]. URL https://www.equinor.com/en/what-we-do/digitalisation-in-our-dna/
volve-field-data-village-download.html (accessed 5.27.21).
Dinov, I.D. 2018. Data Science and Predictive Analytics. Springer International Publishing,
Cham. https://doi.org/10.1007/978-3-319-72347-1.
Duffaut, K. and Landrø, M. 2007. Vp/Vs ratio versus differential stress and rock consolidation—A
comparison between rock models and time-lapse AVO data. Geophysics 72: C81–C94.
https://doi.org/10.1190/1.2752175.
Dvorkin, J. and Mavko, G. 2014. V S predictors revisited. The Leading Edge 33: 288–296.
https://doi.org/10.1190/tle33030288.1.
Eberli, G.P., Baechle, G.T., Anselmetti, F.S. and Incze, M.L. 2003. Factors controlling elastic
properties in carbonate sediments and rocks. The Leading Edge 22: 654–660. https://doi.
org/10.1190/1.1599691.
 Compressional and Shear Sonic Log Determination 85

Elkatatny, S., Tariq, Z., Mahmoud, M., Mohamed, I. and Abdulraheem, A. 2018. Development
of new mathematical model for compressional and shear sonic times from wireline log data
using artificial intelligence neural networks (White Box). Arab J. Sci. Eng. 43: 6375–6389.
https://doi.org/10.1007/s13369-018-3094-5.
Ernst, P., Wehenkel, D., Govindaraju, L. and Rao, R.S. 2006. Extremely randomized trees,
Artificial Neural Network in Hydrology. Kluwer.
Folkestad, A. and Satur, N. 2008. Regressive and transgressive cycles in a rift-basin:
Depositional model and sedimentary partitioning of the Middle Jurassic Hugin Formation,
Southern Viking Graben, North Sea. Sediment Geol. 207: 1–21. https://doi.org/10.1016/j.
sedgeo.2008.03.006.
Gamal, H., Alsaihati, A. and Elkatatny, S. 2022. Predicting the rock sonic logs while drilling
by random forest and decision tree-based algorithms. J. Energy Resour. Technol. 144.
https://doi.org/10.1115/1.4051670.
Gassmann, F. 1951. Elastic waves through a packing of spheres. Geophysics 16: 673–685.
https://doi.org/10.1190/1.1437718.
Geurts, P., Ernst, D. and Wehenkel, L. 2006. Extremely randomized trees. Machine Learning
63(1): 3–42. https://doi.org/10.1007/s10994-006-6226-1.
Gordon, A.D., Breiman, L., Friedman, J.H., Olshen, R.A. and Stone, C.J. 1984. Classification
and regression trees. Biometrics 40: 874. https://doi.org/10.2307/2530946.
Greenberg, M.L. and Castagna, J.P. 1992. Shear-wave velocity estimation in porous rocks:
theoretical formulation, preliminary verification and applications. Geophys. Prospect.
40: 195–209. https://doi.org/10.1111/j.1365-2478.1992.tb00371.x.
Hamada, G. and Joseph, V. 2020. Developed correlations between sound wave velocity and
porosity, permeability and mechanical properties of sandstone core samples. Petroleum
Research 5: 326–338. https://doi.org/10.1016/j.ptlrs.2020.07.001.
Hatampour, A. and Ghiasi-Freez, J. 2013. A fuzzy logic model for predicting dipole shear
sonic imager parameters from conventional well logs. Pet. Sci. Technol. 31: 2557–2568.
https://doi.org/10.1080/10916466.2011.603005.
LeCompte, B., Majekodunmi, T., Staines, M., Taylor, G., Zhang, B., Evans, R. and Chang, N.
2021. Machine learning prediction of formation evaluation logs in the Gulf of Mexico.
In: Offshore Technology Conference. OTC. https://doi.org/10.4043/31093-MS.
Nourafkan, A. and Kadkhodaie-Ilkhchi, A. 2015. Shear wave velocity estimation from
conventional well log data by using a hybrid ant colony–fuzzy inference system: A
case study from Cheshmeh–Khosh oilfield. J. Pet. Sci, Eng. 127: 459–468. https://doi.
org/10.1016/j.petrol.2015.02.001.
Onalo, D., Adedigba, S., Khan, F., James, L.A. and Butt, S. 2018. Data driven model for
sonic well log prediction. J. Pet. Sci. Eng. 170: 1022–1037. https://doi.org/10.1016/j.
petrol.2018.06.072.
Otchere, D.A., Arbi Ganat, T.O., Gholami, R. and Lawal, M. 2021a. A novel custom ensemble
learning model for an improved reservoir permeability and water saturation prediction.
J. Nat. Gas Sci. Eng. 91: 103962. https://doi.org/10.1016/j.jngse.2021.103962.
Otchere, D.A., Arbi Ganat, T.O., Gholami, R. and Ridha, S. 2021b. Application of supervised
machine learning paradigms in the prediction of petroleum reservoir properties:
Comparative analysis of ANN and SVM models. J. Pet. Sci. Eng. 200: 108–182.
https://doi.org/10.1016/j.petrol.2020.108182.
Otchere, D.A., Ganat, T.O.A., Ojero, J.O., Taki, M.Y. and Tackie-Otoo, B.N. 2021c. Application
of gradient boosting regression model for the evaluation of feature selection techniques
in improving reservoir characterisation predictions. J. Pet. Sci. Eng. 109244. https://doi.
org/10.1016/J.PETROL.2021.109244.
86 Data Science and Machine Learning Applications in Subsurface Engineering

Otchere, D.A., Abdalla Ayoub Mohammed, M., Ganat, T.O.A., Gholami, R. and Aljunid
Merican, Z.M. 2022a. A novel empirical and deep ensemble super learning approach
in predicting reservoir wettability via well logs. Applied Sciences 12: 2942. https://doi.
org/10.3390/app12062942.
Otchere, D.A., Ganat, T.O.A., Nta, V., Brantson, E.T. and Sharma, T. 2022b. Data analytics
and Bayesian Optimised Extreme Gradient Boosting approach to estimate cut-offs from
wireline logs for net reservoir and pay classification. Appl. Soft. Comput. 120: 108680.
https://doi.org/10.1016/j.asoc.2022.108680.
Phadke, S., Bhardwaj, D. and Yerneni, S. 2000. Marine Synthetic Seismograms Using Elastic
Wave Equation. 2000 SEG Annual Meeting, Calgary, Alberta.
Pickett, G.R. 1963. Acoustic character logs and their applications in formation evaluation.
Journal of Petroleum Technology 15: 659–667. https://doi.org/10.2118/452-PA.
Rasouli, V., Pallikathekathil, Z.J. and Mawuli, E. 2011. The influence of perturbed stresses near
faults on drilling strategy: A case study in Blacktip field, North Australia. J. Pet. Sci. Eng.
76: 37–50. https://doi.org/10.1016/j.petrol.2010.12.003.
Rojas, E. 2008. Vp-Vs ratio sensitivity to pressure, fluid, and lithology changes in tight gas
sandstones. First Break 26. https://doi.org/10.3997/1365-2397.26.1117.27907.
Safaei-Farouji, M., Hasannezhad, M., Rahimzadeh Kivi, I. and Hemmati-Sarapardeh, A. 2022.
An advanced computational intelligent framework to predict shear sonic velocity with
application to mechanical rock classification. Sci. Rep. 12: 5579. https://doi.org/10.1038/
s41598-022-08864-z.
Simm, R. and Bacon, M. 2014. Seismic Amplitude. Cambridge University Press. https://doi.
org/10.1017/CBO9780511984501.
Sneider, J.S., Clarens, P. de and Vail, P.R. 1995. Sequence stratigraphy of the middle to upper
jurassic, viking graben, North sea. Norwegian Petroleum Society Special Publications
5: 167–197. https://doi.org/10.1016/S0928-8937(06)80068-8.
Suleymanov, V., Gamal, H., Glatz, G., Elkatatny, S. and Abdulraheem, A. 2021.
Real-time prediction for sonic slowness logs from surface drilling data using machine
learning techniques. In: SPE Annual Caspian Technical Conference. SPE. https://doi.
org/10.2118/207000-MS.
Tarafder, S., Badruddin, N., Yahya, N. and Egambaram, A. 2021. EEG-based drowsiness
detection from ocular indices using ensemble classification. pp. 21–24. In: 2021 IEEE 3rd
Eurasia Conference on Biomedical Engineering, Healthcare and Sustainability (ECBIOS).
IEEE, https://doi.org/10.1109/ECBIOS51820.2021.9510848.
Tarafder, S., Badruddin, N., Yahya, N. and Nasution, A.H. 2022. Drowsiness detection using
ocular indices from EEG signal. Sensors 22: 4764. https://doi.org/10.3390/s22134764.
Tosaya, C. and Nur, A. 1982. Effects of diagenesis and clays on compressional velocities in
rocks. Geophys. Res. Lett. 9: 5–8. https://doi.org/10.1029/GL009i001p00005.
Waluyo, W., Uren, N.F. and McDonald, J.A. 1995. Poisson’s ratio in transversely isotropic
media and its effects on amplitude response: An investigation through physical modelling
experiments. pp. 585–588. In: SEG Technical Program Expanded Abstracts 1995. Society
of Exploration Geophysicists. https://doi.org/10.1190/1.1887420.
Zoveidavianpoor, M., Samsuri, A. and Shadizadeh, S.R. 2013. Adaptive neuro fuzzy inference
system for compressional wave velocity prediction in a carbonate reservoir. J. Appl.
Geophy. 89: 96–107. https://doi.org/10.1016/j.jappgeo.2012.11.010.
 Chapter 5
Data-Driven Virtual Flow
Metering Systems
Ramez Abdalla* and Philip Jaeger

1. Introduction
Proper estimation of multiphase flowrates in oil and gas production systems is
an essential tool of monitoring and optimising the production systems. Hence,
one of the routine tests of wells is production testing. It is usually conducted
as a schedulable test to monitor liquid rates, water cuts, and gas oil ratio
(GOR). The production testing is easily conducted using the test separator
to compare the actual production rate with the theoretical one. It is the
most common form of production and reservoir surveillance. However, this
technique has its limitations. The main limitation of this test is the insufficient
resolution or repeatability to identify trends in liquid and water-cut rates over
short periods of time. Another potential problem could be the duration issue.
It is often the case in low-flow rate and deep wells, which require several
time-consuming whole or complete liquid holdup periods. Later, an alternative
solution to the production testing has been developed. This solution is called
multiphase physical flow meters (MPFMs). This technology depends on the
idea of indirectly estimating multiphase flowrates without separating the
phases. This is done by tracking supplementary measurements of fluid phase
properties such as velocity and phase fractions inside the device. These meters
are usually installed at the wellhead, so that the multiphase flowrates of a
particular well can be tracked in real-time. One disadvantage of MPFMs is
that they have higher capital costs (CAPEX) and operating costs (OPEX), and
they require frequent production calibration.

Clausthal University of Technology, Clausthal, Germany.

* Corresponding author: [email protected]
88 Data Science and Machine Learning Applications in Subsurface Engineering

Subsequently, the technology of virtual flow metering (VFM) arises as an

attractive technology in the oil and gas industry. It depends on either analytical
or data-driven models for real-time calculations of phases’ production. In
this chapter, data driven virtual flow metering solutions are presented. At the
beginning, VFM classification is introduced based on modelling paradigms.
They are classified as first principle and data-driven VFMs. Then, an overview
of the applications of data-driven modelling in VFM systems is presented.
Through discussing these applications, the model’s features used in the works,
the predicted variables, the input data for the training, and the respective paper
are emphasised.
In addition, the data driven virtual flow meter components and
methodology to develop this type of flow meter is discussed. The methodology
follows the cross-industry standard process for a data mining framework. This
is a framework for project planning that many industrial projects have used.
Its main components are data understanding and preparation, modelling,
and evaluation. Subsequently, an implementation of this methodology is
presented to estimate flow rate and water-cut prediction from the electrical
submersible pump’s sensor data. Finally, we introduce a field experience with a
data-driven VFM system. In this section, the real operational experience
reported in the literature using the data-driven models are discussed.

2. VFM Key Characteristics

The near wellbore area, wells, pipelines, and production chokes must
be simulated to be able to formulate systems. To come up with precise
estimates of flowrates, models are combined with observations like inlet and
outlet pressures and temperatures, choke openings, etc. Depending on the
measurement data available, the production system can either be represented
as a whole from the reservoir to the processing plant, or it can be divided
into sub models. Figure 5.1 shows a schematic graph of well sub models and
relevant data. Also, optimisation algorithms may be used to modify flowrates
and other tuning parameters. This is mainly to stabilise estimation models by
reducing the discrepancy between model predictions and actual measurements.
The concept behind these systems makes them have general characteristics
in common whatever the system sub modules are. One of these characteristics
is that they are dependent on instrument and system sensitivity to changes in
flow and phase fractions. They are also dependent on tuning/calibration to
extend to a “calibrated range” and adapt to a changing operating condition.
They can be classified into pure data-driven or thermo-hydraulic modelling
VFMs, or hybrids. All of them, either data-driven or mechanistic, should take
into account pressure and temperature drop in the well, and pressure drop over
choke and phase fractions of the flowing stream.
 Data-Driven
Many conservation equations in the Virtual Flow
mechanistic Metering
VFM Systems
s take 89
a dynamic
form. However, because of the steady state or quasi-steady state nature of the
2 3
Chk
P/T P/T
pos

𝑄𝑄𝑄𝑄

P/T

Flow line
length

𝑄𝑄𝑄𝑄 = 𝑓𝑓𝑓𝑓(𝑃𝑃𝑃𝑃1, 𝑇𝑇𝑇𝑇1, 𝑃𝑃𝑃𝑃2, 𝑇𝑇𝑇𝑇2, 𝑃𝑃𝑃𝑃3, 𝑇𝑇𝑇𝑇3, 𝐶𝐶𝐶𝐶ℎ𝑘𝑘𝑘𝑘𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝, . . 𝑥𝑥𝑥𝑥𝑖𝑖𝑖𝑖)

Fig. 5.1. Schematic graph of well sub models.

Many conservation equations in the mechanistic VFMs take a dynamic

Fig. form. However,
.5.1: becausegraph
Schematic of the steady
of wellstate
subor quasi-steady
models. state nature of the
optimisation problem formulation, an optimisation solver can only discover
a solution for a single point in time or can use the solution from the previous
step as
optimis a first
ation estimateformulation,
problem for the prediction of the current
an optimis ationtime
solverstep. Sometimes,
can only discover
due to the nature of the problem, the conservation equations adopt
a solution for a single point in time or can use the solution from the previous steady state
forms or do not take time as a variable parameter. Although it is feasible to
step frame
as a first estimate for the prediction of the current time step. Sometimes,
the VFM as an optimisation problem dynamically, such a strategy is
due to the nature
not applied in VFM of the problem,
literature. the conservation
The primary equations
reason for this adopt
could be the cost steady
of
statecomplex
forms orcomputations
do not takeoftime as a variable
the dynamic parameter.
optimisation Although
for VFM it isAlso,
systems. feasible
to frame the VFM
that would not as an itoptimis
make ation
suitable for problem
real-time dynamically, such a strategy is
applications.
not applied In addition
in VFMtoliterature.
dynamic optimisation,
The primaryKalman
reason filter approaches
for this couldandbeother
the cost
state estimation methods may be employed to develop
of complex computations of the dynamic optimis ation for VFM systems. a dynamic VFM. TheAlso,
major difficulty of this method could be the high level of knowledge required
that would not make it suitable for real- time applications.
for setup and use of this application, as well as the difficulty of tuning the
Inactual
additionfieldto
datadynamic optimis
in a reliable way. ation, Kalman filter approaches and other
state estimation methods may be employed
On the other hand, the data-driven to develop
VFM technique a dynamic
is based VFM. The
on gathering
majorfield
difficulty
data andofmathematically
this method could adaptingbeitthe high
to the level of knowledge
production required
system’s physical
for setup and use of this application, as well as the difficulty of tuning the actual
parameters, such as well bore and choke geometry, flow-line wall thickness,
field etc.,
datawithout providing
in a reliable way.an explicit description of those parameters. Over the
On the other hand, strategy
past few years, this has gained
the data-driven VFMa lottechnique
of popularity not just
is based on for oil
gathering
and gas applications but also for many other applications. The data-driven
field data and mathematically adapting it to the production system’s physical
parameters, such as well bore and choke geometry, flow-line wall thickness,
etc., without providing an explicit description of those parameters. Over the
past few years, this strategy has gained a lot of popularity not just for oil and
90 Data Science and Machine Learning Applications in Subsurface Engineering

model can conduct quick and precise real-time metering if the model has been
properly trained and the exposed conditions fall within the training range.
This method can build models more affordably than mechanistic models since
it does not require much in-depth production engineering domain expertise. In
the following section, we are presenting various data-driven VFM applications
on different oil wells systems.

3. Data Driven VFM Main Application Areas

Virtual sensors are convenient replacements of physical sensors that use
available data during well-known conditions in instrumented wells to predict
other measurements (Vinogradov and Vorobev, 2020). For instance, wellhead
pressure (WHP) and wellhead temperature (WHT) are related to flow line
pressure (FLP) for a specific choke diameter; therefore, it is convenient to
establish a machine learning model among these four variables so that it can
serve as a replacement whenever needed. Since most of the oil and gas wells
exhibit a non-stationary process, where the boundary conditions may change
along with the life of the field, it is probably safe to recognise that the virtual
sensor model may not be valid throughout the life of the well, but for a specific
period of time. Finally, a data-driven VFM would be very beneficial when
there is sufficient measured data, including frequent well tests (e.g., 8–12 per
year), permanent wellhead and flow line sensors (pressure, temperature). In
the upcoming subsections, we are introducing different attempts of VFM for
various production systems.

3.1 Virtual Sensing in ESP Wells

Electric submersible pumps (ESPs) are currently widely employed on many
artificially lifted wells with high water cut and offshore oil wells due to their
simple structure and high efficiency (Takacs, 2018). Among all the artificial
lift systems, ESP is preferred because it can produce high volumes in higher
temperatures and reach deeper depths. The development of sensors and data
acquisition systems make it possible for ESP systems to continuously record
the intake pressure and temperature, pump head, discharge pressure, and
temperature, motor temperature and current, leakage current, vibration, etc.
These data would be recorded at regular intervals and transmitted to surface
remote terminal units (RTUs) (Carpenter, 2019). Figure 5.2 shows a schematic
graph of sub-modules and sensors deployed on the electrical submersible
pump well system.
Based on the advancement in communication, well technology, and field
equipment, a lot of approaches are made for the estimation of flow rates in
real time. In the following, a summary of the applications of soft sensing both
mechanistic and surrogate modelling is given.
 Data-Driven Virtual Flow Metering Systems 91
time. In the following, a summary of the applications of soft sensing both
mechanistic and surrogate modelling is given.
𝐼𝐼𝐼𝐼𝑣𝑣𝑣𝑣𝑝𝑝𝑝𝑝𝑣𝑣𝑣𝑣 , 𝑉𝑉𝑉𝑉𝑣𝑣𝑣𝑣𝑝𝑝𝑝𝑝𝑣𝑣𝑣𝑣

𝑷𝑷𝑷𝑷𝒄𝒄𝒄𝒄 ,
𝑷𝑷𝑷𝑷𝒘𝒘𝒘𝒘𝒘𝒘𝒘𝒘 𝑷𝑷𝑷𝑷𝒎𝒎𝒎𝒎
𝒁𝒁𝒁𝒁𝒄𝒄𝒄𝒄

𝑄𝑄𝑄𝑄

𝐿𝐿𝐿𝐿2

ESP

Motor
𝑇𝑇𝑇𝑇𝑚𝑚𝑚𝑚

𝐿𝐿𝐿𝐿1

𝑞𝑞𝑞𝑞𝑟𝑟𝑟𝑟
• 𝑃𝑃𝑃𝑃෩
𝑟𝑟𝑟𝑟
𝑃𝑃𝑃𝑃𝑤𝑤𝑤𝑤𝑓𝑓𝑓𝑓

Fig. 5.2. A schematic diagram of an ESP well.

Fig. 3.2 Virtual

.5.2: ASensing for SRP
schematic Wellsof an ESP well.
diagram
Beam pumping, or the sucker-rod lift method, is the oldest and most widely
used type of artificial lift. A sucker-rod pumping system is made up of several
Table some
components, 5.1: of
Applications
which operate of on
thetheVFM on ESP
surface and wells.
other parts operate
underground,
Author down in the
Model Summary well. The surface-pumping unit, which drives the
underground pump, consists of a prime mover (usually an electric motor) and,
Camilleri
normally, a beamIn these
fixed to papers,
a pivotalESP
postmodels
which iswith different
called a Sampsonmodifications
post and
et the
al. beam
(2011, are discussed, as
is called a walking beam. well as field case studies where the ESP’s
Several sensors
2015, 2016, can provide
first principle measurements
models serve asofvirtual
sucker-rod
flowpump
meters.operations.
A hybrid
The main measurements are loads on the pump
2016b, 2017) method has been used to measure the flow rate without which form what are
called dynamomter cards which are diagnostic cards that plot the load on the
needing a test separator or multi-phase flowmeter. It is
top rod (polished rod) in relation to the polished rod position as the pumping
statedeach
unit moves through that stroke
the pump
cycle.absorbs
The plotanofamount of power
the polished equating
rod Load vs
Position is known as the Surface Card. Subsequently, a wave equation solution in
that is generated by the motor. The drop of the pressure
is used to derive
thethe downhole
tubing card measurements
provides from the surface card.
of theThe downhole
average card of
density
is a plot of thethe
Load vs Position on the pump’s plunger. Also, on
fluid, which is then converted to a water cut. Obtaining the surface,
continuous measurements for wellhead pressures, temperatures and power
a reference for validating the calculation, a comparison has
measurements of the motor are reported. In addition, the normal frequent tests
been carried
data for any production outare
system between
provided.the calculated
Those are fluidand
levelthe measured
depth using
rates on
acoustic transducer anda production
shale oil well equipped
of multi phaseswith an aESP
using [15, 16,test
separator 14,or17,
18].
92 Data Science and Machine Learning Applications in Subsurface Engineering

Table 5.1. Applications of the VFM on ESP wells.

Author Model Summary

Camilleri In these papers, ESP models with different modifications are discussed, as well
et al. (2011, as field case studies where the ESP’s first principle models serve as virtual flow
2015, 2016, meters. A hybrid method has been used to measure the flow rate without needing
2016b, a test separator or multi-phase flowmeter. It is stated that the pump absorbs an
2017) amount of power equating that is generated by the motor. The drop of the pressure
in the tubing provides measurements of the average density of the fluid, which is
then converted to a water cut. Obtaining a reference for validating the calculation, a
comparison has been carried out between the calculated and the measured rates on a
shale oil well equipped with an ESP (Camilleri, 2011, 2015, 2016a, 2016b, 2017).
Haouche The VFM model is a combination of three main units: Reservoir unit, Electrical
et al. Submersible Pump unit, and Production Tubing unit. A density correction factor is
(2012a, used to take into account the effects of gas on the operational performance of the
2012b) submersible pump (Adrien, 2012; Haouche, 2012).
Binder A Moving Horizon Estimator is applied for flowrate estimation in a well with an ESP.
et al. As input, the bottomhole, the downhole, and the pump pressure sensors are considered
(2015) together with pump parameters. The method showed an accurate performance and
was suggested to be used for industrial applications (Binder et al., 2015).
Zhu et al. In this research, singular spectrum analysis (SSA) is used on a raw production
(2016) dataset without any pre-processing or transformation of the original series. They
investigated the decomposition of the original series into a summation of the
principal independent and interpretable components such as slowly varying trends,
cycling components, and random noise (Zhu et al., 2016).
Krikunov A hybrid physical-machine learning prediction model was developed. It utilis ed a
et al. range of motor frequencies. A numeric optimis ation model was created to suggest
(2019) multi-well operating modes (Krikunov et al., 2019).
Zhu et al. In these studies, a mechanistic model was developed to predict the pump- boosting
(2020, pressure. The objectives were to forecast oil-water emulsion rheology and how it
2021) will affect ESP boosting pressure and describe the pump leaking impact under the
conditions of a gas-liquid flow (Zhu et al., 2020, 2021).
Sabaa et al. This study aims to develop artificial neural network models to predict flow rates
(2022) of ESP artificially lifted wells. Each data set included measurements for wellhead
parameters, fluid properties, ESP downhole sensor measurements, and variable
speed drive (VSD) sensor parameters. The models consisted of four separate neural
networks to predict oil, water, gas, and liquid flow rates (Sabaa et al., 2022).

production test. Figure 5.3 shows the sub-modules of sucker rod pumped well
with relevant tests, dynamometer, and fluid level test example.
The applications of virtual sensing on the sucker rod pump systems known
so far are limited. This may be due to its limited ability in producing high fluid
rates. Some attempts have been made in this regard. Those applications had
various objectives. The first objective is predicting multi phase flow rates
or the dynamic fluid level in the annulus using dynamomter cards and well
head pressure and temperature as inputs. The second objective is inferring the
dynamometer cards using electrical power data.
underground pump, consists of a prime mover (usually an electric motor) and,
normally, a beam fixed to a pivotal post which is called a Sampson post
and the beam is called a walking beam.
Several sensors can provide measurements
Data-Driven of Metering
Virtual Flow sucker- Systems
rod pump
93
operations.

Dynamometer test

Fluid Level test

Fig. 5.3. A schematic graph of an SRP well.

Fig. Virtual
3.2.1 5.3:Flow
A schematic
Meter ongraph of an SRP
Rod Pumping well.
Systems
The challenge of predicting oil, gas and water flow is a function that describes
theThe
multi phase
main flow rates. As
measurements arealoads
solution, data-driven
on the pump which algorithms are usedare
form what to
find a relation between the pump operational parameters and
called dynamomter cards which are diagnostic cards that plot the load on the the produced
oil, gas, (and polished
top rod water. Peng rod) et al.in(2019)
relationhave used
to the the deep
polished rodautoencoders to
position as the
derive
pumping features from through
unit moves dynamometer cards to
each stroke generate
cycle. a predictive
The plot model rod
of the polished for
production. Combining
Load vs Position is knownthis model
as the with
Surface pump
Card. and production
Subsequently, a wavedata leads
equation
to abstract features that show good accordance with the
solution is used to derive the downhole card from the surface card. Thehistory data (Peng
et al., 2019).
downhole card is a plot of the Load vs Position on the pump’s plunger. Also, on
Soft sensing,
the surface, whichmeasurements
continuous replaces the traditional detection
for wellhead method
pressures, to model
temperatures
the dynamic liquid level of the sucker-rod pumping system
and power measurements of the motor are reported. In addition, the normal (Haitao et al.,
2014), proposes a method to calculate the dynamic fluid level. It takes the
submerged pressure as a common solution node to analyse both the plunger
load variation which is contributed by the pump dynamometer card and the
pressure distribution in the annulus. Li et al. (2013) presented the simulated
annealing based on a Gaussian regression modeling.

3.2.2 Virtual Sensing of the Dynamometer Card

Through the years, many researchers have been studying the relation between
electrical parameters and surface cards (Zhang and Tang, 2008). Some theory
formulas also can be built to calculate the card from the electrical parameters.
Nevertheless, some parameters in the formulas cannot be quantified or
94 Data Science and Machine Learning Applications in Subsurface Engineering

measured and some assumptions of the values made the card calculation
inaccurate and unstable. Thus, a machine learning model for dynamometer
card calculation in the rod pumping lift process is used to formulate the
complicated process. Deep neural networks are used to find good weight
combinations in these examples that eventually allow the model to come
up with rules from the input data (electrical parameter) to the target data
(dynamometer cards).
Such a study includes extracting power features and constitutes
an eigenvector in chronological sequence for one period. Afterwards,
dynamometer card data and shape curve image are extracted according to
coordinates and load data. Subsequently, power features and dynamometer
diagram features are normalized by rows and then mapped between 0 and 1.
Finally, a sequence-to-sequence algorithm to infer dynamometer cards from
the power curve features (Peng et al., 2019; Zhu et al., 2021).

3.3 Virtual Sensing for Gas Lifted Wells

Gas lift (GL) is a method of artificial lift that uses an external source of
high-pressure gas for supplementing the formation of gas to lift the well
fluids. The principle of the GL is that gas injected into the tubing reduces the
density of the fluids in the tubing and the bubbles have a ‘scrubbing’ effect on
the liquids. Both factors act to lower the flowing bottomhole pressure (BHP)
at the bottom of the tubing.
One use of the case for VFM modelling is to compute uncorrelated
estimates of the GL rate by using the GL flow control valve performance
model.
Figure 5.4 shows the control volumes of the gas lifted well. In this system,
the inputs are manifold pressure, valve and casing pressure, while the output
is GL rates.
Applications of this research by Ziegel et al. (2014) discuss the design of
a data-driven model that would support and optimize production operations
during normal or gas coning. The data-driven model had a limited set of inputs
such as production choke, GL choke, resulting GL and reservoir pressure
estimation. Therefore, several predictive models will be built to create a final
data-driven model (Ziegel et al., 2014).
Iman et al. (2020) developed a data-driven approach to find an optimal
operating envelope for GL wells. The process involves building a multilayer
perceptron neural network model for generating instantaneous predictions of
multiphase flow rates and other quantities of interest, such as GOR, WCT,
using real-time sensor data at the surface, historical performance and sporadic
test data. The models were developed for generating short-term (30 days)
forecast of cumulative oil, water, gas and liquid production, multiphase flow
the fluids in the tubing and the bubbles have a ‘ scrubbing’ effect on the
liquids. Both factors act to lower the flowing bottomhole pressure (BHP) at the
bottom of the tubing.
Data-Driven Virtual Flow Metering Systems 95
One use of the case for VFM modelling is to compute uncorrelated estimates
of the GL rate by using the GL flow control valve performance model.
Production
Manifold (MP)
Casing Head
Surface
Surface CHP Flowline
gas injection
gas-injection
valve
valve

Test
Manifold

BHP
BHT

PayPay
ZoneZone PayZone
Pay Zone

Fig. 5.4. A schematic graph of a GL well.

rates, WCT, GOR, and reservoir pressure. Using time-series forecasting

Fig.models,.5.4: A schematic
a sensitivity graphwas
analysis of aperformed
GL well. to generate short-term well
response for a selected number of combinations of choke settings and gas
injection rates (Al Sebaiti et al., 2020).
Fig. 5.4 shows the control volumes of the gas lifted well. In this system, the
Khan et al. (2020) used AI techniques to derive a robust correlation for
inputs are manifold pressure, valve and casing pressure, while the output is
forecasting production rates in gas-lift assisted wells. The AI techniques used
GL in
rates.
this research included artificial neuro-fuzzy inference systems, artificial
Applications
neural network of (ANN),
this research by Ziegel
functional et. al.and
networks, (2014) [45] vector
support discuss
machines.the
design
Theyofcollected
a data-driven
test datamodel that would
from several support
GL wells and and
used optimize production
ANN to develop an
operations during normal or
equation to forecast oil flow rate. gas coning. The data-driven model had a limited
set of inputs suchthey
Initially, as production
applied wide choke, GL choke,and
data analytics resulting GL and
then input datareservoir
to the
pressure
modelsestimation.
that were compared Therefore, several
to each other andpredictive models will
to other empirical be built
models. Theyto
could
create predict
a final oil rates with
data-driven model an accuracy
[45]. exceeding 98% (Khan et al., 2020).
Iman et. al. (2020) developed a data-driven approach to find an optimal
3.4 Virtual
operating Sensing
envelope for forGL Gas Wells
wells. Theand Plunger
process Liftedbuilding
involves Wells a multilayer
perceptron
Conventional neural network
plunger model
lifting is for generating
a transient instantaneous
process predictions
that consists of cyclicof
multiphase
openingsflow and rates andof
closings other quantities
a gas of interest,
well. Because suchcomplex
of their as GOR, behaviour,
WCT, using
real-time sensor data
using traditional at the surface,
physics-based historical
models performance
to simulate and sporadic
the coupled test
behaviour
of The
data. reservoir
modelsand werewellbore
developed performance is computationally
for generating short-term (30 rigorous and
days) forecast
challenging.oil,Therefore,
of cumulative water, gas machine
and liquidlearning
production,methodology
multiphase would helpWCT,
flow rates, in
formulating the plunger lifted well system including plunger arrival time,
tubing and casing pressure, and instantaneous gas flow rate. The sequence
of stages in one cycle. The motor valve has only two states, namely closed
or open. The length of each state depends on the pre-programmed triggers,
forecasting production rates in gas-lift assisted wells. The AI techniques used in
this research included artificial neuro-fuzzy inference systems, artificial neural
network (ANN), functional networks, and support vector machines. They
96 Data Science and Machine Learning Applications in Subsurface Engineering
collected test data from several GL wells and used ANN to develop an
equation to forecast oil flow rate.
Shock Spring

Plunger sensor

Catcher

Plunger

Bumper Spring
𝑞𝑞𝑞𝑞𝑟𝑟𝑟𝑟
• 𝑃𝑃𝑃𝑃෩
𝑟𝑟𝑟𝑟
𝑃𝑃𝑃𝑃𝑤𝑤𝑤𝑤𝑓𝑓𝑓𝑓

Fig. 5.5. A schematic graph of a plunger lifted well.

which are also called controller set-points. The main triggers to close a well
Fig. include
5.5:a A schematic
fixed timer for graph of a plunger
the after-flow stage, thelifted well.between the gas
difference
and the calculated critical flow rates, as well as various relations between
casing, tubing, and line pressures. The main triggers to open a well include a
Initially, they
fixed applied
timer wide
for total “offdata analytics
time”, andaverage
calculated then input data
plunger risetovelocity,
the models
and that
different relations
were compared to eachbetween
other andthe pressure
to othermeasurements. FigureThey
empirical models. 5.5 shows
couldthe
predict
control volumes of a plunger lifted
oil rates with an accuracy exceeding 98% [28] . well.
As aforementioned, virtual flow metering includes physics-based and
data-driven methods. When it comes to the plunger lift application, since
the process is extremely transient, the application of physics-based methods
3.4. Virtual Sensing for Gas Wells and
is extremely complex and unviable (Akhiiartdinov et al., 2020). Regarding
Plunger
Lifteddata-driven
Wellsmodels, Andrianov (2018) (GarcAa et al., 2010) and Shoeibi
Omrani etplunger
Conventional al. (2018)lifting
(Loh and
is aOmrani, 2018)process
transient demonstrated
that the application
consists of cyclic
of ANNs to simulate the transient behaviour of severe slugging and
openings and closings of a gas well. Because of their complex behaviour,
liquid-loaded gas wells. On the other hand, Akhiiartdinov et al. (2020) have
using attempted
traditional physics-based
to model a VFM on models to simulate
plunger lifted wells. In the coupled
this study, behaviour of
the objective
reservoir
was toand wellbore
optimize the “on”performance
and “off” periodsis ofcomputationally
the control valve, whichrigorous
serve and
challenging. Therefore,
as parameters machine
for building learning
the response surface.methodology would help in

3.5 Miscellaneous Applications for Identifying Flow Regimes

As aforementioned, there are common characteristics of VFMs. One of the
main characteristics is estimation of the fraction of each phase or, in other
 total ”off time”, calculated average plunger rise velocity, and different
relations between the pressure measurements. Fig. 5.5 shows the control
volumes of a plunger lifted well.
As aforementioned, virtual flow metering includes physics- based and
data-driven methods. When it comes to the plunger lift application, since the
Data-Driven Virtual Flow Metering Systems 97
process

Fig. 5.6. Multiple-phase flow regimes.

words, identifying the flow flow

Fig .6: Multiple-phase regime.
regimesTherefore, various applications arise
in this area. Their objective is to construct a classification model to predict
various flowis regimes based on
extremely transient, theflow measurements.
application of physics-basedTheir work
methods focuses on the
is extremely
identification of multiple-phase
complex flow regimes
and unviable [4]. Regarding by implementing
data- driven models, Andrianovdeep(2018)learning
[23] and Shoeibi Omrani et al. (2018) [32] demonstrated the application of
algorithms ANNsin addition to commonly used machine learning algorithms
to simulate the transient behaviour of severe slugging and liquid-
(Alhashem, loaded
2020;gasArteaga-Arteaga
wells. on the otherethand,
al., 2021; Manikonda
Akhiiartdinov et al.,
et al. (2020) [4] 2018;
have Mask
et al., 2019).attempted
Figureto5.6 shows
model a VFMthe relation
on plunger between
lifted thestudy,
wells. In this pressure drop versus
the objective
was to optimize the “ on” and ”off” periods of the control valve, which serve
flow rate with different flow regimes.
as parameters for building the response surface.

4. Methodology of Building Data-driven VFMs

4.1 Data Collection and Preprocessing
Data collection and preprocessing are two main steps in constructing any
data-driven model. Usually, the data is collected from several locations in the
field or historical data from other relevant wells or fields. After collecting the
data set that will be used to train the data-driven model, this data set usually
requires processing since it may be noisy, corrupted, or has missing values,
deviations, and irrelevant data. Therefore, the preprocessing phase of the data
set is essential to clean and authenticate it before training the model.
In the preprocessing step, the data set can undergo some modifications
and more information can be gathered from it. Applying feature engineering
techniques in the preprocessing step will support constructing the data-driven
model to locate the complex interactions between the original data and the
output or eliminate unnecessary features which increase the computational
costs. One common method of feature engineering is dimensional reduction
using principal component analysis (PCA), feature selection methods, and
linear or nonlinear combinations of raw features.
In general, feature engineering may help the data-driven algorithm to find
complex relationships between the original data and the output variable or
remove redundant features which lead to lower computational cost during
training and prediction steps.
98 Data Science and Machine Learning Applications in Subsurface Engineering

In most of the cases, subject matter experts (SME) construct informative

features for further algorithm training. Creating good features using the input
data which can describe the multiphase flow transport process may help to
obtain better predictions.

4.2 Model Development

Model Development is the phase in which the implemented algorithm
constructs a map of the input and output features. This mapping process is
also called the training or learning phase of the model. In this phase, the model
will modify the available parameters to precisely predict the final output. The
technique the parameters
16 ■ Saunders Template are modified differs depending on the algorithm
implemented in building the model. The next step in the model development
is verifying and testing the model on different data sets to see how the model
will perform on untrained data.
Additionally, the validation step is important to adjust the
hyper-parameters. Hyper parameters are selected before training to decrease
noise and deviations that reduce the prediction precision of the model. K-fold
4.2.Model
cross-validation is a Development
new technique in which the data set is divided into
Model
training andDevelopment
testing data is thenthethephase
modelin hyper
which parameters
the implemented algorithm
are assigned. Then
constructs
the model willabe map of theon
trained input and output
K-folds features.
and the Thiserror
average mapping process is also
is calculated for the
called the training or learning phase of the model. In this phase, the model will
model. The K-fold will be repeated with several hyper parameters to find the
modify the available parameters to precisely predict the final output. The
best ones with the lowest average error.
technique the parameters are modified differs depending on the algorithm
The final
implemented phase is to confirm
in building theThe
the model. performance
next step in of
thethe validated
model modelison a
development
separate dataand
verifying set.testing
In thethe
final
modelstage, there are
on different two
data possible
sets outcomes:
to see how the modelfirst,
will the
errorsperform
on theontraining
untrainedand testing sets are large which is called underfitting,
data.
and second, the errors
Additionally, the on the training
validation step are small but tolarge
is important on the testing
adjust hyper- set
which is called Hyper
parameters. overfitting. Thus,are
parameters validating and testing
selected before trainingthe model to
to decrease assign
noise
and hyper
the best deviations that reduce
parameters the prediction
to figure out theprecision of the model.
finest outcome is theK-fold cross-goal.
optimal
validation
A typical is a for new
choice techniqueofin folds
the number which totheuse
data set is divided
would be k = into
10. training
Figure 5.7
and testing data then the model hyper parameters are assigned. Then the
shows a 10-fold cross-validation.
model will be trained on K-folds and the average error is calculated for the
model. The K-fold will be (Run on)
Training Set
Training folds Test folds

st
1 iteration 𝐸𝐸𝐸𝐸1

2nd iteration 𝐸𝐸𝐸𝐸2

1 K E
E= ? i
k i=1
3rd iteration 𝐸𝐸𝐸𝐸3

…….
10th iteration 𝐸𝐸𝐸𝐸10

Fig. 5.7. 10-fold cross-validation.

Fig. 5.7: 10-fold cross- validation.

 Data-Driven Virtual Flow Metering Systems 99

A common algorithm used for building VFM models is the ANN which
could create a function originating from the inputs that connect them to the
final output. Generally, an ANN consists of an input layer to read the input
features which usually are pressure, temperature, choke opening, or other
production system parameters in the case of a VFM.
The second component of the ANN is the hidden layer where non
linear functions are created to build the connection between the inputs and
the prediction and the final component is the output layer where the results
achieved in the hidden layers go through an activation function to reach the
final prognosis. The ANN showed significant results when applied to a steady
state flow unlike the case of transient movement.
Al-Qutami et al. (2017) showed in three publications the application of
ANN in constructing a VFM model (Al-Qutami et al., 2017, 2017, 2017).
Their work provides a significant sighting to improve VFM models. The latest
methodology applied in their research was an ANN model was trained using
the Levenberg-Marquardt optimisation algorithm and K-fold cross-validation
to specify the number of neurons. This work was able to conclude that the gas
flow rate is most sensitive to choke opening and that the bottom hole pressure
is the most critical feature for the flow rate prediction.
Loh and Omrani (2018) implemented feed-forward neural networks on
simulated and field data to predict oil and gas flow rates. The results showed
that the NN model was effective at steady state flow while the model produced
transient flow imprecise predictions. Furthermore, they conducted a sensitivity
study on the input features that showed that the NN model was able to perform
well even in case of noisy inputs. Finally, they introduced a new back-allocation
technique to predict the flow rate using separator measurements. Alajmi et al.
(2015) introduced an NN model to predict the oil flow rate through the choke
and they added an empirical correlation for critical choke flow to the input
features. The study demonstrated a noticeable enhancement in the NN model
performance compared to the mechanistic models.
To conclude in this section, we have reviewed the main building blocks
of creating VFMs. It started with data gathering, preprocessing, and feature
engineering. We have also presented the validation techniques. Finally, we
presented some of NN applications to VFM as an example for the algorithms
used for data-driven VFMs.

5. Field Experience with a Data-driven VFM System

In the pursuit of VFMs, several industrial and research works were published.
In this section, we focus on highlighting the industrial applications that show
a bench marking or a commercial tool. Among the available industrial works
available Denny et al. (2013) reported the importance of integrating VFM
100 Data Science and Machine Learning Applications in Subsurface Engineering

in an electrical submersible pump model. GarcAa et al. (2010) reported

another solution for monitoring production and injection rates using an NN
model instead of fiscal meters on individual wells. NN is trained by sensors,
well-tests, and simulation data. Their work results showed a 4% error in the
overall wells flow-rates’ predictions compared to the fiscal meters.
Moreover, Olivares et al. (2012) describe the implementation of an NN
model and nodal analysis joint with sporadic and real-time data to determine and
observe daily flow rates (Olivares et al., 2012). The new applied methodology
proved to be more efficient than the back allocation system. Al-Jasmi et al.
(2013) report the application of a flow-rate prediction NN model on a mature
carbonate reservoir in the Middle East (Cramer and Goh, 2009; Cramer
et al., 2011; Goh et al., 2008; Law et al., 2018; Poulisse et al., 2006). Those
papers describe Shell’s Data-Driven model FieldWare Production Universe
(FW PU), which was implemented in several case studies to estimate flow
rates and optimise the production operation. In the six study cases, FW PU
was able to enhance productivity and reduce the amount of gas used in the GL
which significantly affected the cost.
Bello et al. (2014) also used LR with feature extraction PCA which
showed significant results for oil and gas flow-rate predictions. Grimstad
et al. (2016) used pressure drop, choke, and inflow performance models from
prosper to train a B-spline surrogate model. The results showed reasonable
performance when compared to OLGA.

References
Adrien, M.H., Younes, T., Deffous, J-F., Couput, A.J-P., Caulier, R. and Vrielynck, B. 2012.
Smart metering: An online application of data validation and reconciliation approach.
Paper presented at SPE Intelligent Energy International Utrecht, The Netherlands, March.
pp. SPE–149908–MS.
Akhiiartdinov, A., Pereyra, E., Sarica, C. and Jose Severino, J. (eds.). 2020. Data analytics
application for conventional plunger lift modelling and optimisation, volume Day 1 Tue.,
10 November. SPE Artificial Lift Conference and Exhibition – Americas.
AlAjmi, Mohammed D., Abdulraheem, A., Mishkhes, A.T. and Al-Shammari, M.J. (eds.).
2015. Profiling downhole casing integrity using artificial intelligence, volume Day 1 Tue.,
03 March SPE Digital Energy Conference and Exhibition.
Alhashem, M. (ed.). 2020. Machine learning classification model for multiphase flow regimes in
horizontal pipes, volume Day 2 Tue., 14 January IPTC International Petroleum Technology
Conference.
Al-Jasmi, A., Goel, H.K,. Nasr, H., Querales, M., Rebeschini, J., Villamizar, M.A., Carvajal,
G.A., Knabe, S., Rivas, F. and Saputelli, L. 2013. Short-term production prediction in
real time using intelligent techniques. Paper presented at EAGE Annual Conference and
Exhibition at London, UK. In: All Days, SPE. pp. SPE–164813–MS.
Al-Qutami, T.A., Ibrahim, R., Ismail, I. and Ishak, M.A. 2017a. Development of soft sensor to
estimate multiphase flow rates using neural networks and early stopping. International
Journal of Smart Sensing and Intelligent Systems 10(1): 1–24.
 Data-Driven Virtual Flow Metering Systems 101

Al-Qutami, T.A., Ibrahim, R. and Ismail, I. 2017b. Hybrid neural network and regression
tree ensemble pruned by simulated annealing for virtual flow metering application.
pp. 304–309. In: 2017 IEEE International Conference on Signal and Image Processing
Applications (ICSIPA).
AL-Qutami, T.A., Ibrahim, R., Ismail, I. and Ishak, M.A. 2017c. Radial basis function network
to predict gas flow rate in multiphase flow. pp. 141–146. In: Proceedings of the 9th
International Conference on Machine Learning and Computing, ACM.
Al Sebaiti et al. 2020. Robust data-driven well performance optimisation assisted by machine
learning techniques for natural flowing and gas-lift wells in Abu Dhabi, volume Day 4
Thu., 29 October, of SPE Annual Technical Conference and Exhibition. D041S046R002.
Arteaga-Arteaga, H.D., Mora-Rubio, A. Florez, F., Murcia-Orjuela, N., Diaz-Ortega, C.E.,
OrozcoArias, S., delaPava, M., Bravo-OrtAz, M.A., Robinson, M., Pablo Guillen-Rondon,
P. and Tabares-Soto, R. 2021. Machine learning applications to predict two-phase flow
patterns. Peer J. Comp. Sc. 7: e798.
Bello, O., Ade-Jacob, S. and Kun Yuan, K. (eds.). 2014. Development of hybrid intelligent
system for virtual flow metering in production wells, volume All Days. SPE Intelligent
Energy International Conference and Exhibition.
Binder, Benjamin J.T., Pavlov, A., Tor, A. and Johansen, T.A. 2015. Estimation of flow rate and
viscosity in a well with an electric submersible pump using moving horizon estimation.
This work is funded by the Research Council of Norway and Statoil through the petromaks
project no. 215684: Enabling high-performance safety-critical offshore and subsea
automatic control systems using embedded optimization (emopt). IFACPapersOnLine
48(6): 140–146. 2nd IFAC Workshop on Automatic Control in Offshore Oil and Gas
Production OOGP 2015.
Camilleri, L. and Zhou, W. (eds.). 2011. Obtaining real-time flow rate, water cut, and reservoir
diagnostics from ESP gauge data, volume All Days. SPE Offshore Europe Conference and
Exhibition, 2011.
Camilleri, L., El Gindy, M., Rusakov, A. and Adoghe, S. (eds.). 2015. Converting ESP real-time
data to flow rate and reservoir information for a remote oil well, volume Day 2 Wed.,
16 September. of SPE Middle East Intelligent Oil and Gas Symposium.
Camilleri, L., El Gindy, M. and Rusakov, A. (eds.). 2016a. ESP real-time data enables well testing
with high frequency, high resolution, and high repeatability in an unconventional well,
volume All Days. of SPE/AAPG/SEG Unconventional Resources Technology Conference.
Camilleri, L., El-Gindy, M. Rusakov, A., Bosia, F., Salvatore, P. and Rizza, G. (eds.). 2016b.
i Testing the Untestable Delivering Flowrate Measurements with High Accuracy on a
Remote ESP Well, volume Day 2 Tue., 08 November. Abu Dhabi International Petroleum
Exhibition and Conference, 2016.
Camilleri, L., El Gindy, M., Rusakov, A., Ginawi, I., Abdelmotaal, H., Sayed, E., Edris, T. and
Karam, M. (eds.). 2017. Increasing production with high frequency and high-resolution
flow rate measurements from ESPs, volume Day 2 Tue., 25 April SPE Gulf Coast Section
Electric Submersible Pumps Symposium, 2017.
Carpenter, C. 2019. Analytics solution helps identify rod-pump failure at the wellhead. Journal
of Petroleum Technology 71(05): 63–64.
Cramer, R. and Goh, K-C. (eds.). 2009. Data driven surveillance and optimis ation for gas,
subsea and multizone wells, volume All Days. SPE Digital Energy Conference and
Exhibition, 2009.
Cramer, R., Griffiths, W.N., Kinghorn, P., Schotanus, D., Brutz, J.M. and Mueller, K. (eds.).
2011. Virtual measurement value during start up of major offshore projects, volume All
Days. IPTC International Petroleum Technology Conference, 2011.
102 Data Science and Machine Learning Applications in Subsurface Engineering

Denney, T., Wolfe, B. and Zhu, D. 2013. Benefit evaluation of keeping an integrated
model during real-time ESP operations. In: All Days, SPE Digital Conference, SPE,
pp. SPE–163704–MS.
Garcia, A., Almeida, I., Singh, G., Purwar, S., Monteiro, M., Carbone, L. and Herdeiro, M. 2010.
An Implementation of On-Line Well Virtual Metering of Oil Production.
Goh, K-C., Dale-Pine, B., Yong, I.H.W., Peter, V. and Lauwerys, C. (eds.). 2008. Production
surveillance and optimisation for multizone smart wells with data driven models, volume
All Days. SPE Intelligent Energy International Conference and Exhibition, 2008.
Grimstad, B., Gunnerud, V., Sandnes, A., Shamlou, S., Skrondal, I.S., Uglane, V., Ursin-Holm,
S. and Foss, B. (eds.). 2016. A simple data-driven approach to production estimation and
optimization, volume All Days. SPE Intelligent Energy International Conference and
Exhibition, 2016.
Haitao, Y. et al. 2014. Real time calculation of fluid level using dynamometer card of sucker rod
pump well, volume All Days. IPTC International Petroleum Technology Conference, Dec.
IPTC-17773-MS.
Haouche, M., Tessier, A., Deffous, Y. and Authier, J-F. (eds.). 2012. Virtual flow meter pilot:
based on data validation and reconciliation approach, volume All Days. SPE International
Production and Operations Conference and Exhibition, 2012.
Khan, M.K., Tariq, Z. and Abdulraheem, A. 2020. Application of artificial intelligence to
estimate oil flow rate in gas-lift wells. Natural Resources Research 29(6): 4017–4029.
Krikunov, D., Kosyachenko, S., Lukovkin, D., Kunchinin, A., Tolmachev, R. and Chebotarev,
R. (eds.). 2019. AI-based ESP optimal control solution to optimize oil flow across multiple
wells, volume Day 2 Tue., 22 October. SPE Gas, 2019.
Law, H.Y., Phua, P.H., Briers, J. and Kong, J. (eds.). 2018. Extending virtual metering to
provide real time exception based analytics for optimising well management and chemical
injection, volume Day 2 Wed., 21 March. Offshore Technology Conference Asia, 2018.
Li, X., Gao, X., Cui, Y. and Li, K. 2013. Dynamic liquid level modelling of sucker-rod pumping
systems based on Gaussian process regression. pp. 917–922. In: 2013 Ninth International
Conference on Natural Computation (ICNC).
Loh, K. and Omrani, P.S. 2018. Deep Learning and Data Assimilation for Real-Time Production
Prediction in Natural Gas Wells. ArXiv, abs/1802.05141.
Manikonda, K., Hasan, A.R., Obi, C.E., Islam, R., Sleiti, A.K., Abdelrazeq, M.W. and Rahman,
M.A. 2018. Application of machine learning classification algorithms for two-phase
gas-liquid flow regime identification. Paper presented at Abu Dhabi International
Petroleum Exhibition and Conference. In: Day 4 Thu., 18 November, SPE. D041S121R004.
Mask, G., Wu, X. and Ling, K. 2019. An improved model for gas liquid flow pattern prediction
based on machine learning. In: Day 2 Wed., 27 March. Paper presented at the International
Petroleum Technology Conference, Beijing China, IPTC. D021S026R005.
Olivares, G., Escalona, C. and Gimenez, E. 2012. Production monitoring using artificial
intelligence, APLT asset. Paper presented at SPE Intelligent Energy International Utrecht,
The Netherlands. March. In: All Days, SPE. pp. SPE–149594–MS.
Peng, Y., Xiong, C., Zhang, J., Zhang, Y., Gan, Q., Xu, G., Zhang, X., Zhao, R., Shi, J., Liu,
M., Wang, C. and Chen, G. 2019a. Innovative deep autoencoder and machine learning
algorithms applied in production metering for sucker-rod pumping wells, volume Day
1 Mon., 22 July Unconventional Resources Technology Conference. SPE/AAPG/SEG.
D013S011R004.
Peng, Y., Zhao, R., Zhang, X., Shi, J., Chen, S., Gan, Q., Li, G., Zhen, X. and Han, T. (eds.).
2019b. IInnovative Convolutional Neural Networks Applied in Dynamometer Cards
Generation. Paper presented at the SPE Western Regional Meeting, San Jose, California,
USA, April 2019. doi: https://doi.org/10.2118/195264-MS.
 Data-Driven Virtual Flow Metering Systems 103

Poulisse, H., Van Overschee, P., Briers, J., Moncur, C. and Goh, K-C. 2006. Continuous well
production flow monitoring and surveillance. Paper presented at SPE Intelligent Energy
International Amsterdam, The Netherlands, 11–13 April.
Sabaa, A., El Ela, M.A., El-Banbi, A.H. and Sayyouh, Mohamed H.M. 2022. Artificial Neural
Network model to predict production rate of electrical submersible pump wells. SPE
Production & Operations, pp. 1–10. Sept.
Takacs, G. 2018. Chapter 1 - Introduction. pp. 1–10. In: Gabor Takacs (ed.). Electrical Submersible
Pumps Manual (Second Edition). Houston, Texas: Gulf Professional Publishing,
Vinogradov, D. and Vorobev, D. (eds.). 2020. Virtual flowmetering novyport field examples,
volume Day 3, Wed., 28 October. SPE Russian Petroleum Technology Conference.
Zhang, S. and Tang, Y. 2008. Indirect measurement of dynamometer card of pumping unit.
pp. 952–4955. In: 2008 7th World Congress on Intelligent Control and Automation.
Zhu, D., Alyamkin, S., Sesack, L., Bridges, J. and Letzig, J. (eds.). 2016. Electrical submersible
pump operation optimization with time series production data analysis, volume All Days.
SPE Intelligent Energy International Conference and Exhibition.
Zhu, D., Luo, X., Zhang, Z., Li, X., Peng, G., Zhu, L. and Jin, X. 2021. Full reproduction
of surface dynamometer card based on periodic electric current data. SPE Production &
Operations 36(03): 594–603.
Zhu, K., Wang, L., Du, Y., Jiang, C. and Sun, Z. 2020. Deeplog: Identify tight gas reservoir using
multi-log signals by a fully convolutional network. IEEE Geoscience and Remote Sensing
Letters, title=Prediction of Subsurface NMR T2 Distributions in a Shale Petroleum System
Using Variational Autoencoder-Based Neural Networks 17(4): 568–571.
Ziegel, P., Shirzadi, S., Wang, S., Bailey, R., Griffiths, P., Ghuwalela, K., Ogedengbe, A. and
Johnson, D. 2014. A data-driven approach to modelling and optimisation for a North sea
asset using real-time data. Paper presented at the SPE Intelligent Energy Conference &
Exhibition, Utrecht, The Netherlands, April 2014. doi: https://doi.org/10.2118/167850-MS.
 Chapter 6
Data-driven and Machine Learning
Approach in Estimating Multi-zonal
ICV Water Injection Rates in a Smart
Well Completion
Daniel Asante Otchere1,2,* and
Mohammed Ayoub Abdalla Mohammed 3

1. Introduction
The ‘smart’or ‘intelligent’well is considered one of the highly developed types
of nonconventional wells. This statement refers to the cutting-edge versions
of wells that have been introduced in recent years. The progressive nature of
these wells is evident in their ability to gather and analyse data, monitor and
control the production process, and adjust production in response to changing
reservoir conditions. The term ‘smart’ or ‘intelligent’ has been coined to
highlight the high level of automation and the use of advanced technology in
these wells. A typical smart well features a customised completion with packers
or sealing components that partition the wellbore with downhole sensors
and pressure control valves fitted on the production tubing across separate
reservoir intervals in a heterogeneous formation. Downhole sensors allow for
continuous monitoring of temperature and pressure across the reservoir and

1
Centre of Research for Subsurface Seismic Imaging, Universiti Teknologi PETRONAS, 32610, Seri
Iskandar, Perak Daril Ridzuan, Malaysia.
2
Institute for Computational and Data Sciences, Pennsylvania State University, University Park, PA,
USA.
3
Chemical and Petroleum Engineering, UAE University, Sheik Khalifa Street at Tawam R/A, Maqam
District, Al Ain, United Arab Emirates.
* Corresponding author: [email protected]
 Machine Learning Approach in Estimating Multi-zonal ICV Rates 105

control valves, which can be used to calculate approximately zonal flow rates.
In contrast, downhole valves like Inflow Control Valves (ICVs) are flexible
in controlling zonal flow rates and are used as a control variable in optimising
wells to enhance recovery. A smart well can be multilateral with an ICV
controlling each lateral or a single bore well with an ICV controlling each
zone. Most of the new oil and gas field developments include smart wells.
They achieve the desired output while lowering capital and operating costs.
Smart well completion is a technology used to optimise oil and gas reservoir
production. It involves collecting, transmitting, and analysing completion,
production, and reservoir data, allowing for remote selective zonal control.
The utilisation of this technology contributes to the improvement of production
and ultimate recovery, as well as the decrease in capital and operating
expenditures. Smart well completion systems are meticulously designed
to cater to the global demand for intelligent completions, even in the most
difficult conditions. These systems are comprised of a variety of components,
such as permanent monitoring and downhole control systems, zonal isolation
and interval control devices, distributed temperature sensing systems, surface
control and monitoring systems, data acquisition and management software,
as well as system accessories. Multi-zonal reservoirs present a unique set of
challenges for oil and gas production, and smart well completion technology
is well suited to meet these challenges. By dividing the well completion into
multiple production zones that can be controlled independently, intelligent
well completion allows for selective control of production from different
zones in the reservoir. This enables effective management of water injection,
gas and water breakthrough, and individual zone productivity, which can help
to increase ultimate recovery and reduce capital and operating expenditures.
Intelligent well completion technology can also improve the management
of water injection and gas breakthroughs, which can reduce water production
and increase ultimate recovery. By remotely controlling the ICVs, it is
possible to manage the flow of fluids from different zones in the reservoir,
leading to improved production efficiency and reduced costs. In addition to
the benefits mentioned, intelligent well completion technology also allows for
the reduction of the number of wells required for field development, which can
lead to a decrease in drilling and completion costs. Furthermore, it allows for
water management through remote zonal control, reducing surface handling
facilities’ size and complexity.
Operators of a multi-zone intelligent injection well face a problem
in estimating and managing fluid distribution through zonal ICVs in
each reservoir zone. Based on available wellbore, fluid, and well-string
data, including the setting of the zonal ICVs, geometry size, production
string, injection fluid properties, wellhead P/T data, and zonal reservoir
pressure/injectivity data, a simple yet precise empirical correlation has been
106 Data Science and Machine Learning Applications in Subsurface Engineering

developed to estimate the volume of injection fluids to each reservoir zone.

However, these equations can become out of sync and sometimes require
recalibration. New downhole pressure regimes, new fluids (water/gas), and
wellhead pressure are all situations that necessitate recalibration.
As a Production Technologist, one of the most important works is
to estimate the oil production/water injection from/to multiple reservoir
zones based on operational parameters such as bottomhole pressure, tubing
differential pressure, and wellhead pressure. Most of these workflows or
empirical correlations assume physics in some sort, which involves many
assumptions. Hence, this research is being proposed to use a data-driven
approach. This approach uses artificial intelligence (AI) techniques where
the computer learns physics based on data patterns without assumptions and
complex physics. This concept proposes advanced data analytics techniques,
allowing new insights and understanding of the operational parameters
(Otchere et al., 2022b). Advanced data analytics and machine learning will
be used to build a model and predict water injection volumes into multiple
reservoirs using daily production or injection data from an actual field. The
main objectives of this work are to;
1. Study and understand the key operational parameters that are necessary to
estimate zonal rates through the use of advanced data analytics techniques.
2. Create a reliable method using machine learning models to calculate
production rates in multi-zonal reservoirs.
Given the success of AI in the oil and gas business, this initiative has
tremendous research possibilities. Current research applies machine learning
techniques to optimise ICV settings to increase production. In this project,
eight machine learning models will be employed to estimate the injected
water volumes into each reservoir when ICV settings are out of sync, giving
wrong zonal estimates. This approach will replace recalibrating the empirical
correlation to estimate injected water into multiple zones, resulting in
interrupted field production. This approach will be reapplied in oil production
wells with various input parameters upon acceptable results. This research
could provide a simple and accurate data-driven approach to estimating each
reservoir unit’s volume of production fluids based on operational parameters.
The method can also provide real-time estimates of produced volumes, aiding
reservoir management plans. The current results submitted along this proposal
could be extended to an advanced formulation of the reservoir management
plan, daily production operational changes to meet critical targets and the
operator’s policy of developing domestic oil resources responsibly.
 Machine Learning Approach in Estimating Multi-zonal ICV Rates 107

2. Brief Overview of Intelligent Well Completion

A smart well completion system improves production by gathering,
transmitting, and analysing data related to completions, production, and
reservoirs. This allows for remote control of specific zones, ultimately
leading to increased efficiency in the reservoir. The system helps to boost
production through the blending of production from different zones, enables
better management of water injection and breakthroughs, and reduces both
capital and operating expenses by minimising the number of wells needed and
optimising production without costly interventions. The smart well system is
designed to operate in various environments and can include a combination
of devices for zonal isolation, interval control, downhole control, permanent
monitoring, and temperature sensing, as well as software for data acquisition
and management.

2.1 ICV Setting and Determination

Multiple production simulations should be conducted to optimise the design
and sizing of ICVs. The ICV size has a significant impact on the production
rate, which can vary widely across different fields. To achieve specific
goals, such as maximising oil recovery or minimising water production, it
is essential to determine the optimal configuration for the ICVs. This can be
accomplished by designing the ICVs in a way that ensures each setting yields
a unique production rate. This approach greatly influences the optimisation
process, as the resulting data provides valuable insights into the optimal ICV
configuration for achieving the desired outcome.
In the domain of ICV design, the process for determining the optimal design
is usually contingent on the production capability of each lateral. However, in
practice, several oil companies have streamlined the process by employing the
average field production rate as a basis for design, rather than individual lateral
rates. This approach is often deemed more expedient, particularly for logistical
purposes, including maintenance and surveillance. For this research, ICV size
was selected based on individual zonal flow at a particular setting, commingled
flow from two zones iteratively and commingled flow from all three zones
whiles monitoring downhole pressures. The ICV settings were discretised so
that each setting resulted in production rates corresponding to zonal pressures.
Several factors taken into account in the design of the ICV include:
1. Simulations were run to establish the minimum and maximum production
rates for each zone where all zones were commingled and when a single
zone was produced sequentially.
2. The discretisation of the ICV region was accomplished by establishing
predefined intermediate settings (ranging from 0 to 10) that correspond
108 Data Science and Machine Learning Applications in Subsurface Engineering

to the minimum and maximum production rates. The ICV region was
segmented into these settings to provide a quantitative representation of
the valve’s opening, where 0 denotes complete closure and 10 denotes
complete openness.
3. In accordance with downhole pressure monitoring, the maximum
production rate can be achieved by adjusting the ICV (intelligent
completion valve) to its highest setting range (7–10), whereas the
minimum ICV setting range (0–3) results in a production rate of zero.
4. If production rates significantly differed between zones, different sized
ICVs were used appropriately, although this may not be a viable technique.
5. The new ICV settings were tested on different base cases of all zones
producing and one zone producing individually. If more than half of
the ICV settings in a zone gave the same production rate, the maximum
production rate was reduced. Also, the cases simulated were to capture
instances where a particular zone experienced production issues
(gas/water breakthroughs) and had to be shut in. Although the data
was applied to an injection, this approach was simulated considering a
production well to capture various scenarios and production issues.

2.2 Literature Review of ICV Innovations and Machine

Learning Applications
In literature, ICVs have been widely studied as a means of optimising
production in oil and gas reservoirs. These valves are used to manage the
inflow of fluid into a well, allowing for selective control of production from
different zones in the reservoir. This selective control enables effective
management of water injection, gas and water breakthrough, and individual
zone productivity, which can help increase ultimate recovery and reduce
capital and operating expenditures.
AI has been successfully implemented in the petroleum industry, including
reservoir characterisation (Otchere et al., 2021), seismic interpretation
(Otchere et al., 2022c), reservoir engineering (Otchere et al., 2022a) and
many other sectors. Machine learning techniques have also been applied to
the optimisation of ICV valves. Researchers have used machine learning
algorithms to model the relationship between ICV settings and production
rates and to predict the optimal ICV settings for a given set of conditions.
These models can take into account various factors such as wellbore geometry,
fluid properties, and reservoir characteristics.
Various techniques have been suggested to enhance the productivity
of smart wells. For example, Mubarak et al. (2008) employed an intensive
production test to reduce water production from a trilateral smart well, where
the laterals were found to be sensitive to ICV settings. Different settings
 Machine Learning Approach in Estimating Multi-zonal ICV Rates 109

were observed to be practical for different sources of water production.

Similarly, Jalali et al. (1998) improved the delivery rate of a smart gas well
by unchoking the bottom layer gradually after producing the top layer without
any restriction. In addition, Yeten et al. (2002) employed a gradient-based
approach to optimise the cumulative oil recovery from smart wells. Naus
et al. (2006) put forward a methodology to determine the changes in flow
rate due to ICV settings and tested it in two reservoirs with the objective
of achieving maximum ultimate recovery. Furthermore, Brouwer and Jansen
(2002) evaluated the impacts of smart completions with several well targets
and limitations, while Alhuthali et al. (2010) used smart wells and optimal
rate control for waterflood optimisation. Yeten (2003) proposed a general
methodology that uses genetic algorithms, hill climbing, and artificial neural
networks (ANNs) to optimise well type, trajectory, location, and ICV settings.
Another study, conducted by Behrouz et al. (2016), is the use of an intelligent
well technology to develop a new method for selecting and ranking candidate
wells and fields, determining interval control valve (ICV) size, and optimising
ICV settings. An efficient ICV setting optimisation in an intelligent well was
proposed, which maximised cumulative oil, minimised water production
or conducted both. Real case studies were considered to demonstrate the
effectiveness and robustness of the proposed methodology, resulting in
a considerable improvement in the objective function using the developed
methodology.
Huang et al. (2011) aimed to develop an intelligent well completion
system in China by optimising downhole monitoring, data transmission,
intelligent optimisation, and production control systems. The authors
developed an ICV system with proprietary internet protocol (IP) and created
the first set of intelligent completion system (ICS) in China for measurable
and optimisable oil well exploitation. Results showed that the system is stable
and reliable, resulting in a 10.5% increase in oil production. Malakooti et
al. (2020) proposed an integrated control and monitoring (ICM) algorithm
that can enhance production from multi-zone I-wells. The algorithm employs
a two-level optimisation process to maximise either projected zonal or oil
production reliability. The authors validated the algorithm by subjecting it to a
commercial transient wellbore simulator. The simulation results demonstrate
that the ICM algorithm can optimise oil production while reducing the
number of flow tests required. Furthermore, the study highlights that the
exact estimation of zonal properties is not essential to achieve maximum oil
production in multi-zone I-wells. The algorithm’s efficacy suggests that it
may be applicable to optimising multi-well flow rate allocation and the start-
up process of multi-zone I-wells.
In literature, researchers have also used other machine learning techniques
such as reinforcement learning, decision tree, and fuzzy logic to optimise the
110 Data Science and Machine Learning Applications in Subsurface Engineering

ICV valves. These techniques have shown promising results in finding the
optimal ICV settings and improving the performance of oil and gas reservoirs.
Overall, the literature review suggests that combining ICV data and machine
learning techniques can be an effective approach for optimising production
in oil and gas reservoirs. Machine learning algorithms can be used to model
the relationship between ICV settings and production rates and to predict the
optimal ICV settings for a given set of conditions. This can help to increase
ultimate recovery, reduce capital and operating expenditures, and improve the
performance of oil and gas reservoirs.

3. Methodology
In the context of ICV design, the process usually relies on the production
capacity of individual laterals. Notwithstanding, some oil companies opt for
a simplified approach by utilising the mean production rate of the field rather
than individual lateral rates.
1. Data acquisition: For this research, data is to be acquired from the
production technology team in the subsurface department. The data
required from multiple wells are SPFM WI Rate, Annulus Pressure,
Tubing Pressure, ICV dP, ICV position, Annulus Temperature for the
separate reservoir zones (lower, middle and upper), MPFM Rate, THT
(tubing head temperature), Annulus Pressure, THP (tubing head pressure),
Tubing dP (differential pressure). Having some ICV mappings in the
wells of interest is also essential. The data from the mapping helps gain
insight and train the machine learning model on accurate data. The data
were grouped per hour and within a two-year range.
2. Data preprocessing and preparation: Based on the collected data,
preprocessing techniques are used to clean the data. Missing values and
data variables were removed to prepare the data for further analysis. The
total number of data used, based on the descriptive statistics after the
removal of missing data, is 3,767 data points.
3. Exploratory data analysis: The most critical and time-consuming part
of any machine learning project is the exploratory data analysis. It is
the first direct encounter with the data towards understanding it. Initial
data investigations were carried out to discover patterns in data, spot
anomalies, and check assumptions with the help of summary statistics,
heatmaps, and pair plot visualisation.
4. Explainable AI and feature selection: In this step, the various input
features are inspected and checked for their importance. Based on the
results from the data analysis, feature selection is to be performed using
explainable AI techniques (model agnostic metrics). The approach is
 Machine Learning Approach in Estimating Multi-zonal ICV Rates 111

different from the standard methods that operate on correlation. Since

correlation does not mean causation, the model agnostic metric is used to
indicate the variables that have a causal effect on estimating zonal rates.
5. Machine Learning: Several advanced machine learning models were
used to estimate the zonal rates for each reservoir from the data provided.
A holdout cross-validation of 10% was applied to segregate the data into
training (3,390), and test (377) sets. The performance of these models will
be compared using several error metrics and visualisation techniques. The
model with the least errors, high accuracy, and high probability of fit will
be selected and deployed.
The aim is to assess when the empirical correlation goes out of sync and if
the model performance deteriorates over time, assuming all conditions remain
the same. A suitable machine learning model that can estimate zonal rates
can be used to estimate the required injection volumes needed for specific
zones. This model can help production technologists plan the optimum ICV
positions and tubing head injected volumes required to sustain pressures in
the relevant zones. This approach will also limit the need for strategic ICV
mapping during pressure build up (PBU) and delays in bringing production
up to the required optimal level before shut-in. A schematic of the workflow
is illustrated in Fig. 6.1.

Fig. 6.1. Detailed schematic workflow of the methodology.

3.1 A Brief Overview of Models Used in This Study

The summary analysis of the reviewed models used in estimating zonal rates
in this study is summarised in Table 6.1.
112 Data Science and Machine Learning Applications in Subsurface Engineering

Table 6.1. Summary of models used in this study.

Base learner Description Developed or

Implemented
by Authors
Bootstrap aggregation Bootstrap aggregation of trees is used to reduce the Breiman
(Bagging) variance for algorithms that have high variance. (1996)
Ridge Regression Ridge regression is a robust technique against Kuhn and
multicollinearity. The algorithm minimises standard Johnson
errors by applying some bias to the model estimates, (2013)
resulting in a more reliable prediction. The algorithm
calculates the difference between the means of the
standardised dependent and independent variables and
dividing by their standard deviations.
Least absolute Adds a penalty term to the regression equation that Tibshirani
shrinkage and forces the sum of the absolute values of the regression (1996)
selection operator coefficients to be less than a specific value. This has
(Lasso) Regression the effect of shrinking the coefficients of less important
variables to zero, effectively removing them from the
model.
KNeighbors It is based on the idea that data points with similar Silverman
Regressor features tend to have similar output values. The KNN and Jones
regressor predicts the output value for a new data (1989)
point by looking at the K-nearest training data points
and averaging their output values. The value of K
determines the number of neighbours considered in the
prediction, and the distance metric used to measure the
similarity between data points can be chosen based on
the problem at hand.
Decision Tree It uses induction and pruning techniques to build Gordon et al.
hierarchical decision boundaries and remove (1984)
unnecessary structures from the decision tree to battle
overfitting.
Extreme Gradient An implementation of gradient boosted decision trees of Chen and
Boosting (XGBoost) the first and second-order to maximise the loss function Guestrin
adding an extra regularisation term to adjust the final (2016)
weights acquired to avoid overfitting.
Extremely Potentially performs better than the random forest, Geurts et al.
Randomised Trees although it uses a simpler algorithm to construct the (2006)
(Extra tree) decision trees used as ensemble members.
Random Forest An advanced decision tree that is robust against Breiman
overfitting and offers easy interpretability. (2001)

3.2 Criteria for Model Evaluation

Successfully training a model and achieving a good convergence does not
signify the end of the model process. In model regression analysis, the
model’s error is the difference between the actual data points and the best fit
 Machine Learning Approach in Estimating Multi-zonal ICV Rates 113

line produced by the algorithm. There is not a single statistic that can analyse
all forms of data. The metrics are impacted by a number of components,
including but not limited to the presence of outliers, the machine learning
algorithm used, the ease with which derivatives can be discovered, and the
prediction confidence. Therefore, the model errors were evaluated using the
following standards to determine whether the models utilised in this research
were appropriate;
1. Mean Absolute Error (MAE): The MAE is a well-established evaluation
metric used to quantify the accuracy of predictive models. It is defined
as the L1 loss, representing the sum of absolute differences between
the predicted output and the target variable. By measuring the average
magnitude of errors without taking into account the direction of errors, the
MAE provides a reliable measure of model performance. Additionally,
this metric is particularly sensitive to relative errors, making it a useful
tool for assessing performance in applications where errors of all sizes are
equally important. The MAE is also robust against global scaling of the
predicted output and the presence of outliers, making it a versatile metric
for evaluating model performance in various settings. Mathematically,
the MAE is written as;

∑
n
yi − ŷi
MAE = i =1
(6.1)
n
2. Root Mean Squared Error (RMSE): This performance evaluation metric is
popular because it is interpretable as the standard deviation of the model’s
prediction errors and specifies the closeness of predicted data to actual
data. This function is written as follows;

∑
n
( yi − yˆi ) 2
RMSE = i =1 (6.2)
n
3. Akaike Information Criterion (AIC): This evaluation standard was
developed using a frequentist probability framework that assigns a model
a score based on its maximum likelihood estimation. This technique
assesses models’ quality and accuracy, resulting in a better model fit of
the data. This criterion is expressed as;
AIC = 2K – 2(log – likelihood) (6.3)
In order to assess the accuracy of prediction models based on the number
of input features, various evaluation metrics are employed. However, except
for R2, a lower value for the performance evaluation criteria indicates
better model performance. If a conclusion cannot be drawn based on these
114 Data Science and Machine Learning Applications in Subsurface Engineering

parameters, a ranking rule is employed to determine precedence. Specifically,

the ranking order is established as follows: MAE, RMSE, AIC, and R2.

4. Results and Discussion

4.1 Explainable AI
All 21 input features were used to train eight models, and their results are
illustrated in Fig. 6.2. From the results, the worst performing models are
the Lasso and Ridge algorithms. This result indicates that linear models are
inappropriate for this study due to the high dimensional input data.
The RF model was used for the multi-output prediction of the smart well
injection into the upper, middle, and lower reservoir under study. The same
training and testing datasets used in training all the models were used in this
approach. The Permutation Feature Importance (PFI) analysis was used to
identify the relevant features that can help predict the volume of water being
injected in multiple zones. Model agnostic metrics are useful since machine
learning models are designed to be interpretable to help understand how a
model generated a given prediction.
Figure 6.3 visualises the PFI values as absolute values for both train and
test data. Many input variables have a significantly low importance score from
the results. This result implies that a small number of variables can capture
the predictive value of these input variables. The PFI plot demonstrates that

Fig. 6.2. Evaluation of models using all features.

 Machine Learning Approach in Estimating Multi-zonal ICV Rates 115

Fig. 6.3. PFI of initial input features.

permuting a variable reduces the multi-zonal prediction accuracy. According

to this analysis, some of the characteristics are more significant than others. It
is essential to state that this analysis was done for both the training and testing
dataset to help account for features that may help with the generalisation
power of the model. It was noticed that the top essential features are the same
for both training and testing datasets. Features such as WI_Rate, the zonal
ICV position, zonal annulus temperature, the tubing, and ICV differential
pressures were considered essential. This result suggests that these features
will contribute to the generalisation power of the model. If a feature is deemed
necessary for the train set but not for the testing, this feature will probably
cause the model to overfit. An example of this feature is the UZ_ICV_dP.
Again, when all logs, including the new features, were inputted, the mean test
accuracy was 0.998.
As such, sensitivity analysis were performed to identify the number of
input features relevant to this study. Using the six high-performing models
above, the top 10 and seven features were analysed against all features using
accuracy. From the results in Fig. 6.4, the models scored high accuracy when
seven input features were used. The features exhibiting causal relations to
the three targets are WI_Rate, UZ (upper zone)_Tub_Pres, MZ (middle zone)
_Tub_Pres, LZ (lower zone)_Tub_Pres, UCV_(upper control valve) Position,
MCV (middle control valve)_Position, LCV (lower control valve)_Position.
116 Data Science and Machine Learning Applications in Subsurface Engineering

Fig. 6.4. Sensitivity analysis on feature selection based on PFI analysis.

Per domain knowledge, these features relate to the total injected volume, the
zonal ICV position, which indicates that a particular zone is closed or open,
and the tubing pressures of each zone.

4.2 Model Evaluation

Further evaluation was conducted on the seven features selected as relevant
to identify the top-performing models for further optimisation. Significantly,
the ET and RF models exhibited superior performance compared to the other
models, even though none of them demonstrated overfitting when evaluated
on the holdout data. As a result, we conducted additional analysis to assess
the prediction errors of the models. Figure 6.5 illustrates the accuracy and

Fig. 6.5. Comparison of RF and ET models using the seven relevant features to all input features.
 Machine Learning Approach in Estimating Multi-zonal ICV Rates 117

consistency of all the recommended models, based on the error measurements.

MAE, RMSE, and R2 results indicate that the seven features resulted in massive
improvements in model performance. Although the RF model achieved a
higher test accuracy, the ET model performed better comparatively in terms
of MAE and RMSE. In a broader perspective, the seven features resulted in a
40% and 47% reduction in the MAE and RMSE, respectively.

4.3 Sensitivity Analysis

Based on the test data, Figs. 6.6 to 6.8 show the kernel density estimation
(KDE) of the expected and actual zonal rates. The actual values are depicted
in red, while the expected values are indicated in blue. The zonal injection
rates predicted by the ET model are much closer to the observed test data. This
suggests that the ET model is capable of accurately representing the diverse
range of the data, which makes it appropriate for use in predicting zonal
injection rates for other wells in the same field. In addition, the assessment
metrics discussed earlier are confirmed by the sensitivity analysis.
Figures 6.9 to 6.11 show the joint plot and regression combination to
visualise the predicted zonal and actual test data zonal rates and error margins.
The joint plot function is a distribution plot that shows univariate histogram
KDE curves and bivariate graphs. Analysis of the regression plot and 95%
confidence interval illustrated in these figures indicate that the ET model
using the seven PFI selected features performed accurately. The ET model
showed a higher confidence interval and matched the range of injected
volumes based on the histogram and KDE plot. However, the confidence

Fig. 6.6. Kernel density estimate demonstrating the similarity between the estimated and actual test
upper zonal rates.
118 Data Science and Machine Learning Applications in Subsurface Engineering

Fig. 6.7. Kernel density estimate demonstrating the similarity between the estimated and actual test
middle zonal rate.

Fig. 6.8. Kernel density estimate demonstrating the similarity between the estimated and actual test
lower zonal rate.

level tends to deteriorate as high volumes of water were predicted for the
upper zone, which is analysed further. For the ET model, most data after
200,000 bbls (barrels) were widely scattered due to insufficient high injected
volume data into that zone. The ET model, however, successfully captured the
 Machine Learning Approach in Estimating Multi-zonal ICV Rates 119

Fig. 6.9. Cross-plot of ET predicted vs actual upper zonal rates.

Fig. 6.10. Cross-plot of ET predicted vs actual middle zonal rates.

trends despite insufficient data from this analysis. Although the lower zone
experienced similar data inadequacy in high injected volumes, the model
did not experience low confidence in its prediction. This observation can be
attributed to the variation in the training data and the wide spread of test data
for the lower zone compared to the upper zone with a high concentration of
low injection volumes.
120 Data Science and Machine Learning Applications in Subsurface Engineering

Fig. 6.11. Cross-plot of ET predicted vs actual lower zonal rates.

4.4 Model Deployment

The excellent results achieved for the multi-zonal water injection rates by the
ET model using the model agnostic selected features gave confidence in its
deployment. Data from another well with calculated ICV zonal rates when the
ICV curves used to determine the rates were in sync was used. The data had
3,767 data points representing about five months of hourly data. Figure 6.12
shows the results of the predicted rates for the injected well the model was
deployed on. The accuracy of the combined zonal rate prediction to the actual
water injection is apparent and confirms the accuracy of the model. The zonal
rates, illustrated in Figs. 6.13 to 6.15, will show the predicted rates and how

Fig. 6.12. Total water injection rate vs Combined predicted zonal rates.
 Machine Learning Approach in Estimating Multi-zonal ICV Rates 121

Fig. 6.13. Comparison of actual and predicted upper zonal ICV rates against ICV position.

Fig. 6.14. Comparison of actual and predicted middle zonal ICV rates against ICV position.

Fig. 6.15. Comparison of actual and predicted lower zonal ICV rates against ICV position.
122 Data Science and Machine Learning Applications in Subsurface Engineering

they make predictions based on ICV positions. This is to confirm that all the
predicted zonal rates went to the actual zone where the ICV was opened.

5. Conclusions
In summary, eight supervised machine learning models were employed to
predict multi-zonal rates for a smart water injection well in an oil and gas
production setting. The machine learning models were trained and evaluated
using PFI, a model-agnostic metric, to identify relevant features. Seven of the
features were relevant and used as input to further train the top six models
from the previous evaluation where all features were used. The Extra Trees
model achieved the highest precision and consistency of 708 bbls RMSE and
68 bbls MAE compared to the test data. The Extra Trees predicted zonal rates
were also visualised using KDE and joint plots to confirm their accuracy.
Upon satisfactory results, the Extra Trees model was deployed on a new
well with five months of hourly data, and the combined predicted zonal
rates matched the total injected rate. Additionally, the model could predict
zonal rates in instances where the ICVs were in the closed and fully opened
positions. Overall, the results of this research demonstrate the potential of
machine learning in predicting multi-zonal rates in oil and gas production and
highlight the use of the Extra Trees model as a robust and effective tool for
this task.
This research provides preliminary results for the approach employed in
water injection wells. When used in producer wells, this method could provide
a simple and accurate data-driven approach to estimating each reservoir
unit’s volume of produced fluids based on downhole parameters. The method
can also provide real-time estimates of produced volumes, aiding advanced
formulation of a reservoir management plan, daily production operational
changes to meet critical targets, and the operator’s policy of developing
domestic oil resources responsibly. This project’s success will not end ICV
mapping but will drastically reduce the number of mappings performed in the
life of the field. One limitation of this study is that the model is as good as the
data used to train it. When new data is introduced that falls out of the range
of the model, ICV mapping will be required, and the model will be retrained
with the new data to fine-tune its prediction. This limitation, however, does
not downplay the immense advantages this study provides. This approach is
recommended to save time and money during production using a smart well
completion.

Code Availability
The Jupyter Notebook used in this study is hosted at https://github.com/
ascotjnr/Smart-Well-Completion.
 Machine Learning Approach in Estimating Multi-zonal ICV Rates 123

Acknowledgement
The authors wish to thank the University Teknologi Petronas and the Centre
for Subsurface Seismic Imaging and Hydrocarbon Prediction for supporting
this work.

References
Alhuthali, A.H., Datta-Gupta, A., Yuen, B. and Fontanilla, J.P. 2010. Field applications
of waterflood optimisation via optimal rate control with smart wells. SPE Reservoir
Evaluation & Engineering 13: 406–422. https://doi.org/10.2118/118948-PA.
Behrouz, T., Rasaei, M.R. and Masoudi, R. 2016. A novel integrated approach to oil production
optimisation and limiting the water cut using intelligent well concept: using case studies.
Iranian Journal of Oil and Gas Science and Technology 5: 27–41. https://doi.org/10.22050/
IJOGST.2016.13827.
Breiman, L. 1996. Bagging predictors. Mach Learn. 24: 123–140. https://doi.org/10.1007/
bf00058655.
Breiman, L. 2001. Random forests. Mach. Learn. 45: 5–32. https://doi.
org/10.1023/A:1010933404324.
Brouwer, D.R. and Jansen, J.D. 2002. Dynamic optimisation of water flooding with smart
wells using optimal control theory. In: European Petroleum Conference. SPE, Aberdeen.
https://doi.org/10.2118/78278-MS.
Chen, T. and Guestrin, C. 2016. XGBoost: A scalable tree boosting system. pp. 785–794.
In: Proceedings of the ACM SIGKDD International Conference on Knowledge Discovery
and Data Mining. ACM, New York, NY, USA. https://doi.org/10.1145/2939672.2939785.
Geurts, P., Ernst, D. and Wehenkel, L. 2006. Extremely randomised trees. Mach. Learn.
63: 3–42. https://doi.org/10.1007/s10994-006-6226-1.
Gordon, A.D., Breiman, L., Friedman, J.H. Olshen, R.A. and Stone, C.J. 1984. Classification
and regression trees. Biometrics 40: 874. https://doi.org/10.2307/2530946.
Huang, Z., Li, Y., Peng, Y., Shen, Z., Zhang, W. and Wang, M. 2011. Study of the intelligent
completion system for liaohe oil field. Procedia Eng. 15: 739–746. https://doi.org/10.1016/j.
proeng.2011.08.138.
Jalali, Y., Bussear, T. and Sharma, S. 1998. Intelligent completion systems: the reservoir
rationale. In: All Days. SPE. https://doi.org/10.2118/50587-MS.
Kuhn, M. and Johnson, K. 2013. Applied Predictive Modeling. New York, NY: Springer New
York, https://doi.org/10.1007/978-1-4614-6849-3.
Malakooti, R., Ayop, A.Z., Maulianda, B., Muradov, K. and Davies, D. 2020. Integrated
production optimisation and monitoring of multi-zone intelligent wells. J. Pet. Explor.
Prod. Technol. 10: 159–170. https://doi.org/10.1007/s13202-019-0719-5.
Mubarak, S.M., Pham, T.R., Shamrani, S.S. and Shafiq, M. 2008. Case study: the use of
downhole control valves to sustain oil production from the first maximum reservoir contact,
multilateral, and smart completion well in ghawar field. SPE Production & Operations
23: 427–430. https://doi.org/10.2118/120744-PA.
Naus, N.M.J.J., Dolle, N. and Jansen, J.-D. 2006. Optimisation of commingled production using
infinitely variable inflow control valves. SPE Production & Operations 21: 293–301.
https://doi.org/10.2118/90959-PA.
Otchere, D.A., Arbi Ganat, T.O., Gholami, R. and Lawal, M. 2021. A novel custom ensemble
learning model for an improved reservoir permeability and water saturation prediction.
J. Nat. Gas Sci. Eng. 91: 103962. https://doi.org/10.1016/j.jngse.2021.103962.
124 Data Science and Machine Learning Applications in Subsurface Engineering

Otchere, D.A., Abdalla Ayoub Mohammed, M., Ganat, T.O.A., Gholami, R. and Aljunid
Merican, Z.M. 2022a. A novel empirical and deep ensemble super learning approach
in predicting reservoir wettability via well logs. Applied Sciences 12: 2942. https://doi.
org/10.3390/app12062942.
Otchere, D.A., Ganat, T.O.A., Nta, V., Brantson, E.T. and Sharma, T. 2022b. Data analytics
and Bayesian Optimised Extreme Gradient Boosting approach to estimate cut-offs from
wireline logs for net reservoir and pay classification. Appl. Soft Comput. 120: 108680.
https://doi.org/10.1016/j.asoc.2022.108680.
Otchere, D.A., Tackie-Otoo, B.N., Mohammad, M.A.A., Ganat, T.O.A., Kuvakin, N.,
Miftakhov, R., Efremov, I. and Bazanov, A. 2022c. Improving seismic fault mapping
through data conditioning using a pre-trained deep convolutional neural network: A
case study on Groningen field. J. Pet. Sci. Eng. 213: 110411. https://doi.org/10.1016/J.
PETROL.2022.110411.
Silverman, B.W. and Jones, M.C. 1989. E. Fix and J.L. Hodges (1951): An important contribution
to nonparametric discriminant analysis and density estimation. International Statistical
Review 57: 233–247.
Tibshirani, R. 1996. Regression shrinkage and selection via the lasso. Journal of the Royal
Statistical Society. Series B (Methodological) 58: 267–288.
Yeten, B., Durlofsky, L.J. and Aziz, K. 2002. Optimisation of smart well control. In: SPE
International Thermal Operations and Heavy Oil Symposium and International Well
Technology Conference. Calgary.
Yeten, B. 2003. Optimum Deployment of Nonconventional Wells (PhD). Stanford University,
California.
 Chapter 7
Carbon Dioxide Low Salinity Water
Alternating Gas (CO2 LSWAG)
Oil Recovery Factor Prediction in
Carbonate Reservoir
Using Supervised Machine Learning Models
Eric Thompson Brantson,1,* Zainab Ololade Iyiola,1
Yao Yevenyo Ziggah,2 Alexander Ofori Mensah,1
Daniel Asante Otchere,3,4 Efua Eduamba Abakah-Paintsil1
and Emmanuel Karikari Duodu1

1. Introduction
With rising global energy demand and diminishing oil reserves, enhanced oil
recovery (EOR) from existing brownfields is becoming increasingly important
(Sheng, 2011). As production from the petroleum reservoir increases, the
reservoir’s inherent primary energy depletes, resulting in insufficient reservoir
pressure to drive the oil to the surface. As a result, secondary recovery methods
(water or gas injection) are needed to boost production output. The oil
recovered through both primary and secondary processes ranges from about
1
Department of Petroleum and Natural Gas Engineering, School of Petroleum Studies, University of
Mines and Technology, Tarkwa, Ghana.
2
Department of Geomatic Engineering, Faculty of Geosciences and Environmental Studies,
University of Mines and Technology, Tarkwa, Ghana.
3
Centre of Research for Subsurface Seismic Imaging, Universiti Teknologi PETRONAS, 32610, Seri
Iskandar, Perak Daril Ridzuan, Malaysia.
4
Institute for Computational and Data Sciences, Pennsylvania State University, University Park, PA,
USA.
* Corresponding author: [email protected]
126 Data Science and Machine Learning Applications in Subsurface Engineering

20%–40% of the original oil in place (OOIP) (Stalkup, 1984). Following the
application of primary and secondary oil recovery techniques, two-thirds of
the OOIP remains in the reservoir (Gbadamosi et al., 2019).
According to Brantson et al. (2019), tertiary oil recovery methods are
capable of retrieving a greater amount of oil than primary or secondary
recovery methods. In this study, EOR synergy methods involving the injection
of low salinity water (LSW) and carbon dioxide (CO2) gas were employed
to enhance oil recovery at the field scale. The use of LSW in field trials has
demonstrated a significant improvement in oil recovery (Robertson, 2007;
Webb et al., 2004; Vledder et al., 2010), making it a favourable option
over conventional chemical EOR methods in terms of chemical costs,
environmental impact, and field process implementation (Dang et al., 2014).
On the other hand, CO2-EOR has the potential to recover an additional
15%–20% of the remaining oil in place (Ahmed, 2018). The two main
mechanisms behind CO2 injection schemes are oil swelling and viscosity
reduction (AlQuraishi et al., 2019). It is noted that all reservoir lithologies,
including siliciclastic, carbonate, and others, are appropriate for CO2-EOR
provided they have an enough seal to contain hydrocarbons and interconnected
pore space for fluid accumulation and flow (Verma, 2015). Therefore,
estimating hydrocarbon’s ultimate recovery factor provides more detailed
insights into oilfield development strategies (Roustazadeh et al., 2022). The
methods of material balance, decline curve analysis, and dynamic numerical
simulation are used to estimate recovery factors, but they are time-consuming.
Over the years, there has been increasing awareness and concern over
the continuous buildup of greenhouse gases in the atmosphere. This is due to
human activities such as burning of fossil fuels, deforestation, and industrial
processes that emit greenhouse gases such as carbon dioxide (CO2), methane,
and nitrous oxide. The accumulation of these gases has led to an increase
in global temperatures, causing significant changes in the climate, which
poses a threat to the entire world. One of the consequences of climate change
is rising sea levels, which have been observed to be a direct result of the
increasing levels of CO2 and other greenhouse gases in the atmosphere. The
rising sea levels have led to devastating consequences, such as flooding of
coastal areas, displacement of communities, loss of habitats, and destruction
of infrastructure (Santos et al., 2014). Therefore, it is crucial to find ways to
mitigate the impact of greenhouse gas emissions on the environment and the
world at large.
In recent years, there has been considerable attention given to the potential
use of CO2-EOR as a long-term anthropogenic CO2 storage application
(Mandadige et al., 2016). CO2-EOR is a technique used in the oil and gas
industry to extract more oil from wells by injecting CO2 into the reservoir.
This technique not only increases the amount of recoverable oil but also has
 CO2 LSWAG Oil Recovery Factor Prediction in Carbonate Reservoir 127

the added benefit of storing the CO2 underground. This can be an effective way
to mitigate the impact of greenhouse gas emissions on the environment and
provide a solution to the problem of global warming. Therefore, it is important
for researchers, policymakers, and stakeholders to continue exploring the
potential of CO2-EOR as a long-term anthropogenic CO2 storage application,
while also working towards reducing greenhouse gas emissions and finding
sustainable solutions to climate change.
The dominant cause of observed anthropogenic global warming has been
unrestricted CO2 emissions from fossil fuel combustion due to the rising usage
of fossil fuels in producing electricity and other manufacturing processes
(Kharecha et al., 2008). Over the last 35 years, the oil and gas industry has
made numerous technological advancements and operational practices for
injecting CO2 for enhanced oil recovery (Yong et al., 2016). CO2 flooding has
emerged as one of the most promising EOR methods because it uses readily
available, naturally occurring CO2 from reservoirs (Sun et al., 2017). Typical
incremental oil recovery by CO2 flooding ranges between 5%–25% (Wu
et al., 2021). The CO2 injection method has traditionally been used in reservoirs
with an oil gravity of less than 25 (Stosur, 2003).
In the quest to find the optimal use of CO2 in enhanced oil recovery for
secondary and tertiary modes, several techniques have been employed. These
techniques include miscible and immiscible CO2 flooding, CO2 huff-n-puff,
and CO2-foam injection, among others (Christensen et al., 1998). Amongst
these techniques, the combined forms of EOR methods have proven to
optimise significantly oil recovery synergistically (Afzali et al., 2018; Teklu
et al., 2016). A lot of CO2-LSWAG experiments (AlQuraishi et al., 2011;
Dang et al., 2014; Naderi and Simjoo, 2019; Pourafshary and Moradpour,
2019; Zolfaghari et al., 2013) in both carbonate and sandstone reservoirs have
been performed which reported an incremental oil recovery from the initial oil
in place but not much on reservoir scale.
Hybrid low-salinity gas flooding has garnered significant attention from
researchers in recent years (Jiang et al., 2010; Kulkarni et al., 2004) due to its
many advantages, such as low cost, low minimum miscibility pressure, and
environmental friendliness. CO2 is frequently used as the injection gas in this
hybrid approach because of these benefits (Ma and James, 2022). The injection
of LSW changes the gas solubility in water, affecting gas/oil interactions
and ultimately enhancing oil recovery. Despite the potential benefits, there
are varying observations regarding the effectiveness of this hybrid method,
with some researchers reporting improved recovery. Others, however, have
found no improvement in recovery over continuous gas injection (CGI).
Furthermore, CO2 partitioning between oil and water in the reservoir (Dang
et al., 2014) can lead to different recovery rates depending on the reservoir
conditions.
128 Data Science and Machine Learning Applications in Subsurface Engineering

Low salinity CO2-EOR was extensively described in the works of Zekri

et al. (2015) for carbonate cores and Teklu et al. (2016) in carbonate cores and
low permeability sandstone cores. Numerous studies on CO2 LSWAG have
been conducted in sandstone reservoirs with strongly water-wet sandstone
cores and little clay content, which are unfavourable conditions for the LSW
process (Dang et al., 2014) unlike carbonate reservoirs which are complex
and have received relatively little attention. It was reported that by combining
the low-salinity waterflooding method and the CO2-EOR method, a new
LSW-alternating-CO2 (LS-WACO2) EOR method emerges. It was concluded
that the solubility of CO2 in LSW was the main reason for the increased
residual oil mobilisation more than in ordinary WAG. In the work of Dang
et al. (2014), a hybrid LSW alternating miscible CO2 flooding injection was
modelled in a 1D core. Their research indicates that the hybrid approach
is effective in resolving late production problems associated with water
alternating gas WAG injection. Pourafshary and Moradpour (2019) conducted
a comprehensive review of Dang et al.’s (2014) work and recommended further
field-scale simulation studies based on experimental work to fully evaluate the
useful benefits of the hybrid methods developed by Dang et al. (2014). Their
recommendation was based on the promising results obtained by Dang et al.
(2014), and highlights the need for more comprehensive studies to advance
the understanding of this novel method. Furthermore, only very limited
simulation studies have been done with regard to CO2 LSWAG injection in
carbonate reservoirs (Al-Shalabi et al., 2014; Hamoud and Pranoto, 2016).
In 2011, Aleidan and Mamora (2011) conducted a study that aimed to
compare the effects of various water/CO2 injection schemes on oil recovery
from carbonate core samples. The study analysed the impact of SWAG and
WAG injection schemes at different salinities. The researchers discovered that
when they switched to LSW, there was an increase in oil recovery for both
injection methods. This increase was as a result of the hybrid method and the
synergy between gas and LSW injections. Gas dissolution in brine led to a
reduction in fluid mobility, which accounted for the incremental oil recovery.
Jiang et al. (2010) also conducted a study to investigate the impact of water
salinity on the performance of WAG during miscible flooding in sandstones
using highly viscous crude oil. The study aimed to examine the effect of
low-salinity water on the effectiveness of WAG. The researchers discovered
that in LSW, a higher gas solubility regulated the mobility ratio of water
and viscous oil. This regulation occurred because the gas solubility in brine
increases as salinity decreases, leading to a reduction in gas mobility and an
increase in the viscosity of the brine. Consequently, the mobility of water and
viscous oil became more balanced, which resulted in improved oil recovery
rates. In addition to the findings mentioned previously, the research conducted
by the team also showed that LSW is more effective than high-salinity water in
 CO2 LSWAG Oil Recovery Factor Prediction in Carbonate Reservoir 129

improving oil recovery rates during high viscous crude oil sandstones miscible
flooding. This is a critical finding that highlights the importance of considering
the impact of water salinity on WAG performance during miscible flooding.
The team’s results demonstrate that the use of LSW can significantly increase
oil recovery rates, providing further evidence of the potential benefits of LSW in
enhanced oil recovery techniques. These findings have important implications
for the petroleum industry, as they provide valuable insights into how water
salinity can affect WAG performance, and how this knowledge can be used to
optimise oil recovery rates in reservoirs with highly viscous crude oil.
In their experimental study, Al-Abri et al. (2019) investigated the effect
of hybrid injection of immiscible CO2 and smart water on sandstone core
samples. The study found that the synergy between gas injection and the
various ions present in the water samples led to a significant improvement
in oil recovery. To conduct the study, the researchers utilized three synthetic
brines, each containing 5000 ppm of MgCl2, NaCl, and KCl, respectively.
The research revealed that while MgCl2-containing water had the highest
solubility of CO2 in brine, it resulted in the lowest oil recovery among the
three tests. Furthermore, the multicomponent ion exchange of the smart water
altered the rock’s wettability, making it more water-wet, thus improving oil
recovery without the need for gas injection.
The effectiveness of hybrid injection of LSW and miscible CO2 for EOR
in carbonates was investigated through a simulation study by Al-Shalabi
et al. (2016) using the UTCOMP reservoir simulator. The study aimed to
compare the performance of CGI with that of the hybrid method. The results
showed that the CGI method achieved a high recovery of 98.9%, while the
hybrid approach only increased it to 99.7% by controlling viscous fingering.
Therefore, the study concluded that the hybrid method may not be suitable
for conditions where gas miscibility is the primary mechanism for EOR. The
study’s findings could be useful in optimising the selection of EOR techniques
for carbonate reservoirs to achieve maximum recovery with minimum effort
and cost.
Consideration of the initial rock wettability is critical in hybrid LSW/gas
methods. The alteration of wettability from oil-wet to water-wet is particularly
significant in sandstone, as it positively affects the performance of LSW.
Conversely, if the rock is initially water-wet, the LSW and hybrid methods
will not be effective. Ramanathan et al. (2015) conducted an experimental
study of seawater alternating gas (SeaWAG) and LSWAG injection for oil
recovery from water-soaked sandstone. The study found that LSWAG had a
lower recovery factor than SeaWAG due to the initial high water saturation of
the rocks. In contrast, recovery by WAG increased from 76% to over 97% in
an aged oil-wet core when low-salinity brine replaced seawater.
130 Data Science and Machine Learning Applications in Subsurface Engineering

AlQuraishi et al. (2017) conducted a study that found that the low-salinity
alternating miscible CO2 method was not effective for clay-free sandstones,
but did result in a recovery value of 35.1% of the OOIP when clays were
present. Yang et al. (2005) conducted research and discovered that CO2 has
the ability to decrease the oil and brine interfacial tension (IFT) under constant
temperature and pressure conditions. This decrease in IFT can contribute to
additional oil recovery through the hybrid method. Further studies conducted
by Teklu et al. (2016) and Ramanathan et al. (2016) showed that the decrease
in IFT was less than 10 dynes/cm. However, a study by Kumar et al. (2016)
reported comparatively high IFT in the presence of CO2. Despite these mixed
findings, it is important to note that the change in IFT is relatively small
and is not considered to be the primary mechanism in the hybrid LSW/gas
approach for increasing oil recovery. Other mechanisms, such as the reduction
of residual oil saturation and improvement of displacement efficiency, are
believed to play a more significant role in the hybrid method. Nonetheless, the
discovery of the potential of CO2 to reduce IFT remains an important aspect of
research in the oil and gas industry.
The issue associated with pure gas flooding is unfavourable mobility
resulting in viscous fingering and a reduction in volumetric sweep efficiency.
With less gas needed for EOR projects, the WAG approach helps to solve this
significant problem. Additionally, conventional CO2 WAG methods typically
cause a delay in oil production, which the current study can alleviate with CO2
LSWAG. By overcoming the issue of late production that conventional WAG
commonly faces, CO2 LSWAG speeds up the synergy of these several process
mechanisms. According to previous studies (Dang et al., 2014; Pourafshary
and Moradpour, 2019; Sheng, 2014), the primary oil recovery mechanism in
CO2-LSWAG has been proposed to be wettability alteration to a more water-
wet condition.
However, the modeling studies of CO2-LSWAG in a 1D homogeneous
model and at field scale in sandstone reservoirs (Dang et al., 2014) had been
done and characterised by the expensive computational cost (Belazreg and
Mahmood, 2020; Jaber et al., 2019). There is little or no application of fast
and reliable machine learning models to forecast the performance of CO2
LSWAG as shown in Table 7.1, for secondary and tertiary modes. Despite
most CO2-LSWAG studies indicating improvement in oil recovery in
Table 7.1, some studies (Jiang et al., 2010; Ramanathan et al., 2015) with core
samples came out with negative or neutral outcomes due to initial wettability
being strongly water-wet which is not favourable for effective low salinity
water injection. Furthermore, references can be made to the reviewed work
by Ma and James (2022) based on CO2-LSWAG laboratory experiments with
no machine learning models reported.
 Table 7.1. CO2LSWAG based on field, experimental, numerical simulation and machine learning models for oil recovery factor predictions.

Reference Type of Porous Media Type of Injection Injection Scheme Ultimate Oil Recovery, % Experiment/ Machine
Fluid OOIP Field/ Numerical Learning
Simulation Model
Al Quraisha et Berea and Bentheimer LSW + CO2 LSW + HSW in 82.40% Experiment No ML

CO2 LSWAG Oil Recovery Factor Prediction in Carbonate Reservoir

al. (2017) Sandstone secondary mode, CO2 model
HSW + CO2 WAG in tertiary mode
Al-Shalabi et al. Carbonate Reservoir LSWI + CO2 Tertiary mode First Coreflood = 84.97% PHREEQC No ML
(2014) Injection Second Coreflood = 93.65% Simulation and model
Experiment
Dang et al. Sandstone CO2 +LSW Secondary to tertiary Incremental Oil Recovery of 4.5 Field, Experiment and No ML
(2014) modes – 9% of OOIP compared to Numerical Simulation model
CO2 HSWAG
Naderi and Sandstone CO2 + LSW Secondary to tertiary 16% of the OOIP on top of Compositional No ML
Simjoo (2018) modes LSW and 30% of the OOIP on numerical simulation model
top of HSW
Zolfghari et al. Sandstone CO2 + LSW Tertiary Mode 92% Experiment No ML
(2013) model
Teklu et al. Low Permeability LSW + CO2 Secondary to tertiary Coreflood experiment 1 = Experiment No ML
(2016) Carbonate modes 81.9% model
Medium Permeability Coreflood experiment 2 =
Berea Sandstone 83.4%
Ultra-Low Permeability Coreflood experiment 3 =
Three Forks mudstone 81.5%
Core Disc
Saxena (2017) Sandstone LSWI + CO2 Tertiary mode 39% OOIP Compositional
numerical simulation No ML
and Field model

131
ML = Machine Learning, LSW = Low Salinity Water, OOIP = Original Oil In Place
132 Data Science and Machine Learning Applications in Subsurface Engineering

Al-Jifri et al. (2021) develop two new empirical equations for predicting
oil recovery factor in waterflooded heterogeneous reservoirs based on these
parameters of water injection rate, permeability anisotropy, water viscosity,
and reservoir heterogeneity with no proxy correlations currently existing for
CO2-LSWAG which this study strive to achieve. Furthermore, Roustazadeh
et al. (2022) developed three regression-based models including the support
vector machine (SVM), extreme gradient boosting (XGBoost), and stepwise
multiple linear regression (MLR) and various combinations of three databases
to construct machine learning (ML) models and estimate the oil and/or gas
recovery factor (RF). The following authors (Aliyuda et al., 2020; Alpak et al.,
2019; Chen et al., 2020; Ibrahim et al., 2022; Sharma et al., 2010; Tahmasebi
et al., 2020) also applied ML algorithms in predicting hydrocarbon recovery
factors from different reservoirs but not with regards to the present study with
CO2 LSWAG.
In this study, a carbonate field model based on CO2-LSWAG flooding
was simulated using a compositional simulator with geochemical models
incorporated, and then Multivariate Adaptive Regression Splines (MARS) and
Group Method of Data Handling (GMDH) machine learning methods were
used to develop proxy models for prediction of oil recovery factor. Therefore,
this study advocates for the injection of low-salinity water alternating
CO2 as an EOR technique due to its high recovery factor for improving
microscopic and macroscopic displacement efficiencies. The use of machine
learning proxy models as prediction tools will also enhance the efficient
full-field implementation of this technique in reducing the computational time
associated with numerical simulations (Amar et al., 2021; Kalam et al., 2021)
in carbonate reservoirs.
The structure of this paper is organised as follows. Section 2 is the
methodology that describes the use of a compositional simulator coupled with
fluid flow and geochemical modelling, as well as machine learning tools.
Section 3 is the results and discussion that analyses the currently proposed
method and its comparison to the use of a conventional simulator to optimise
operational conditions. Section 4 concludes the study and spells out the major
findings drawn from the present study.

2. Methodology
2.1 Modelling of CO2-LSWAG
This paper’s numerical method used a compositional simulator to generate
various scenarios of the CO2-LSWAG in a carbonate reservoir with SO4
 CO2 LSWAG Oil Recovery Factor Prediction in Carbonate Reservoir 133

and carboxylate ion exchange. The assumptions made in modelling the

CO2-LSWAG in CMG compositional simulator under miscible conditions are:
1. Carboxylate ion exchange was solely responsible for the wettability
alteration.
2. Multiphase multicomponent flow equations and EOS flash calculations
were fully coupled to geochemical reactions.
3. Relative permeability sets of oil and water are altered by a scaled
carboxylate ion exchange equivalent fraction.
4. Carboxylate ion exchange is also dependent on mineral dissolution and
precipitation reactions.

2.2 Geochemical Reactions of CO2-LSWAG

Carbonates with SO42– and carboxylate ion exchange interpolation was carried
out on sets of relative permeabilities. Equations (7.2) and (7.3) show the
partitioning of CO2 between oil and water in the reservoir (Dang et al., 2014).
The aqueous reactions considered in CO2 LSWAG modelling are:
CO2(g) → CO2(aq) (7.1)
–
CO2(aq) + H2O ↔ H* + HCO 3 (7.2)
–
H+ + OH ↔ H2O (7.3)
CaCH3COO+ ↔ CH3COO– + Ca2+ (7.4)
2–
CaSO4 ↔ Ca + SO 4
2+
(7.5)
2–
MgSO4 ↔ Mg2+ + SO 4 (7.6)
The mineral reaction considered in CO2-LSWAG modelling for calcite
and dolomite are:
–
CaCO3 + H+ ↔ Ca2+ + HCO 3 (7.7)
–
CaMg(CO3)2 + 2H+ ↔ Ca2+ + 2HCO 3 + Mg2+ (7.8)
The ion exchange considered in CO2-LSWAG modelling is given as:
2–
SO 4 + 2CH3COO– X ↔ 2CH3COO– + SO4 – X2 (7.9)

2.3 Machine Learning Methods

The machine learning techniques of MARS and GMDH were used in this
paper to predict Oil Recovery Factor (ORF) by varying the low salinity ionic
134 Data Science and Machine Learning Applications in Subsurface Engineering

Fig. 7.1. Machine learning workflow.

concentrations of the water injected for the CO2 LSWAG. Figure 7.1 shows
the general workflow used for all the supervised machine-learning techniques.
The datasets used were from the simulation results, which were then divided
into training and testing data. The general workflow for all the supervised
algorithms used in this study for CO2-LSWAG ORF prediction as a single
objective problem is shown in Fig. 7.1.

2.3.1 Multivariate Adaptive Regression Splines (MARS)

The multivariate nonparametric regression technique known as MARS
was introduced by Friedman in 1991. This method utilises piecewise linear
segments, or splines, to capture complex nonlinear relationships between
input and target variables. The mathematical relationship between the basis
function terms and the dependent variable in MARS is expressed in Eq. (7.10):
M
= CO + ∑ Ck β k ( X ),
ORF (7.10)
k =1

where ORF is the oil recovery factor as the target variable. There are several
parameters that must be taken into account while simulating the interaction
between two or more variables in a reservoir system. These parameters include
the intercept (Co), the number of basis function terms (M), and the vector of
the kth basis function’s unknown coefficients (Ck), where k is a value between
1 to M. The basis functions themselves are denoted by βk and X is used to
represent the input variables for the reservoir parameters.
According to Friedman’s (1991) findings, the MARS technique uses
multivariate spline functions as basis functions, which are represented in
Eq. (7.11):
Zn
βk ( X ) = ∏  p
z =1
zn ⋅ ( X m( z , n ) − szn )  ,
+
(7.11)
 CO2 LSWAG Oil Recovery Factor Prediction in Carbonate Reservoir 135

In this context, we use the variable Zn to represent the number of splits

that resulted in the value of βk . The variable pzn is utilized as a sign indicator
that can take on values ±1. The notation [.]+ is used to signify taking only
the positive part. The function m (z, n) labels the predictor variable, while szn
represents the corresponding knot value of variable Xm(z,n).
The MARS algorithm is a popular approach for developing models that
can capture nonlinear relationships between variables. The algorithm typically
consists of two steps, the forward and backward passes. In the forward pass,
the algorithm investigates potential knots and begins with a model consists
of just the constant term. This means that the initial model assumes that there
is no relationship between the predictor variables and the response output
variable. As the algorithm progresses, it gradually adds basis functions (BFs)
to the model, which allows it to capture the nonlinearity between the predictor
variables and the response output variable. The backward pass is carried out
after the forward pass and is used to eliminate estimators that are irrelevant to
the output variable. This is done to streamline the developed model eliminating
the least significant BFs according to the generalised cross-validation (GCV)
criterion given in Eq. (7.12). The GCV criterion is a widely used method for
selecting the best subset of predictors that minimise the prediction error of
the model. By removing the non-influential predictors, the model becomes
more interpretable and easier to understand, while still accurately capturing
the nonlinear relationships between the predictor variables and the response
output variable.
1 N

N
∑ (O − P )
i i
2

GCV (ϕ ) = i=1
2
, (7.12)
 M (ϕ ) 
1−
 N 
To achieve a balance between the model size and its fit to the dataset,
a tuning parameter denoted as φ is utilised. The overall field measurements
training dataset is represented by N, while Oi and Pi denote the observed and
predicted values of the dataset, respectively. By adjusting the value of the
tuning parameter, one can regulate the trade-off between the complexity of
the model and its accuracy in capturing the dataset’s patterns. M(φ) is defined
in Eq. (7.13) as the effective number of parameters employed in the model:
M(φ) = (φ + 1) + d.φ, (7.13)
The parameter d serves as a penalty or smoothing factor in the
non-parametric MARS algorithm, where its value determines the number
of basis functions and the smoothness of the estimated functions. A higher
value of d leads to fewer basis functions and smoother estimates, whereas a
136 Data Science and Machine Learning Applications in Subsurface Engineering

lower value results in a larger model with more basis functions. For further
information on the selection of d values and a comprehensive explanation of the
non-parametric MARS algorithm, please refer to Friedman’s work from 1991.
The format for the MARS equations expressed in terms of max(.) for
univariate linear regression of the pth piece of variable xk as the basis function
is presented in Eq. (7.14). This mathematical expression demonstrates
how BFp () is represented within the MARS framework. To ensure clarity,
appropriate synonyms have been used, and sentences have been rearranged
without altering the intended meaning.
BFP(xk) = max(0, xk – ap) or BFP(xk) = max(0, ap – xk). (7.14)
In mathematical notation, the term max(.) indicates that solely the positive
portion of the input is preserved while assigning a zero value to the negative
portion, as presented in equation (7.15). Piecewise linear functions can be
represented in the form of max (xk – ap), with the knot point specified at a
particular value ap.
 xk − a p , xk ≥ a p
max(0, xk − a p ) =
 , (7.15)
0, otherwise
To create a local linear regression with continuous knots, the max (.)
function is utilized. Through recursive spline fitting and splitting, the knots
are best chosen. The Eq. (7.14) BFp (xk) will only be zero when the second
term in the equation exceeds zero, which is a crucial point to remember.
Additionally, when modeling two variable interactions, the basis function
BFps () can be expressed using two univariate basis functions for xk and xi, as
shown in Eq. (7.16):
BFPs(xk , xi) = BFP(xk ) × BFs(xi). (7.16)

2.3.2 Group Method of Data Handling (GMDH)

In this study, GMDH is used to develop an explicit mathematical model
with degrees of the polynomials to predict the oil recovery factor for
CO2-LSWAG under the miscible condition in a carbonate reservoir. Ivakhnenko
(1966) cyberneticist proposed a GMDH polynomial neural network (PNN))
algorithm for constructing high-order regression-type polynomials. However,
due to the self-organising nature of the GMDH algorithm (selecting its own
input parameters automatically based on the parameter’s contribution to the
final output), the GMDH modelling approach is able to overcome many
artificial neural network (ANN) limitations as an alternative method (Lv
et al., 2023). GMDH employs optimisation criteria to determine network
 CO2 LSWAG Oil Recovery Factor Prediction in Carbonate Reservoir 137

connectivity, network size, and the coefficient for the optimum model with
model reduction with less human intervention (Ivakhnenko, 1971). The input
and output relationship are expressed in polynomial form, with the model
automatically selecting the most influential parameters (Farlow, 1984). In this
study, GMDH is used to develop a mathematical model for predicting the CO2-
LSWAG oil recovery factor. Ivakhnenko (1971) applied the Kolmogorov-
Gabor polynomials theory function to find the output parameter, which is
expressed in Eq. (7.17):
n n n n n n
ao + ∑ ai xi + ∑∑ ai a j xi x j + ∑∑∑ ai a j ak xi x j xk +  (7.17)
ORF GMDH =
=i 1 =i 1 =j 1 =i 1 =j 1=
k 1

where, aij…k, xij…k, ORF, and n represent polynomial coefficients, input

parameters, the predicted values and the number of variables, respectively.
From Eq. (7.17) the polynomial variables are obtained by employing least
square minimisation analysis as expressed in Eq. (7.18):
N
=δ j2 ∑ (ORF
i=1
i
GMDH
− yi ) 2 (7.18)
GMDH
where N, yi, and ORFi represent the number of training data points, actual
values, and predicted values, respectively.

2.3.3 Performance Metrics

The accuracy of the prediction from the machine learning methods was
measured using the coefficient of determination (R 2), mean squared error
(MSE) and Pearson’s correlation coefficient (r) to understand the target
against variations of individual independent variables.
The R2 is a statistical index with the best model prediction having an
optimum value closer to 1, MSE gets closer to zero for an accurate model,
and r can be –1 or +1, indicating a parametric effect on output (Brantson
et al., 2019). Equations (7.19), (7.20), and (7.21) illustrate the mathematical
expressions used for R2 MSE and r, respectively:
2
 n 
 ∑ ( ORF observed ) (
− ORF observed mean × ORFpredicted − ORF predicted mean ) 
R2 =  i=1  (7.19)
 n 
∑ ( ORF ) × ( ORF )
2 2
 observed − ORF observed mean predicted − ORF predicted mean 
 i=1 
1 n
( ORFmeasured − ORFpredicted )
2
=MSE ∑
n i =1
(7.20)
138 Data Science and Machine Learning Applications in Subsurface Engineering

∑ ( ORF
observed ) (
− ORF observed mean × ORFpredicted − ORF predicted mean )
r= i=1
(7.21)
n n

∑ ( ORF ) × ∑ (ORF )
2 2
observed − ORF observed mean predicted − ORF predicted mean
i 1 =i 1

where, n = total number of oil recovery factor data points observed,

ORFobserved = oil recovery factor observed data points, ORFobserved mean = mean
of the oil recovery factor observed data points, and ORFpredicted mean = mean of
the oil recovery factor predicted data points.

2.3.4 Dataset Standardisation

The CO2-LSWAG modeling process involves the use of data with varying
physical units. These physical units could include meters, grams, seconds,
or any other units that are relevant to the data being used. In order to ensure
that the data is comparable and can be used to generate accurate models,
normalisation is applied. Normalisation is a technique that scales the data
to a common range of values. In this particular study, both the input and
output vectors were normalised in the interval [–1, 1]. This range was chosen
because it is a common range of values used in normalisation techniques.
The normalisation process was implemented to ensure that there is a constant
variation in the data as expressed in Eq. (7.22). By normalising the data, the
range of values for both input and output vectors is restricted to a specific
range, thus making it easier to compare and analyse the data. It is important
to note that normalisation is not always necessary, but it is often used when
working with data that has varying physical units. By normalising the data, it
is possible to remove any biases that may arise due to differences in physical
units. This can help to ensure that the models generated from the data are
accurate and reliable.
( y − ymin ) × (xi − xmin )
=yi ymin + max , (7.22)
(xmax − xmin )
The resulting normalised data is denoted by yi, where the maximum and
minimum values (ymax and ymin) of the normalised data are set at 1 and –1,
respectively, the measured values (xi) are scaled to a range of –1 to 1, with
xmin and xmax the measured data serving as the lower and upper bounds of this
range. This method of normalisation is commonly used.

3. Results and Discussion

3.1 Numerical Model Description
This study employed an inverted five-spot injection carbonate model to
simulate CO2-LSWAG oil recovery at the field scale. The reservoir had
 CO2 LSWAG Oil Recovery Factor Prediction in Carbonate Reservoir 139

dimensions of approximately 7400 × 7400 × 28 ft3 along the Y, X, and Z

directions. A total of 12,150 grid blocks were utilised, comprising of 45 blocks
in both X and Y directions and six blocks in the Z direction. The simulation
utilised flexible grids and precise corner point geometry to accurately
represent the geological features and description of the reservoir, particularly
for heterogeneities and wells. This approach was preferred over conventional
cartesian grids. The reservoir exhibited heterogeneity with variations in
porosity and permeability. Porosity values ranged between 9.5%–29%,
while permeability differed in all directions with the vertical permeability
lower than the horizontal permeability. The active fluid phases used in the
model were water and oil, which are commonly found in reservoirs. The
porosity and permeability variations depicted in Figs. 7.2a and 7.2b provide
valuable insights into the characteristics of the reservoir model. Additionally,
Fig. 7.3 depicts the oil and water relative permeability curves used for the
CO2-LSWAG simulation for high and low salinity as a function of water
saturation. The aqueous or mineral reaction structure selected was for

(a)

(b)
Fig. 7.2. Petrophysical properties of the model (a) Porosity model (b) Permeability model.
140 Data Science and Machine Learning Applications in Subsurface Engineering

Fig. 7.3. Low and high salinity water and oil relative permeability curves (set #1 is for the original
relative permeability plot while set #2 is for the LSW relative permeability plot).

2–
carbonates with SO 4 simple relative permeability interpolation. The
2–
component wizard used linear interpolation for SO 4 concentration to effect
the change.
Upon completion of the reservoir simulation model, well completion
procedures were implemented. In this study, we determined that the inverse
five-spot pattern was the most optimal approach for the reservoir simulation.
To achieve maximum effectiveness, it was ensured that all six grid layers
were perforated for both production and injection wells, with vertical wells
for efficient oil extraction. The placement of the injection well in the central
location allowed for efficient fluid sweep throughout the reservoir, moving
the oil towards the production wells situated at the corners, resulting in a high
areal sweep efficiency. The sweeping motion of the injected fluid played a
significant role in improving oil recovery in the reservoir. The implementation
of the inverse five-spot pattern and vertical wells in all six grid layers of the
model enabled efficient extraction of the oil.
 CO2 LSWAG Oil Recovery Factor Prediction in Carbonate Reservoir 141

3.2 Input and Target Dataset

A simulation study was carried out to obtain a simulation dataset for input
and output for training and testing to establish a mapping relationship.
Table 7.2 shows the reservoir parameters that were used in running the
carbonate simulation model. Additionally, Table 7.3 shows the initial run
whose result was used as the base case for the sensitivity run for varied low
salinity water ranges between 200–3000 ppm. The total dataset obtained
within the low salinity water ranges was 1385 sample points, with 1,107 used
for training and the rest used for testing the model. Furthermore, Fig. 7.4a
shows the average pressure for the base case and CO2-LSWAG scenarios.

Table 7.2. Reservoir parameters for the carbonate model.

Parameter Value
Reference pressure, psi 3337
Reference depth, ft 8596
Water-Oil contact depth, ft 8950
Permeability, mD 29.76–269.40
Porosity, % 9.5–29.
Top of Reservoir Sand, ft 8596
Swcon 0.076
Soirw 0.1434
1-Sorw 0.75
1-Soirw 0.8566
pH 5.22
Water Injection Rate, bbl/day 7,000
Gas Injection Rate, MMft3/day 30
Injector Bottomhole Pressure, psi 6,710
Formation Water Salinity, ppm 90,044.10
Equation of State (EOS) Peng Robinson

Table 7.3. Base case for the formation and injected water.

Species Formation Water, ppm Base Case Injected Water, ppm

Ca2+ 18492 26
Mg 2+
2320 77
Na+ 68520.1 660
HCO3– 100 0
SO42– 612 46.8
142 Data Science and Machine Learning Applications in Subsurface Engineering

Recovery factor prediction can be a function of multiple variables of

the reservoir characteristics to be determined. The operational controllable
input parameters for the MARS and GMDH models are average reservoir
pressure (ARP), cumulative gas injection (CGI), cumulative low salinity
water injection (CLSWI), gas cumulative production (GCP), oil cumulative
production (OCP), gas injection rate (GIR), low salinity water injection rate
(LSWIR), and the target is the ORF for CO2-LSWAG as shown in Table 7.4.
The LSW and CO2 injection rate employed in the carbonate model for the
simulation period is shown in Fig. 7.4b.
For constructing the proxy machine learning models, the oil saturation
maps obtained for the 15 years of CO2-LSWAG injection oil recovery are
shown in Figs. 7.5a, 7.5b, and 7.5c for 5, 10, and 15 years, respectively. It
was observed that CO2 injection lowers the viscosity of the crude oil after it
becomes miscible with it. Figure 7.5 shows the initial and final oil saturation of
the reservoir, depicting how the CO2-LSWAG injection lowered the viscosity
to recover more oil. It can be seen from the final oil saturation plot in Fig. 7.5c
of the CO2-LSWAG that the injection scheme lowered the viscosity of the
reservoir oil, enabling more oil to be recovered.

Table 7.4. Dataset ranges for CO2-LSWAG oil recovery factor prediction.

Variable/Unit Minimum Average Maximum Standard Data Type

Deviation
Average reservoir 3349.11 5187.38 6679.61 1298.05 Input
pressure (psi)
Cumulative gas 0.00 12.12 25.64 5.28 Input
injection (MMMSCF)
Cumulative low 0.00 6.83 18.48 4.24 Input
salinity water injection
(MMbbl)
Gas cumulative 0.00 7.61 17.02 3.16 Input
production
(MMMSCF)
Oil cumulative 0.00 13.73 15.96 3.17 Input
production (MMbbl)
Gas injection rate 0.00 13.81 30.00 4.77 Input
(MMSCFD)
Low salinity water 0.00 5.09 7.17 1.58 Input
injection rate
(Mbbl/day)
Oil recovery factor (%) 0.00 72.38 84.07 16.68 Output
 CO2 LSWAG Oil Recovery Factor Prediction in Carbonate Reservoir 143

(a)

(b)
Fig. 7.4. Simulation Parameters: (a) Average pressure base case and CO2 LSWAG, (b) LSW and CO2
injection rate.

3.3 MARS Modelling

MARS modelling combines both recursive spline fitting and partitioning
that maintains the positive component of the combination while having
less influence on non-contributing properties (Brantson et al., 2018). As
shown in Table 7.4, all the inputs and target will be loaded into the MARS
algorithm for developing the trained model for spline fitting optimisation
computations. Also, the MARS model automatically selects the most
influential input parameters that affect the target response leaving out the
non-contributing variables. Furthermore, the maximum basis functions,
maximum interactions and degrees of freedom for knot optimisation used
144 Data Science and Machine Learning Applications in Subsurface Engineering

(a)

(b)

(c)
Fig. 7.5. Oil saturation maps for the simulation period: (a) oil saturation at 5 years, (b) oil saturation
at 10 years, (c) oil saturation at 15 years.
 CO2 LSWAG Oil Recovery Factor Prediction in Carbonate Reservoir 145

Table 7.5. Basis functions and corresponding equations of MARS model.

Basis Functions (BF) Equations

BF1 max (0, OCP – 1.03286 × 107)
BF2 max (0, 1.03286 × 107 – OCP)
BF4 max (0, 4.09049 × 108 – GCP)
BF5 max (0, GCP – 5.33908 × 109)
BF7 max (0, GCP – 4.43216 × 109)
BF10 max (0, 3802.39 – ARP)
BF12 max (0, 3.97977 × 109 – CGI)
BF13 max (0, OCP – 2.79576 × 106)
BF15 max (0, LSWIR – 7042.81)
BF17 max (0, OCP – 7.16527 × 106)
BF19 max (0, OCP – 1.52833 × 107)

are 20, 1 and 3, respectively. Hence the relative importance of the five
input variables used are oil cumulative production (100%), gas cumulative
production (2.22%), cumulative gas injection (2.36%), the water injection rate
(1.63%), and average pressure (1.46%). The relevance factor computed for
all the parameters chosen automatically has a positive impact on the ORF as
the target. The ideal developed MARS model for CO2-LSWAG is expressed
in Eq. (7.23). Also, Table 7.5 indicates the basis functions and corresponding
equations of the ORF MARS model written in Eq. (7.23):
= 25.7653 + 1.84934 ×10−6 × BF1 −1.57532 ×10−6 × BF 2 + 1.00632 ×10−8 × BF 4
ORF
− 2.33007 ×10−9 × BF5 + 2.36461 × 10−9 × BF 7 + 0.00204046 × BF10
(7.23)
− 3.65011 × 10−9 × BF12 + 3.67347 ×10−6 × BF13 + 0.0775678 × BF15
− 3.02324 ×10−7 × BF17 − 4.68672 ×10−7 × BF19

Figures 7.6a to 7.6d show the cross plots for the MARS training and
testing models used in this study. The results show both the training and
testing data points are close to the ideal line, which is an indication of the
MARS model’s robustness. Three independent datasets (low salinities of
573.86 ppm, 1250.51 ppm, and 2949.15 ppm) were used to test the MARS
model to verify its accuracy and reliability. Furthermore, the CO2-LSWAG
recovery factor prediction can also be assessed from Table 7.6, showing the
statistical performance. The boxplot in Fig. 7.6e shows the residuals for both
MARS training and testing sets. It can be observed that the residuals do not
vary significantly from zero. It can be stated that the training model established
can predict the testing datasets within acceptable accuracy predictions.
It is shown that the mean and median lie close to zero for a good model.
Figure 7.6e also shows the MARS model residual distributions and outliers.
146 Data Science and Machine Learning Applications in Subsurface Engineering

(a)

(b)
Fig. 7.6 contd. ...
 CO2 LSWAG Oil Recovery Factor Prediction in Carbonate Reservoir 147

...Fig. 7.6 contd.

(c)

(d)
Fig. 7.6 contd. ...
148 Data Science and Machine Learning Applications in Subsurface Engineering

...Fig. 7.6 contd.

(e)
Fig. 7.6. MARS training and testing models: (a) Crossplot for the MARS training model,
(b) Crossplot for the MARS testing for 573.86 ppm, (c) Crossplot for the MARS testing for 1250.51
ppm, (d) Crossplot for the MARS testing for 2949.15 ppm, (e) Boxplot residuals for MARS training
and testing.

Table 7.6. Train and test results for MARS model.

Models Training Testing Testing Testing

(573.86 ppm) (1250.51 ppm) (2949.15 ppm)
MSE 0.0564 0.2023 0.5541 0.4518
R2 0.9998 0.9996 09993 0.9995

3.4 GMDH Modelling

In building the GMDH model, the most influential input parameters are usually
automatically selected that have an influence on the target variable. Three
input variables that have the most influence on the target are CLSWI, OCP,
and LSWIR. The number of layers and neurons used are 2 and 2, respectively,
for GMDH model computations. The final architecture obtained was after a
series of optimisation processes by observing the network performance.
Figures 7.7a to 7.7d show the cross plots for both the GMDH training
and testing models used in this study. It can be seen that both the training
and testing data points are close to the ideal line, which is an indication of
the GMDH model’s robustness. Furthermore, the CO2-LSWAG recovery
factor prediction can also be assessed from Table 7.7, showing the statistical
CO2 LSWAG Oil Recovery Factor Prediction in Carbonate Reservoir 149

(a)

(b)
Fig. 7.7 contd. ...
150 Data Science and Machine Learning Applications in Subsurface Engineering

...Fig. 7.7 contd.

(c)

(d)
Fig. 7.7 contd. ...
 CO2 LSWAG Oil Recovery Factor Prediction in Carbonate Reservoir 151

...Fig. 7.7 contd.

(e)
Fig. 7.7. GMDH training and testing models: (a) Crossplot for the GMDH training model,
(b) Crossplot for the GMDH testing for 573.86 ppm, (c) Crossplot for the GMDH testing for
1250.51 ppm, (d) Crossplot for the GMDH testing for 2949.15 ppm, (e) Boxplot residuals for GMDH
training and testing.

Table 7.7. Training and testing results for the GMDH model.

Models Training Testing Testing Testing

(573.86 ppm) (1250.51 ppm) (2949.15 ppm)
MSE 0.1674 0.0239 0.0219 0.0231
R2 0.9999 0.9999 0.9999 0.9999

performance. Also, it can be observed from Table 7.7 that the testing data
performed better than the MARS model testing. The boxplot in Fig. 7.7e shows
the residuals for both GMDH training and testing sets. It can be observed that
the residuals do not vary significantly from zero. It can be stated that the
training model established can predict the testing datasets within acceptable
accuracy predictions. It can also be seen that the mean and median lie close to
zero for a good model. Figure 7.7e also shows the residual distributions and
least outliers for GMDH testing.
Figure 7.8 shows the GMDH topology used to develop the ORF equations.
The ORF was formed from three variables from the input layer. The OCP
variable combines with CLSWI and LSWIR in the input layer to form two
variables in the hidden layer before combining them to build the target layer.
152 Data Science and Machine Learning Applications in Subsurface Engineering

The summary of the GMDH model’s equations for two layers is expressed
in Eqs. (7.24) to (7.26) as:
Layer #1
Number of neurons: 2
A = ao + a1 × OCP + a2 × CLSWI + a3 × OCP × CLSWI + a4 × (OCP)2
+ a5 × (CLSWI)2 (7.24)
a0 = 0.03088 a1 = 0.94845 a2 = 0.04723
a3 = 0.05133 a4 = 0.02253 a5 = 0.00212
B = –bo + b1 × LSWIR + b2 × OCP – b3 × OCP × LSWIR – b4 × (LSWIR)2
+ b5 × (OCP)2 (7.25)
b0 = 0.00396 b1 = 0.00360 b2 = 0.99753
b3 = 0.00431 b4 = 0.00041 b5 = 0.01058
Layer #2
Number of neurons: 1
ORF = –co + c1 × B + c2 × A + c3 × A × B – c4 × (B)2 + c5 × (OCP)2 (7.26)
c0 = 0.00038 c1 = 0.61360 c2 = 0.37942
c3 = 409.55682 c4 = 205.30898 c5 = 204.23896
where A and B are virtual independent inputs or nodal variables used in the
GMDH neural network.

Fig. 7.8. Schematic for the GMDH topology.

 CO2 LSWAG Oil Recovery Factor Prediction in Carbonate Reservoir 153

3.5 Numerical Simulator and Machine Learning Computational Time

Figure 7.9 shows the simulation time for the CMG commercial numerical
simulator and the machine learning algorithms using Intel core i5 – 11400H,
4.5 GHz 6Core, and 16 GB of RAM. It can be seen that the numerical
simulation took a long runtime to execute. Also, this significant reduction
in simulation time indicates the robustness of the machine-learning proxy
models (MARS and GMDH). Furthermore, the proxy models can serve as
substitutes for direct simulation of CO2-LSWAG for design parameters
within the ranges considered. In summary, using a GPU (graphics processing
unit) could further improve the computational time of this machine-learning
algorithm. In comparison to using a GPU, this method may offer even greater
benefits in terms of speed and efficiency when running multiple simulations
at the same time.

Fig. 7.9. Simulation time for machine learning and numerical models.

4. Conclusion
The problem addressed in this study is the lack of attention and research
on using machine learning techniques to use CO2-LSWAG for EOR in
carbonate reservoirs. The quality and composition of the water used in most
EOR processes are not adequately considered, and there is a need for fast
computational methods for predicting future recovery factors. We used
multiphase multicomponent flow equations, geochemical modelling, and
compositional simulation datasets to build proxy models for predicting
oil recovery factors in carbonate reservoirs for CO2-LSWAG EOR. Two
supervised machine learning models (MARS and GMDH) were used to
154 Data Science and Machine Learning Applications in Subsurface Engineering

analyse the data generated by these simulations under various low-salinity

conditions. The conclusion from the study are as follows:
1. The study showed that implementing CO2-LSWAG at a field scale can
significantly increase the recovery of oil by mainly reducing viscous
fingering.
2. The GMDH model gave better testing predictions as compared to the
MARS prediction model with derived mathematical correlations for the
CO2 LSWAG oil recovery factor.
3. The machine learning models predicted the outputs from the commercial
simulator by significantly reducing simulation time and generalising
without compromising the required accuracy.
4. It is recommended that CO2-LSWAG should be implemented on a field
scale based on the current study to increase recovery while reducing
operational costs.

Acknowledgment
We express our appreciation to the anonymous reviewers whose valuable inputs
have significantly contributed to the success of this research. Additionally, our
gratitude goes to the Computer Modelling Group (CMG) for providing us
with the commercial software that was instrumental to the completion of this
project. Lastly, we acknowledge the indispensable support of the University of
Mines and Technology, Tarkwa, Ghana, GNPC School of Petroleum Studies,
and the Petroleum and Natural Gas Engineering Department.

References
Afzali, S., Rezaei, N. and Zendehboudi, S.A. 2018. Comprehensive review on enhanced oil
recovery by water alternating gas (WAG) injection. Fuel 227: 218–246. https://doi.
org/10.1016/j.fuel.2018.04.015.
Ahmed, T. 2018. Reservoir Engineering Handbook (4th Edn.). Elsevier Inc.
Al-Abri, H., Pourafshary, P., Mosavat, N. and Al Hadrami, H. 2019. A study of the performance
of LSWA CO2 EOR technique on improvement of oil recovery in sandstones. Petroleum
5(1): 58–66. https://doi.org/10.1016/j.petlm.2018.07.003.
Alaidan, A. and Mamore, D. 2010. SWACO2 and WACO2 efficiency improvement in carbonate
cores by lowering water salinity. In: Canadian Unconventional Resources and International
Petroleum Conference. https://doi.org/10.2118/137548-MS.
Aleidan, A. and Mamora, D. 2011. Miscible CO2 injection in highly heterogeneous carbonate
cores: experimental and numerical simulation studies. In: SPE Middle East Oil and Gas
Show and Conference. https://doi.org/10.2118/141469-MS.
Aliyuda, K., Howell, J. and Humphrey, E. 2020. Impact of geological variables in controlling
oil-reservoir performance: an insight from a machine-learning technique. SPE Reservoir
Evaluation & Engineering 23(04): 1314–1327. https://doi.org/10.2118/201196-PA.
 CO2 LSWAG Oil Recovery Factor Prediction in Carbonate Reservoir 155

Al-Jifri, M., Al-Attar, H. and Boukadi, F. 2021. New proxy models for predicting oil recovery
factor in waterflooded heterogeneous reservoirs. Journal of Petroleum Exploration and
Production 11(3): 1443–1459. https://doi.org/10.1007/s13202-021-01095-4.
Alpak, F.O, Araya–Polo, M. and Onyeagoro, K. 2019. Simplified dynamic modeling of
faulted turbidite reservoirs: a deep-learning approach to recovery-factor forecasting for
exploration. In: SPE Reservoir Evaluation & Engineering 22(04): 1240–55. https://doi.
org/10.2118/197053-PA.
AlQuraishi, A.A. and Shokir, E.M. El-M. 2011. Experimental investigation of miscible
CO2 flooding, Journal of Petroleum Science and Technology 29(19). https://doi.
org/10.1080/10916461003662976.
AlQuraishi, A.A., Amao, A.M., Al-Zahrani, N.I., AlQarni, M.T. and AlShamrani, S.A. 2019.
Low salinity water and CO2 miscible flooding in berea and bentheimer sandstones. Journal
of King Saud University – Engineering Sciences 31(3): 286–295. https://doi.org/10.1016/j.
jksues.2017.04.001.
Al-Shalabi, E.W., Sepehrnoori, K. and Pope, G. 2014. Geochemical investigation of the
combined effect of injecting low salinity water and carbon dioxide on carbonate reservoirs.
Energy Procedia 63: 7663–7676. https://doi.org/10.1016/j.egypro.2014.11.800.
Al-Shalabi, E., Sepehrnoori, K. and Pope, G. 2016. Numerical modelling of combined low
salinity water and carbon dioxide in carbonate cores. Journal of Petroleum Science and
Engineering 137: 157–171. https://doi.org/10.1016/j.petrol.2015.11.021.
Amar, M.N., Ghahfarokhi, A.J., Ng, C.S.W. and Zeraibi, N. 2021. Optimization of WAG in
real geological field using rigorous soft computing techniques and nature-inspired
algorithms. Journal of Petroleum Science and Engineering 206(109038): 1–13. https://doi.
org/10.1016/j.petrol.2021.109038.
Belazreg, L. and Mahmood, S.M. 2020. Water alternating gas incremental recovery factor
prediction and WAG pilot lessons learned. Journal of Petroleum Exploration and
Production Technology 10: 249–269. https://doi.org/10.1007/s13202-019-0694-x.
Brantson, E.T., Ju, B., Omisore, O.O., Wu, D., Aphu, E.S. and Liu, N. 2018a. Development
of machine learning predictive models for history matching tight gas carbonate reservoir
production profiles. Journal of Geophysics and Engineering 15(5): 7–12. https://doi.
org/10.1088/1742-2140/aaca44.
Brantson, E.T., Ju, B., Omisore, B.O., Wu, D., Selase, A.E. and Liu, N. 2018b. Development
of machine learning predictive models for history matching tight gas carbonate reservoir
production profiles. Journal of Geophysics and Engineering 15(5): 2235–2251.
Brantson, E.T., Ju, B., Ziggah, Y.Y., Akwensi, P.H., Sun, Y., Wu, D. and Addo, B.J. 2019.
Forecasting of horizontal gas well production decline in unconventional reservoirs using
productivity, soft computing and swarm intelligence models. Natural Resources Research
28: 717–756.
Chen, Y., Zhu, Z., Lu, Y., Hu, C., Gao, F., Li, W., Sun, N. and Feng, T. 2020. Reservoir recovery
estimation using data analytics and neural network based analogue study. In: SPE/IATMI
Asia Pacific Oil & Gas Conference and Exhibition. https://doi.org/10.2118/196487-MS.
Christensen, J.R., Stenby, E.H. and Skauge, A. 1998. Review of WAG field experience.
In: Proceedings of the International Petroleum Conference and Exhibition of Mexico.
https://doi.org/10.2118/90589-MS.
Dang, C.T.Q, Nghiem, L.X, and Chen, Z. 2014. CO2 low salinity water alternating gas: a new
promising approach for enhanced oil recovery. In: SPE Improved Oil Recovery Symposium,
1–19. https://doi.org/10.2118/169071-MS.
Farlow, S.J. 1984. Self-Organizing Methods in Modeling GMDH Type Algorithms. New York:
Marcel-Dekker. CRC Press.
156 Data Science and Machine Learning Applications in Subsurface Engineering

Friedman, J.H. 1991. Estimating Functions of Mixed Ordinal and Categorical Variables using
Adaptive Splines (pp. 1–42). Department of Statistics, Stanford Univ.
Gbadamosi, A.O., Radzuan, J., Manan, M.A., Agi, A. and Adeyinka, S.Y. 2019. An overview of
chemical enhanced oil recovery: recent advances and prospects. Int. Nano Lett. 9: 3–10.
https://doi.org/10.1007/s40089-019-0272-8.
Hamouda, A.A. and Pranoto, A. 2016. Synergy between low salinity water flooding and CO2 for
EOR in chalk reservoirs. In: Proceedings of the SPE EOR Conference at Oil and Gas West
Asia, 2016; Society of Petroleum Engineers. TX, USA: Richardson.
Ibrahim, A.F., Alarifi, S.A. and Elkatatny, S. 2022. Application of machine learning to predict
estimated ultimate recovery for multistage hydraulically fractured wells in niobrara
shale formation. Computational Intelligence and Neuroscience 1–10. http://dx.doi.
org/10.1155/2022/7084514.
Ivakhnenko, A.G. 1966. Group method of data handling a rival of the method of stochastic
approximation. Soviet Automatic Control 1(13): 43–71.
Ivakhnenko, A.G. 1971. Polynomial theory of complex system. IEEE Transactions on System,
Man and Cybernetics 1(4): 364–378. 10.1109/TSMC.1971.4308320.
Jaber, A.K., Alhuraishawy, A.K. and AL-Bazzaz, W.H. 2019. A data-driven model for
rapid evaluation of miscible CO2-WAG flooding in heterogeneous clastic reservoirs.
In: Proceedings of the SPE Kuwait Oil & Gas Show and Conference, 13–16. https://doi.
org/10.2118/198013-MS.
Jiang, H., Nuryaningsih, L. and Adidharma, H. 2010. The effect of salinity of injection brine
on water alternating gas performance in tertiary miscible carbon dioxide flooding:
experimental study. SPE Western Regional Meeting. https://doi.org/10.2118/132369-MS.
Kalam, S., Khan, R.A., Khan, S., Faizan, M., Amin, M., Ajaib, R. and Abu-Khamsin, S.A.
2021. Data-driven modelling approach to predict the recovery performance of low-salinity
waterfloods. Natural Resources Research 30: 1697–1717. https://doi.org/10.1007/s11053-
020-09803-3.
Kharecha, P.A. and Hansen, J.E. 2008. Implications of peak oil for atmospheric CO2 and climate.
Global Biogeochem 22: 6–10. https://doi.org/10.48550/arXiv.0704.2782.
Kulkarni, M.M. and Rao, D.N. 2004. Experimental investigation of various methods of tertiary
gas injection. Paper Presented at the SPE Annual Technical Conference and Exhibition.
https://doi.org/10.2118/90589-MS.
Kumar, H., Shehata, A. and Nasr-El-Din, H. 2016. Effectiveness of low salinity and CO2 flooding
hybrid approaches in low permeability sandstone reservoirs. In: SPE Trinidad and Tobago
Section Energy Resources Conference. https://doi.org/10.2118/180875-MS.
Lv, Q., Zhou, T., Zheng, R., Nakhaei-Kohani, R., Riazi, M., Hemmati-Sarapardeh, A.,
Li, J. and Wang, W. 2023. Application of group method of data handling and gene
expression programming for predicting solubility of CO2-N2 gas mixture in brine. Fuel
332(6): 126025, 10.1016/j.fuel.2022.126025.
Ma, S. and James, L.A. 2022. Literature review of hybrid CO2 low salinity water-alternating-
gas injection and investigation on hysteresis effect. Energies 15(21): 7891. https://doi.
org/10.3390/en15217891.
Mandadige, S.P., Ranjith, P.G., Tharaka, D.R., Ashani, S.R., Koay, A. and Choi, X. 2016. A
review of CO2-enhanced oil recovery with a simulated sensitivity analysis. Energies
9(7): 7–22. https://doi.org/10.3390/en9070481.
Naderi, S. and Simjoo, M. 2019. Numerical study of low salinity water alternating CO2 injection
for enhancing oil recovery in a sandstone reservoir: coupled geochemical and fluid flow
modeling. Journal of Petroleum Science and Engineering 173: 279–286. https://doi.
org/10.1016/j.petrol.2018.10.009.
 CO2 LSWAG Oil Recovery Factor Prediction in Carbonate Reservoir 157

Pourafshary, P. and Moradpour, N. 2019. Hybrid EOR methods utilizing low-salinity water.
Enhanc. Oil Recovery Process. New Technol. 8: 25.
Ramanathan, R., Shehata, A. and Nasr-El-Din, H. 2015. Water alternating CO2 injection process-
does modifying the salinity of injected brine improve oil recovery? In: Proceedings of the
OTC Brasil, Rio de Janeiro, Brazil, 27 October; Offshore Technology Conference: Rio de
Janeiro, Brazil. https://doi.org/10.4043/26253-MS.
Ramanathan, R., Shehata, A. and Nasr-El-Din, H. 2016. Effect of rock aging in oil recovery
during water-alternating-CO2 injection process. In: SPE Improved Oil Recovery
Conference. https://doi.org/10.2118/179674-MS.
Robertson, E.P. 2007. Low-Salinity Waterflooding to Improve Oil Recovery-Historical Field
Evidence. Paper Presented at the Annual Technical Conference and Exhibition, 11–14,
https://doi.org/10.2118/109965-MS.
Roustazadeh, A., Ghanbarian, B., Shadmand, M.B., Taslimitehrani, V. and Lake, L.W. 2022.
Estimating Oil and Gas Recovery Factors via Machine Learning: Database-Dependent
Accuracy and Reliability. Preprint. http://dx.doi.org/10.48550/arXiv.2210.12491.
Santos, R., Loh, W., Bannwart, A. and Trevisan, O. 2014. An overview of heavy oil properties
and its recovery and transportation methods. Brazilian Journal of Chemical Engineering
31(3): 571–576. http://dx.doi.org/10.1590/0104-6632.20140313s00001853.
Saxena, K. 2017. Low Salinity Water Alternate Gas Injection Process for Alaskan Viscous Oil
EOR [Master’s Thesis, University of Alaska]. University of Alaska Fairbanks. http://hdl.
handle.net/11122/7638.
Sharma, A., Srinivasan, S. and Lake, L.W. 2010. Classification of oil and gas reservoirs based
on recovery factor: a data-mining approach. In: SPE Annual Technology Conference &
Exhibition 1: 50–70. https://doi.org/10.2118/130257-MS.
Sheng, J.J. 2011. Modern Chemical Enhanced Oil Recovery: Theory and Practice. Elsevier
Publishing Corporation.
Sheng, J.J. 2013. Enhanced Oil Recovery Field Case Studies. Waltham, Mass.: Gulf Professional
Publishing.
Stalkup, F.I. 1984. Miscible Displacement (SPE Monograph Series). Society of Petroleum
Engineers.
Stosur, G.J. 2003. EOR: Past, Present, and What the Next 25 Years May Bring. Paper Presented
at the SPE International Improved Oil Recovery Conference in Asia Pacific, Kuala Lumpur,
Malaysia. https://doi.org/10.2118/84864-MS.
Sun, X., Zhang, Y., Chen, G. and Gai, Z. 2017. Application of nanoparticles in enhanced oil
recovery: a critical review of recent progress. Energies 10(3): 1–6. https://doi.org/10.3390/
en10030345.
Tahmasebi, P., Kamrava, S., Bai, T. and Sahimi, M. 2020. Machine learning in geo-and
environmental sciences: From small to large scale. Advances in Water Resources
142(11): 103619. http://dx.doi.org/10.1016/j.advwatres.2020.103619.
Teklu, T.W., Alameri, W., Graves, R.M., Kazemi, H. and AlSumaiti, A.M. 2016. Low salinity
water-surfactant-CO2 EOR. Journal of Petroleum Science Engineering 3(3): 309–320.
https://doi.org/10.1016/j.petlm.2017.03.003.
Verma, M.K. 2015. Fundamentals of carbon dioxide enhanced oil recovery (CO2-EOR): A
supporting document of the assessment methodology for hydrocarbon recovery using
CO2-EOR associated with carbon sequestration. U.S. Geological Survey 15: 1–19.
https://doi.org/10.3133/ofr20151071.
Vledder, P., Fonseca, J.C., Wells, T., Gonzalez, I. and Ligthelm, D. 2010. Low Salinity Water
Flooding: Proof of Wettability Alteration on a Field Wide Scale. Presented at the SPE
Improved Oil Recovery Symposium, 24–28. https://doi.org/10.2118/129564-MS.
158 Data Science and Machine Learning Applications in Subsurface Engineering

Webb, K.J., Black, C.J.J. and Al-Ajeel, H. 2004. Low Salinity Oil Recovery Log-Inject-
Log. Paper presented at the SPE/DOE Symposium on Improved Oil Recovery, 17–21.
https://doi.org/10.2118/89379-MS.
Wu, D., Brantson, E.T. and Ju, B. 2021. Numerical simulation of water alternating gas flooding
(WAG) using CO2 for high-salt argillaceous dolomite reservoir considering the impact of
stress sensitivity and threshold pressure gradient. Acta Geophysica 69(4): 1349–1365.
https://doi.org/10.1007/s11600-021-00601-w.
Yang, D., Tontiwachukwuthikul, P. and Gu, Y. 2005. Interfacial tensions of the crude oil +
reservoir brine + CO2 systems at pressures up to 31 MPa and temperatures of 27°C and
58°C. Journal of Chemical Engineering Data 50(4): 1242–1249. https://doi.org/10.1021/
je0500227.
Yong, T., Zhengywan, S., Jibo, H. and Fulin, Y. 2016. Numerical simulation and optimization
of enhanced oil recovery by the in situ generated CO2 huff-n-puff process with compound
surfactant. Journal of Chemistry 206: 13. https://doi.org/10.1155/2016/6731848.
Zekri, A., Al-Attar, H., Al-Farisi, O., Almehaideb, R. and Lwisa, E.G. 2015. Experimental
investigation of the effect of injection water salinity on the displacement efficiency
of miscible carbon dioxide WAG flooding in a selected carbonate reservoir. Journal of
Petroleum Exploration and Production Technology 5: 363–373. https://doi.org/10.1007/
s13202-015-0155-0.
Zolfaghari, H., Zebarjadi, A., Shahrokhi, O. and Ghazanfari, M.H. 2013. An experimental study
of CO2-low salinity water alternating gas injection in sandstone heavy oil reservoirs. Iranian
Journal of Oil and Gas Science and Technology 2(3): 37–47. https://doi.org/10.22050/
ijogst.2013.3643.
 Chapter 8
Improving Seismic Salt Mapping
through Transfer Learning Using
A Pre-trained Deep Convolutional
Neural Network
A Case Study on Groningen Field
Daniel Asante Otchere,1,2,* Abdul Halim Latiff,1 Nikita Kuvakin,3
Ruslan Miftakhov,3 Igor Efremov 3 and Andrey Bazanov 3

1. Introduction
The interpretation of seismic data is a critical aspect of geological exploration.
It enables geologists and engineers to identify and delineate subsurface
structures such as faults, reservoirs, and geological formations. However,
traditional approaches to seismic interpretation, such as manual picking and
horizon tracking, are laborious and time-consuming, particularly in areas
with complex geology and numerous faults. Furthermore, these methods are
prone to noise and other stratigraphic challenges, making them less reliable
and accurate (Otchere et al., 2022c). Getting a reliable velocity model using

1
Centre of Research for Subsurface Seismic Imaging, Universiti Teknologi PETRONAS, 32610, Seri
Iskandar, Perak Daril Ridzuan, Malaysia.
2
Institute for Computational and Data Sciences, Pennsylvania State University, University Park, PA,
USA.
3
GridPoint Dynamics, 77 Hopton Road, London, SW162EL, United Kingdom.
* Corresponding author: [email protected]
160 Data Science and Machine Learning Applications in Subsurface Engineering

seismic salt imaging requires consistent interpretation analysis. However, the

following restrictions may apply to traditional methods:
1. Clearly delineated salt boundaries at lithological transition zones.
2. Mapping complex salt build-ups require a great deal of time.
3. Instead of using automatic picking, a significant amount of manual
horizon interpretation is required.
Salts are essential subsurface structures that significantly impact the
geology of a region, and salt tectonics play a critical role in the storage of CO2
gas, which is an essential part of the global energy transition journey (Duffy
et al., 2022). The study of salts and their geological properties is of immense
importance for the energy industry and our understanding of the Earth’s
subsurface processes. The application of deep learning has revolutionised the
oil and gas industry in the last decade, providing a new and efficient approach
to seismic interpretation (Otchere et al., 2022c), reservoir engineering (Otchere
et al., 2022a), and subsurface characterisation (Otchere et al., 2022b). This
technology has significantly contributed to mapping and imaging subsurface
structures, particularly salt structures, which are notoriously difficult to
interpret due to their complex geometries and stratigraphic complexities.
Compared to traditional seismic interpretation methods, deep learning-based
algorithms offer several advantages, such as higher accuracy and faster
processing times. Additionally, the ability of deep learning models to learn
from large datasets and identify complex patterns makes them ideal for
identifying subtle subsurface features that conventional techniques may miss.
Consequently, applying deep learning-based algorithms in the oil and gas
industry has grown tremendously in recent years. It is expected to continue
to play a significant role in future subsurface characterisation, interpretation,
and modelling.
Seismic edge-detection algorithms have been widely used in the oil
and gas industry to detect and interpret stratigraphic features in seismic
data volumes. These algorithms have become increasingly popular due to
their ability to extract subtle features in seismic data that may be indicative
of subsurface structures (Chopra and Marfurt, 2007). They are particularly
useful in areas where the subsurface is complex and difficult to interpret, such
as in areas with numerous faults or where stratigraphic layering is intricate.
Furthermore, edge-detection algorithms have been successfully integrated
with machine learning techniques to improve the accuracy of interpretation
results and reduce the time required for interpretation (Otchere et al., 2022c).
These advancements have opened up new possibilities for the efficient
and effective mapping and characterisation of subsurface structures, with
 Improving Seismic Salt Mapping through Transfer Learning 161

significant implications for the oil and gas industry. However, deploying the
method to new fields gives inaccurate mapping interpretation. One way of
boosting deep learning model performance in a new field is through transfer
learning.
In recent years, there has been a growing interest in using transfer
learning to improve the accuracy of salt segmentation in seismic images.
Transfer learning can significantly enhance our ability to identify and map
subsurface salt structures, which can significantly impact the development of
new energy resources and climate change mitigation. Using the specialised
information learned in one context to solve the problems in another is known
as transfer learning, which is a promising approach to address disparities in
model prediction and ground truth (Li et al., 2020; Pan and Yang, 2010). The
idea is to get insights into a problem using a plethora of available interpreted
data when the initial model prediction of a comparable situation for which less
information is not desirable. This calls for attention to the distinction between
the model prediction and ground truth, reflected in the different distributions of
their features and boundaries. In particular, the feature-based transfer learning
strategy known as “domain adaptation” is a potential option (Ben-David et al.,
2007; Pan et al., 2011). To do this, it first determines a feature space in which
source domain data retain their intrinsic structure while minimising distribution
disparity across domains. Transfer component analysis is a pioneering
technique described by (Pan et al., 2011). Minimisation of maximum mean
discrepancy (MMD) has emerged as a popular method for domain adaptation
in machine learning. This method aims to learn transferable features between
different domains while preserving the variation in the source domain. By
representing the feature space as a reproducing kernel Hilbert space (RKHS),
MMD minimisation enables the computation of distances between probability
distributions, facilitating knowledge transfer from the source to the target
domain. In this way, MMD-based domain adaptation has been successfully
applied to various problems, including image classification, object detection,
and natural language processing. An implementation with multiple kernels
has been developed using this method (Lixin Duan et al., 2012).
Deep learning models have revolutionised the field of machine learning
by enabling computers to automatically learn complex representations of data.
The ability to learn features from data has made deep learning particularly
useful in domains where large amounts of data are available, such as computer
vision and natural language processing. Recently, there has been an increasing
interest in combining deep learning with transfer learning, which aims to
leverage knowledge from related tasks to improve performance on new tasks.
To this end, Ghifary et al. (2014) introduced a deep adaption neural network
(DaNN) that includes a multi-modal dependency (M-MD) term in the loss
162 Data Science and Machine Learning Applications in Subsurface Engineering

function to minimise MMD while maintaining variation in the source domain.

This approach has shown promising results in various applications, including
image classification and speech recognition.
Domain adaptation has become an increasingly popular area of research
in recent years as more and more problems arise with data distribution
differences between the source and target domains. Among the various
techniques developed to address these issues, domain-adversarial neural
networks (DANN) have shown great potential (Ganin et al., 2016). The key
idea behind DANN is to build a network that can predict the source domain
problem while avoiding the discrimination between the source and target
domains. This is achieved by using an adversarial loss function that helps
minimise the difference between the feature representations of the source
and target domains. The use of DANNs has been demonstrated in several
applications, including image recognition, natural language processing, and
speech recognition.
Shi et al. (2019) defined the challenge of interpreting and extracting salt
boundaries as a 3D image segmentation problem and assessed a deep CNN
technique using an encoder-decoder architecture. The authors designed a data
generator to train the model that extracts randomly positioned sub-volumes
from a large-scale 3D training data set, followed by data reinforcement. The
binary classification of sub-volumes into salt and non-salt was used as ground
truth labels by thresholding the velocity model. These were then fed into the
network, resulting in numerous iterations of network training. This approach
enabled the network to learn salt features effectively, leading to more accurate
salt boundary detection in seismic volumes. Their model, when validated,
automatically extracted minor salt properties from 3D seismic sections.
Liu et al. (2019) also used a Squeeze-Extraction Feature Pyramid Network
(Se-FPN) for salt deposits image segmentation. They used SeNet as a backbone
to implicitly learn to suppress unimportant regions while highlighting key
features. The authors suggested an enhanced FPN to combine multiscale
information, where hypercolumns were added to the network to combine
information from multiple scales. The proposed Se-FPN, when deployed in
the TGS Salt Identification Challenge, got high-quality segmentation with a
Mean Intersection over Union (MIU) of 0.86. To overcome the discrepancies
between model fault prediction and ground truth (Yan et al., 2021), used
transfer learning to improve the prediction performance of a CNN pre-trained
with synthetic labels. The authors retrained the model using actual field
seismic interpretations. They used random sample consensus (RANSAC) to
automatically obtain and categorise actual seismic samples. This approach
resulted in an improvement in fault detection accuracy. The existing literature
provides a theoretical foundation for our objective, which is to leverage
 Improving Seismic Salt Mapping through Transfer Learning 163

transfer learning to enhance the detection of salt bodies and evaluate the
effectiveness of the pre-trained CNN. These prior investigations demonstrate
that transfer learning can be used to improve the accuracy of CNNs for image
recognition tasks, even when the target dataset differs from the source dataset.
The domain shifts (actual-predicted discrepancy) between the predicted and
ground truth can be reduced to an acceptable level in a predetermined feature
space using domain adaptation through transfer learning.
This study will demonstrate the application of transfer learning using
actual field interpretation to a CNN pre-trained with synthetic labels to
generate salt probability models that can be used as a valuable property in the
seismic imaging and velocity modelling phases. Transfer learning and deep
learning techniques’ object and edge detection have demonstrated promising
success in various domains, making them an appealing approach for seismic
salt mapping. The use of these techniques can potentially enhance the detection
of subsurface salt bodies and improve the accuracy of salt mapping, resulting
in more efficient and effective exploration and production of hydrocarbon
reservoirs. Hence, the main contributions of this research are:
1. Improved salt segmentation accuracy: Transfer learning can significantly
enhance our ability to identify and map subsurface salt structures in
seismic images. Using semantic segmentation to improve salt probability
volume by retraining the model on a labelled dataset, it can learn to
recognise salt structures with greater accuracy and efficiency, which
is critical for the energy industry and our understanding of the Earth’s
subsurface processes.
2. Cost and time savings: Salt segmentation in seismic images is
time-consuming and labour-intensive, requiring significant expertise and
resources. By using transfer learning to automate this process, we can save
time and reduce costs while improving the accuracy of the segmentation.
3. Enhanced data analysis: Accurate segmentation of salt structures in
seismic images is essential for various applications, including hydrocarbon
exploration, CO2 storage, and geohazard assessment. By improving the
accuracy of salt segmentation, we can enhance our understanding of the
subsurface geology and make more informed decisions about energy
exploration and production.
4. Potential for future research: Using transfer learning for salt segmentation
in seismic images is a relatively new area of research with significant
potential for future exploration and development. By demonstrating the
effectiveness of this approach, we can inspire new research in the field
and help to advance our understanding of subsurface processes and the
energy transition journey.
164 Data Science and Machine Learning Applications in Subsurface Engineering

These contributions will help reduce the time and resources seismic
interpreters spend interpreting seismic features. The findings in this study
are of great importance for the energy industry and the broader scientific
community, as they provide new insights into the use of transfer learning for
improving the accuracy of salt segmentation in seismic images. The results of
this study will also highlight the potential of this approach for developing new
energy resources and mitigating climate change by improving our ability to
identify and map subsurface salt structures accurately.

2. Method
2.1 Collection and Description of Data
For this work, the Groningen seismic field data was selected as the case study.
The system unit was a Windows 10 Operating System (OS) with the next
generation AMD Radeon Pro™ and the highest-performing NVIDIA Quadro®
professional graphics capable of 2-petaFLOPS tensor performance. A single
NVIDIA RTX 8000 GPU and quadruple Intel (R) Xeon (R) W-2223 i5 CPU
running at 3.60 GHz, with 32.0 GB of DDR4-2666 MHz DRAM, were used
to run the study. This research used a pre-trained DCNN model trained using
synthetic and real data. The proposed approach consisted of two main parts,
generating a salt probability model and improving the model output through
transfer learning, as illustrated in Fig. 8.1. The optimised DCNN model for
salt body probability prediction can be reused to map new salt bodies, thus
enhancing interpretation efficiency. The model estimation section of the
workflow represents an end-to-end model that automatically can map salt
bodies after selecting several parameters.

2.2 Deep Convolutional Neural Network in Salt Mapping

and Post-processing
For a long time, seismic data has become increasingly important in oil and
gas exploration due to the valuable information it provides about subsurface
structures. Seismic data is usually obtained by sending a wave signal into the
ground and measuring the reflected signals that bounce back from different
subsurface layers. This data is processed and interpreted to create seismic
images of the subsurface structures. With the advancement in exploration
technology, 3D seismic data volumes have become more common, providing a
more detailed and accurate representation of the subsurface geology. However,
the large volume of data generated by this technology presents a significant
challenge for interpretation and analysis, necessitating the development of more
efficient and effective techniques. To reduce the computation cost associated
Fig. 8.1. Illustration of workflow using a pre-trained transfer learned DCNN seismic salt mapping model.
Improving Seismic Salt Mapping through Transfer Learning
165
166 Data Science and Machine Learning Applications in Subsurface Engineering

with training a neural network directly on 3D seismic data, the Geoplat AI

software was employed in this study to import the entire 3D seismic volume
and the interpreted salts of the Groningen field. The software’s pre-trained
DCNN, which is based on a computer vision-based model, was used. The
model was trained on an extensive and diverse 3D synthetic seismic dataset
image library created by fragmenting labelled images. These synthetic data
were carefully designed to capture complex geological factors and imaging
quality. These images were segmented using the DCNN model, which was
initially trained to assign a label of 1 to salt-labelled images and a label of 0 to
non-salt-labelled images, as illustrated in Fig. 8.2. The accurate segmentation
of a synthetic volume fragment is shown on the left, while the predicted output
created by the pre-trained neural network is shown on the right.
The pre-trained DCNN model’s outstanding performance prompted
its application to the Groningen Field’s actual dataset for automatic salt
probability volume creation. The selected fragment size for the prediction was
based on visually identifying salt tops and bottoms in the seismic volume.
The resultant inline, crossline, and depth sections of the Groningen Field
are displayed in Fig. 8.3. The ResU-net architecture framework, which has
been pre-trained, is used to segment the 3D seismic images obtained from
the seismic volume. Figure 8.4 illustrates the application framework of the
DCNN for the creation of the salt probability volume. An additional signal
gain was applied to handle low-quality seismic data, which amplifies and
restores the regions with amplitude expression coefficient loss.

Fig. 8.2. (a) Accurate segmentation of a synthetic volume fragment, and (b) projected result produced
using a pre-trained neural network.
Fig. 8.3. Seismic volume showing Inline 8768, Xline 8367, and depth 2800.
Improving Seismic Salt Mapping through Transfer Learning
167
168 Data Science and Machine Learning Applications in Subsurface Engineering

Fig. 8.4. Illustration of the application framework of the DCNN in creating the salt probability
volume.

2.2.1 Simplified Architecture of Residual U-net

The ResU-net architecture used in this study serves as the backbone of the
DCNN architecture employed. The ResU-net capitalises on the power of deep
residual learning to overcome encoder challenges by transferring information
between layers, resulting in a deep neural network that effectively deals with
degradation issues. Figure 8.5 depicts a simplified schematic illustration
of the ResU-net, which utilises complete pre-activation deep residual
units to train the model. Using bypass connections inside networks aids
information propagation without degradation, enhancing the convolutional
neural network’s architecture by reducing parameters while maintaining or
improving the transfer learning function’s efficiency.

Fig. 8.5. A simplified end-to-end U-network architecture for 3D salt mapping.

 Improving Seismic Salt Mapping through Transfer Learning 169

2.3 Transfer Learning Application

Semantic segmentation can help improve probability findings tailored to the
interpretation interval’s peculiarities. The failure of the default probability
volume to deliver quantitative results underpins the necessity for the transfer
learning technique. Concerning post-processing, there is the possibility of
fine-tuning the DCNN to handle the problem of thorough tracking of faults
peculiar to a specific section of the seismic area. The DCNN allows it to
address the difficulty of automating the interpretation of highly complicated
faults. Hence, expert transfer learning based on manually interpreted salt
horizons was incorporated to improve the accuracy of the prediction in places
where there is minimal definitive evidence of salt and also to include the
existing interpreted salts in the Groningen Field.
In applying the transfer learning process, one set of salt body interpretations
on Inlines 9090, 7820, and 9108 was used as the training segment. In contrast,
a different set on Inline 8271 was used as the validation section (Fig. 8.6). A
different set of interpreted salt on crosslines 7960, 9034, and 8188 was used
as part of the transfer learning process, with the latter used for validation
(Fig. 8.7). The training and validation segments were chosen irrespective of
each other and ease of interpreting salt horizons or otherwise. After selecting
the interpreted salt sections for the transfer learning process, several iterations
and output weights were specified with the results compared. The iterations
and weights are systematically chosen based on the erroneous probabilities
in the probability volume created using new weights and the dice coefficient
value on training and validation segments. Where underfitting is observed, the
number of iterations is increased to improve the model performance whiles
also critically avoiding overfitting.

2.4 Criteria for Model Evaluation

In assessing the performance of the DCNN predicted salt bodies, the models’
errors were assessed using the Dice Similarity Coefficient (DSC), commonly
known as the Sørensen–Dice index, is a numerical tool based on spatial overlap
used for determining how similar two sets of data are. This index is likely the
most widely used tool for validating artificial intelligence-based (AI) image
segmentation predictions. The validation challenge becomes an assessment
of segmentation reproducibility and spatial overlap index. It was sometimes
referred to as the percentage of explicit agreement (Zou et al., 2004). DSC
values range from 0 to 1, with 0 indicating no spatial overlap and 1 indicating
 170 Data Science and Machine Learning Applications in Subsurface Engineering

Fig. 8.6. Expert interpreted salt bodies showing: (a) Inline 9090, (b) Inline 7820, (c) Inline 8271 and (d) Inline 9108.
 Improving Seismic Salt Mapping through Transfer Learning 171

Fig. 8.7. Expert interpreted salt bodies showing: (a) Crossline 9034, (b) Crossline 7960, and
(c) Crossline 8188.
172 Data Science and Machine Learning Applications in Subsurface Engineering

Fig. 8.8. Illustration of DSC spatial overlap metric.

complete overlap between two sets of binary image segmentation results, A and
B target regions (Fig. 8.8). The DSC is defined as (Yao et al., 2020);
DSC(A, B) = 2(A ∩ B)/(|A| + |B|) (8.1)
where ∩ represents the intersection of the two target regions |A| and |B| is the
cardinality of the set. When there is a much higher number of background voxels
than the number of target voxels, the DSC can be conceptualised as a special
case of the kappa statistic, typically applied in situations involving reliability
analyses. This concept has been shown previously by Zou et al. (2004).

3. Results and Discussion

3.1 Calculated Salt Body Volume
In Fig. 8.9, the low level of accuracy in predicting salt probability bodies is
noted. However, the possibility of using DCNN to map salt bodies on the
base of seismic data has been illustrated. This low accuracy predicted salt
probability volume could give an indication of the location and boundaries
of the salt bodies. While the calculated salt body volume provides a broad
indication of the presence of salt in the Groningen Field, it does not provide
precise information regarding the exact position of the top, base, and wall
sides of salt bodies. This is because the calculated volume is based on the
probabilistic output of the DCNN model. As such, it may contain sections
of “false probability” where the analysis of the reflectivity pattern does not
suggest the presence of a salt body.
 Improving Seismic Salt Mapping through Transfer Learning 173

Fig. 8.9. Comparison of predicted salt probability volume (left) and expert interpretation (right) of
(a) Crossline 8188, (b) Inline 9090, and (c) Crossline 9034.

Nonetheless, these areas are still shown on the attribute as having a

high probability of salt occurrence, indicating the potential presence of salt.
Further analysis and interpretation by geoscientists are required to accurately
determine the position and extent of salt bodies within the field. This event is
because “false probability” sections occur when there is no obvious evidence
of a salt body. The occurrence of probabilities, which are bodies traced in a
distinct direction, is projected to have a salt probability. These probabilities
are a fragmented portrayal of some extensive salt bodies and the requirement
for vertical extent rectification. These false probabilities occur because
synthetic and field labels used to pre-trained DCNN do not capture all the
intrinsic characteristics in every field. Some of these false probabilities are
salt probability zone projections tracked in the perpendicular direction. One
approach for enhancing the probability attribute in such cases is by adding
manually interpreted salt labels to retrain the pre-trained DCNN. Due to the
unsatisfactory salt body prediction by the pre-trained model, transfer learning
is recommended to help improve the model performance by tuning and
recalculating the model weights to the type of salt bodies in this study. This
approach will decrease the number of “false probability” results where the
analysis of the reflectivity pattern does not suggest salt being present.
174 Data Science and Machine Learning Applications in Subsurface Engineering

3.2 Semantic Segmentation – Transfer Learning Application

A high similarity between the transfer learned salt body detection and the
interpreted salt bodies in the seismic sections could be noticed in the final salt
probability volume, as shown in Figs. 8.11 and 8.10. The neural network was

Fig. 8.10. Transfer learned salt probability volume and expert interpretation (solid black line)
showing: (a) Crossline 9034, (b) Crossline 7960, and (c) Crossline 8188.
Fig. 8.11. Transfer learned salt probability volume and expert interpretation (solid black line) showing: (a) Inline 9090, (b) Inline 7820, (c) Inline 8271 and
(d) Inline 9108.
Improving Seismic Salt Mapping through Transfer Learning
175
176 Data Science and Machine Learning Applications in Subsurface Engineering

able to successfully retrain itself on the new labelled data set by readjusting its
weights in the deep learning process. As a result, the accuracy in the prediction
of salt probability bodies has increased compared to the initial results obtained
before the retraining of the CNN. The results showed a statistically significant
improvement when iterations were increased in steps of 100 from 100 to
10,000. The best combination of input data and iterations was determined
by comparing the pair-wise DSC metrics of all the iterations and increasing
the number of input labelled data. DSC results sequentially increased with
successive increments in labelled training data and iterations until they reached
their maximum based on case-to-case analysis. There was no evidence of a
learning curve phenomenon, and segmentations seemed to have little effect
after iterations 10,000 and input labelled data set size above 5.
This research demonstrated the functionality of DSC as a straightforward
criterion for validating the consistency and spatial overlap accuracy of
manual and automated segmentations. The corresponding DSC values for
iteration 10,000 using seven input labelled data for training and validation
(four inlines and three crosslines) are shown in Fig. 8.12. The probabilistic
fractional segmentation showed a wide range of spatial overlap (0.6–0.95)
with the corresponding estimated ground truth (expert interpreted sections).
This result gives an indication that further improvements can be achieved.
In the inline sections, the reproducibility was significantly higher based on a
higher number of labelled training inputs. This improvement motivated the
increment of the labelled data size in subsequent crossline investigations to
develop a suitable matching approach to register these salt body boundaries

Fig. 8.12. Learning transfer metrics showing Dice Similarity Coefficient results at each iteration.
 Improving Seismic Salt Mapping through Transfer Learning 177

and enhance the visual representation of salt body mapping. However,

increasing the training labelled data for the crossline section did not yield any
significant improvements.

3.3 Sensitivity Analysis of Model and Expert Interpretations

Image segmentation is a critical task in many fields, including geophysics,
where it can be used to identify and extract features of interest from seismic
data. The accuracy and consistency of segmentation results are essential for
ensuring the reliability of subsequent analysis and interpretation. One approach
to improving segmentation performance is to use transfer learning, which
involves fine-tuning a pre-trained CNN on new data. The transfer learned
CNN based on the statistical validation analysis based on DSC was expanded
to different seismic sections from the same field compared to the ground
truth of expert interpretations. The results in Fig. 8.13 indicate no significant
difference in DSC between the predicted salt probability volume and manual
interpretation. This result suggests that the repeated segmentations through
transfer learning did not introduce significant bias. This observation is an
important finding because it indicates that transfer learning can be used to
effectively extend a model’s performance to new data without compromising
its accuracy or consistency.
To highlight the critical points for reinterpretations, we recommend that
future studies focus on the following areas:
1. Transfer learning can be a practical approach for extending the
performance of a segmentation model to new data.
2. The dice similarity index can be a valuable metric for evaluating
segmentation performance. However, it is recommended to include other
metrics, like Intersection over Union (IoU), to confirm the final results.
3. The use of transfer learning did not introduce significant bias into the
segmentation results.
4. Further validation studies can help confirm segmentation results’
reliability and consistency using transfer learning.
In summary, our study demonstrates the potential of transfer learning for
improving the accuracy and consistency of image segmentation in geophysics.
Using the DSC as our evaluation metric, we validated the performance of our
transfer learned CNN on new data and confirmed the absence of significant
bias. These findings provide valuable insights for future research and the
development of automated geophysical segmentation tools.
178 Data Science and Machine Learning Applications in Subsurface Engineering

Fig. 8.13. Transfer learned salt probability volume and expert interpretation (solid black line) showing
the deployed sections of Inline 8271, crossline 8188 and depth 2048.

4. Conclusions
The seismic interpretation field has faced significant challenges due to the
limited availability of labelled seismic data. These challenges have created a
bottleneck in developing deep learning algorithms to aid seismic interpretation.
Accurate and high-quality seismic data are critical to successfully applying
machine learning algorithms to interpret subsurface geological structures.
Without sufficient labelled data, deep learning models are unable to learn the
patterns and correlations necessary to accurately predict subsurface features,
such as salt bodies. In order to solve this issue, a CNN that had previously
been trained on synthetic labels was considered a potential source of relevant
data for developing an adequate salt segmentation model. The application
of transfer learning was required to reduce the visual discrepancy across the
seismic volume because of the inevitable disparity between the model and
the actual data. The application of this process in the Groningen field has
demonstrated the effectiveness of integrating DNNs and transfer learning
for salt body detection. In its conceptualisation, DSC is a particular instance
of the kappa statistic, a well-known and pragmatic reliability and agreement
index. It offers the metric to analyse and measure the model’s performance
and the adjustments necessary for the retraining process. The transfer learning
process might involve several repetitions. Depending on the data quality and
the geological complexity, it may be necessary to complete several iterations
 Improving Seismic Salt Mapping through Transfer Learning 179

to achieve a representative model for the deployed seismic field. The transfer
learning technique in this study achieved a Dice similarity index of 0.99 and
0.92 on the training and validation sections, respectively. This result shows
that the transfer learned model can automatically capture subtle salt bodies
from 3D seismic with minimal manual input.
The proposed method advocates for continued research and deep learning
techniques implementation in seismic interpretation. This activity involves
manual seismic interpretation and often requires much expert knowledge. It
would be beneficial to have labelled data of interpreted seismic from different
fields. Despite a large portion of training data coming from synthetic data, the
successful knowledge transfer still requires some field interpretation samples
from actual seismic volumes. Once labelled data is made available, the time
involved in fully interpreting salt boundaries or other attributes on seismic
volumes will be drastically reduced because of the generalisation of the DCNN
model and can accurately map and interpret different seismic volumes.

Data and Software Availability

The data used in this study is available at https://geo.public.data.uu.nl/vault-
nam-geological-model/Publication%5B1605778324%5D/original/. To access
the software used in this article, contact the developers at https://geoplat.
ai/#demo.

Acknowledgement
The authors wish to extend their genuine gratitude to the University Teknologi
Petronas and the Centre for Subsurface Seismic Imaging for their support and
valuable contributions to this research and to Geoplat AI for providing the
software for this work.

References
Ben-David, S., Blitzer, J., Crammer, K. and Pereira, F. 2007. Analysis of representations for
domain adaptation. pp. 137–144. In: Advances in Neural Information Processing Systems
19. The MIT Press. https://doi.org/10.7551/mitpress/7503.003.0022.
Chopra, S. and Marfurt, K.J. 2007. Volumetric curvature attributes for fault/fracture
characterisation. First Break 25: 35–46. https://doi.org/10.3997/1365-2397.2007019.
Duffy, O.B., Hudec, M., Peel, F., Apps, G., Bump, A., Moscardelli, L., Dooley, T., Bhattacharya,
S., Wisian, K. and Shuster, M. 2022. The Role of Salt Tectonics in the Energy Transition: An
Overview and Future Challenges. Earth Arxiv Preprint. https://doi.org/10.31223/X5363J.
Ganin, Y., Ustinova, E., Ajakan, H., Germain, P., Larochelle, H., Laviolette, F., Marchand, M.
and Lempitsky, V. 2016. Domain-adversarial training of neural networks. Advances in
Computer Vision and Pattern Recognition 189–209. https://doi.org/10.1007/978-3-319-
58347-1_10.
180 Data Science and Machine Learning Applications in Subsurface Engineering

Ghifary, M., Kleijn, W.B. and Zhang, M. 2014. Domain adaptive neural networks for object
recognition. pp. 898–904. In: Pacific Rim International Conference on Artificial
Intelligence. Springer, Cham. https://doi.org/10.1007/978-3-319-13560-1_76.
Li, W., Gu, S., Zhang, X. and Chen, T. 2020. Transfer learning for process fault diagnosis:
Knowledge transfer from simulation to physical processes. Comput. Chem. Eng.
139: 106904. https://doi.org/10.1016/j.compchemeng.2020.106904.
Liu, B., Jing, H., Li, J., Li, Y., Qu, G. and Gu, R. 2019. Image segmentation of salt deposits
using deep convolutional neural network. pp. 3304–3309. In: 2019 IEEE International
Conference on Systems, Man and Cybernetics (SMC). IEEE. https://doi.org/10.1109/
SMC.2019.8913858.
Lixin Duan, Tsang, I.W. and Dong Xu. 2012. Domain transfer multiple kernel learning. IEEE
Trans. Pattern Anal. Mach. Intell. 34: 465–479. https://doi.org/10.1109/TPAMI.2011.114.
Otchere, D.A., Abdalla Ayoub Mohammed, M., Ganat, T.O.A., Gholami, R. and Aljunid
Merican, Z.M. 2022a. A novel empirical and deep ensemble super learning approach
in predicting reservoir wettability via well logs. Applied Sciences 12: 2942. https://doi.
org/10.3390/app12062942.
Otchere, D.A., Ganat, T.O.A., Nta, V., Brantson, E.T. and Sharma, T. 2022b. Data analytics
and Bayesian Optimised Extreme Gradient Boosting approach to estimate cut-offs from
wireline logs for net reservoir and pay classification. Appl. Soft Comput. 120: 108680.
https://doi.org/10.1016/j.asoc.2022.108680.
Otchere, D.A., Tackie-Otoo, B.N., Mohammad, M.A.A., Ganat, T.O.A., Kuvakin, N.,
Miftakhov, R., Efremov, I. and Bazanov, A. 2022c. Improving seismic fault mapping
through data conditioning using a pre-trained deep convolutional neural network: A
case study on Groningen field. J. Pet. Sci. Eng. 213: 110411. https://doi.org/10.1016/J.
PETROL.2022.110411.
Pan, S.J. and Yang, Q. 2010. A survey on transfer learning. IEEE Trans. Knowl. Data Eng.
22: 1345–1359. https://doi.org/10.1109/TKDE.2009.191.
Pan, S.J., Tsang, I.W., Kwok, J.T. and Yang, Q. 2011. Domain adaptation via transfer
component analysis. IEEE Trans. Neural Netw. 22: 199–210. https://doi.org/10.1109/
TNN.2010.2091281.
Shi, Y., Wu, X. and Fomel, S. 2019. SaltSeg: Automatic 3D salt segmentation using a deep
convolutional neural network. Interpretation 7: SE113–SE122. https://doi.org/10.1190/
INT-2018-0235.1.
Yan, Z., Zhang, Z. and Liu, S. 2021. Improving performance of seismic fault detection by fine-
tuning the convolutional neural network pre-trained with synthetic samples. Energies
(Basel) 14: 3650. https://doi.org/10.3390/en14123650.
Yao, A.D., Cheng, D.L., Pan, I. and Kitamura, F. 2020. Deep learning in neuroradiology: a
systematic review of current algorithms and approaches for the new wave of imaging
technology. Radiol. Artif. Intell. 2: e190026. https://doi.org/10.1148/ryai.2020190026.
Zou, K.H., Warfield, S.K., Bharatha, A., Tempany, C.M.C., Kaus, M.R., Haker, S.J., Wells, W.M.,
Jolesz, F.A. and Kikinis, R. 2004. Statistical validation of image segmentation quality
based on a spatial overlap index1. Acad. Radiol. 11: 178–189. https://doi.org/10.1016/
S1076-6332(03)00671-8.
 Chapter 9
Super-Vertical-Resolution
Reconstruction of Seismic
Volume Using A Pre-trained Deep
Convolutional Neural Network
A Case Study on Opunake Field
Daniel Asante Otchere,1,2,* Abdul Halim Latiff,1 Nikita Kuvakin,3
Ruslan Miftakhov,3 Igor Efremov 3 and Andrey Bazanov 3

1. Introduction
Seismic surveys are essential for subsurface structure exploration and analysis
in the oil and gas and civil engineering sectors (Farfour et al., 2021; Talwani
and Kessinger, 2003). Reflection seismic surveys are acquired to map potential
subsurface features and structural and stratigraphic hydrocarbon traps (Haldar,
2018). The subsurface rock’s acoustic impedance changes produce reflections
due to the Earth’s reaction to synthetically generated acoustic waves near
the surface (Aminzadeh and Dasgupta, 2013; Selley and Sonnenberg, 2015).
Although the seismic energy travels in the form of elastic waves, most imaging
procedures presume that the recorded waves are exclusively composed of
compressional waves (Gray, 2014). Seismic image resolution is the ability
to accurately and delicately depict subsurface structures from the Earth’s
1
Centre of Research for Subsurface Seismic Imaging, Universiti Teknologi PETRONAS, 32610, Seri
Iskandar, Perak Daril Ridzuan, Malaysia.
2
Institute for Computational and Data Sciences, Pennsylvania State University, University Park, PA,
USA.
3
GridPoint Dynamics, 77 Hopton Road, London, SW162EL, United Kingdom.
* Corresponding author: [email protected]
182 Data Science and Machine Learning Applications in Subsurface Engineering

reflected energy, as it determines the level of detail that can be observed in

seismic images (Roy Chowdhury, 2011). However, several factors, including
subsurface structure complexity, noise and interference in the data, low sensor
resolution, seismic energy wavelength, and spatial sampling interval of the
data acquisition system, frequently limit the resolution of seismic images (Ker
et al., 2010). These constraints impede the efficiency of precisely interpreting
the subsurface structures, resulting in costly exploration and production
activity errors.
Artificial intelligence (AI) has been increasingly used in the oil and gas
industry to improve its activities in subsurface characterisation, reservoir
engineering, and production (Otchere et al., 2022a). Recent advances in AI,
notably computer vision, have shown promise in enhancing seismic image
quality and limiting the impact of noise and interference. Through learning
patterns and features in the seismic images and using this knowledge to fill
in missing or impaired sections of new images, AI approaches can be utilised
to improve the spatial resolution of seismic images. These approaches have
the potential to overcome conventional methods’ limitations and offer more
accurate and detailed images of subsurface structures.
Several studies have applied AI techniques to improve the resolution of
seismic images to improve subsurface interpretation activities. For example,
Li et al. (2022) used a deep convolutional neural network (DCNN) to enhance
low-resolution and noisy seismic images, demonstrating improved image
quality and resolution for fault detection. (Otchere et al., 2022b) employed
a DCNN to improve the seismic resolution of the Groningen field, resulting
in a more accurate interpretation of subsurface faults. Picetti et al. (2019)
utilised a generative adversarial network (GAN) to process low-quality
seismic migrated images and improve the resolution of features. These studies
demonstrate the potential of AI in overcoming the limitations of traditional
methods and improving the resolution of seismic images.
In summary, seismic image resolution plays a vital role in subsurface
exploration and analysis. However, seismic imaging has inherent limitations
regarding image resolution and quality, which might affect the accuracy and
reliability of the resulting interpretation. AI techniques have the potential
to improve seismic image resolution and quality using advanced machine
learning algorithms, such as denoising and super-resolution techniques, and
overcome the limitations of conventional techniques. Further research is thus
required to optimise and validate the potential of AI completely and to explore
its potential for automating the interpretation of seismic images.
This study will demonstrate the application of a pre-trained DCNN on
the Opunake field for seismic image resolution improvement and to improve
the signal-to-noise ratio (SNR) that can be used as a valuable property in the
seismic imaging and velocity modelling phases. Two image resolution DCNN
 Super-Vertical-Resolution Reconstruction of Seismic Volume 183

techniques will be compared to identify the most appropriate technique for

vertical resolution improvements. The reported successes of using synthetic
data to pre-train a DCNN due to the unavailability of data and deploying it on
a real field make AI a viable candidate for vertical and lateral seismic image
resolution.

2. Brief Overview
Seismic imaging is a well-known geophysical technique for visualising
underlying structures and features in the Earth’s subsurface. It is extensively
used in the oil and gas sector to discover and map hydrocarbon reservoirs and
in civil engineering for site characterisation and underground utility mapping.
Despite its widespread use, Seismic imaging has intrinsic limitations in terms
of image resolution and quality, which affects the accuracy and dependability
of the ensuing interpretation. Several factors can limit the resolution of seismic
images in the oil and gas industry (Deng et al., 2021; Sun et al., 2022; Yan and
Chunqin, 2008):
1. Seismic wavelength: The resolution of a seismic image is directly related
to the wavelength and finite size of the seismic source transmitted into the
subsurface. Shorter wavelengths result in higher-resolution images, while
longer wavelengths result in lower-resolution images.
2. Seismic source frequency and depth: The frequency of the seismic
source (e.g., air gun, vibrator) also influences image resolution. Higher
frequencies produce higher-resolution images but consume more energy
and may not penetrate the subsurface as deeply. The resolution of a
seismic image often decreases as depth increases due to the increased
absorption of seismic energy as it passes through the subsurface.
3. Geology: The composition and structure of the Earth’s subsurface can
also affect the resolution of the seismic image. Harder, more consolidated
materials produce clearer images than softer, more porous materials.
Seismic energy can be scattered by complex geology, such as folds or
faults, resulting in a lower-resolution image.
4. Data processing: The resolution of an image can be influenced by the
quality of the data processing techniques employed. Image resolution can
be improved using advanced processing techniques such as inversion and
migration. Some techniques, such as filtering, can help to minimise noise
and improve resolution, while others can reduce the resolution.
5. Receiver spacing: The spacing between the seismic receivers that collect
data can also impact the image’s resolution. Closer spacing results in
higher-resolution images, but it also increases the cost of the survey.
184 Data Science and Machine Learning Applications in Subsurface Engineering

Table 9.1. Summary of articles reviewed on seismic image resolution and image resolution using AI
in oil and gas, civil engineering, and medical fields.

Author(s) Title Summary

Oil and Gas Field
Jo et al. (2022) Machine learning- The authors used a Convolution U-net algorithm to
based vertical enhance seismic image resolution. The researchers
resolution showed that using a spectral enhancement CNN
enhancement U-net model trained with time-variant wavelets
considering the seismic produced better outcomes for the seismically
attenuation attenuated area, leading to a considerable increase in
vertical image resolution.
Li et al. (2022) Deep Learning The authors used CNN to improve the resolution
for Simultaneous of seismic images and denoise them. Compared to
Seismic Image conventional methods, the CNN obtained better
Super-Resolution and performance in enhancing detailed depiction of
Denoising subsurface structural and stratigraphic structures.
Wang and Applying machine The authors utilised CNN to improve the resolution
Nealon (2019) learning to 3D seismic of marine seismic images for post-stack structural
image denoising and seismic image augmentation and noise reduction.
enhancement They demonstrated that their method considerably
improved seismic image resolution while preserving
the varied scales of geologic structures, from high-
resolution faults and diffractors to deep subsalt
deposits, for better analysis.
Sun et al. Random noise The authors recommended applying GANs to seismic
(2022) suppression and profiles to improve low resolution and suppress
super-resolution random noise. Their efforts resulted in GANs
reconstruction obtaining an ideal SNR and significantly enhancing
algorithm of seismic the resolution of the reconstructed cross-sectional
profile based on GAN seismic image, allowing for more vivid observations
of geological structures such as fractures.
Lan et al. Seismic Data This study employs a wavelet transform and a
(2023) Denoising Based on residual neural network (ResNet) to denoise and
Wavelet Transform and enhance seismic images. This approach demonstrated
the Residual Neural that the DWT-Resnet could isolate the effective signal
Network from the noise using wavelet decomposition, which
enhances the data quality going into the network and
increases the resolution and accuracy, resulting in
excellent noise suppression outcomes.
Orozco-del- A genetic algorithm for This study employs a genetic algorithm (GA) for
Castillo et al. filter design to enhance kernel optimisation to improve the internal properties
(2014) features in seismic of salt bodies and the sub-salt stratigraphy. The GA
images was statistically compared to other algorithms and
proved more appropriate for automatic interpretation.

Table 9.1 contd. ...

 Super-Vertical-Resolution Reconstruction of Seismic Volume 185

...Table 9.1 contd.

Author(s) Title Summary

Medical Field
Cengiz et al. Classification of breast The authors suggested utilising the Wavelet
(2022) cancer with deep Transform to denoise the input images and CNN
learning from noisy to classify them. Based on PSNR measurements,
images using wavelet the authors discovered that CNN produced good
transform classification results.
Gondara (2016) Medical Image Using a limited sample size, the author demonstrated
Denoising Using that denoising autoencoders built using convolutional
Convolutional layers can effectively denoise medical images.
Denoising The results show that heterogeneous images can
Autoencoders be merged to enhance sample size for improved
denoising performance.
Chen et al. Brain MRI super- In this study, the authors used a 3D Densely
(2018) resolution using Connected Super-Resolution Network (DCS) high-
3D deep densely resolution elements of structural brain MR images.
connected neural Using a dataset of 1,113 patients, they discovered that
networks their neural network surpasses bicubic interpolation
and other deep learning algorithms in reconstructing
4× resolution-reduced images. Including a CNN
increased image resolution significantly and offered a
more realistic depiction of anomalies in the scans.
Shi et al. DeSpecNet: A CNN- The use of a deep CNN to despeckle optical
(2019) based method for coherence tomography (OCT) images is proposed
speckle reduction in this paper. The researchers discovered that their
in retinal optical method resulted in significant improvements in
coherence tomography visual quality and quantitative indices and high
images generalisation ability for various retinal OCT images.
The CNN surpassed state-of-the-art approaches in
suppressing speckles and showing subtle details
while retaining edges.
Civil Engineering Field
(Mukherjee et High-resolution Using high-resolution aboveground magnetic data,
al., 2022) imaging of subsurface the authors demonstrate the possibility of imaging
infrastructure using pipelines and other subterranean infrastructure using
deep learning artificial AI-based methods. Their findings showed that their
intelligence on drone advanced algorithms and processes resulted in a
magnetometry 100-fold increase in efficiency compared to
commonly available open-source deep learning
AI workflows and software. Superior resolution
and interpretability are also proven above the
conventional geophysical inversion. The method
could be extended to other geophysical data types
at different scales and resolutions, such as seismic,
electromagnetic, and gravity.
Table 9.1 contd. ...
186 Data Science and Machine Learning Applications in Subsurface Engineering

...Table 9.1 contd.

Author(s) Title Summary

Jiang et al. Edge-Enhanced GAN This study proposes a GAN-based edge-enhancement
(2019) for Remote Sensing network (EEGAN) for robust satellite image SR
Image Super-resolution reconstruction and a noise-insensitive adversarial
learning technique. This study showed that combining
the recovered intermediate image and improved
edges produces results with excellent credibility and
clear content. Extensive trials on Open-Source Data
sets revealed that the reconstruction performance was
superior to state-of-the-art SR techniques.

Overall, the resolution of a seismic image is a complex combination

of these and other factors and can vary significantly based on the specific
circumstances of the survey. One promising approach to mitigate these
limitations is denoising algorithms, which can filter out unwanted noise
and enhance the signal-to-noise ratio of the seismic data. For example,
wavelet-based denoising methods effectively reduce noise in seismic images
(Yan and Chunqin, 2008).
Through the use of advanced machine learning and deep learning
algorithms, AI has the potential to increase the resolution and quality of
seismic images greatly. In addition to improving image resolution and quality,
AI can also help automate the interpretation of seismic images, reducing the
time and cost of subsurface mapping and exploration. For example, CNNs
have been used to classify and identify subsurface features in seismic images
(An et al., 2021). Another approach is using super-resolution reconstruction
techniques, which can enhance the resolution of the seismic image beyond
the limitations of the seismic source size. These techniques typically involve
multiple lower-resolution images to reconstruct a higher-resolution image and
have been successfully applied to various imaging modalities (Deng et al.,
2021; Sun et al., 2022).
To summarise, AI has proven effective at improving seismic image
resolution and reducing the effects of noise and interference. These techniques
have also been used in the medical field, with similar results in enhancing
resolution and minimising noise in medical imaging. However, there are still
drawbacks to using AI in these applications, such as the requirement for vast
volumes of training data and the possibility of overfitting. More studies are
required to improve the efficiency of AI for these applications. The use of AI
in seismic imaging is still in its early stages, and further research is required
to optimise and validate these methods.
 Super-Vertical-Resolution Reconstruction of Seismic Volume 187

3. Methodology
3.1 Regional Geological Overview of the Opunake Field
The Opunake Field is located in the Taranaki region of New Zealand, on
the North Island’s western side. The Taranaki Basin is a large sedimentary
basin formed approximately 50–60 million years ago during the early to
mid-Cenozoic Era. It is known for its active volcanoes and tectonic activity,
as shown in Fig. 9.1. These structural activities have shaped the geology of
the area (Stagpoole and Nicol, 2008). A complex mixture of sedimentary

Fig. 9.1. Location map of Taranaki Basin showing simplified structural elements modified after
Douglas (2005) and Rajabi et al. (2016).
188 Data Science and Machine Learning Applications in Subsurface Engineering

(sandstone, shale, and limestone), volcanic, and metamorphic rock formations

deposited by rivers and oceans at the time make up the Opunake Field’s
geology (Nodder, 1993).
The Opunake Field contains several layers of volcanic rock created by
volcanic eruptions in the late Cenozoic Era and lies on top of the sedimentary
rocks. These volcanic rocks include basalt, andesite, and rhyolite, which
erupted from nearby volcanoes, including Mount Taranaki, the dominant
stratovolcano in the region (Cronin et al., 2021). In addition to sedimentary
and volcanic rocks, the Opunake Field also contains metamorphic rocks,
formed when the sedimentary and volcanic rocks were subjected to high
pressures and temperatures during tectonic activity. These metamorphic rocks
include schist, gneiss, and quartzite, in which the forces of heat and pressure
have altered (Nodder, 1993).
The region’s tectonic activity has resulted in faulting and folding, which
have an impact on the geology of the Opunake Field. The field is located
on a number of anticlines and synclines, which are tectonic folds in the
Earth’s crust (Nicol et al., 2007). Several significant faults, including the
Pohokura and Mangahewa faults, which have shaped the region’s geology,
border the Taranaki Basin (Stagpoole and Nicol, 2008). The Moki and Kapuni
Formations are part of the Taranaki Group, a sequence of sedimentary rocks
that were deposited in the Taranaki Basin during the early Cenozoic era. The
main rock formations in the field are sandstones, siltstones, and mudstones of
the Moki Formation, which were deposited during the Oligocene and Miocene
periods (Higgs et al., 2012). The younger Kapuni Formation, comprised of
sandstones, siltstones, and mudstones with interbedded coal seams, was then
deposited on top of the Moki Formation (Voggenreiter, 1993). The Taranaki
Group is divided into several formations based on their lithology and age.

3.2 Local Geological Overview of the Opunake Field

The Opunake Field is an oil and gas field located in the Kapuni Formation,
the uppermost formation in the Taranaki Group. The field primarily comprises
sandstones and siltstones, the primary reservoirs for oil and gas. The field is
also rich in coal, mined and used for energy production (McBeath, 1977).
The field is located in an area with high tectonic activity and has
experienced several earthquakes and volcanic eruptions in the past. These
events have significantly impacted the geology of the field, including the
folding and faulting of the rock formations and the alteration of the sedimentary
layers. The geological history of Opunake Field can be divided into three
main periods: the Paleozoic, the Mesozoic, and the Cenozoic.
During the Paleozoic period, the area was part of a shallow sea that
covered much of New Zealand. The sea was home to a variety of marine life,
 Super-Vertical-Resolution Reconstruction of Seismic Volume 189

and over time, the remains of these organisms were buried and compacted
into sedimentary rocks. These rocks, which include sandstone, shale, and
limestone, are the oldest in the area and can be found at the base of the
Opunake Field.
During the Mesozoic period, the area continued to experience tectonic
activity, which caused the land to rise and form mountains. These mountains,
now long gone, were made up of sandstone, shale, and limestone. Later, these
rocks were buried in sedimentary layers made up of shale, sandstone, and coal.
During the Cenozoic period, the area experienced further tectonic activity
and experienced several periods of erosion. These activities caused the land
to be reshaped, exposing the sedimentary layers to the surface. The area is
now home to a variety of rock formations, including sandstone, shale, and
limestone, as well as coal and oil deposits.
Overall, the geological history of Opunake Field is a complex and
dynamic geological system reflecting the millions of years of tectonic activity,
volcanic eruptions, and sedimentary processes in the Taranaki region. These
activities have resulted in a wide range of rock formations and sedimentary
layers that have formed over the years. These rocks and sediments provide a
rich resource for geological study and exploration. The Opunake Field is a
valuable oil and gas production resource that has also played an important role
in developing New Zealand’s energy industry (McBeath, 1977).
This study was performed using a software environment consisting of a
system unit running on a Windows 10 Pro-64-bit operating system, CUDA
Toolkit 10.2, and the Pytorch 2.0 framework. The experimental hardware
environment of the unit is the next-generation AMD Radeon Pro™ and
the highest performing NVIDIA Quadro® professional graphics capable of
2-petaFLOPS tensor performance. This study was run on a single NVIDIA
RTX 8000 GPU, quadruple Intel (R) Xeon (R) W-2223 CPU running at
3.60 GHz, 32.0 GB of DDR4-2666 MHz DRAM. Throughout the course of
this research, the pre-trained DCNN model of the Geoplat AI software was
utilised. The model was pre-trained using both synthetic and real data.

3.3 Deep Convolutional Neural Network in Seismic Image Resolution

Seismological data are obtained through the reflection of seismic wave signals,
and after a series of processing steps, the seismic amplitude data are obtained
and prepared for interpretation. The amount of seismic data, including
3D volume data, is constantly increasing as exploration technology advances.
A network’s calculation cost rises if it is directly trained using 3D data.
The Geoplat AI software was used to import the entire 3D seismic volume
of the Opunake Field for this study. The software includes a pre-trained
DCNN model that is based on computer vision. The model was trained on
190 Data Science and Machine Learning Applications in Subsurface Engineering

Fig. 9.2. Seismic volume showing Inline 2700, Xline 5500 and depth 2040.

a large and varied 3D synthetic seismic dataset image library that considers
various complex geological factors and imaging quality. The model’s accuracy
was verified and successfully applied to the Opunake Field to generate a
conditioned probability volume. The widely known ResU-net architectural
framework, which was pre-trained, was used to segment the 3D seismic
images obtained from the Opunake Field. A signal gain was added to amplify
and restore regions with amplitude expression coefficient loss. This additional
gain aided in the processing of low-quality seismic data. Figure 9.2 displays
the inline, crossline, and depth sections of the Opunake Field.

3.3.1 Simplified Architecture of Residual U-net

The Simplified Architecture of Residual U-Net (ResU-Net) is a convolutional
neural network designed for image segmentation tasks. It is an extension of
the U-Net architecture, known for its ability to perform well on various image
 Super-Vertical-Resolution Reconstruction of Seismic Volume 191

segmentation tasks. The ResU-Net architecture consists of an encoder and a

decoder, connected by a series of skip connections. The encoder is responsible
for extracting features from the input image, while the decoder is responsible
for reconstructing the image from the features. The skip connections allow
the decoder to access features from multiple scales, which helps improve the
model’s accuracy.
One key difference between the ResU-Net and U-Net architecture is
using residual blocks in the encoder and decoder. A residual block is a block
of layers that learns the residual mapping between the input and the output
rather than the actual mapping. This residual block helps reduce the number of
parameters in the model and improves its learning ability. The mathematical
formulation of the ResU-Net architecture can be described as follows:
Let X be the input image and Y be the output image (i.e., the segmentation
map). The encoder consists of a series of convolutional layers and residual
blocks, which can be written as:
E(X) = F(X) + X (9.1)
where F(X) is the residual mapping learned by the residual block.
The decoder consists of a series of upsampling layers and residual blocks,
which can be written as:
D(E(X)) = G(E(X)) + E(X) (9.2)
where G(E(X)) is the residual mapping learned by the residual block.
The model’s final prediction is obtained by applying a convolutional
layer to the decoder output:
Y' = Conv(D(E(X))) (9.3)
A schematic framework of the ResU-Net architecture is shown in
Fig. 9.3 below.
In the simplified schematic framework of a ResNet, the input image
is passed through a series of convolutional layers and residual blocks. The
convolutional layers extract features from the input image, while the residual
blocks are responsible for learning the residual mapping between the input
and the output. The output of the residual block is then added to the original
input to produce the final output of the block. The output of the final residual
block is passed through a series of upsampling layers, increasing the output’s
resolution. This upsampled output is then passed through a final convolutional
layer, which produces the final prediction of the model. Overall, the
residual blocks and upsampling layers form a U-shaped structure, with the
convolutional layers forming the encoder and the upsampling layers forming
the decoder. The skip connections between the encoder and decoder allow
192 Data Science and Machine Learning Applications in Subsurface Engineering

Fig. 9.3. A simplified end-to-end U-network architecture.

the model to access features from multiple scales, which helps improve the
model’s accuracy.
The key difference between a standard CNN and a residual neural network
lies in how information flows through the network. In a standard CNN, the
layers learn the actual mapping between the input and the output. In a residual
neural network, the layers learn the residual mapping between the input and
the output, which is then added to the original input to produce the final output.
This residual network helps reduce the number of parameters in the model and
improves its learning ability. Overall, the ResU-Net architecture combines
the ability of the U-Net to capture multi-scale features with the ability of
residual blocks to reduce the number of parameters and improve learning.
This technique augments the convolutional neural network architecture and
makes it a powerful tool for image resolution tasks.

3.4 Training and Testing Process

The methodology for training and testing the two CNN models in this study
can be broken down into several steps:
1. Data preparation: The first step was to collect and prepare the dataset
that would be used to train and test the models. A dataset of low-quality
seismic volumes and their corresponding high-quality versions from the
Opunake field, which were affected by various imaging issues such as
noise and resolution degradation, was used in this research. The dataset
was divided into two sections: training and testing. The models were
trained using the training set, and their performance was evaluated using
the test set.
2. Model building: The conventional CNN model was constructed using
a standard architecture with several convolutional and pooling layers,
 Super-Vertical-Resolution Reconstruction of Seismic Volume 193

followed by a fully connected layer. The GAN model, on the other hand,
was built using a generator and a discriminator network. The generator
network was used to generate new high-quality seismic volumes. In
contrast, the discriminator network was used to differentiate between the
generated and the original high-quality seismic volumes.
3. Model training: The models were trained using the training set. The
conventional CNN model was trained to improve low-quality seismic
volumes by lowering noise and increasing resolution. The GAN model
was trained to generate new high-quality seismic volumes that were
similar to the original high-quality seismic volumes.
4. Model evaluation: Once the models were trained, they were evaluated
using the test set. The performance of the models was evaluated using
three commonly used image quality metrics: PSNR, SSIM, and SNR.
PSNR measures the ratio between the maximum possible power of a
signal and the power of the noise that corrupts the signal. SSIM measures
the structural similarity between two images, while SNR measures the
strength of a signal relative to the amount of noise present in the signal.
5. Model comparison: The performance of the models was compared based
on the results of the evaluation. The PSNR, SSIM, and SNR values for
each model were calculated for the test set and compared to determine
which model performed better.
6. Model selection: Based on the comparison of the two models, the one that
performed better regarding the image quality metrics was selected as the
model of choice.
It is important to note that this process of training and testing models
using a dataset is an iterative process. The model architecture, parameters and
dataset itself can be fine-tuned and optimised. Additionally, visual inspection
of the enhanced images can also be used to evaluate the results.

3.5 Criteria for Model Evaluation

In estimating the performance of the DCNN image resolution, the models’
errors were assessed using the following metrics.
1. SNR measures the strength of a signal concerning the volume of total
noise present in the signal. In image processing, SNR is generally used
to measure the quality of an image. SNR is a simple metric and provides
a single number that can be used to compare the quality of an image.
However, it cannot give a comprehensive evaluation of an image’s
quality as it only considers the values and does not take into account the
visual perception. It is common to use other metrics in conjunction with
194 Data Science and Machine Learning Applications in Subsurface Engineering

SNR to get a complete evaluation of image quality, such as the SSIM and
PSNR, which are more robust and take into account both luminance and
structural information. The formula for SNR is:
SNR = 10 × log10(Peak signal power/Mean square error) (9.4)
2. PSNR measures the quality of a reconstructed image compared to a
reference image. It is defined as the ratio of the peak signal power to
the power of the noise in the reconstructed image. This metric is
mathematically expressed as;
PSNR = 10 × log10(MAX 2/MSE) (9.5)
where MAX is the maximum possible pixel value of the image, and MSE
is the mean squared error between the conditioned and original volume.
3. SSIM is a measure of the similarity between two images. It considers
the luminance, contrast, structure, mean, standard deviation and
cross-covariance of the image pixels and gives an understanding of
the structural similarity between two images. SSIM is mathematically
written as;

SSIM ( x, y ) =
(2 × µ x × µ y + C1) × ( 2 × Σ xy + C2 )
(9.6)
(( µ x2 ) (
× µ y 2 + C1 × Σ x2 + Σ y 2 + C2 ))
where µx and µy are the pixel values’means of images x and y, respectively,
Σx and Σy are the pixel values’ standard deviations of images x and y,
respectively, Σxy is the pixel values’ covariance of images x and y and C_1
and C_2 are constants.

4. Results and Discussion

4.1 Conditioned Seismic Volume
In Figs. 9.4, 9.5, and 9.6, you will note the high level of improvement in
the conditioned image resolution volumes for inlines, crosslines, and depth,
respectively. In this study, we employed two CNN models, named the mean
resolution volume and super-resolution volume based on their performance,
to increase the quality of seismic images. The results showed massive
improvements visually, especially in the deeper parts with low quality.
Specifically, the super-resolution volume model achieved visually better
results. While the use of deep learning in seismic image processing has
demonstrated significant improvements in image quality, it is important to
note that the models are not without limitations. One of the key limitations is
the presence of false geological structures, which can be introduced during the
 Super-Vertical-Resolution Reconstruction of Seismic Volume 195

Fig. 9.4. Inline 2000 comparison of (a) original, (b) mean resolution conditioned, and
(c) super-resolution conditioned volumes.
196 Data Science and Machine Learning Applications in Subsurface Engineering

Fig. 9.5. Xline 3000 comparison of (a) original, (b) mean resolution conditioned, and
(c) super-resolution conditioned volumes.
 Super-Vertical-Resolution Reconstruction of Seismic Volume 197

Fig. 9.6. Depth 1400 comparison of (a) original, (b) mean resolution conditioned, and
(c) super-resolution conditioned volumes.

image enhancement process. Additionally, the process of enhancing the image

can lead to the loss of some important geological features, which can impact
the accuracy of subsequent interpretations.
The success of these models can be attributed to their ability to leverage
the power of deep learning and computer vision in geophysics and seismic
image processing. The models were able to extract more information from the
input data, which led to significant improvements in image quality. Further
investigation using PSNR and SSIM will be employed to critically evaluate
the models’ performance. PSNR is generally more commonly used and can
provide good results if the images have a similar dynamic range and colour
balance. Meanwhile, SSIM is considered a better metric of perceived image
quality. This metric will enable us to accurately assess which model is better
in terms of image quality improvement.

4.2 Model Evaluation

In discussing the results of image quality assessment metrics such as PSNR,
SSIM, and SNR, it is important to consider the context of the images being
compared and the specific application for which they will be used. It is
198 Data Science and Machine Learning Applications in Subsurface Engineering

important to understand that these metrics each provide a different perspective

on the quality of the enhanced or denoised image.
PSNR is a simple metric that measures the ratio of the maximum possible
power of a signal to the power of the noise that corrupts the signal. It is
generally used to compare the original image to the processed image in order
to assess the quality of image enhancement and denoising methods. A higher
PSNR value indicates a higher quality image with less noise. From the results
in Fig. 9.7, the mean resolution model achieved the highest PSNR compared
to the super-resolution model. This metric is easy to understand and compute.
However, it cannot give a comprehensive evaluation of an image’s quality as it
only considers the values and does not take into account the visual perception.
SSIM takes into account the mean, standard deviation, and
cross-covariance of the image pixels and gives an understanding of the
structural similarity between two images. Based on the data it takes into
account when computing, SSIM is a more reliable metric for evaluating
image quality than PSNR, but it requires more computation than PSNR. This
metric is based on the idea that the human visual system is highly adapted to
recognise small changes in structural information, such as edges and textures,
rather than small changes in pixel intensities. It is typically used as a measure
of image quality in signal processing applications. However, it cannot give a
comprehensive evaluation of an image’s quality as it only considers the values
and does not take into account the visual perception. As such, it is imperative
to combine it with other metrics like the PSNR. A higher SSIM value
indicates a higher quality image with less noise. As such, the mean resolution
model achieved the highest SSIM compared to the super-resolution model,
which is shown in Fig. 9.8. The SSIM is usually measured between 0 and 1.
However, for this study, the original seismic volume was used as a standard
for comparison. As such, the initial SSIM was 1, whiles the improvements
gained by the models were estimated between 0 and 1 to capture the increased
SSIM. The mean resolution model achieved an improved SSIM of 0.4,
whiles the super-resolution resulted in an SSIM of 0.31. This result shows
that not all high resolution images should be classified as the best model.
Some enhancements may result in the loss of important geological structures
through false improvements, as typified by the SSIM results.
When evaluating image enhancement or denoising methods, PSNR,
SSIM, and SNR can be used together to provide a more comprehensive image
quality evaluation. For example, a method that provides a high PSNR value
and a high SSIM value is likely to produce an image that is both visually vivid
and accurate in terms of preserving the original information. Additionally,
a high SNR value means that the signal is stronger than the noise, and it
indicates that the denoising method performed well in removing the noise.
It is also important to note that even though the PSNR and SSIM values may
 Super-Vertical-Resolution Reconstruction of Seismic Volume 199

Fig. 9.7. Comparison of the models PSNR improvements.

Fig. 9.8. Comparison of the models’ SSIM improvements.

be high, the visual perception of the image can be different. That is why it is
good practice to use multiple metrics to get a more comprehensive evaluation
and to consider the specific application for which the image will be used.
Hence, when there is the need to make an image resolution very high, the
super-resolution model performs better, especially in the shallower part of the
seismic volume. However, when the model is needed to enhance the seismic
volume for interpretation, the mean resolution model is the better of the 2 due
to the fact that more geological structures and features were preserved while
exhibiting fewer falsely enhanced areas or less confident resolution in the
deeper parts of the subsurface.
Figure 9.9 presents the spectral analysis results performed on the
original, mean, and super-resolution volumes along inline 2000. Our
findings indicate that the original volumes had an SNR of 70.9, which was
significantly improved upon by the mean resolution volume achieving an
SNR of 212.5. Interestingly, despite the improvement in spatial resolution, the
200 Data Science and Machine Learning Applications in Subsurface Engineering

Fig. 9.9. Comparison of spectral analysis for (a) original seismic, (b) super-resolution volume, and
(c) mean resolution volume.
 Super-Vertical-Resolution Reconstruction of Seismic Volume 201

Fig. 9.10. Super-resolution conditioned volume showing Inline 2700, crossline 5500, and depth 2040.

super-resolution volume only recorded an SNR of 83.6, suggesting that while

super-resolution techniques can improve the resolution of seismic data, they
may not always result in better SNR. Based on the spectral analysis results,
the mean volume outperformed the super-resolution volume due to its ability
to preserve low-frequency information, leading to a better SNR. Additionally,
the super-resolution volume amplifies the noise in the data, resulting in a
lower SNR. This is shown in Figs. 9.10 and 9.11 representing the pictorial
depiction of the Super-resolution conditioned volume and the Mean resolution
conditioned volume respectively.
Our study highlights the potential of using the mean resolution (GAN)
techniques to enhance the seismic data quality and emphasises the importance
of evaluating the effectiveness of super-resolution (CNN) techniques on
a case-by-case basis. Overall, our findings provide valuable insights for
the seismic imaging community in their efforts to enhance the quality of
subsurface images for accurate interpretation and analysis.
202 Data Science and Machine Learning Applications in Subsurface Engineering

Fig. 9.11. Mean resolution conditioned volume showing Inline 2700, crossline 5500, and depth 2040.

5. Conclusions
Insufficient labelled samples have hindered the progress of deep learning
application in seismic interpretation. In this study, we aimed to enhance an
image using AI by building and comparing two different CNN models. The
first model was a traditional CNN, while the second model was a more recent
architecture known as a GAN.
The conventional CNN (super-resolution) model was trained using a
dataset of low-quality images and their corresponding high-quality versions.
The model was then used to enhance the test images by increasing their
resolution and reducing noise. On the other hand, the GAN (mean resolution)
model was trained using a dataset of high-quality images. The generator
network of the GAN was used to generate new high-quality images, which
were then compared with the original test images.
 Super-Vertical-Resolution Reconstruction of Seismic Volume 203

The performance of both models was evaluated using three commonly

used image quality metrics: PSNR, SSIM, and spectral analysis. PSNR
measures the ratio between the maximum possible power of a signal and the
power of the noise that corrupts the signal. SSIM measures the structural
similarity between two images, while SNR measures the strength of a signal
relative to the amount of noise present in the signal.
The evaluation results showed that both models could enhance the
images to a certain extent. However, the GAN model performed better than
the conventional CNN model in terms of all three metrics. The GAN model
achieved a higher PSNR, SSIM, and SNR value than the conventional CNN
model, indicating that the generated images were of higher quality and
contained less noise.
In conclusion, the study successfully demonstrated the use of AI for image
enhancement using CNN models. Both traditional CNN and GAN models
were able to enhance the images. However, the GAN model performed better
in terms of all three image quality metrics. The GAN model showed superior
performance in terms of PSNR, SSIM, and SNR, which means it was able
to generate images with less noise and more structural similarity. Based on
these results, it can be concluded that the GAN model is a better choice for
image enhancement tasks. The proposed methodology aims to encourage
the utilisation and exploration of deep learning techniques in seismic
interpretation.

Data and Software Availability

The data used in this study is available by request through the corresponding
author. To access the software used in this article, contact the developers at
https://geoplat.ai/#demo.

Acknowledgement
The authors thank the University Teknologi Petronas and the Centre for
Subsurface Seismic Imaging for supporting this work and to Geoplat AI for
providing the software for this work.

References
Aminzadeh, F. and Dasgupta, S.N. 2013. Fundamentals of Petroleum Geophysics. pp. 37–92.
https://doi.org/10.1016/B978-0-444-50662-7.00003-2.
An, Y., Guo, J., Ye, Q., Childs, C., Walsh, J. and Dong, R. 2021. Deep convolutional neural
network for automatic fault recognition from 3D seismic datasets. Comput. Geosci.
153: 104776. https://doi.org/10.1016/j.cageo.2021.104776.
204 Data Science and Machine Learning Applications in Subsurface Engineering

Cengiz, E., Kelek, M.M., Oğuz, Y. and Yılmaz, C. 2022. Classification of breast cancer with
deep learning from noisy images using wavelet transform. Biomedical Engineering/
Biomedizinische Technik 67: 143–150. https://doi.org/10.1515/bmt-2021-0163.
Chen, Y., Xie, Y., Zhou, Z., Shi, F., Christodoulou, A.G. and Li, D. 2018. Brain MRI super
resolution using 3D deep densely connected neural networks. 2018 IEEE 15th
International Symposium on Biomedical Imaging (ISBI 2018) April, pp. 739–742.
https://doi.org/10.1109/ISBI.2018.8363679.
Cronin, S.J., Zernack, A.v., Ukstins, I.A., Turner, M.B., Torres-Orozco, R., Stewart, R.B., Smith,
I.E.M., Procter, J.N., Price, R., Platz, T., Petterson, M., Neall, V.E., McDonald, G.S.,
Lerner, G.A., Damaschcke, M. and Bebbington, M.S. 2021. The geological history and
hazards of a long-lived stratovolcano, Mt. Taranaki, New Zealand. New Zealand Journal of
Geology and Geophysics 64: 456–478. https://doi.org/10.1080/00288306.2021.1895231.
Deng, M.-D., Jia, R.-S., Sun, H.-M. and Zhang, X.-L. 2021. Super-resolution reconstruction of
seismic section image via multi-scale convolution neural network. E3S Web of Conferences
303: 01058. https://doi.org/10.1051/e3sconf/202130301058.
Douglas, A. 2005. Slow slip on the northern Hikurangi subduction interface, New Zealand.
Geophys. Res. Lett. 32: L16305. https://doi.org/10.1029/2005GL023607.
Farfour, M., Gaci, S., El-Ghali, M. and Mostafa, M. 2021. A review about recent seismic
techniques in shale-gas exploration. pp. 65–80. In: Methods and Applications in Petroleum
and Mineral Exploration and Engineering Geology. Elsevier. https://doi.org/10.1016/
B978-0-323-85617-1.00012-6.
Gondara, L. 2016. Medical image denoising using convolutional denoising autoencoders.
pp. 241–246. In: 2016 IEEE 16th International Conference on Data Mining Workshops
(ICDMW). IEEE. https://doi.org/10.1109/ICDMW.2016.0041.
Gray, S.H. 2014. Seismic imaging. pp. S1-1-S1-16. In: Encyclopedia of Exploration Geophysics.
Society of Exploration Geophysicists. https://doi.org/10.1190/1.9781560803027.entry4.
Haldar, S.K. 2018. Exploration geophysics. pp. 103–122. In: Mineral Exploration. Elsevier.
https://doi.org/10.1016/B978-0-12-814022-2.00006-X.
Higgs, K.E., King, P.R., Raine, J.I., Sykes, R., Browne, G.H., Crouch, E.M. and Baur, J.R.
2012. Sequence stratigraphy and controls on reservoir sandstone distribution in an Eocene
marginal marine-coastal plain fairway, Taranaki Basin, New Zealand. Mar. Pet. Geol.
32: 110–137. https://doi.org/10.1016/J.MARPETGEO.2011.12.001.
Jiang, K., Wang, Z., Yi, P., Wang, G., Lu, T. and Jiang, J. 2019. Edge-enhanced GAN for remote
sensing image super-resolution. IEEE Transactions on Geoscience and Remote Sensing
57: 5799–5812. https://doi.org/10.1109/TGRS.2019.2902431.
Jo, Y., Choi, Y., Seol, S.J. and Byun, J. 2022. Machine learning-based vertical resolution
enhancement considering the seismic attenuation. J. Pet. Sci. Eng. 208: 109657.
https://doi.org/10.1016/j.petrol.2021.109657.
Ker, S., Marsset, B., Garziglia, S., le Gonidec, Y., Gibert, D., Voisset, M. and Adamy, J. 2010.
High-resolution seismic imaging in deep sea from a joint deep-towed/OBH reflection
experiment: Application to a mass transport complex offshore Nigeria. Geophys. J. Int.
182: 1524–1542. https://doi.org/10.1111/j.1365-246X.2010.04700.x.
Lan, T., Zeng, Z., Han, L. and Zeng, J. 2023. Seismic data denoising based on wavelet transform
and the residual neural network. Applied Sciences 13(1): 655. https://doi.org/10.3390/
app13010655.
Li, J., Wu, X. and Hu, Z. 2022. Deep learning for simultaneous seismic image super-resolution
and denoising. IEEE Transactions on Geoscience and Remote Sensing 60. https://doi.
org/10.1109/TGRS.2021.3057857.
 Super-Vertical-Resolution Reconstruction of Seismic Volume 205

McBeath, D.M. 1977. Gas-condensate fields of the Taranaki basin, New Zealand. New Zealand
Journal of Geology and Geophysics 20: 99–127. https://doi.org/10.1080/00288306.1977.
10431594.
Mukherjee, S., Bell, R.S., Barkhouse, W.N., Adavani, S., Lelièvre, P.G. and Farquharson,
C.G. 2022. High-resolution imaging of subsurface infrastructure using deep learning
artificial intelligence on drone magnetometry. The Leading Edge 41: 462–471. https://doi.
org/10.1190/TLE41070462.1.
Nicol, A., Mazengarb, C., Chanier, F., Rait, G., Uruski, C. and Wallace, L. 2007. Tectonic
evolution of the active Hikurangi subduction margin, New Zealand, since the Oligocene.
Tectonics 26(4): 1–24. https://doi.org/10.1029/2006TC002090.
Nodder, S.D. 1993. Neotectonics of the offshore Cape Egmont Fault Zone, Taranaki Basin, New
Zealand. New Zealand Journal of Geology and Geophysics 36: 167–184. https://doi.org/1
0.1080/00288306.1993.9514566.
Orozco-del-Castillo, M.G., Ortiz-Alemán, C., Urrutia-Fucugauchi, J., Martin, R., Rodriguez-
Castellanos, A. and Villaseñor-Rojas, P.E. 2014. A genetic algorithm for filter design
to enhance features in seismic images. Geophys. Prospect. 62: 210–222. https://doi.
org/10.1111/1365-2478.12026.
Otchere, D.A., Abdalla Ayoub Mohammed, M., Ganat, T.O.A., Gholami, R. and Aljunid
Merican, Z.M. 2022a. A novel empirical and deep ensemble super learning approach
in predicting reservoir wettability via well logs. Applied Sciences 12: 2942. https://doi.
org/10.3390/app12062942.
Otchere, D.A., Tackie-Otoo, B.N., Mohammad, M.A.A., Ganat, T.O.A., Kuvakin, N.,
Miftakhov, R., Efremov, I. and Bazanov, A. 2022b. Improving seismic fault mapping
through data conditioning using a pre-trained deep convolutional neural network: A
case study on Groningen field. J. Pet. Sci. Eng. 213: 110411. https://doi.org/10.1016/J.
PETROL.2022.110411.
Picetti, F., Lipari, V., Bestagini, P. and Tubaro, S. 2019. Seismic image processing through the
generative adversarial network. Interpretation 7: SF15–SF26. https://doi.org/10.1190/INT-
2018-0232.1.
Rajabi, M., Ziegler, M., Tingay, M., Heidbach, O. and Reynolds, S. 2016. Contemporary
tectonic stress pattern of the Taranaki Basin, New Zealand. J. Geophys. Res. Solid Earth
121: 6053–6070. https://doi.org/10.1002/2016JB013178.
Roy Chowdhury, K. 2011. Seismic Data Acquisition and Processing. Dordrecht: Springer,
pp. 1081–1097. https://doi.org/10.1007/978-90-481-8702-7_52.
Selley, R.C. and Sonnenberg, S.A. 2015. Methods of exploration. pp. 41–152. In: Elements of
Petroleum Geology. Elsevier. https://doi.org/10.1016/B978-0-12-386031-6.00003-5.
Shi, F., Cai, N., Gu, Y., Hu, D., Ma, Y., Chen, Y. and Chen, X. 2019. DeSpecNet: A CNN-based
method for speckle reduction in retinal optical coherence tomography images. Phys. Med.
Biol. 64: 175010. https://doi.org/10.1088/1361-6560/AB3556.
Stagpoole, V. and Nicol, A. 2008. Regional structure and kinematic history of a large subduction
back thrust: Taranaki Fault, New Zealand. J. Geophys. Res. Solid Earth 113. https://doi.
org/10.1029/2007JB005170.
Sun, Q.F., Xu, J.Y., Zhang, H.X., Duan, Y.X. and Sun, Y.K. 2022. Random noise suppression and
super-resolution reconstruction algorithm of seismic profile based on GAN. J. Pet. Explor.
Prod. Technol. 12: 2107–2119. https://doi.org/10.1007/S13202-021-01447-0/FIGURES/7.
Talwani, M. and Kessinger, W. 2003. Exploration geophysics. pp. 709–726. In: Encyclopedia
of Physical Science and Technology. Elsevier. https://doi.org/10.1016/B0-12-227410-
5/00238-6.
206 Data Science and Machine Learning Applications in Subsurface Engineering

Voggenreiter, W.R. 1993. Structure and evolution of the Kapuni Anticline, Taranaki Basin, New
Zealand: Evidence from the Kapuni 3D seismic survey. New Zealand Journal of Geology
and Geophysics 36: 77–94. https://doi.org/10.1080/00288306.1993.9514556.
Wang, E. and Nealon, J. 2019. Applying machine learning to 3D seismic image denoising and
enhancement. Interpretation 7: SE131–SE139. https://doi.org/10.1190/INT-2018-0224.1.
Yan, F. and Chunqin, Z. 2008. Seismic data denoising based on second wavelet transform.
pp. 186–189. In: 2008 International Conference on Advanced Computer Theory and
Engineering. IEEE. https://doi.org/10.1109/ICACTE.2008.118.
 Chapter 10
Petroleum Reservoir Characterisation
A Review from Empirical to
Computer-Based Applications
Ebenezer Ansah,1,* Anthony Ewusi,2 Eric Thompson Brantson3
and Jerry S.Y. Kuma2

1. Introduction
In the early days, geological studies entailed reservoir description by
integrating geologic, geophysical, and well logging data (Yu et al., 2011).
Given this, practical problems related to geology and engineering highlighted
the importance of integrated reservoir characterisation. That led to immense
research and application of various reservoir characterisations in the oil and
gas industries.
According to Ma (2011), reservoir characterisation is the study of the
properties of a reservoir using various field specialisations (geological,
geophysical, petrophysical, and engineering), including uncertainty analysis
of geological, engineering data, and spatial variations. There exists a
correlation among these sparse data sources, and analysing them can bring
out a great understanding of the reservoir. The most used data for reservoir
characterisation are seismic data (2D, 3D, and 4D), well logs, and core data.
Recent advancement has seen increased development in the application of

1
Department of Petroleum Geosciences and Engineering, School of Petroleum Studies, University of
Mines and Technology, Tarkwa, Ghana.
2
Department of Geological Engineering, Faculty of Geosciences and Environmental Studies,
University of Mines and Technology, Tarkwa, Ghana.
3
Department of Petroleum and Natural Gas Engineering, School of Petroleum Studies, University of
Mines and Technology, Tarkwa, Ghana.
* Corresponding author: [email protected]
208 Data Science and Machine Learning Applications in Subsurface Engineering

microseismic in the field of reservoir characterisation. In the characterisation

process, well logs provide a high vertical resolution, while seismic data
provide a high lateral resolution since it covers a large area. The core analysis
gives a precise insight into the in-situ behaviour of the reservoir (highest
resolution), but its acquisition is expensive and needs huge analysis effort,
making it limited. Thus, all the available data sources complement each other
to effectively attain a complete reservoir description.
To highlight a good reservoir model, the correlation in the various data sets
is insufficient since the data have some level of uncertainty, imprecision, and
incompleteness (Verma et al., 2012). Therefore, Artificial Intelligence (AI)
techniques have been one of the widely used modelling practices to handle
such problems relating to the data set used for reservoir characterisation
(Verma et al., 2012). Given the training dataset, a good model can be trained
for complex problems. Literature has highlighted that AI methodologies have
the strength of ameliorating models by integrating with other computing
techniques (Anifowose et al., 2017; Nikravesh, 2003). These computing
techniques such as genetic algorithms, and fuzzy logic help to correct some
various limitations (such as overfitting, parameter selection, etc.) of the AI
algorithms.
This paper provides a structured overview of the application of AI
in reservoir characterisation. Several reviews highlighting the use of AI in
reservoir properties can be found in (Anifowose et al., 2015; Anifowose
et al., 2017; Nikravesh and Aminzadeh, 2001; Otchere et al., 2021a; Saikia
et al., 2020; Yu et al., 2011). In this paper, the emphasis is on the empirical
approach, AI application and its enhancement with other computational
models, and recent AI advancement for reservoir property prediction. The
paper will also address some challenges and perspectives on the way forward
for reservoir characterisation.
The following is how the paper is structured. The second section provides
an overview of the empirical correlation used for petrophysical reservoir
characterisation. Section 3 highlights the application of fractals analysis in
reservoir characterisation. Section 4 is the application of AI in petrophysical
reservoir characterisation with an emphasis on the application of artificial
neural network (ANN) and support vector machine (SVM). Section 5 looks at
the review of lithology and facies analysis while Section 6 is the contribution
of seismic data in reservoir property prediction. Section 7 looks at the
application of hybrid AI in reservoir characterisation. Section 8 highlights the
summary of the individual AI algorithms used in reservoir characterisation
with key consideration of the data input, activation function, output, and the
statistical measures for the AI algorithms. Section 9 addresses the challenges
and the way forward for reservoir characterisation. Section 10 concludes the
study and spells out the major keys drawn from the present review.
 Petroleum Reservoir Characterisation 209

2. Empirical Models for Petrophysical Property Prediction

Reservoir study using petrophysical analysis is one of the useful tools because
it provides the basic input data for integrating reservoir characterisation
and reservoir resource evaluation. The empirical models for reservoir
characterisation are represented by the correlational relationship between
the input and output variables represented by mathematical expressions.
These empirical models are vital due to their potency to correlate measurable
variables to important variables which may be difficult to determine. In
the light of this, Reza et al. (2017) highlighted the importance of well log
(input variable) analysis and its application in earth science and relation to
reservoir property prediction (output variables). They further commented
on the application in the various stages of an oilfield and their variation to
different methods used to determine reservoir properties. The petrophysical
analysis embodies the prediction of the reservoir parameters (which include
but are not limited to porosity (Phi), permeability (K), water saturation (Sw),
and lithology). To date, the reliable way of determining the true reservoir
properties is laboratory measurement (porosity, permeability, grain density,
and saturation) on core samples and well-test interpretation. However, these
core samples are only limited to small intervals due to the cost involved in
acquiring them. Therefore, the determination of the reservoir properties for
the well interval relies on the interpretation of the vast petrophysical well logs
acquired. The interpretation requires a systematic step in processing the well
log data. Due to this, the results of each step always have an impact on the
subsequent step.

2.1 Porosity and Permeability Prediction Models

According to Mohaghegh et al. (1996), the traditional method for estimating
reservoir parameters has adopted a linear or nonlinear relationship among
individual log. They highlighted in their study that the traditional approach
is limited to assumptions, constants, and heterogeneity of the reservoir but
Nelson (1994) in his paper reviewed some of these problems in detail. The
traditional approach of estimating Phi and K from core samples are the use
of models such as the Kozeny-Carman equation (Kozeny, 1927; Carman,
1937 and 1956), Wyllie and Rose model (Wyllie and Rose, 1950), and
Neutron-Density relation for Phi estimation (Gaymard and Poupon, 1968),
Timur correlation for K estimation (Timur, 1968) from wireline logs. The
empirical relation used to estimate Phi from log response (density and neutron
log) yields a total porosity (PhiT) of the formation. This PhiT is not effective
to use in reservoir characterisation due to reservoir heterogeneity and issues
such as the presence of an isolated pore network. Mavko and Nur (1997)
underlined the fact that the Kozeny-Carman’s relation observed low Phi
210 Data Science and Machine Learning Applications in Subsurface Engineering

where there is a decrease in K with a rapid decrease in Phi. To account for this,
Schwartz and Kimminau (1987) introduced a consolidated model to survey
the effect of multiple accumulations of cementation on Phi. Bourbie et al.
(1987), also proposed a solution by introducing a variable power on the Phi in
the Kozeny-Carman equation (Eq. (10.1)).
K = Bϕnd2 (10.1)
where K is the permeability, ϕ is the porosity, d is the diameter, and n is
the variable power introduced by Bourbie et al. (1987). The n varies from
3 and 7–8 for higher porosities and lower porosities, respectively. With all
the solutions, Mavko and Nur (1997), proposed a solution by accounting for
the porosity percolation threshold. Their research included the percolation
threshold in the Kozeny-Carman relation (Eq. (10.2)) to accurately fit the
observed permeability to a well-sorted material and also extend the scope of
the model.
K = B(ϕ – ϕc)d2 (10.2)
where ϕc is the minimum threshold porosity. Note, the (ϕ – ϕc) replacing ϕ in
the Kozeny-Carman relation caters for the percolation threshold.

2.2 Saturation Prediction Models

Various models have been developed to estimate Sw which mostly depend
on the geology of the formation to be evaluated. The application of these
models can lead to a different prediction of Sw. The common model used
for determining Sw is Archie’s equation (10.3), Simandox equation (10.4),
Waxman Smits’ equation (10.5), Indonesian equation (10.6) and Dual water
equation (10.7) correlation. An overview of the several water saturation
models is presented in Table 10.1. Archie’s is the most widely used model and
it is applied to clean sand and carbonates having zero shaly volume (Shedid
and Saad, 2017). Archie (1941) used resistivity data to establish a model to
determine the Sw model for clean sandstone and carbonate reservoirs on an
assumption that water is the conductive element in the reservoir. Yet, in 2016,
Mehana and El-Monier (2016) indicated that Archie’s assumption is not valid
in shale formation since it has a high mineral and organic content. Also, Shedid
and Saad (2017) highlighted the same disadvantage of Archie’s model. They
highlighted that the presence of clay minerals in a formation complicates the
results from Archie’s and may give a misleading result.
Following this, several empirical models were built to estimate Sw in
shaly and shaly sand formation. All aforementioned models are associated
with various uncertainties. These uncertainties can also be dependent on the
accuracy and precision of the well logs and core measurements. Note that,
 Petroleum Reservoir Characterisation 211

Table 10.1. Characteristics of various water saturation models.

Sw model Characteristics
Archie’s model Archie’s equation (Archie 1942) shows a relationship between water saturation
(Archie 1942) to the true permeable formation resistivity, the formation porosity and the
formation water resistivity. The challenge, therefore, arises due to the presence
of shale in the reservoir which is a conductive medium and hence is against
the original assumptions of Archie’s equation, which was a clean sandstone
reservoir (Archie 1942). The presence of shale causes a disparity in the reading
of the total resistivity of the reservoir and brings about an overshot in the water
saturation predicted by Archie’s equation (Archie 1942).
Simandoux The Simandoux model was developed to study the volumetric effects of
(Simandoux, reducing clay volume on the conductivity of the rock matrix and the overall
1963) saturation.
The Simandoux model is applied in fine siltstone of clay-rich formation
regardless of the specific distribution form of clay or clay applied in shaly
sandstone.
Simandoux experimented on only four synthetic samples using one type of clay
(montmorillonite) with a constant value of porosity. Hence, the model leads to
optimistic results when the porosity is less than 20%, and it cannot be relied on
in low porosity situations.
Also, the model does not show a volumetric balance between sandstone
volume and the clay volumes with consideration of the lack of shale formation
factor in the clay term making the correlation of clay effect in the model too
large and hence reducing the amount of water saturation estimated
(Sam-Marcus et al., 2018)
Waxman-Smits The Waxman–Smits model is based on laboratory measurements of resistivity,
(Waxman-Smits porosity and saturation of real rocks. The major assumptions of the Waxman–
1968) Smits model about clay formation and its properties are as follows: Clay
surface conductivity is assumed to share a directly proportional relationship
with the factor Qv (defined as the milli-equivalents of exchangeable clay
counterions per unit volume of pore space), and the F* term replicated in both
the sandstone resistivity term and the shale resistivity term.
This model served as the premise of the widely used dual water model. The
Waxman–Smit’s equation is often used as a standard against other methods,
due to its high experimental backing, but the determination of CEC (Cation
Exchange Capacity) is a time-consuming experiment and this is the major
limitation of the Waxman–Smits model.
Indonesian The Indonesian equation relics as a benchmark for field-based models that
(Poupon and work reliably with log-based analysis regardless of special core analysis data.
Leveaux, 1971) The Indonesian equation also does not particularly assume any specific shale
distribution. The Indonesian model also has an extra feature as the only model
considered the saturation exponent (n). According to Shedid and Saad (2017),
the results from the Indonesian predictor have been obtained with a simpler
equation, which is more convenient for quick interpretation.
Dual water Dual water is an improved form of Waxman-Smits which contains irreducible
(Clavier et al. water saturation and free water. In this method, it is proposed that the
1977) contribution of clay minerals to the resistivity of reservoir rock is caused by the
presence of free water within the pore spaces and the bound water within the
clay matrix.
Dual water was developed to account for the conductivity at the surface of a
clay mineral within the volume of shaly sandstone.
212 Data Science and Machine Learning Applications in Subsurface Engineering

these uncertainties have direct implications on the economics of any project

but often petrophysicists do not quantify them. To assess the implication,
Bower and Fitz (2000) quantified the sensitivity associated with the dual
water model to determine the most cost-effective way to reduce ambiguity
and deliver a framework for good estimation. Figure 10.1 provides a flowchart
of the influence of petrophysical properties on the Net Present Value (NPV)
for a project. Also, a balance in capillary pressure and gravitational forces
plays a vital role in the distribution of Sw above a Free Water Level (FWL)
(Lian et al., 2016). Therefore, the availability of capillary pressure data from
Special Core Analysis (SCAL) can help estimate the Sw at any given height
(Brantson, 2022; Harrison and Jing, 2001; Olakunle et al., 2015; Rudyk and
Al-Lamki, 2015). Possibly, a good saturation height model can be used to
estimate the Sw value away from the cored well location provided that the
geology and sedimentological setting of the predicted environment is known
(Hu and Chen, 2012; Lian et al., 2016; Luo et al., 2013).

( a * Rw)
Sw = n
(10.3)
Rt * Phiem

1 φ m * Swn Vsh * Sw
= + (10.4)
Rt a * Rw Rsh
*
1 φ m * SwT n  Rw 
= T * 1 + B * Qv  (10.5)
Rt a * Rw  SwT 

1  φm Vcl (1− (Vcl / 2) 

=  +  * Swn / 2 (10.6)
Rt  a * Rw Rcl 
 
*
1 φ m * SwT n  1 Swb  1 1 
= T * +  −  (10.7)
Rt a  Rw SwT  Rwb Rw  
where m is the cement factor, m* presents the cement factor used for Dual
water and Waxman-Smiths, n is saturation exponent, a represents Tortuosity
factor, Vsh is wet shale volume, Sw represents effective water saturation, SwT
is Total water saturation, Swb is Bounded water saturation, Rw represents
Formation water resistivity, Rwb also represents Bound water resistivity, Rt is
Input resistivity curve, Qv represents cation exchange capacity per unit total
pore volume, B is Equivalent conductance of clay cations and T representing
Formation temperature in degree centigrade.
 Petroleum Reservoir Characterisation 213

Fig. 10.1. Influent of petrophysical properties on the net present value (NPV). Modified after Bowers
and Fitz (2000).

3. Fractal Analysis in Reservoir Characterisation

Many numerical fashions have been used to acquire reservoir properties which
have facilitated information on the relationship between reservoir properties
and the components of reservoir decision programs. Fractals have emerged
as one of the effective tools for quantifying complicated reservoir structures
and are extensively used to model pore shape and flow phenomena in
conventional and unconventional reservoirs (Bian et al., 2020). Hewett (1986)
initially developed thoughts of fractal geometry to reservoir descriptions and
evaluations. Hewett and Bahrens (1990) confirmed how the distribution of
reservoir properties may be regarded as a consciousness of a random function
by advocating two uses of condition simulation. They utilised the flow process
and properties of various scales in the reservoir and assessed the uncertainties
associated with the final results using incomplete data sampling.
Sahimi (1993) handled the maximum essential elements of flow in porous
rock, from modelling to fractals, percolation and current developments.
Heiba et al. (1992) modelled the percolation theory of two-phase relative
permeability, whilst Katz and Thompson confirmed that the porous rock
214 Data Science and Machine Learning Applications in Subsurface Engineering

interface of numerous sandstones is fractal. Shook et al. (1992), upon defining

scaling as the extrapolation of the outcome of one scale to another scale, Li
and Lake (1995) implemented the concept to develop a scaling procedure
for a heterogeneous rock. Costa (2006) developed a fractal predictive model
for permeability by incorporating the concept of fractal geometry into the
fundamental Kozeny-Carman equation. The developed model was then used
to predict the permeability of a variety of porous materials and discovered
that the version was more adaptable to experimental results than other
models. Zheng and Li (2015) developed a fractal permeability model for
sandstone reservoirs that is entirely based on a fractal capillary model and
Darcy’s law. The version is made up of a fractal Kozeny-Carman constant that
represents the microstructural parameters of the pores. Yu and Cheng (2002)
created the fractal permeability model for heterogeneous porous media and
delivered the set of rules for the fractal size of the pore dimension and the
fractal size of tortuosity primarily based on the box-counting method. Bian
et al. (2020), analyse how microporous systems affect the permeability of
clay silt reservoirs according to the relationship between permeability and
pore structure parameters. In addition, Dong (2020) advanced a predictive
permeability model for low-permeability sandstone that takes into
consideration the microstructural parameters and tortuosity. Upon comparing
the two predicted model outputs to the core measured permeability, a relative
average error of 8% was obtained. The reviewed literature has proven that
fractal theory has been validated to be another good prospect for enhancing
reservoir characterisation and can also be used to assess field-wide project
performance.

4. Application of Artificial Intelligence in Petrophysical

Property Prediction
In recent years, the use of technology has grown widely and the application
of such technology has helped solve problems in the petroleum sector.
Computer-based intelligence like ANNs and Machine Learning (ML) has
helped solve petroleum-related problems such as the nonlinearity problems
(Ouenes, 2000; Kaydani et al., 2012) associated with the traditional
correlations. It has higher capabilities for recognising the irregularities
in the multiresolution of the individual variables that are not tended to by
conventional relationships. Also, in using computer-based intelligence,
there are no required assumptions to be made (Otchere et al., 2021a), and its
operation is based on numerical models which can identify irregularities and
help lessen noise present in the dataset (Saikia et al., 2020). Olson (1998),
highlighted in his study that, the application of computer-based intelligence
for formation evaluation is important to establish a reliable linear relationship
 Petroleum Reservoir Characterisation 215

that exists between the estimated reservoir elastic properties from the wireline
logs and corresponding sampled fluid information. The application of
computer-based intelligence provides a different perspective on the
implementation of well logs and also brings to light more details contained
in these input logs (e.g., the relationship between Gamma Ray (GR), Deep
resistivity (ILD), and Density (RHOB) with formation K). In this section,
AI techniques such as ANN, SVM and fuzzy logic (FL) will be discussed by
providing the main theoretical framework and some applications in reservoir
characterisation.
The application of ML algorithms has been applied to several quantitative
analyses of well logs for the estimation of reservoir properties (Helle
et al., 2001; Huang et al., 1996; Huang and Williamson, 1997). This learning
approach has proved to be simple and has provided an accurate solution for
the valuation of reservoir properties using several well logs. Nikravesh et al.
(2003) noted that the computational process for the estimation of the reservoir
parameters seems to be more reliable because they are independent of the
uncertainties which come with the logging process.
For the various AI concept and theories, Otchere et al. (2021a) have
presented an in-depth report on the mathematical structure of some of the
AI algorithms such as ANN, SVM, and Relevance Vector Machine (RVM).
Likewise, Saika et al. (2020) and Otchere et al. (2021a) provides a detailed
overview of the structure of AI algorithms and their advantages and limitations.

4.1 Artificial Neural Networks (ANNs)

In the early 1940s, McCulloch and Pitts (1943) idealised a mathematical
model as a representation of the human biological neural system. This idea
of the mathematical model is termed the ANN. The neurons of the human
body are dedicated to information processing. These neurons are made up
of cell body branches tree-like axons (which are signal transmitters) and
dendrites (which are signal receivers). The cell body also includes a nucleus
which produces chemicals required for the neuron. The chemical produced
by the nucleus is then transferred between a neuron and its adjacent target
cell through the synaptic connections found at the end of the axon branch
(which are termed axon terminals). In the ANN architecture, the input signals
(x1, x2,…, xn); which serves as the chemical product or the electrical signal) are
transmitted along the axons to the synaptic connections. At this point, weights
(w1, w2,…, wn) will be applied to the signals from the axon. These weights
control the direction and influence of the cells. The product of the weight
and signal (w1x1, w2x2, …, wnxn) will be transferred from the dendrites to the
neuron cell. At the neuron cell, a summation of all the products of the weights
and signals from the dendrites is made. A signal is then fired along the axon
216 Data Science and Machine Learning Applications in Subsurface Engineering

if the summation value exceeds a threshold. During this stage, an activation

function (f) performs a transformation on the signal received from the axon
and influences the frequency of the final signal. Some common activation
functions employed in the design of an ANN architecture are sigmoid,
gaussian, tangent hyperbolic, piecewise linear, and others (Zendehboudi
et al., 2018). A summary of a single-neuron ANN architecture is demonstrated
in Fig. 10.2. Also, Fig. 10.3 presents another conceptual model of an ANN
and Ensemble architecture indicating the various stages of operation. The
mathematical model of the single neuron computation with an x input is
shown below
n
=j ∑w x
i=1
i i +a (10.8)

where xi is the input signal from the axon terminal arriving at synapse i; wi
represents the weight of the synapse i; a is the bias term, and j is the summation
of all input received from the dendrites. The output signal after is given by
the equation:
y = f (j) (10.9)
where y represents the neuron output signal transmitted through an axon to
another perceptron, and f represents the activation transform function.
The various categories of the ANN and the activation function employed
on the input variable give room for the possible design of an ANN as shown in
Fig. 10.4. Zendehboudi et al. (2018) presented in detail some of the possible
and popular ANN designs.

Fig. 10.2. Mathematical framework of a single neural ANN showing the follow of data.
 Petroleum Reservoir Characterisation 217

(a)

(b)
Fig. 10.3. Example of (a) conceptual model for the BPNN modelled by Mohaghegh et al. (1996), and
(b) conceptual model framework for the Ensemble Learning Framework modelled after Anifowose
et al. (2017).

In Fig. 10.3a, the ANN was designed for estimating petrophysical

properties. The model is a three-layer BPNN with 11 inputs, 28 hidden layers,
and 1 output neuron. The model utilised the sigmoid transfer functions in
the hidden and output layers. The authors used a small learning rate of 0.1
and a learning momentum of 0.6 to solve the problem of nonlinearity and
complexity in the input data. In Fig. 10.3b, for the ensemble work frame, level
218 Data Science and Machine Learning Applications in Subsurface Engineering

Fig. 10.4. Types of ANN based on different criteria (modified after Zendehboudi et al., 2018).

1_stratification starts by applying machine learning to identify the training

and testing dataset; with level 2, a base learner is created to cater for
uncertainties existing in the dataset; level 3 produces an output on the various
uncertainties handled in level 2; following level 3, the performance of the
trained model is accessed using the training and testing data; finally, the output
is combined using suitable combination rules (level 4) to develop an overall
compromise-based ensemble result (level 5).

4.1.1 ANN Application in Petrophysical Reservoir Prediction

Mohaghegh et al. (1996) introduced a new application of ANN to characterise
the heterogeneity of a reservoir. They encompassed the application of a
three-layer feed-forward Back Propagation Neural Network (BPNN) approach
(Fig. 10.3a). The authors proposed using a General Regression Neural Network
(GRNN) to first identify the optimal network design to achieve a steady state
in a shorter time. The GRNN lacks reproducibility with different datasets
when trained with a different dataset. Therefore, the GRNN was used by the
authors to identify the optimal network design structure and the BPNN was
used to obtain the final results. Upon the high reservoir heterogeneity, a good
fit was achieved by comparing it with the actual laboratory measurement from
core logs. Hamada and Elshafel (2010) presented a successful implementation
of the ANN to determine the Phi and K of the sand gas reservoir by exploiting
 Petroleum Reservoir Characterisation 219

the transverse relaxation time of Nuclear Magnetic Resonance (NMR) T2 and

well logs. They used the ANN model which involved the NMR T2 spin value
as well as RHOB and resistivity logs to predict the Phi and K of the test wells.
The ANN paradigm showed a good correlation with the core Phi and K values
and that of the NMR predicted Phi and K for the test wells. Singh et al. (2016)
employed the well-known BPNN to predict Phi for a gas field. Three well
logs (RHOB, sonic (DT), and deep resistivity log) were used as the input
variable for Phi estimation using one hidden layer. The authors correlated the
predicted Phi to that of the Phi estimated using an empirical approach which
entails the use of Neutron (NPHIL) and RHOB log. A correlation of 0.97 was
obtained which indicates a good prediction. The robustness of the model was
validated by testing against other offset wells using core measurement. The
model reached a good correlation coefficient of 0.994 for the Phi estimated.
Hele and Bhatt (2002) also achieved excellent results when the authors
used ANN to predict fluid saturation using wireline data while Elshafei and
Hamada (2009) successfully estimated reservoir properties achieving an
excellent correlation to core data using ANN. Al-Bulushi et al. (2009) applied
ANN to estimate a nonlinear Sw using open-hole log data for a sandstone
reservoir. From their results, they established a good Root Mean Square Error
(RMSE) and correlation using the wireline logs. Zabihi et al. (2011) similarly
designed an ANN model to predict K damage caused by the accumulation of
barite sulfate scale in core voids with good precision during core flooding. The
model took into account data from a case study to examine the K reduction
owing to the build-up of barium sulfate in Malaysia and Berea sandstone
cores. Results confirmed a high accuracy prediction of the modelled ANN
to the measured permeability values. Nyein et al. (2018) developed an ANN
paradigm for Sw prediction that included five (5) input parameters (GR, ILD,
RHOB, photoelectric factor (PEF), and NPHI) and multiple hidden layer
nodes. Likewise, the Mean Square Error (MSE) for the model’s training and
testing was 0.0109 and 0.012, respectively. This model achieved a decent
correlation with core information. Likewise, the authors also utilized ANN
in predicting Phi and classified the various lithology from the well logs.
The model applied for estimating Phi inculcated six wireline logs (GR, ILD,
RHOB, NPHI, PEF, and DT) as inputs with two latent nodes. The resulting
Phi estimated matched with core data with MSEs of 0.001412 and 0.00154
for training and testing, respectively. The good prediction highlights the
importance of using impactful data input to train the ANN algorithm for a
specified output variable (Table 10.2).
 Table 10.2. Literature on the application of ANN in petrophysical reservoir prediction.

220 Data Science and Machine Learning Applications in Subsurface Engineering

Author(s) Input Variables Predicted Model Used Activation Output Statistical Measures Remarks
Output Functions
Mohaghegh et Digitized well logs Porosity, General Sigmoid Application of Not Noticed (NN) For the benchmark
al. (1996) Permeability, Regression Neural Transfer the feed-forward ANN algorithm,
Porosity, and Network (GRNN), Function ANN proved to consideration for
Saturation Three-Layer be successful in automatic parameter
Feed Forward the prediction of selection would have
Backpropagation the petrophysical helped to improve
properties the predictive
performance of the
model.
Wong et al. Well logs (GR, ILD, Porosity Support Vector Gaussian Results showed that MSE Normalised Although the model
(2005) LLS, RXO, RHOZ, Machines (SVMs), Radial Basis SVM produced good Data achieved a good
NPHI, PEF, DT) Back Propagation Kernel results as compared SVM = 0.026347 result, a more robust
Neural Network Function to BPNN by taking BPNN = 0.074090 feature extraction or
(BPNN) into consideration MSE Raw Data selection technique
the choice of the SVM = 23.7126 could have a better-
activation function BPNN = 66.6817 improved prediction
accuracy
Al-Bulushi et Well logs (GR, Water Artificial Neural Tan-Sigmoid Results showed ANN Proper
al. (2009) NPHI, RHOZ, ILD), Saturation Network (ANN) the robustness of RMSE = 0.025, R hyperparameter tuning
Cation Exchange with Resilient saturation prediction = 0.91 is a problem and with
Capacity data Propagation, for the test field and SHF the SHF a clear water
Multiple Linear other fields RMSE = 0.079 leg should be defined
Regression, in other to increase its
Saturation Height performance.
Function (SHF)
Zabihi (2012) Case Study data Permeability Multilayer Neural Hyperbolic Results indicated A Total Average With a proper
(Core Length, Network with Tangent and the proposed Neural Absolute Deviation hyperparameter
Core Diameter, Back Propagation Logistic Network predicted (TAA) of 1.06% tunning algorithm,
Average Porosity, Algorithm Sigmoid permeability with for testing and a the efficiency of
Permeability, high efficiency and TAAD of 2.03% for this algorithm will
Temperature, a minimal error of performance. improve.
Differential Pressure) about 2%
Al-Anazi and Well logs (Gr, NPhi, Porosity and Support Vector Gaussian SVR outperformed SVM attained a Although SVM
Gate (2012) DT, RHOB, ILD) Permeability Regression (SVR) radial basis MLP with the sparse 95% confidence performed well, a
and Multilayer and sigmoid dataset median MSE of more robust feature
Perception (MLP) kernel function the prediction selection tool, such as
Neural Network as compared to the embedded model,
MLPNN may have enhanced
prediction accuracy
Gholami et al. Well Logs (DT, GR, Permeability General Independent After comparing SVM attained an R The hold-out cross-
(2012) NPHI, RHOZ, PEF, Regression Neural Component results for SVM and (0.96) and RMSE validation approach

Petroleum Reservoir Characterisation 221

ILD, LLS), Core data Network (GRNN) Analysis GRNN, the SVM (0.28) between used has irregular data
and Support Vector (ICA), loss was more accurate core and predicted splitting, which can
Machine (SVM) Function, than the prediction permeability lead to an incomplete
kernel type k, using the GRNN model evaluation
Radial Basis
Function
(RBF)
Table 10.2 contd. ...
...Table 10.2 contd.

222 Data Science and Machine Learning Applications in Subsurface Engineering

Author(s) Input Variables Predicted Model Used Activation Output Statistical Measures Remarks
Output Functions
Sebtosheikh Seismic Inversion, Lithology Support Vector Kernel Results indicate a NN With the SVM, in
and Salehi Well logs Machine (SVM): Function good prediction of a case where the
(2015) Statistical the SVM after the quantity of features
Learning Theory training set size for an individual data
(SLT), Structural reduced point exceeds the
Risk Minimization number of training
(SRM), Empirical datasets, the SVM
Risk Minimization will underperform
(ERM)
Anifowose et Porosity and Porosity and Extreme Learning Sigmoidal Results indicate The proposed ELM The improved
al. (2015) Permeability datasets Permeability Machine (ELM), Activation that the ELM had average R2 ensemble learning
were obtained from Functional Function, Sine outperformed the = 0.95 and 0.92; model adopted has
two major oil and Network, Random Activation other machine RMSE = 0.18 and been reported to be
gas reservoirs with Forest (RF), Function learning algorithms 0.56; MAE = 0.08 efficient in handling
diverse geological Multivariate and Hardlim with a higher and 0.48 for porosity most classification
formations Regression Activation correlation and permeability, and regression issues
Function coefficient and lower respectively. with greater success
prediction errors in the related sciences
than the other
algorithm used
Srisutthiyakorn 2D/3D Binary Permeability Multilayer Neural Tan-sigmoid Results using MNN CNN had the lowest Although CNN has
(2016) Segmented Images, Network (MNN), Function, and CNN produced testing MSE of proven to be efficient
Porosity, Specific Convolutional Rectified a good prediction 2.4307*E05 and and represent complex
Surface Area Neural Network Linear of permeability MNN has the lowest nonlinear functions,
(CNN), Gradient Function especially in the testing MSE of without a large dataset
Descent (GD) case of 2D CNN in 2.3999*E05 in the its efficiency becomes
multiscale. case of the 2D in void.
multiscale.
Al-Mudhafar Well logs (NPhi, Lithofacies Probabilistic GBM Results indicated PNN attained a PNN comparatively is
(2017a) Vsh, Sw and Neural Network Binomial and a 95.81% correct 95.81% accuracy for insensitive to outliers,
Permeability (PNN), AdaBoost Loss prediction of lithology prediction. but with a good
Generalised Function lithofacies whiles Permeability feature selection tool,
Boosted GBM led to much GBM the performance of
Regression Model more accurate R2 = 0.9953 and the algorithm would
(GBM), Multiple permeability RMSE = 28.43 have improved.
Linear Regression prediction compared MLR

Petroleum Reservoir Characterisation 223

(MLR) to MLR R2 = 0.9551 and
RMSE = 88.42
Al-Mudhafar Well logs (NPhi, Sw, Lithofacies Kernel Support Non-zero KSVM had good KSVM lead to Although LDA can
(2017b) Vsh) Vector Machine Lagrangian accuracy for attaining a 99.55% recognize patterns
(KSVM) Multipliers the lithofacies accuracy for with known prior
and Linear classification lithofacies prediction assumptions, KSVM
Discriminate has proven to be
Analysis (LDA) efficient in identifying
distinct features with
the decision function
defined by assisting
vectors

Table 10.2 contd. ...

...Table 10.2 contd.

224 Data Science and Machine Learning Applications in Subsurface Engineering

Author(s) Input Variables Predicted Model Used Activation Output Statistical Measures Remarks
Output Functions
Moghadasi et Well logs (GR, Porosity and Artificial Neural Levenberg- Results indicated Classical Neural Even though the
al. (2017) NPHI, DT, Calliper, Permeability Network (ANN), Marquardt that the statistical Network bagging ANN
ILD, RHOZ, Sw, Principal Optimization ANN had an Porosity approach can optimize
Vsh), lithology, Core Component Algorithm improved quality as R2 = 73.7%, MSE = the ANN model,
data Analysis (PCA)- compared to those 0.0089 its performance is
ANN, Statistical obtained from both Permeability heavily dependent on
ANN based on PCA-based ANN R2 = 72.2%, MSE = the decision criteria,
bagging approach and the classical 0.0074 which is often difficult
ANN models. PCA based ANN to determine
Porosity
R2 = 79.8%, MSE =
0.0085
Permeability
R2 = 74.1%, MSE =
0.0092
Bagged ANN
Porosity
R2 = 94%, MSE =
0.001
Permeability
R2 = 85%, MSE =
0.0034
Iturraran- Reflection 2D Porosity (Phi), Backpropagation Gamma Test Results show a good NA With the ANN having
Viveros and Seismic data, Seismic Saturation ANN, Conjugate correlation between the advantage of
Munoz-Garcia Attribute, Well logs (Sw), and Gradient ANN, the predicted not knowing the
(2018) (SP, ILD, DT, GR) Volume of BFGS ANN parameters with the process responsible
Clay (Vsh) core data obtained for generating
from the location the data but still
performing better, it
always has the issue
of hyperparameter
tuning and also
dimensionality issues.
Zhong and Carr Core data, Well logs Porosity Support Vector Particle Swarm PSO-MKF-SVM PSO-MKF-SVM The hybridised model
(2019) (GR, RHOZ, Vsh) Machine (SVM), Optimization displayed better R = 0.9560, R2 = has highlighted a good
Multilayer (PSO) and performance over 0.9140, RMSE = accuracy based on
Perceptron Mixed Kernel different methods 1.6505, AAE = prediction, reliability
Neural Network Function with MLPNN having 1.4050 and MAE = and computation time

Petroleum Reservoir Characterisation 225

(MLPNN) (MKF) the worst. 2.717 than the conventional
algorithms
Gogoi and 2D Post-Stack The volume Multilayer Feed Sigmoid Results indicate a Vshale With the ability
Chatterjee Seismic data, Well of shale Forward Neural Activation reliable prediction Cross-correlation of to approximate
(2019) logs (DTs, RHOZ, (Vsh), Water Network (MLFN) Function using the MLFN 0.88 was attained complex function,
GR, LLS, ILD, Saturation model by MLFN between its hyperparameters
NPHI) (Sw) the actual and the selection becomes
predicted with a challenging and
0.031 error always rely on a trial
Sw and error approach
A cross-correlation
of 0.85 was attained
by MLFN between
the actual and the
predicted with a 0.05
error
Table 10.2 contd. ...
...Table 10.2 contd.

226 Data Science and Machine Learning Applications in Subsurface Engineering

Author(s) Input Variables Predicted Model Used Activation Output Statistical Measures Remarks
Output Functions
Soleimani et al. Reflection Seismic Porosity Multi-layer Gaussian Results indicate the PNN had an average With PNN having
(2020) data, Well Logs Feed-Forward Function supremacy of the Validation error of a good validation
(RHOZ, GR, DT) Network (MLFN), Neural Network 4.6 as compared to error, it is inefficient
Radial Basis compared with the 4.8 for MLFN in terms of memory
Function (RBF), other methods used usage.
Probabilistic
Neural Network
(PNN),
Geostatistics,
Regression
Analysis
Okon et al. Depth interval, GR, Porosity, Multiple-input and Levenberg- Results indicate a MIMO-FFBPNN
(2021) RHOZ, Resistivity Permeability Multiple-outputs Marquardt, reliable prediction Porosity
logs and Water Feed-forward Bayesian using the MIMO- R= 0.9243 and MSE
Saturation Back-propagation Regularisation FFBPNN model = 1.7243
Neural Network and Scaled MIMO-FFBPNN
(MIMO-FFBPNN) Conjugate Permeability
Gradient R= 0.9810 and MSE
= 0.0003
MIMO-FFBPNN
Water Saturation
R= 0.9631 and MSE
= 0.0049
Mohamed et al. T2 logarithmic Permeability Feed-forward Bayesian Results show that The model had the
2022 mean, T2 peak, T2 Backpropagation Regularisation the Elman Neural highest prediction
components range Neural Network, Network had the of 0.99 correlation
and T2 component Cascade-forward best prediction of coefficient and 0.88
range index Neural Network, permeability R2 as compared
Elman Neural with the other
Network, Pattern methods
Recognition
Neural Network
and Distributed
Delay Neural
Network

Petroleum Reservoir Characterisation 227

228 Data Science and Machine Learning Applications in Subsurface Engineering

4.2 Support Vector Machine (SVM)

The concept of support vector machines was first introduced in the early
1960s but later in the 1990s, its application became popular since it was able
to outperform the classification problems of the ANN approach (Zendehboudi
et al., 2018). SVMs are ML algorithms with kernel functions used for
regression (SVR; Fig. 10.5a) and classification (SVC; Fig. 10.5b) analysis
as shown in Fig. 10.5. SVMs have contributed widely to various applications
such as regression, classification and forecasting. Some major advantage of
SVMs is their ability to handle noisy data and high dimensional data (Furey
et al., 2000). Also, the simplest structure of SVMs is capable of classifying
two linearly-conceivable sets of data classes taking into consideration linear
hyperplanes with maximum margins from both data classes.
Considering a classification SVM with 2-dimensional space class (class 1
and 2) that can be handled as a linear classification as shown in Fig. 10.5b.
Assuming an n input data represented by a = {a1, a2,..., an} a tagged variable yn
which falls between –1 and +1 and can help classify the input variable. In this
case, the ith input can fall into class 1 when yn = +1 and class 2 when yn = –1.
Given a vector w which is perpendicular to the decision boundary as shown in
Figure, the decision function can be expressed as:
n
c ( a, w, e )= w.a + e= ∑wa
i=1
i i +e (10.10)

where c is the decision function, a represents the input vector, w denotes a

vector perpendicular to the separating line, and e is a bias.

Fig. 10.5. SVM architecture showing: (a) regression SVM/RVM in a high dimensional space; and
(b) supervised support vector creating a linear relationship between two non-linear data making use
of optical hyperplane.
 Petroleum Reservoir Characterisation 229

SVM classification problems correspond to the maximisation of the

margins separating the two classes (Zendehboudi et al., 2018). To help
minimise this, a Lagrange multiplier (αi ≥ 0) is introduced to convert from a
constrained to an unconstrained problem. Also, SVM can classify any given
input in higher dimensional space utilising a transformation such as sigmoid,
polynomial, and radial basis functions in the kernel when the classes are
not linearly separable using a linear hyperplane (Zendehboudi et al., 2018).
Considering the kernel function, a vector xi and xj are introduced into the
objective function. Since the optimisation depends on the dot product of the
sample vector, the conversion to feature space can be abridged to find a kernel
function K(xi, xj). The mathematical expression for the objective function can
be expressed as;
n
1 n n
G
= ∑α
=i 1
i − ∑∑ yiαi K ( xi , x j ) y jα j
2=i 1 =j 1
(10.11)

4.2.1 Machine Learning (ML) Application in Petrophysical Reservoir Prediction

With the use of SVMs, Al-Anazi and Gate (2012) analysed the contribution
of limited well log data proportion on the prediction of Phi and K in a
heterogeneous sandstone reservoir. For validation, the Support Vector
Regression (SVR) of the SVM was correlated with Multilayer Perception
(MLP) neural network, with the output indicating the supremacy of the SVR
over the MLP.
Wong et al. (2005) also predicted Phi for an offshore Western Australian
heterogeneous reservoir, implementing the use of the Gaussian radial basis
activation function SVM. The model does not define a relative balance between
the complexity of trade-off parameter management and the observed error to
minimise processing time. The authors validated their outcome by comparing
the model to that of BPNN, establishing the benefit of SVM as a substitute
supervised ML paradigm for characterising heterogeneous reservoirs. Gholami
et al. (2012) as well did a comparative study by predicting K for gas wells in
the Southern Pars field using SVM and GRNN. They employed the use of
Independent Component Analysis (ICA) for the selection of dependent logs.
The results showed a good correlation of 0.97 between core and predicted
permeability. Also, upon comparing the results for SVM to that for GRNN, the
SVM reached a faster and more accurate hydrocarbon reservoir permeability
prediction than the GRNN. SVM as reported by Asante-Okyere et al. (2020b)
has seen a greater performance in the field of petrophysical properties.
Table 10.3 provides the applications of SVM in the field of petrophysical
properties prediction.
 Table 10.3. Literature on the application of SVM in petrophysical reservoir prediction.

230 Data Science and Machine Learning Applications in Subsurface Engineering

Author(s) Input Variables Predicted Model Used Activation Output Statistical Measures Remarks
Output Functions
Wong et al. Well logs (GR, Porosity Support Vector Gaussian Results showed that MSE Normalised Data Although the model
(2005) ILD, LLS, RXO, Machines (SVMs), Radial Basis SVM produced good SVM = 0.026347 achieved a good result,
RHOZ, NPHI, Back Propagation Kernel results as compared BPNN = 0.074090 a more robust feature
PEF, DT) Neural Network Function to BPNN by taking MSE Raw Data extraction or selection
(BPNN) into consideration the SVM = 23.7126 technique could have
choice of the activation BPNN = 66.6817 a better-improved
function prediction accuracy
Anifowose et Well Logs (GR, Porosity and Functional Networks Kernel The result showed The hybrid model attained Although SVM can
al. (2011) Phi, RHOB, Sw, Permeability (FN), Type-2 Fuzzy Activation a higher correlation a maximum R2 of 0.96 for handle small datasets,
ILD, CALI) Logic System (T-2 Function coefficient for porosity prediction and R2 the paper has proven
FLS), Support Vector the hybrid model of 0.70 for permeability hybridising the model
Machine (SVM) compared with the prediction. has a greater potential
individual models of handling the issues
of the benchmark SVM
algorithm
Al-Anazi and Well logs (Gr, Porosity and Support Vector Gaussian SVR outperformed SVM attained a 95% Although SVM
Gate (2012) NPhi, DT, Permeability Regression (SVR) radial basis MLP with the sparse confidence median MSE performed well, a
RHOB, ILD) and Multilayer and sigmoid dataset of the prediction as more robust feature
Perception (MLP) kernel compared to MLPNN selection tool, such as
Neural Network function the embedded model,
may have enhanced
prediction accuracy
Gholami et Well Logs (DT, Permeability General Regression Independent After comparing SVM attained an R The hold-out cross-
al. (2012) GR, NPHI, Neural Network Component results for SVM and (0.96) and RMSE validation approach
RHOZ, PEF, (GRNN) and Support Analysis GRNN, the SVM was (0.28) between core and used has irregular data
ILD, LLS), Core Vector Machine (ICA), loss more accurate than the predicted permeability splitting, which can
data (SVM) Function, prediction using the lead to an incomplete
kernel type GRNN model evaluation
k, Radial
Basis
Function
(RBF)
Anifowose et Reflection Permeability Support Vector Kernel Upon comparison, SVM produced a good SVM outperforms
al. (2013) Seismic data, Machine (SVM, Function the SVM model gave prediction for the seismic T2FL when working
Well Logs Type-2 Fuzzy Logic the most accurate plus log input with an with small data size
(T2FL) permeability prediction RMSE of about 0.49, R as compared to T2FL
of about 0.7, and MAE of which work best
about 0.38 with input data with

Petroleum Reservoir Characterisation 231

uncertainties
Sebtosheikh Seismic Lithology Support Vector Kernel Results indicate a good NN With the SVM, in
and Salehi Inversion, Well Machine (SVM): Function prediction of the SVM a case where the
(2015) logs Statistical Learning after the training set quantity of features
Theory (SLT), size reduced for an individual data
Structural Risk point exceeds the
Minimization (SRM), number of training
Empirical Risk datasets, the SVM will
Minimization (ERM) underperform

Table 10.3 contd. ...

...Table 10.3 contd.

232 Data Science and Machine Learning Applications in Subsurface Engineering

Author(s) Input Variables Predicted Model Used Activation Output Statistical Measures Remarks
Output Functions
Al-Mudhafar Well logs (NPhi, Lithofacies Kernel Support Non-zero KSVM had good KSVM lead to attaining Although LDA can
(2017b) Sw, Vsh) Vector Machine Lagrangian accuracy for a 99.55% accuracy for recognize patterns
(KSVM) and Multipliers the lithofacies lithofacies prediction with known prior
Linear Discriminate classification assumptions, KSVM
Analysis (LDA) has proven to be
efficient in identifying
distinct features with
the decision function
defined by assisting
vectors
Zhong and Core data, Porosity Support Vector Particle PSO-MKF-SVM PSO-MKF-SVM The hybridised model
Carr (2019) Well logs (GR, Machine (SVM), Swarm displayed better R = 0.9560, R2 = 0.9140, has highlighted a good
RHOZ, Vsh) Multilayer Optimization performance over RMSE = 1.6505, AAE = accuracy based on
Perceptron Neural (PSO) and different methods with 1.4050 and MAE = 2.717 prediction, reliability
Network (MLPNN) Mixed MLPNN having the and computation time
Kernel worst. than the conventional
Function algorithms
(MKF)
Wang et al Core data, Microfacies Support Vector Radial Basis SVM algorithm Results indicate an 84% Although the paradigm
(2019) Wireline logs Machine (SVM) Kernel attained a good prediction accuracy had a high prediction
(GR, Vsh) Function automatic and attained using the SVM accuracy, it could
quantitative model have been higher if a
identification good feature selection
of depositional tool had been used to
microfacies with identify the relevant
a good prediction parameters for model
accuracy training.
Qiang et al. Well Logs (GR, Porosity, Support Vector Radial Basis SVM Showed good SVM attained an average The implementation of
(2020) SP, Calliper, Saturation, Machine (SVM), Kernel prediction for porosity overall R of 0.84 for all feature selection tools
ILD, DT, DTs, Spatial Sequential Gaussian Function while a hybrid of the input wells may have resulted in
RHOZ, NPHI), distribution Simulation (SGS), GIS and SVM gave the improvement of
Core Data, of acoustic and Gaussian a higher prediction the performance of the
and Post stack impedance Indicator Simulation of spatial variation of model.
Inversion (GIS) reservoir properties for
the reservoir
Xie et al. Well Log (GR, Lithology Randomised N-fold Results show that the RETC had a predictionWith the data input
(2020) CNL) Identification Ensemble Tree Cross- proposed framework accuracy of 89.4% andconsidered for the
Classifier (RETC), Validation, (RETC) has a higher 91.1% in the Daniuduilithology identification,
Random Forest Multivariate prediction accuracy and Hangyinqi Gas Field,
the prediction accuracy
(RF), Gradient Tree Outlier for sandstone lithology respectively would have improved
Boosting (GTB), and than the other AI if another contributing
Xgboosting classifiers. well log such as
photoelectric factor,

Petroleum Reservoir Characterisation 233

and density were used.
Asante- Core data, Well Water Least Square Principal Overall PCA-LSSVM PC-LSSVM The hybridised SVM
Okyere et al. logs (GR, DT, Saturation Support Vector Component had the least prediction Testing (RMSE and AAE) attained a good
(2020a) SP, ILD, NPHI) Machine (LSSVM), Analysis error and outperformed = 0.048 and 0.034 prediction due to
Subtractive (PCA), PCA-ANFIS-SCM ANFIS-SCM the application of
Clustering Method Coupled for the prediction of Testing (RMSE and AAE) the feature selection
(SCM), Adaptive Simulation saturation. = 0.055 and 0.036 technique adopted and
Neuro-fuzzy Annealing its ability to also work
inference System (CSA), perfectly with a small
(ANFIS) Simplex dataset
Algorithm
234 Data Science and Machine Learning Applications in Subsurface Engineering

5. Lithology and Facies Analysis

The study of sedimentary depositional environment and its heterogeneity
(Table 10.4) has been studied by many authors (e.g., Meshri, 1986; Pettijohn,
1983; Yu, 2008) except Qui (2000). The various authors suggested that the
sedimentary depositional environment and its heterogeneity comparison
can be best classified by studying its facies. In the early 1990s, Walker and
James (1992) underlined that facies analysis is essential when studying
basin fills. Lithology and facies analysis help to subdivide the reservoir rock
sequence into respective lithostratigraphic units to help map and correlate
between depositional areas and analyse their source. Due to the complexity
of facies modelling, Tang (2008) outlined the two basic approaches for facies
modelling, which are: the exploitation of empirical equations such as the
Archie equation, multivariate statistics, and neural network. In recent years,
the latter has gained more attention, with which several researchers have
used multivariate statistics and neural networks for facies modelling (e.g.,
Bestagini et al., 2017; Miller et al., 2019; Tran, 2020) and identify lithology
(e.g., Al-Anazi and Gates, 2010; Al-Mudhafar, 2017a; Alzubaidi et al., 2021).

5.1 AI Applications in Lithology and Facies Analysis

Rafiq et al. (2016) proposed an idea of microseismic facies analysis that
enables the segmentation of an eccentric reservoir into discrete facies in
response to microseismic accompanied by the combined exposition of
microseismic with 3D seismic data. This approach was originally proposed
by Eaton et al. (2014) and has been modified by the authors. Their findings
correlated with the principal curving irregularities from 3D seismic data. With
the combination of microseismic facies analysis and field parting acquired
from log and core data, the authors delineated reservoir segments marking
structural and depositional trends. De Matos et al. (2007) proposed two new
semiautomatic alternative methods for seismic facies analysis. They used
Kohonen Self-Organising Maps (SOMs) clustering as an advanced approach
for mapping seismic facies, evaluating their quantity, and selecting seismic
trace peculiarities in each oriented geological segment with the help of the
wavelet transform algorithm. Validating their results, both approaches were
used and utilised the synthetic and real seismic data from the Namorado
deepwater oilfield in Campos Basin, offshore Brazil. For several seismic
data sets, the results confirmed a good prediction of SOM clustering. The
results also showed an improved seismic facies analysis by utilising trace
distinctiveness noticed by the wavelet transform technique.
An integration of wireline logs into seismic lithofacies classification
was conducted by Avseth and Mukerji (2002). The input data for the study
comprised GR, RHOB, seismic velocity, and impedance. The lithofacies
 Table 10.4. Literature on the application of AI in lithology and facies analysis.

Author(s) Input Predicted Model Used Activation Output Statistical Measures Remarks
Variables Output Functions
Avseth and Well logs Lithofacies Rock Physics, MLDA: Mean Results demonstrate All applied models The predictive
Mukerji (GR, Mahalanobis and Covariance. a somewhat better had a success rate of performance of the
(2002) RHOZ, Discriminant NN: Sigmoid forecast utilizing the NN about 80% model would have
DT) Analysis (MLDA), Transfer Function as compared with the been improved if
Probability Density PDF and the MLDA, additional inputs such as
Function (PDF), yet the MLDA ended photoelectric factor and
Classical Neural up being viable for the spontaneous potential
Network (NN) grouping of well log were considered
into discrete lithofacies
Tang (2008) Well Logs Electrofacies Probabilistic Radial Basis PNN had a good PNN attained a Although the PNN
(GR, SP, Neural Network function classification prediction accuracy achieved a good
DT, NPHI, (PNN) performance above 70% classification, comparing
RHOZ, PE, it with other statistical

Petroleum Reservoir Characterisation 235

ILD) classification tools
would have indicated the
prediction strength of the
algorithm as compared
to other models.
Al-Mudhafar Well logs Lithofacies Probabilistic GBM Results indicated PNN attained a PNN comparatively is
(2017a) (NPhi, Vsh, and Neural Network Binomial and a 95.81% correct 95.81% accuracy for insensitive to outliers,
Sw Permeability (PNN), AdaBoost loss prediction of lithofacies lithology prediction. but with a good feature
Generalised Function whiles GBM led to Permeability selection tool, the
Boosted much more accurate GBM performance of the
Regression Model permeability prediction R2 = 0.9953 and algorithm would have
(GBM), Multiple compared to MLR RMSE = 28.43 improved.
Linear Regression MLR
(MLR) R2 = 0.9551 and
RMSE = 88.42
Table 10.4 contd. ...
...Table 10.4 contd.

236 Data Science and Machine Learning Applications in Subsurface Engineering

Author(s) Input Predicted Model Used Activation Output Statistical Measures Remarks
Variables Output Functions
Al-Mudhafar Well logs Lithofacies Kernel Support Non-zero KSVM had good KSVM lead to Although LDA can
(2017b) (NPhi, Sw, Vector Machine Lagrangian accuracy for the attaining a 99.55% recognize patterns
Vsh) (KSVM) Multipliers lithofacies classification accuracy for with known prior
and Linear lithofacies prediction assumptions, KSVM has
Discriminate proven to be efficient
Analysis (LDA) in identifying distinct
features with the
decision function defined
by assisting vectors
Wang et al Core data, Microfacies Support Vector Radial Basis SVM algorithm attained Results indicate Although the paradigm
(2019) Wireline Machine (SVM) Kernel Function a good automatic an 84% prediction had a high prediction
logs (GR, and quantitative accuracy attained accuracy, it could have
Vsh) identification using the SVM been higher if a good
of depositional model feature selection tool had
microfacies with a good been used to identify the
prediction accuracy relevant parameters for
model training.
Ameur- Well Logs Lithofacies Multilayer MLPNN MLPNN had a good MLPNN MLPNN is insensitive
Zaimeche et (GR, Perceptron Sigmoid lithofacies prediction RMSE and MAE to outliers present in
al. (2020) RHOZ, K, Neural Network Activation with the lowest RMS = 0.39 and 0.23, the input data and also
TH), Core (MLPNN), Function and highest R-value respectively work well for complex
data Classical Cluster CCA as compared with R (testing) = 0.92 function as compared
Analysis (CCA) Partitioning and the classical cluster CCA with CCA which is
Hierarchical analysis RMSE and MAE sensitive to outliers.
= 1.04 and 0.54,
respectively
R = 0.68
Xie et al. Well Log Lithology Randomised N-fold Cross- Results show that the RETC had a With the data input
(2020) (GR, CNL) Identification Ensemble Tree Validation, proposed framework prediction accuracy considered for the
Classifier (RETC), Multivariate (RETC) has a higher of 89.4% and 91.1% lithology identification,
Random Forest Outlier prediction accuracy in the Daniudui and the prediction accuracy
(RF), Gradient Tree for sandstone lithology Hangyinqi Gas Field, would have improved if
Boosting (GTB), than the other AI respectively another contributing well
and Xgboosting classifiers. log such as photoelectric
factor, and density was
used.
Asante- Lithology Lithofacies Gradient-Boosted K-mean, Results show improved Comparing the The K-mean cluster
Okyere et al. data and Machine (GBM) Gaussian Mixture performance in terms of various model GBM algorithm employed
(2020b) Well logs Models classification accuracy had a Training and is based on the cluster
(RHOB, by the K-mean -based Testing accuracy of centres of the data with
GR, SP, GBM 80.97% and 80.96%, which if the data cluster
CN, and respectively while does not show a defined
ILD) K-mean GBM had circular cluster, its

Petroleum Reservoir Characterisation 237

82.89% and 81.42% classification becomes
for training and difficult.
testing, respectively
Chai et al. Seismic Seismic Deep Learning BridgeNet and The accuracy of NN An ANN with a higher
(2022) data facies (DL) Neural ResNet projected facies kernel size outperforms
Network along the training one with a shorter kernel
and testing portions size. Furthermore, the
indicates that the results BridgeNet with shortcut
are congruent with connections outperforms
geologic sedimentation, the BridgeNet without
confirming the shortcut connections.
generalization
capabilities of the
enhanced flexible
BridgeNet. DL
238 Data Science and Machine Learning Applications in Subsurface Engineering

classification was based on the six different facies group from the North Sea
turbidite reservoir, taking into consideration the clay content, grain size, and
bedding configuration. After data filtering, the authors used a neural network
and various multivariate statistical techniques to classify the lithofacies. The
Mahalanobis Linear Discriminant Analysis (MLDA), Probability Density
Function (PDF) classifier, and a Multilayer Feed-forward design with
Back-propagation (MLFN-BP) mass adaptation methodology were used based
on their ability to establish a link between facies and their physical properties.
The results emphasised the use of PDF for the identification of new facies
using holdout cross-validation; however, the MLFN-BP classifier was the best
algorithm for dealing with the multidimensional cluster boundary. In terms of
model performance, Otchere et al. (2021a), suggested there would have been
an improvement if a stronger selection technique was employed to select only
pertinent input variables and outlined a selection bias for their model. To also
increase the performance of their model, the rock physics model would have
been the best conditioning tool for the input well log data. Ameur-Zaimeche
et al. (2020) conducted additional research on the feed-forward neural network
using the MLP to reconstruct the lithofacies breaks in the Sif Fatima oil field
in Algeria. Linking MLP to cluster analysis, their outcomes showed that MLP
is better suited for forecasting the non-cored lithofacies.
Tang (2008) defined carbonate lithofacies with some well logs using a
Probabilistic Neural Network (PNN) by analysing the multidimensional
correlation between the variables. Because of its ability to analyse the
multidimensionality that exists between well logs and distinct facies, the PNN
outperformed discriminate analysis and multi-logistic statistical algorithms.
To assess the reliability of the PNN lithofacies prediction, an integration of two
log lithology indicators and the model was used to map zonally using simple
kriging. There was a good correlation between the zonal facies map with the
conceptual reservoir model which indicates a good reservoir modelling and
flow simulator. Al-Mudhafar (2017b) used well logs to investigate the use
of the Kernel Support Vector Machine (KSVM) to model the distribution of
lithofacies. Because of its ability to recognize distinct classes with a decision
function defined by additional subgroups of supporting vectors, the author
chose KSVM. To further validate the model, it was cross-validated using the
known lithofacies. The validation yielded a 99.55% accuracy using the KSVM
algorithm. Also in recent years, there has been increasing documentation of
improved facies and lithology analysis (e.g., Asante-Okyere et al., 2020b;
Otchere et al., 2022b; Shen et al., 2019; Xie et al., 2021).
 Petroleum Reservoir Characterisation 239

6. Seismic Guided Petrophysical Property Prediction

Integration of seismic and well log data for petrophysical property prediction
requires bringing the source data to a common dimension (depth and time),
specifically with both source in depth (well log) and time (seismic). This is
usually achieved by using a seismic-to-well tie (Saikia et al., 2020) by matching
the seismic response at the well location to that of the synthetic trace extracted
from the well log. The synthetic trace is generated by convolving the reflection
coefficient with a synthetic wavelet and using the reflection behaviour of the
geological layers. Upon generating the synthetic wavelet, information on the
depth of the synthetic trace and its equivalent time for the real seismic trace
would be established. Once the best connection (stretch and squeeze) is attained,
the depth-to-time relation is established (Saikia et al., 2020).
The use of seismic-guided reservoir property prediction is one of the
improved methods to help estimate reservoir properties (Table 10.5). In fields
where there are not many wells and interpreting the reservoir only from the
well logs becomes difficult, this method adopts statistical tools to estimate the
reservoir property for the entire field. The information gained from the well
log is not used proficiently for a larger field since this information will be
confined to a specific location. Due to this difference in well log information,
integrating the seismic data helps to give an estimation of the varying reservoir
property of the field at large. In recent years, different AI algorithms have
been implemented to help in the estimation of reservoir properties with the
use of seismic data and well logs.
Gogoi and Chatterjee (2019) predicted reservoir properties from well logs
and seismic data by employing the MLF algorithm in Tipam sandstone and
Barail arenaceous sandstone reservoirs. The data set used includes complete
acoustic impedance from 2D post-stacked seismic data, as well as RHOB,
Phi, and shear impedance. The MLF algorithm generated a single-layer shale
volume and Sw from numerous hidden layers. Upon analysis, the outcome
showed a minimal estimation error for both reservoirs, suggesting that
MFL can be used to predict such properties. The MLF has been established
to be effective in the estimation of sand or shale reservoir properties by
Iturraran-Viveros and Munoz-Garcia (2018). They investigated the use of
seismic traits and wireline logs as input factors for describing petrophysical
properties for the Tenerife field at a seismic scale, with the option of sands or
shales. The gamma test, a multidimensional smoothing tool, was used to help
optimise the input variable selection for training the ANN. It thus supported
a decent assessment of Phi and Sw and brought about a decent connection
with core, well log just as seismic amplitudes for each facies. Soleimani
et al. (2020) assessed the performance of the MLF by comparing it with PNN,
RBF, Multiple Linear Regression (MLR), and Geostatistical Method (GM)
 Table 10.5. Literature on the application of AI in seismic guided petrophysical property prediction.

240 Data Science and Machine Learning Applications in Subsurface Engineering

Author(s) Input Variables Predicted Model Used Activation Output Statistical Remarks
Output Functions Measures
De Matos et al. 3D Seismic data Seismic Self-Organizing Gaussian- AN improvement in NN Even though SOM
(2007) Facies Maps (SOMs), distribution seismic facies prediction provides good lithofacies
Wavelet Transform, function was obtained by using classification with
K-mean Partitive the wavelet transform seismic data as input, the
Clustering technique to detect the addition of well log data
trace singularities would help improve the
classification performance.
Anifowose et Reflection Permeability Support Vector Kernel Upon comparison, SVM produced a SVM outperforms T2FL
al. (2013) Seismic data, Machine (SVM, Function the SVM model gave good prediction for when working with small
Well Logs Type-2 Fuzzy the most accurate the seismic plus log data size as compared
Logic (T2FL) permeability prediction input with an RMSE to T2FL which work
of about 0.49, R of best with input data with
about 0.7, and MAE uncertainties
of about 0.38
Sebtosheikh Seismic Lithology Support Vector Kernel Results indicate a good NN With the SVM, in a case
and Salehi Inversion, Well Machine (SVM): Function prediction of the SVM where the quantity of
(2015) logs Statistical Learning after the training set size features for an individual
Theory (SLT), reduced data point exceeds the
Structural Risk number of training
Minimization datasets, the SVM will
(SRM), Empirical underperform
Risk Minimization
(ERM)
Rafiq et al. Post stack Microseismic Cluster Analysis Principal Results indicate NN Although cluster analysis
(2016) 3D Seismic Facies Curvature integrating microseismic provided great insight,
data, seismic Anomaly with seismic data it is heavily affected by
attributes, Well provides valuable sampling error and would
logs, Core insight for characterizing appear to be legitimate
data Hydraulic unconventional reservoirs upon analysis.
fracture
treatment log
Srisutthiyakorn 2D/3D Binary Permeability Multilayer Neural Tan- Results using MNN and CNN had the lowest Although CNN has
(2016) Segmented Network (MNN), sigmoid CNN produced a good testing MSE of proven to be efficient
Images, Porosity, Convolutional Function, prediction of permeability 2.4307*E05 and and represent complex
Specific Surface Neural Network Rectified especially in the case of MNN has the lowest nonlinear functions,
Area (CNN), Gradient Linear 2D CNN in multiscale. testing MSE of without a large dataset its
Descent (GD) Function 2.3999*E05 in the efficiency becomes void.
case of the 2D in
multiscale.

Petroleum Reservoir Characterisation 241

Iturraran- Reflection Porosity Backpropagation Gamma Results show a good NN With the ANN having the
Viveros and 2D Seismic (Phi), ANN, Conjugate Test correlation between the advantage of not knowing
Munoz-Garcia data, Seismic Saturation Gradient ANN, predicted parameters with the process responsible
(2018) Attribute, Well (Sw), and BFGS ANN the core data obtained for generating the data but
logs (SP, ILD, Volume of from the location still performing better, it
DT, GR) Clay (Vsh) always has the issue of
hyperparameter tuning
and also dimensionality
issues.
Table 10.5 contd. ...
...Table 10.5 contd.

242 Data Science and Machine Learning Applications in Subsurface Engineering

Author(s) Input Variables Predicted Model Used Activation Output Statistical Remarks
Output Functions Measures
Gogoi and 2D Post-Stack The volume Multilayer Feed Sigmoid Results indicate a reliable Vshale With the ability
Chatterjee Seismic data, of shale Forward Neural Activation prediction using the A cross-correlation to approximate
(2019) Well logs (DTs, (Vsh), Water Network (MLFN) Function MLFN model of 0.88 was attained complex function, its
RHOZ, GR, Saturation by MLFN between hyperparameters selection
LLS, ILD, (Sw) the actual and the becomes challenging and
NPHI) predicted with a always rely on a trial and
0.031 error error approach
Sw
A cross-correlation
of 0.85 was attained
by MLFN between
the actual and the
predicted with a
0.05 error
Soleimani et Reflection Porosity Multi-layer Gaussian Results indicate the PNN had an average With PNN having a
al. (2020) Seismic data, Feed-Forward Function supremacy of the Neural Validation error of good validation error, it
Well Logs Network (MLFN), Network compared with 4.6 as compared to is inefficient in terms of
(RHOZ, GR, Radial Basis the other methods used 4.8 for MLFN memory usage.
DT) Function (RBF),
Probabilistic
Neural Network
(PNN),
Geostatistics,
Regression
Analysis
Qiang et al. Well Logs (GR, Porosity, Support Vector Radial SVM Showed good SVM attained an The implementation of
(2020) SP, Calliper, Saturation, Machine (SVM), Basis prediction for porosity average overall R of feature selection tools
ILD, DT, DTs, Spatial Sequential Kernel while a hybrid of GIS 0.84 for all the input may have resulted in
RHOZ, NPHI), distribution Gaussian Function and SVM gave a higher wells the improvement of the
Core Data, of acoustic Simulation prediction of spatial performance of the model.
and Post stack impedance (SGS), and variation of reservoir
Inversion Gaussian Indicator properties for the
Simulation (GIS) reservoir

Petroleum Reservoir Characterisation 243

244 Data Science and Machine Learning Applications in Subsurface Engineering

by estimating Phi and compressional wave velocity from seismic attribute

and well logs in an Iranian oil field. The PNN-Gaussian function selection
documented the most reduced error, most noteworthy correlation, and least
computational time as compared to the others by using three unique datasets
for testing, training, and validation.
Evaluation of SVM to ANN and Type-2 Fuzzy Logic (TFL) was
conducted by Anifowose et al. (2013) to help improve the forecast of K in a
carbonate reservoir. The prediction was done by the use of a 3D seismic signal
and wireline logs as input parameters. The results showed that combining
seismic signatures and wireline logs enhanced the training performance of the
SVM and ANN models. Sebtosheikh and Salehi (2015) led a test impact of a
limited training dataset in estimating the lithology from transformed seismic
characteristics and directed log information in a heterogeneous carbonate
reservoir using the SVM technique, the outcomes of which demonstrated its
robustness in the face of limited data. It was also discovered that reducing
the training and testing data sizes did not affect the presentation of SVM in
lithological prediction. To improve the prediction of the spatial distribution
of acoustic impedance, Phi, and saturation for a gas field, Qiang et al. (2020)
proposed a new model to help improve the reservoir quality prediction from
the integration of well logs, core and seismic inversion data. The authors
employed the use of Sequential Gaussian Simulation (SGS) and Gaussian
Indicator Simulation (GIS), which are both statistical algorithms to estimate
the spatial distribution of the reservoir properties. In order to improve the
robustness of the model, it was calibrated to post-stack seismic inversion to
establish a relationship utilising SVM, Radial Basic Function (RBF) kernel,
and well logs. Finally, SVM produced a good relationship between the thin
lithofacies layers and seismic waveforms, resulting in a better prediction of
the facies’ spatial distinction. Also, there was a good correlation of Phi and
Sw prediction utilising SVM to facies classification with superior reservoir
zones. To show the robustness of SVM, Wang et al. (2019) applied SVM to
identify depositional microfacies from wireline logs. The SVM model was
trained using integrated depositional microfacies obtained via core sample
observation and well logs. The authors developed a quantifiable discrimination
SVM model and compare it to standard results, which aided in the prediction
of depositional microfacies in uncored wells with up to 84% accuracy.

7. Hybrid Models of AI for Petrophysical Property Prediction

AI has proven to improve the accuracy of reservoir property prediction as
compared to conventional mathematical approaches, yet the need to further
develop its effectiveness is still sought after (Otchere et al., 2021a). The
various improvements to AI model accuracy and reducing processing time
 Petroleum Reservoir Characterisation 245

are to incorporate different models to handle the various limitations such as

issues of dimensionality, overfitting, and local minima. In order to curb this
limitation, various literature has applied various models for feature selection
(Asante-Okyere et al., 2020; Dorrington et al., 2004; Lim et al., 2004; Zerrouki
et al., 2014) and improved its optimisation (Amiri et al., 2015; Anifowose
et al., 2011; Anifowose et al., 2017; Saemi et al., 2007) to improve the
modelling ability of AI algorithms.
Moghadasi et al. (2017) investigated the use of bagged ANN, ANN-based
Principal Component Analysis (PCA), and BPNN to predict permeability
and porosity. Seven well logs were utilised as model parameters from the
11 wells used, and the predicted output was trained using core permeability
and porosity. PCA was used as a feature extraction method to help identify
the relevant variables. In order to average the predicted multiple datasets to
enhance the level of stability and lower uncertainty, the bootstrap sampling
bagging technique was used to generate multiple training sets for the prediction.
Based on the models generated, the bagged ANN ensemble model prediction
documented the highest R2 of 0.94 (94%) and 0.85 (85%), respectively, and
the lowest MSE of 0.001 and 0.0034 for Phi and K. Srisutthiyakorn (2016)
developed a Multilayer Neural Network (MNN) and Convolutional Neural
Network (CNN) paradigms in predicting permeability from geometry
and feature extraction from rock images. Their paradigm employed slope
descent-based back-propagation with Bayesian regularisation for training.
MNN was modelled using porosity, Euler number, integral of the mean
curvature, and specific surface area. From their research, the feed-forward
network gave rise to a higher testing MSE (mean squared error) compared to
the Bayesian regularisation network and CNN.
Xie et al. (2020) developed a coarse-to-fine basis that incorporates
outlier detection, and multi-class classification with a very randomized
tree-based classifier to resolve the issue of repeatability of most machine
learning algorithms. Comparisons with some ML classifiers, such as random
forest, gradient tree boosting, and xgboosting, were used to help improve the
model’s competency. In sandstones, the model outperformed the baseline
classifiers in prediction accuracy. Asante-Okyere et al. (2020a) adopted a
hybridised model to precisely estimate reservoir water saturation. This model
adopted PCA as a feature extraction technique to boost the effectiveness of
the optimised least square support vector machine (LSSVM) and adaptive
neuro-fuzzy inference system-subtractive cluster method (ANFIS-SCM).
During training and testing the PC-LSSVM model achieved a better
performance as compared to the proposed PC-ANFIS-SCM model. For their
outcome, the PC-LSSVM hybridised model had the lowest prediction error
and outperformed the hybridised PC_ANFIS-SCM model with the valuation
of water saturation. Asante-Okyere et al. (2020b) proposed a hybrid model for
246 Data Science and Machine Learning Applications in Subsurface Engineering

lithology classification that combines gradient-boosted machine and cluster

algorithms. The authors take advantage of the benefit of combining results
from clustering well log data into two and three groups using K-means and
Gaussian mixture models (GMM) to create a more precise gradient-boosted
machine (GBM) lithology model. Their findings demonstrated improvement
in terms of classification accuracy rate by the K-means-based GBM classifier.
Furthermore, when the setup classifiers were tested on the entire dataset,
the GMM-based GBM perceived an improved presentation. The classifiers’
evaluation of the matrices revealed that the cluster-based hybrid GBM models
performed well, which was attributed to an increase in quality in perceiving
mudstone and siltstone, which are the primary lithofacies present in the South
Yellow Ocean’s Southern Basin.
Anifowose et al. (2015) investigated the performance of SVM to determine
the optimal value of the regularization parameter on which it is dependent.
Following further evaluation, a stacked generalization ensemble model of SVM
was suggested to improve reservoir parameter forecasting while accounting
for K and Phi. This study used input variables from six North American
well logs and eight Middle Western well logs for K estimation. Following a
comparison with the standard SVM, the suggested SVM ensemble outscored
the orthodox one. To satisfy the increased demand for the hybrid model to
solve complex industrial problems, Anifowose et al. (2011) hybridised the
functional network and SVM (FN-SVM) to further improve the reliability of
the orthodox SVM. Six wells with variable conventional well logs were used
as inputs for the model to predict K and Phi and were contrasted with the
conventional SVM. With the use of the Least Square Based feature selection
algorithm of FN, the independent well logs were chosen, which helped with
the dimensionality reduction problem and improved the time productivity of
the model. By comparing the models, the FN-SVM had a higher correlation
coefficient and a shorter handling time than the conventional one. Zhong and
Carr (2019) compared a hybrid Particle Swarm Optimisation (PSO) mixed
kernel function-based SVM (PSO-MKF-SVM) to various types of MKF-SVM,
SVM, and ANN in the prediction of Phi when well logs and core data were
insufficient. Gamma Ray (GR), density logs, and three extracted parameters
(slope of GR, slope of density, and Volume of Shale) from six wells were used
as input for the model. SVMs with different core functions have been used
to predict Phi and experienced difficulties in discovering the regularisation
parameter and the global optimum. Notwithstanding, the applied MKF-SVM
showed an improvement in the localisation of the optimal control parameters,
resulting in an increased efficiency based on R2, correlation coefficient (r),
and RMSE. To test their equivalence, the MKF-SVM (RBF-LS-SVM) was
compared to MLPNN and the newly developed PSO-MKF-SVM. As a basis
for comparison, the proposed PSO-MKF-SVM model outperformed the other
 Petroleum Reservoir Characterisation 247

methods with RMSE = 1.6505, R2 = 0.9140, r = 0.9560, Average Absolute

Error (AAE = 1.4050) and Maximum Absolute Error (MAE = 2.717).
Otchere et al. (2021b) proposed a hybrid ensemble model for the
prediction of reservoir permeability and water saturation using well logs from
the Volve Field. The input logs used were GR, calliper (CALI), RHOZ, DT,
NPHI, PEF, rate of penetration (ROP) and resistivity logs from two wells.
The authors applied a pre-processing tool to clean the data and select input
variables which are relevant for training and prediction. The grid search
hyperparameter algorithm was used to select optimal parameters for effective
model training. The model results indicated a high training accuracy for the
traditional XGBoost for permeability while the proposed ensemble model
(RFLasso-XGBoost regression) had the most robust fit for the permeability
prediction of 0.98 test score. Also, for water saturation prediction, the proposed
RFLasso-XGBoost model exhibited the highest correlation coefficient for
both training (0.98) and testing (0.93). Otchere et al. (2022a) proposed a new
application of deep ensemble super learner to establish a relationship between
wireline logs and NMR T2LM and predict NMR T2LM. The predicted log
served as the basis for subsequent input for their developed methodology
and mathematical models to predict reservoir wettability. The core plugs of
sandstone samples from Western Australia were used for the analysis. Results
indicate an excellent performance of the deep ensemble super learner.

8. Summary
Tables 10.2 to 10.5 show a summary of the works done concerning the
application of AI techniques in reservoir characterisation. The summary
comprises the input data, predicted output, AI architecture, activation function,
output performance, and the statistical measure for the AI algorithm. An
overview of the advantages and limitations of some AI algorithms is discussed
in detail by Saikia et al. (2020). It can be observed from Tables 10.2 to 10.5
that all works done considered supervised and unsupervised AI algorithms to
achieve a good prediction for their developed models. Considering this, most
industries produce enough data for their operations but to confirm the actual
ground truth, a small amount of core data are being generated to validate the
model of the empirical relation used. With little core information, training
a supervised model becomes difficult. To put to use the few core dataset, a
semi-supervised algorithm becomes useful. The semi-supervised algorithm
considers using the small dataset in a supervised manner and later applies
an unsupervised model to optimise the final prediction. From Tables 10.2 to
10.5, it can be seen that there is limited work done concerning unsupervised
techniques and hybrid unsupervised techniques in the field of reservoir
characterisation. Furthermore, it has also been observed that with the various
248 Data Science and Machine Learning Applications in Subsurface Engineering

AI algorithms used, conditioning of the input data helps to improve the

prediction performance of the algorithm. The use of rock physics models as a
complementary tool in conditioning the well logs has in recent times gained
much attention and has proven to be useful in well log conditioning. Also,
much work in combining rock physics and AI algorithms has not been done.
In all, it can be inferred from Table 10.1 that, this review work will be a basis
to advance the study of AI applications in reservoir characterisation.

9. Challenges and Perspectives

Yu et al. (2011) highlighted that reservoir characterisation is a multidisciplinary
integration exercise that involves qualitative and quantitative interpretation.
With increased complexity and abundance of data from various geological and
reservoir problems, new technologies and methods are increasingly adopted
in reservoir characterisation. Proper integration substitutes the need to run
expensive and unrelated logs to make available correct estimation faster and
strengthen your confidence as an operator. A good representative reservoir
characterisation presents a good model for the optimisation of the lifetime
performance of a field. Achieving that requires an accurate measurement
available from multiple disciplines to prepare a good quantitative and
qualitative representation of the reservoir. Significant experiments have been
made in this regard, and many challenges remain, some perspectives for the
future are discussed.

9.1 AI Perspective
Literature has highlighted various modifications of the ANN algorithm like
the functional network (FNN), PNN, and Radial Basis Function (RBF)
to help remove the various limitation of the ANN algorithms. This has
led to addressing nonlinearity between various inputs which the empirical
correlations could not highlight. AI has produced improved prediction and
classification in a different task in reservoir characterisation, that is, the use of
hybrid modelling (Amiri et al., 2015; Anifowose et al., 2011; Anifowose et al.,
2017; Saemi et al., 2007) and also ensemble model (Anifowose et al., 2015) to
make available good hyperparameters for modelling.
Feature extraction and selection by AI algorithms is one of the main
areas most literature is addressing to help improve the prediction capability
of these AI tools. Feature selection looks at the individual input for the AI
algorithms by looking at the necessary and unnecessary features which intern
affect the efficiency of the algorithm. Since the conventional AI algorithm
lacks the power of feature extraction, most research has looked at hybridising
AI algorithms to cater to that deficiency. Considering the various feature
selection tools (PCA, forward feature selection (FFS), linear discriminant
 Petroleum Reservoir Characterisation 249

analysis (LDS), etc.), fuzzy-aided hybrid models have proven to provide good
performance in many reviewed literature.
Due to the spatial dimension of the various dataset used to train the AI
algorithm, the need to transform the dataset from a high dimensional space to
a low dimensional space while keeping the high dimensional structure helps
improve the prediction power of the ANN and ML algorithm. This process of
reducing dimensionality is termed feature extraction. With the development of
deep learning algorithms, the problem of feature extraction has been addressed
since the algorithm can extract valuable information from the primary data
input. The literature reviewed here has indicated that the application of soft
computing methods has to do with the trial-and-error method of choosing
the parameter settings to obtain high-performance accuracy. Therefore, future
studies have to employ an automatic selection of parameters for soft computing
methods to avoid long time wasted in looking for optimal parameters.
In recent years, the oil industry has seen a high volume of data generation
by progressive sensors (Saikia et al., 2020) which has led to the development of
advanced modelling and estimation techniques. Such advanced modelling is the
application of a deep learning algorithm (e.g., CNN) to handle feature selection
(Shaheen, 2016) and extract useful messages from the input data through its
hidden neurons (Saikia et al., 2020). The application of deep learning (DL)
has not been well exploited in the domain of reservoir characterisation but has
yielded much success in the field of image, and speech recognition (Bae et al.,
2016; Fujiyoshi et al., 2019; He et al., 2016; Pouyanfar et al., 2018) and other
fields like the medical sector (Chen et al., 2018; Fang et al., 2019). Although
DL has in recent years been applied to solve various learning assignments,
training the model is difficult (Rere et al., 2016) but with improved data
availability, the performance of DL models has been better. Various algorithms
such as Stochastic Gradient Descent (SGD), Conjugate Gradient (CG),
Hessian-free optimization (HFO), and Krylov Subspace Descent (KSD) have
been implemented to curtail this deficiency over the years. These algorithms
have shown some limitations such as several manual tuning schemes for
SGD, slowness of the CG, and more memory consumption of the HFO and
KSD (Rere et al., 2016). To resolve the issues of some of these algorithms,
the hybridisation of metaheuristics to DL will be a good area to address. The
metaheuristic optimisation techniques have been successfully applied to solve
many optimization problems in the research area of engineering, sciences, and
related industries. Research on hybridising DL with metaheuristics algorithms
is scarce in the field of reservoir characterisation. Notwithstanding, its other
important status which is learning from unlabelled data concerning reservoir
characterisation has also not been properly exploited. Likewise, with the
availability of core information coupled with well log data, the application
of semi-supervised deep learning approaches can lead to greater success in
250 Data Science and Machine Learning Applications in Subsurface Engineering

Fig. 10.6. Proposed framework for the semi-supervised learning Deep Neural Network.

reservoir characterisation and has to be exploited. Therefore, future works

will be directed to looking at the application of semi-supervised deep learning
approaches to help improve prediction for reservoir properties, and facies
analysis and also look at improving the DL model with some metaheuristic
algorithms. Figure 10.6 shows a framework for the proposed semi-supervised
deep learning model. In the semi-supervised learning algorithm architecture
(Fig. 10.6), a small amount of labelled data is used for partial training of the
model (supervised learning). After the supervised training, the trained model
and the unlabelled dataset are used to predict the output which produces a
pseudo label. Since these pseudo results may not be accurate, the algorithm
links the pseudo labels data and the labelled training data at the same time
linking the input labelled training data and that of the unlabelled training data.
This linked data again uses the data to train the model to reduce the error and
improve the accuracy of the model.

9.2 Rock Physics Perspective

Geophysical and petrophysical well logs are subjected to noise and various
processing uncertainties, henceforth, a proper pre-processing technique
is required to improve the qualitative and quantitative analysis of the data.
Saikia et al. (2020) highlighted noise, outliers, spurious attribute removal,
and a suitable correlation of different data sources in terms of time, depth,
and frequency as some aspects of pre-processing technique. To handle these
effectively, the application of rock physics has gained much attention in the
society of geophysics. This paradigm has contributed to well log conditions.
Also, rock physics in recent years has gained much attention in the oil
industry to help condition the input well logs in other to improve reservoir
characterisation (Chi and Han, 2009; Mukerji et al., 2001; Saberi, 2013;
Saxena et al., 2018). To capture the heterogeneity and complexity of the
rock’s log response and the uncertainty associated with theoretical relations, a
 Petroleum Reservoir Characterisation 251

Fig. 10.7. Application of Rock Physics for well conditioning and its contribution to reservoir
characterisation.

seismic-guided approach coupled with rock physics has been the best
method. Since geological location has a great influence on the acquired
data (geophysical, core data, and petrophysical), the characteristics (elastic
moduli) of the geology can be addressed by the use of rock physics
models and to understand the geophysical data response by such geology
(Fig. 10.7). In Fig. 10.7, rock physics models are used to condition the various
input logs (such as density and sonic logs) which are then used to better
enhance reservoir characterisation (deterministic inversion, facies probability
mapping, and others). The use of this technique helps in understanding the
geological complexity, improving the well log response, and a considerable
level of accuracy with a smaller number of data samples can be achieved.
Therefore, future works should exploit the use of rock physics models
alongside AI models to condition the various inputs used for the prediction
of reservoir properties to have a more accurate and precise result for future
property modelling.

10. Conclusions
The past years have seen the petroleum industry generate abundant data for
its resource estimation. However, under current economic conditions, most
field data acquisition has been limited thereby relying on the acquired data
for reservoir characterisation taking into account the use of complex models.
Integration of all available data (petrophysics, geology, geophysics, rock
mechanics, and engineering) play an important role in achieving a good
reservoir characterisation. The integration of the available data provides a high
252 Data Science and Machine Learning Applications in Subsurface Engineering

lateral resolution (seismic data), and high vertical resolution (well logs) while
core data provides a physical measurement of the reservoir. Also, integrating
these data sources helps to improve inter-well estimation taking into account
the seismic data. Reservoir characterisation had been challenged by the fact
that the various models have to handle complex nonlinearity present in the
available dataset and also the various uncertainty present in the dataset and
models. These flaws are mitigated by using strong data condition methods
for dimensionality reduction, denoising, and identifying appropriate features
for prediction to achieve accurate and reliable reservoir characterisation. The
conclusion for the reservoir characterisation is highlighted as follows:
1. Most ML paradigms employed for petrophysical seismic characterisation
have been reviewed. These ANN models have been used to improve the
issues facing the empirical approach used to determine petrophysical
parameters. Most literature highlights the use of SVM over ANN models
since the former gives the least error, handles small datasets, and has
faster processing times.
2. There has been an increase in the use of hybridizsd models in the field
of reservoir characterisation. The hybrid models have helped to solve
various problems facing conventional AI algorithms such as parameter
optimisation, weight adjustment, computational time, uncertainty in data,
and feature selections.
3. The selection of most hyperparameters for AI models’ application to
reservoir characterisation in the literature entails a guide technique of trial
and error which is tedious and time-ingesting and often ends in suboptimal
trained models. Hence, advanced studies must discover greater green
strategies for using metaheuristics algorithms for the automatic selection
of optimal hyperparameters and for improving computational time.
4. In a complex application, deep neural networks outperform traditional
neural networks in handling high data complexity with automatic feature
extraction. As a result, this method avoids the need for a separate feature
extraction and selection phase.
5. Finally, while deep neural networks and hybridised metaheuristic models
are widely used in many classification problems, their application has not
been employed much in predicting reservoir properties. Seismic-driven
integration model with deep learning and hybrid metaheuristic models
with DL is yet to be explored in depth.
6. While data condition improves the overall prediction of various AI
algorithms, the use of rock physics to help condition the various log
inputs in this domain is yet to be explored in depth.
 Petroleum Reservoir Characterisation 253

References
Al-Anazi, A. and Gates, I.D. 2010. A support vector machine algorithm to classify
lithofacies and model permeability in heterogeneous reservoirs. Engineering Geology
114(3-4): 267–277. https://doi.org/10.1016/j.enggeo.2010.05.005.
Al-Anazi, A.F. and Gates, I.D. 2012. Support vector regression to predict porosity and
permeability: Effect of sample size. Computers and Geosciences 39: 64–76. https://doi.
org/10.1016/j.cageo.2011.06.011.
Al-Bulushi, N., King, P.R., Blunt, M.J. and Kraaijveld, M. 2009. Development of artificial neural
network models for predicting water saturation and fluid distribution. Journal of Petroleum
Science and Engineering 68(3–4): 197–208. https://doi.org/10.1016/j.petrol.2009.06.017.
Al-Mudhafar, W.J. 2017a. Integrating well log interpretations for lithofacies classification
and permeability modeling through advanced machine learning algorithms. Journal
of Petroleum Exploration and Production Technology 7(4): 1023–1033. https://doi.
org/10.1007/s13202-017-0360-0.
Al-Mudhafar, W.J. 2017b. Integrating kernel support vector machines for efficient rock facies
classification in the main pay of Zubair formation in South Rumaila oil field, Iraq. Modeling
Earth Systems and Environment 3: 1–8. https://doi.org/10.1007/s40808-017-0277-0.
Alzubaidi, F., Mostaghimi, P., Swietojanski, P., Clark, S.R. and Armstrong, R.T. 2021. Automated
lithology classification from drill core images using convolutional neural networks.
Journal of Petroleum Science and Engineering 197: 107933. https://doi.org/10.1016/j.
petrol.2020.107933.
Ameur-Zaimeche, O., Zeddouri, A., Heddam, S. and Kechiched, R. 2020. Lithofacies prediction
in non-cored wells from the Sif Fatima oil field (Berkine basin, southern Algeria): a
comparative study of multilayer perceptron neural network and cluster analysis-based
approaches. Journal of African Earth Sciences 166: 103826. https://doi.org/10.1016/j.
jafrearsci.2020.103826.
Amiri, M., Ghiasi-Freez, J., Golkar, B. and Hatampour, A. 2015. Improving water saturation
estimation in a tight shaly sandstone reservoir using artificial neural network optimised
by imperialist competitive algorithm: A case study. Journal of Petroleum Science and
Engineering 127: 347–358. https://doi.org/10.1016/j.petrol.2015.01.013.
Anifowose, F. and Abdulraheem, A. 2011. Fuzzy logic-driven and SVM-driven hybrid
computational intelligence models applied to oil and gas reservoir characterization.
Journal of Natural Gas Science and Engineering 3(3): 505–517. https://doi.org/10.1016/j.
jngse.2011.05.002.
Anifowose, F.A., Abdulraheem, A., Al-Shuhail, A. and Schmitt, D.P. 2013. Improved
permeability prediction from seismic and log data using artificial intelligence techniques.
pp. 2190–2196. In: SPEMmiddle East Oil and Gas Show and Conference. March. OnePetro.
https://doi.org/10.2118/164465-MS.
Anifowose, F.A., Labadin, J. and Abdulraheem, A. 2015. Ensemble model of non-linear
feature selection-based extreme learning machine for improved natural gas reservoir
characterisation. Journal of Natural Gas Science and Engineering 26: 1561–1572.
https://doi.org/10.1016/j.jngse.2015.02.012.
Anifowose, F.A., Labadin, J. and Abdulraheem, A. 2017. Hybrid intelligent systems in
petroleum reservoir characterisation and modelling: The journey so far and the challenges
ahead. Journal of Petroleum Exploration and Production Technology 7: 251–263.
https://doi.org/10.1007/s13202-016-0257-3.
Archie, G.E. 1952. Classification of carbonate reservoir rocks and petrophysical considerations.
Aapg Bulletin 36(2): 278–298. https://doi.org/10.1306/3D9343F7-16B1-11D7-
8645000102C1865D.
254 Data Science and Machine Learning Applications in Subsurface Engineering

Asante-Okyere, S., Shen, C., Ziggah, Y.Y., Rulegeya, M.M. and Zhu, X. 2020a. Principal
component analysis (PCA) based hybrid models for the accurate estimation of reservoir
water saturation. Computers and Geosciences 145: 104555. https://doi.org/10.1016/j.
cageo.2020.104555.
Asante-Okyere, S., Shen, C., Ziggah, Y.Y., Rulegeya, M.M. and Zhu, X. 2020b. A novel hybrid
technique of integrating gradient-boosted machine and clustering algorithms for lithology
classification. Natural Resources Research 29: 2257–2273. https://doi.org/10.1007/
s11053-019-09576-4.
Avseth, P. and Mukerji, T. 2002. Seismic lithofacies classification from well logs using statistical
rock physics. Petrophysics: The SPWLA Journal of Formation Evaluation and Reservoir
Description 43(02): 70–81.
Bae, H.S., Lee, H.J. and Lee, S.G. 2016. Voice recognition based on adaptive MFCC and deep
learning. pp. 1542–1546. In: 2016 IEEE 11th Conference on Industrial Electronics and
Applications (ICIEA), June. IEEE. 10.1109/ICIEA.2016.7603830.
Bestagini, P., Lipari, V. and Tubaro, S. 2017. A machine learning approach to facies classification
using well logs. pp. 2137–2142. In: Seg. Technical Program Expanded Abstracts 2017.
Society of Exploration Geophysicists. https://doi.org/10.1190/segam2017-17729805.1.
Bian, H., Xia, Y., Lu, C., Qin, X., Meng, Q. and Lu, H. 2020. Pore structure fractal characterization
and permeability simulation of natural gas hydrate reservoir based on CT images. Geofluids
2020: 1–9. https://doi.org/10.1155/2020/6934691.
Bourbié, T., Coussy, O. and Zinszner, B. 1987. Acoustics of Porous Media. Gulf Publ. Co.,
translated by N. Marshall from French, Acoustique des Milieux Poreu. https://doi.
org/10.1121/1.402899.
Bowers, M.C. and Fitz, D.E. 2000. A probabilistic approach to determine uncertainty in
calculated water saturation. In: SPWLA 41st Annual Logging Symposium. June. OnePetro.
Brantson, E.T., Sibil, S., Osei, H., Owusu, E.B., Takyi, B. and Ansah, E. 2022. A new approach
for saturation height modelling in a clastic reservoir using response surface methodology
and artificial neural network. Upstream Oil and Gas Technology 9: 100081. https://doi.
org/10.1016/j.upstre.2022.100081.
Carman, P.C. 1937. Fluid flow through granular beds. Trans. Inst. Chem. Eng. 15: 150–166.
https://doi.org/10.1016/S0263-8762(97)80003-2.
Carman, P.C. 1956. Flow of gases through porous media. Butterworths, London. https://doi.
org/10.1016/S0263-8762(97)80003-2.
Chai, X., Nie, W., Lin, K., Tang, G., Yang, T., Yu, J. and Cao, W. 2022. An open-source package
for deep-learning-based seismic facies classification: benchmarking experiments on the
SEG 2020 Open Data. IEEE Transactions on Geoscience and Remote Sensing 60: 1–19.
10.1109/TGRS.2022.3144666.
Chen, H., Engkvist, O., Wang, Y., Olivecrona, M. and Blaschke, T. 2018. The rise of deep learning
in drug discovery. Drug Discovery Today 23(6): 1241–1250. https://doi.org/10.1016/j.
drudis.2018.01.039.
Chi, X.G. and Han, D.H. 2009. Lithology and fluid differentiation using a rock physics template.
The Leading Edge 28(1): 60–65. https://doi.org/10.1190/1.3064147.
Costa, A. 2006. Permeability‐porosity relationship: A re-examination of the Kozeny‐Carman
equation based on a fractal pore‐space geometry assumption. Geophysical Research
Letters 33(2). https://doi.org/10.1029/2005GL025134.
De Matos, M.C., Osorio, P.L. and Johann, P.R. 2007. Unsupervised seismic facies analysis
using wavelet transform and self-organising maps. Geophysics 72(1): 9–21. https://doi.
org/10.1190/1.2392789.
Dong, S., Xu, L., Dai, Z., Xu, B.I.N., Yu, Q., Yin, S., Zhang, X., Zhang, C., Zang, X.,
Zhou, X. and Zhang, Z. 2020. A novel fractal model for estimating permeability in
 Petroleum Reservoir Characterisation 255

low-permeable sandstone reservoirs. Fractals 28(08): 2040005. https://doi.org/10.1142/

S0218348X20400058.
Dorrington, K.P. and Link, C.A. 2004. Genetic-algorithm/neural-network approach to seismic
attribute selection for well-log prediction. Geophysics 69(1): 212–221. https://doi.
org/10.1190/1.1649389.
Eaton, D., Caffagni, E., Van der Baan, M. and Matthews, L. 2014. Passive seismic
monitoring and integrated geomechanical analysis of a tight-sand reservoir during
hydraulic-fracture treatment, flowback, and production. pp. 1537–1545. In: Unconventional
Resources Technology Conference, Denver, Colorado, 25–27 August. Society of
Exploration Geophysicists, American Association of Petroleum Geologists, Society of
Petroleum Engineers. https://doi.org/10.15530/urtec-2014-1929223.
Elshafei, M. and Hamada, G.M. 2009. Neural network identification of hydrocarbon potential
of shaly sand reservoirs. Petroleum Science and Technology 27(1): 72–82. https://doi.
org/10.2118/110959-MS.
Fang, S.H., Tsao, Y., Hsiao, M.J., Chen, J.Y., Lai, Y.H., Lin, F.C. and Wang, C.T. 2019. Detection
of pathological voice using cepstrum vectors: A deep learning approach. Journal of Voice
33(5): 634–641. https://doi.org/10.1016/j.jvoice.2018.02.003.
Fujiyoshi, H., Hirakawa, T. and Yamashita, T. 2019. Deep learning-based image recognition
for autonomous driving. IATSS Research 43(4): 244–252. https://doi.org/10.1016/j.
iatssr.2019.11.008.
Furey, T.S., Cristianini, N., Duffy, N., Bednarski, D.W., Schummer, M. and Haussler, D. 2000.
Support vector machine classification and validation of cancer tissue samples using
microarray expression data. Bioinformatics 16(10): 906–914. https://doi.org/10.1093/
bioinformatics/16.10.906.
Gaymard, R. and Poupon, A. 1968. Response of neutron and formation density logs in
hydrocarbon bearing formations. The Log Analyst 9(05).
Gholami, R., Shahraki, A.R. and Jamali Paghaleh, M. 2012. Prediction of hydrocarbon reservoirs
permeability using support vector machine. Mathematical Problems in Engineering
2012: 1–18. https://doi.org/10.1155/2012/670723.
Gogoi, T. and Chatterjee, R. 2019. Estimation of petrophysical parameters using seismic
inversion and neural network modelling in Upper Assam basin, India. Geoscience Frontiers
10(3): 1113–1124. https://doi.org/10.1016/j.gsf.2018.07.002.
Hamada, G.M. and Elshafei, M.A. 2010. Neural network prediction of porosity and permeability
of heterogeneous gas sand reservoirs using NMR and conventional logs. Nafta
61(10): 451–465. https://doi.org/10.2118/126042-MS.
Harrison, B. and Jing, X.D. 2001. Saturation height methods and their impact on volumetric
hydrocarbon in place estimates. In: SPE Annual Technical Conference and Exhibition.
September. OnePetro. https://doi.org/10.2118/71326-MS.
He, K., Zhang, X., Ren, S. and Sun, J. 2016. Deep residual learning for image recognition.
pp. 770–778. In: Proceedings of the IEEE Conference on Computer Vision and Pattern
Recognition. IEEE.
Heiba, A.A., Sahimi, M., Scriven, L.E. and Davis, H.T. 1992. Percolation theory of
two-phase relative permeability. SPE Reservoir Engineering 7(01): 123–132. https://doi.
org/10.2118/11015-PA.
Helle, H.B., Bhatt, A. and Ursin, B. 2001. Porosity and permeability prediction from wireline
logs using artificial neural networks: A North Sea case study. Geophysical Prospecting
49(4): 431–444. https://doi.org/10.1046/j.1365-2478.2001.00271.x.
Helle, H.B. and Bhatt, A. 2002. Fluid saturation from well logs using committee neural networks.
Petroleum Geoscience 8(2): 109–118. https://doi.org/10.1144/petgeo.8.2.109.
256 Data Science and Machine Learning Applications in Subsurface Engineering

Hewett, T.A. 1986. Fractal distributions of reservoir heterogeneity and their influence on fluid
transport. In: SPE Annual Technical Conference and Exhibition. October. OnePetro.
https://doi.org/10.2118/15386-MS.
Hewett, T.A. and Behrens, R.A. 1990. Conditional simulation of reservoir heterogeneity with
fractals. SPE Formation Evaluation 5(03): 217–225. https://doi.org/10.2118/18326-PA.
Hu, Y., Yu, X. and Chen, G. 2012. Classification of the average capillary pressure function
and its application in calculating fluid saturation. Petroleum Exploration and Development
39(6): 778–784. https://doi.org/10.1016/S1876-3804(12)60104-9.
Huang, Z., Shimeld, J., Williamson, M. and Katsube, J. 1996. Permeability prediction with
artificial neural network modelling in the Venture gas field, offshore eastern Canada.
Geophysics 61(2): 422–436. https://doi.org/10.1190/1.1443970.
Huang, Z. and Williamson, M.A. 1997. Determination of porosity and permeability in reservoir
intervals by artificial neural network modelling, offshore Eastern Canada. Petroleum
Geoscience 3(3): 245–258. https://doi.org/10.1144/petgeo.3.3.245.
Iturrarán-Viveros, U. and Muñoz-García, M.A. 2018. Porosity and water saturation in sands
or shales using Artificial Neural Networks and seismic attributes in a clastic reservoir in
Colombia. pp. 1282–1285. In: International Geophysical Conference, Beijing, China,
24–27 April 2018. December. Society of Exploration Geophysicists and Chinese Petroleum
Society. https://doi.org/10.1190/IGC2018-314.
Kaydani, H., Mohebbi, A. and Baghaie, A. 2012. Neural fuzzy system development for the
prediction of permeability from wireline data based on fuzzy clustering. Petroleum Science
and Technology 30(19): 2036–2045. https://doi.org/10.1080/10916466.2010.531345.
Kozeny, J. 1927. Uber kapillare Leitung des Wassers im Boden-Aufstieg, Versickerung und
Anwendung auf die Bewasserung, Sitzungsberichte der Akademie der Wissenschaften
Wien. Mathematisch Naturwissenschaftliche Abteilung 136: 271–306.
Li, D. and Lake, L.W. 1995. Scaling fluid flow through heterogeneous permeable media. SPE
Advanced Technology Series 3(01): 188–197. https://doi.org/10.2118/26648-PA.
Lian, P.Q., Tan, X.Q., Ma, C.Y., Feng, R.Q. and Gao, H.M. 2016. Saturation modelling in a
carbonate reservoir using capillary pressure based saturation height function: a case study
of the Svk reservoir in the Y Field. Journal of Petroleum Exploration and Production
Technology 6(1): 73–84. https://doi.org/10.1007/s13202-015-0159-9.
Lim, J.S. and Kim, J. 2004. Reservoir porosity and permeability estimation from well logs
using fuzzy logic and neural networks. In: SPE Asia Pacific Oil and Gas Conference and
Exhibition. OnePetro. https://doi.org/10.2118/88476-MS.
Luo, H., Tang, D. and Tang, Y. 2015. Study on prediction of oil water contact in carbonate
reservoir with capillary pressure data. Editorial Department of Petroleum Geology and
Recovery Efficiency 20(2): 71–73.
Ma, Y.Z. 2011. Uncertainty analysis in reservoir characterization and management: How
much should we know about what we don’t know? In: Ma, Y.Z. and P.R. La Pointe
(eds.). Uncertainty Analysis and Reservoir Modelling, AAPG Memoir. 96: 1–15.
DOI:10.1306/13301404M963458.
Mavko, G. and Nur, A. 1997. The effect of a percolation threshold in the Kozeny-Carman
relation. Geophysics 62(5): 1480–1482. https://doi.org/10.1190/1.1444251.
McCulloch, W.S. and Walter, P. 1943. A logical calculus of the ideas immanent in nervous
activity. The Bulletin of Mathematical Biophysics 5: 115–133. https://doi.org/10.1007/
BF02478259.
Mehana, M. and El-monier, I. 2016. Shale characteristics impact on Nuclear Magnetic
Resonance (NMR) fluid typing methods and correlations. Petroleum 2(2): 138–147.
https://doi.org/10.1016/j.petlm.2016.02.002.
 Petroleum Reservoir Characterisation 257

Meshri, I.D. 1986. On the reactivity of carbonic and organic acids and generation of secondary
porosity: Roles of organic matter in sediment diagenesis. In: Gautier, D.L. (ed.). Roles of
Organic Matter in Sediment Diagenesis. SPEM Special Publication. 38: 123–128.
Miller, R.S., Rhodes, S., Khosla, D. and Nino, F. 2019. Application of artificial intelligence for
depositional facies recognition-permian basin. In: Unconventional Resources Technology
Conference, Denver, Colorado, 22–24 July, Society of Exploration Geophysicists:
pp. 4410–4415. https://doi.org/10.15530/urtec-2019-193.
Moghadasi, L., Ranaee, E., Inzoli, F. and Guadagnini, A. 2017. Petrophysical well log
analysis through intelligent methods. In: SPE Bergen One Day Seminar. April. OnePetro.
https://doi.org/10.2118/185922-MS.
Mohaghegh, S., Arefi, R., Ameri, S., Aminiand, K. and Nutter, R. 1996. Petroleum reservoir
characterisation with the aid of artificial neural networks. Journal of Petroleum Science
and Engineering 16(4): 263–274. https://doi.org/10.1016/S0920-4105(96)00028-9.
Mohamed, E., Elsayed, M., Hassan, A., Mahmoud, M. and El-Husseiny, A. 2022. A machine
learning approach to predict the permeability from nmr t2 relaxation time distribution
for various reservoir rock types. In: ADIPEC. October. OnePetro. https://doi.
org/10.2118/211624-MS.
Mukerji, T., Avseth, P., Mavko, G., Takahashi, I. and González, E.F. 2001. Statistical rock
physics: Combining rock physics, information theory, and geostatistics to reduce
uncertainty in seismic reservoir characterization. The Leading Edge 20(3): 313–319.
https://doi.org/10.1190/1.1438938.
Nelson, P.H. 1994. Permeability-porosity relationships in sedimentary rocks. The Log Analyst
35(03).
Nikravesh, M. and Aminzadeh, F. 2001. Past, present and future intelligent reservoir
characterisation trends. Journal of Petroleum Science and Engineering 31(2–4): 67–79.
https://doi.org/10.1016/S0920-4105(01)00121-8.
Nikravesh, M., Zadeh, L.A. and Aminzadeh, F. 2003. Soft Computing and Intelligent Data
Analysis in Oil Exploration. Elsevier.
Nyein, C.Y., Ghareb, M., Hamada and Ahmed Elsakka. 2018. Artificial Neural Network (ANN)
prediction of porosity and water saturation of shaly sandstone reservoirs. AAPG Asia
Pacific Region, the 4th AAPG/EAGE/MGS Myanmar Oil and Gas Conference: Myanmar:
A Global Oil and Gas Hotspot: Unleashing the Petroleum Systems Potential.
Okon, A.N., Adewole, S.E. and Uguma, E.M. 2021. Artificial neural network model for reservoir
petrophysical properties: Porosity, permeability and water saturation prediction. Modeling
Earth Systems and Environment 7(4): 2373–2390. https://doi.org/10.1007/s40808-020-
01012-4.
Olakunle, I., Chinedu, A., Udoka, N. and Muyiwa, E. 2015. Saturation height modelling in a
partially appraised gas field using analogue field core data: An optimisation case study
of ZAN field in the Niger Delta. In: SPE Nigeria Annual International Conference and
Exhibition. OnePetro. https://doi.org/10.2118/178374-MS.
Olson, T.M. 1998. Porosity and permeability prediction in low-permeability gas reservoirs from
well logs using neura networks. In: SPE Rocky Mountain Regional/Low-Permeability
Reservoirs Symposium. April. OnePetro. https://doi.org/10.2118/39964-MS.
Otchere, D.A., Ganat, T.O.A., Gholami, R. and Ridha, S. 2021a. Application of supervised
machine learning paradigms in the prediction of petroleum reservoir properties: Comparative
analysis of ANN and SVM models. Journal of Petroleum Science and Engineering
200: 108182. https://doi.org/10.1016/j.petrol.2020.108182.
Otchere, D.A., Ganat, T.O.A., Gholami, R. and Lawal, M. 2021b. A novel custom ensemble
learning model for an improved reservoir permeability and water saturation prediction.
258 Data Science and Machine Learning Applications in Subsurface Engineering

Journal of Natural Gas Science and Engineering 91: 103962. https://doi.org/10.1016/j.

jngse.2021.103962.
Otchere, D.A., Abdalla Ayoub Mohammed, M., Ganat, T.O.A., Gholami, R. and Aljunid
Merican, Z.M. 2022a. A novel empirical and deep ensemble super learning approach in
predicting reservoir wettability via well logs. Applied Sciences 12(6): 2942. https://doi.
org/10.3390/app12062942.
Otchere, D.A., Ganat, T.O.A., Nta, V., Brantson, E.T. and Sharma, T. 2022b. Data analytics
and Bayesian Optimised Extreme Gradient Boosting approach to estimate cut-offs from
wireline logs for net reservoir and pay classification. Applied Soft Computing 120: 108680.
https://doi.org/10.1016/j.asoc.2022.108680.
Ouenes, A. 2000. Practical application of fuzzy logic and neural networks to fractured reservoir
characterization. Computers & Geosciences 26(8): 953–962. https://doi.org/10.1016/
S0098-3004(00)00031-5.
Pettijohn, F.J. 1975. Sedimentary Rocks. New York: Harper & Row. Vol. 3, p. 628.
Pouyanfar, S., Sadiq, S., Yan, Y., Tian, H., Tao, Y., Reyes, M.P., Shyu, M.L., Chen, S.C. and
Iyengar, S.S. 2018. A survey on deep learning: Algorithms, techniques, and applications.
ACM Computing Surveys (CSUR) 51(5): 1–36. https://doi.org/10.1145/3234150.
Qiang, Z., Yasin, Q., Golsanami, N. and Du, Q. 2020. Prediction of reservoir quality from log-
core and seismic inversion analysis with an artificial neural network: A case study from
the Sawan Gas Field, Pakistan. Energies 13(2): 486. https://doi.org/10.3390/en13020486.
Qiu, Y.N. 2000. Development of geological reservoir modelling in past decade, Beijing (in
Chinese). Acta Petrolei Sinica. 12(4): 55–62.
Rafiq, A., Eaton, D.W., McDougall, A. and Pedersen, P.K. 2016. Reservoir characterization
using microseismic facies analysis integrated with surface seismic attributes. Interpretation
4(2): 167–181. https://doi.org/10.1190/INT-2015-0109.1.
Rere, L.M., Fanany, M.I. and Arymurthy, A.M. 2016. Metaheuristic algorithms for
convolution neural network. Computational Intelligence and Neuroscience. https://doi.
org/10.1155/2016/1537325.
Reza, C., Mohammed, A.R. and Mahmund, M. 2017. Determination of the petrophysical
parameters using geostatistical method in one of the hydrocarbon reservoirs in South West
of Iran. J. Sci. Eng. Res. 4(12): 44–55.
Rudyk, S. and Al-Lamki, A. 2015. Saturation-height model of omani deep tight gas reservoir.
Journal of Natural Gas Science and Engineering 27: 1821–1833. https://doi.org/10.1016/j.
jngse.2015.11.015.
Saberi, M.R. 2013. Rock physics integration: From petrophysics to simulation. In: 10th Biennal
International Conference and Expositions. Vol. 444.
Saemi, M., Ahmadi, M. and Varjani, A.Y. 2007. Design of neural networks using genetic
algorithm for the permeability estimation of the reservoir. Journal of Petroleum Science
and Engineering 59(1–2): 97–105. https://doi.org/10.1016/j.petrol.2007.03.007.
Sahimi, M. 1993. Flow phenomena in rocks: From continuum models to fractals, percolation,
cellular automata, and simulated annealing. Reviews of Modern Physics 65(4): 1393.
https://doi.org/10.1103/RevModPhys.65.1393.
Saikia, P., Baruah, R.D., Singh, S.K. and Chaudhuri, P.K. 2020. Artificial Neural Networks in the
domain of reservoir characterisation: A review from shallow to deep models. Computers
and Geosciences 135: 104357. https://doi.org/10.1016/j.cageo.2019.104357.
Saxena, V., Krief, M. and Adam, L. 2018. Handbook of Borehole Acoustics and Rock Physics for
Reservoir Characterization. Elsevier.
Schwartz, L.M. and Kimminau, S. 1987. Analysis of electrical conduction in the grain
consolidation model. Geophysics 52(10): 1402–1411. https://doi.org/10.1190/1.1442252.
 Petroleum Reservoir Characterisation 259

Sebtosheikh, M.A. and Salehi, A. 2015. Lithology prediction by support vector classifiers using
inverted seismic attributes data and petrophysical logs as a new approach and investigation
of training data set size effect on its performance in a heterogeneous carbonate reservoir.
Journal of Petroleum Science and Engineering 134: 143–149. https://doi.org/10.1016/j.
petrol.2015.08.001.
Shaheen, F., Verma, B. and Asafuddoula, M. 2016. Impact of automatic feature extraction in
deep learning architecture. pp. 1–8. In: 2016 International Conference on Digital Image
Computing: Techniques and Applications (DICTA). IEEE. 10.1109/DICTA.2016.7797053.
Shedid, S.A. and Saad, M.A. 2017. Comparison and sensitivity analysis of water saturation
models in shaly sandstone reservoirs using well logging data. Journal of Petroleum Science
and Engineering 156: 536–545. https://doi.org/10.1016/j.petrol.2017.06.005.
Shen, C., Asante-Okyere, S., Yevenyo Ziggah, Y., Wang, L. and Zhu, X. 2019. Group
method of data handling (GMDH) lithology identification based on wavelet analysis
and dimensionality reduction as well log data pre-processing techniques. Energies
12(8): 1509. https://doi.org/10.3390/en12081509.
Shook, M., Li, D. and Lake, L.W. 1992. Scaling immiscible flow through permeable media by
inspectional analysis. In Situ (United States) 16(4).
Singh, S., Kanli, A.I. and Sevgen, S. 2016. A general approach for porosity estimation using
artificial neural network method: A case study from Kansas gas field. Studia Geophysica et
Geodaetica 60: 130–140. https://doi.org/10.1007/s11200-015-0820-2.
Soleimani, F., Hosseini, E. and Hajivand, F. 2020. Estimation of reservoir porosity using
analysis of seismic attributes in an Iranian oil field. Journal of Petroleum Exploration and
Production Technology 10(4): 1289–1316. https://doi.org/10.1007/s13202-020-00833-4.
Srisutthiyakorn, N. 2016. Deep-learning methods for predicting permeability from 2D/3D binary-
segmented images. pp. 3042–3046. In: SEG Technical Program Expanded Abstracts 2016).
Society of Exploration Geophysicists. https://doi.org/10.1190/segam2016-13972613.1.
Tang, H. 2008. Improved carbonate reservoir facies classification using artificial neural network
method. In: Canadian International Petroleum Conference. June. OnePetro. https://doi.
org/10.2118/2008-122.
Timur, A. 1968. An investigation of permeability, porosity, and residual water saturation
relationships. In: SPWLA 9th Annual Logging Symposium. June. OnePetro.
Tran, T.V., Ngo, H.H., Hoang, S.K., Tran, H.N. and Lambiase, J.J. 2020. Depositional facies
prediction using artificial intelligence to improve reservoir characterisation in a mature
field of Nam con son basin, offshore Vietnam. In: Offshore Technology Conference Asia.
October. OnePetro. https://doi.org/10.4043/30086-MS.
Verma, A.K., Cheadle, B.A., Routray, A., Mohanty, W.K. and Mansinha, L. 2012. Porosity
and permeability estimation using neural network approach from well log data. pp. 1–6.
In: SPE Annual Technical Conference and Exhibition (May).
Walker, R.G. 1992. Facies, facies models, and modern stratigraphic concepts. Facies Models:
Response to Sea Level Change: p. 454. https://doi.org/10.4116/jaqua.34.19.
Wang, D., Peng, J., Yu, Q., Chen, Y. and Yu, H. 2019. Support vector machine algorithm
for automatically identifying depositional microfacies using well logs. Sustainability
11(7): 1919. https://doi.org/10.3390/su11071919.
Wong, K.W., Fung, C.C., Ong, Y.S. and Gedeon, T.D. 2005. Reservoir characterization using
support vector machines. pp. 354–359. In: International Conference on Computational
Intelligence for Modelling, Control and Automation and International Conference on
Intelligent Agents, Web Technologies and Internet Commerce (CIMCA-IAWTIC’06),
November. IEEE. Vol. 2. 10.1109/CIMCA.2005.1631494.
260 Data Science and Machine Learning Applications in Subsurface Engineering

Wyllie, M.R.J. and Rose, W.D. 1950. Some theoretical considerations related to the quantitative
evaluation of the physical characteristics of reservoir rock from electrical log data. Journal
of Petroleum Technology 2(04): 105–118. https://doi.org/10.2118/950105-G.
Xie, Y., Zhu, C., Zhou, W., Li, Z., Liu, X. and Tu, M. 2018. Evaluation of machine learning
methods for formation lithology identification: A comparison of tuning processes and
model performances. Journal of Petroleum Science and Engineering 160: 182–193.
https://doi.org/10.1016/j.petrol.2017.10.028.
Xie, Y., Zhu, C., Hu, R. and Zhu, Z. 2021. A coarse-to-fine approach for intelligent logging
lithology identification with extremely randomised trees. Mathematical Geosciences
53: 859–876. https://doi.org/10.1007/s11004-020-09885-y.
Xie, Y., Zhu, C., Hu, R. and Zhu, Z. 2021. A coarse-to-fine approach for intelligent logging
lithology identification with extremely randomised trees. Mathematical Geosciences
53: 859–876. https://doi.org/10.1007/s11004-020-09885-y.
Yu, B. and Cheng, P. 2002. A fractal permeability model for bi-dispersed porous media.
International Journal of Heat and Mass Transfer 45(14): 2983–2993. https://doi.
org/10.1016/S0017-9310(02)00014-5.
Yu, X., Ma, Y.Z., Psaila, D. et al. 2011. Reservoir characterisation and modelling: A look back
to see the way forward. In: Ma, Y.Z. and P.R. La Pointe (eds.). Uncertainty Analysis and
Reservoir Modelling, AAPG Memoir. 96: 289–309. DOI:10.1306/13301421M963458.
Yu, X.H. 2008. Hydrocarbon Reservoir Sedimentology of Clastic Sandstone (2nd Edn.) (in
Chinese): Beijing, China: Petroleum Industry Press: p. 551.
Yu, Y., Lin, L., Zhai, C., Chen, H., Wang, Y., Li, Y. and Deng, X. 2019. Impacts of lithologic
characteristics and diagenesis on reservoir quality of the 4th member of the Upper
Triassic Xujiahe Formation tight gas sandstones in the western Sichuan Basin,
southwest China. Marine and Petroleum Geology 107: 1–19. https://doi.org/10.1016/j.
marpetgeo.2019.04.040.
Zabihi, R., Schaffie, M., Nezamabadi-Pour, H. and Ranjbar, M. 2011. Artificial neural network
for permeability damage prediction due to sulfate scaling. Journal of Petroleum Science
and Engineering 78(3–4): 575–581. https://doi.org/10.1016/j.petrol.2011.08.007.
Zendehboudi, S., Rezaei, N. and Lohi, A. 2018. Applications of hybrid models in chemical,
petroleum, and energy systems: A systematic review. Applied Energy 228: 2539–2566.
https://doi.org/10.1016/j.apenergy.2018.06.051.
Zerrouki, A.A., Aifa, T. and Baddari, K. 2014. Prediction of natural fracture porosity from well
log data by means of fuzzy ranking and an artificial neural network in Hassi Messaoud
oil field, Algeria. Journal of Petroleum Science and Engineering 115: 78–89. https://doi.
org/10.1016/j.petrol.2014.01.011.
Zheng, B. and Li, J.H. 2015. A new fractal permeability model for porous media based on
Kozeny-Carman equation. Natural Gas Geoscience 26(1): 193–198. https://doi.
org/10.1155/2022/8088151.
Zhong, Z. and Carr, T.R. 2019. Application of a new hybrid particle swarm optimization-mixed
kernels function-based support vector machine model for reservoir porosity prediction:
A case study in Jacksonburg-Stringtown oil field, West Virginia, USA. Interpretation
7(1): 97–112. https://doi.org/10.1190/INT-2018-0093.1.
 Chapter 11
Artificial Lift Design for Future
Inflow and Outflow Performance for
Jubilee Oilfield
Using Historical Production Data and
Artificial Neural Network Models
Solomon Adjei Marfo,1 Eric Thompson Brantson,2,*
Eric Mensah Amarfio,2 Abakah-Paintsil Efua Eduamba,2
Iyiola Zainab Ololade,2 Alexander Mensah Ofori,2
Ebenezer Ansah3 and Emmanuel Karikari Duodu2

1. Introduction
Forecasting production parameters such as production rate, oil recovery,
or reserves estimation is an essential aspect that enables operators to
determine the economic profitability of a petroleum venture. Inflow and
Tubing Performance Relationships (IPR and TPR) are the two mathematical
techniques used to analyse and predict the performance of a well. AL-Dogail
et al. (2018) used artificial intelligence (AI) (backpropagation network and
fuzzy logic) as another technique to predict the Inflow IPR of a gas reservoir
for effective reservoir management.

1
Department of Chemical and Petrochemical Engineering, GNPC School of Petroleum Studies,
University of Mines and Technology, Tarkwa, Ghana.
2
Department of Petroleum and Natural Gas Engineering, GNPC School of Petroleum Studies,
University of Mines and Technology, Tarkwa, Ghana.
3
Department of Petroleum Geosciences and Engineering, GNPC School of Petroleum Studies,
University of Mines and Technology, Tarkwa, Ghana.
* Corresponding author: [email protected]
262 Data Science and Machine Learning Applications in Subsurface Engineering

The operator of Jubilee Field, Ghana, recently updated its production

figures downward from 87,000 to 70,000 Barrel of Oil Per Day (BOPD)
due to certain technical issues such as increasing water cuts (Renpu, 2011),
reduced gas offtake by Atuabo Gas, and sand contamination of the flowlines
among a host of others (Agyeman, 2020). Initially, field assessments made
by operators of Jubilee Field, Tullow Oil, indicated a production peak rate
of 120,000 BOPD, water injection capacity of 230,000 Barrel of Water Per
Day (BWPD), and gas export and injection capacity of up to 160 MMscf/d
(Schempf, 2011). As a result of these technical problems, this production rate
has been cut down. These myriads of challenges can be solved by means of
implementing an artificial lift system (Boyun et al., 2007).
Artificial lift generally is a means of lowering the bottomhole flowing
pressure so that a well can produce at some desired rate. Traditionally, the
artificial lift had been referred to as downhole pumping or gas lift activity
associated with mature fields, where the average reservoir pressure has
declined such that the reservoir can no longer produce from its own natural
energy (Mohammed et al., 2016). About 90% of all producing wells require
some form of artificial lift to increase the flow of fluids from the wells when
the reservoir no longer has sufficient energy to naturally produce at economic
rates for enhanced financial performance (Flatern, 2015). When reservoir
energy is insufficient or water cut is higher despite sufficient energy, then
the artificial lift method could be used to maintain oil well production under
rational production pressure drawdown (Fleshman et al., 2011). Geographical
and environmental circumstances were considered the dominant factors in the
selection of artificial lifting and some other subordinated factors, including
reservoir pressure, productivity index, properties of reservoir fluid, and inflow
performance according to Neely et al. (1981). Some of these lift methods are
suited for high-rate wells whiles others are useful for low-rate wells. Several
factors are taken into consideration before an artificial lift method is selected
for a specific application (Battia and McAllister, 2014; Woods and Lea, 2017).
To address these challenges and enhance production, this study aims to
investigate the feasibility of screening and designing an optimum artificial lift
method to increase production and predict future IPR and TPR for the Jubilee
Field using historical production data. Also, this study investigated the use
of Radial Basis Function Neural Network (RBFNN) and Backpropagation
neural network (BPNN) models to predict Tubing Head Pressure (THP)
using historical data. A similar application study was carried out by Wu et al.
(2016) and Brantson et al. (2018) in forecasting tight gas carbonate reservoir
production profiles. Furthermore, the application of Artificial Neural Networks
(ANNs) in the oil and gas industry began in the 1980s, however, it has become
rampant over this decade and is now a common industry practice (Baudoin,
2016; Brantson et al., 2019a, 2019b; Chen et al., 2021).
 Artificial Lift Design for Future Inflow and Outflow Performance for Jubilee Oilfield 263

This paper is structured into the following sections. Section 2 dwells on

the method used to forecast future IPR and TPR as well as the ANN prediction
of THP. Section 3 states the results obtained from the implementation of the
IPR, TPR, and ANN. Section 4 summarises the major conclusion obtained
from the study.

2. Methodology
2.1 Artificial Lift Screening Techniques
To apply artificial lift techniques to a field, screening must be done to ensure
that the right artificial lift system is applied to the field in question. Some
factors to be considered during the screening process are location, depth,
estimated production, reservoir properties, and other factors. The screening
process is the initial procedure to evaluate the suitability of a certain artificial
lifting system (Brown, 1982). The goal is to phase out inappropriate systems
progressively, reducing the selection to a few contenders for the next selection
process. It is attempted to choose the optimum system, and to compare
design parameters with existing methodologies and charts (Lea and Nickens,
1999). In this study, field parameters were compared to the parameters on the
screening chart (Takacs, 2015) and the appropriate lift method was selected.

2.2 Inflow Performance Relationship Production Forecast

An inflow performance relationship is a means of evaluating a reservoir’s
deliverability in oil and gas production. An IPR curve is a graphical representation
of the relation between the flowing bottomhole pressure and liquid production
rate (Boyun et al., 2007). IPR curves are presented in a standardised manner,
with the flowing bottomhole pressure on the ordinate of the graph and the
corresponding production rate on the abscissa (Hill et al., 1993).
Equations (11.1) and (11.2) are based on Vogel’s model and they can
be used to generate a family of curves. This is done to improve the future
performance of the well, which can be utilised to assess the good performance
throughout the whole life of a well by the design engineer (Boyun et al., 2007).
 kro 
 
 Bo µo  f
*
J f
=
* (11.1)
J p  kro 
 
 Bo µo  p
264 Data Science and Machine Learning Applications in Subsurface Engineering

J *f p f   pwf  
2
pwf
q= 1 − 0.2 − 0.8   (11.2)
1.8  pf  p  
  f  

where,
p–f = reservoir pressure in a future time.
J *f = future productivity index
J *p = present productivity index
kro = relative permeability to oil
Bo = oil formation volume factor
µo = oil viscosity
pwf = bottomhole pressure
q = flowrate

2.3 Outflow Performance Relationship Production Forecast

Vertical lifting performance (VLP) also known as Outflow Performance
Relationship (OPR) is a description of bottomhole pressure as a function of
flow rate. Different correlations for VLP are provided for outflow performance
analysis.

2.4 PROSPER Procedure for Well Model Set-Up

PROSPER (Production and System Performance) was used to build a
separate model for each component of a well system that contributes to
overall performance and then enables each model subsystem to be verified by
performance matching. The program, therefore, guarantees the exactness of
the calculation. When the model system is modified to real data, PROSPER is
reliably employed to simulate different scenarios based on surface production
data and to anticipate the reservoir pressure.
PROSPER has five main sections. They are the Options Summary, PVT
Data, IPR Data, Equipment Data, and Analysis Summary. The following
steps were used to build the model, perform sensitivity analysis, and design
an optimum gas lift model.
Step 1: Options Summary
Well characteristics, fluid type, well completion, and desired lift method are
defined in this section as shown in Table 11.1. The Black Oil model is selected
since it is suitable for a variety of applications and hydrocarbon systems.
 Artificial Lift Design for Future Inflow and Outflow Performance for Jubilee Oilfield 265

Table 11.1. Options summary data.

Data Remarks
Options The options menu is used to define the characteristics of the well. Well
characteristics such as fluid type, well completion, and desired lift method are
defined here.
PVT Model The Black Oil model was selected. This model is suitable for usage in a wide
variety of applications and hydrocarbon fluid systems. A minimum of GOR
solution, oil viscosity and water formation salinity are required. Data from PVT
are imputed, and correlations have been chosen that best corresponds to the
location or oil type.
Fluid Oil and Water
Description
Temperature Rough approximation
Model
Flow Type Tubing flow
Well Type Producer
Prediction Pressure and temperature offshore model
Well Cased hole
Completion
Reservoir Single branch reservoir
Type
Artificial Gas lift
Lift Design

Step 2: PVT Data

Data such as Solution Gas Oil Ratio (GOR), water salinity, and impurities
present in the oil are imputed into the simulation program as shown in
Table 11.2. Data for matching actual field data to the simulation program is
also imputed in this menu. The bubble point and oil viscosity correlation were
selected. Glazo and Beal et al. correlations were the best correlation models
with the least standard deviation.

Table 11.2. PVT data.

Input Data Values

Solution GOR (scf/bbl) 1 243
Oil Gravity (°API) 36
Gas Gravity 0.878
Water Salinity (ppm) 100000
Mole percent H2S (%) 0
Mole percent CO2 (%) 0
Mole percent N2 (%) 0
266 Data Science and Machine Learning Applications in Subsurface Engineering

Step 3: Equipment Data

Under this section of the PROSPER software, well parameters such as well
deviation, surface equipment, downhole equipment, geothermal gradient, and
average heat capacities are keyed in.

2.4.1 Deviation Survey Data Input

The measured depth of the reservoir and the true vertical depth were entered.
Because all the wells under consideration are vertical, top, and bottom
perforations were used. The data used can be seen in Table 11.3.
Table 11.3. Deviation survey data.

Parameters Values
Datum at Christmas Tree (ft) 0
Measured Depth (ft) 12519
True Vertical Depth (ft) 12300

2.4.2 Surface Equipment Data Input

Information entered in this section includes the Christmas tree, manifold, or
choke. An ambient temperature of 60°F and an overall heat transfer coefficient
of 8 BTU/hr/ft2/F were keyed in. Data is found in Table 11.4.

Table 11.4. Surface equipment data.

Parameters Values
Manifold/Christmas Tree (True Vertical Depth) 0
Temperature of Surroundings (ft) 60 °F
Overall Heat Transfer 8 BTU/hr/ft2/F

2.4.3 Downhole Equipment Data Input

This section allows the user to input the tubing and casing inside and outside
diameters. Also, the roughness coefficient of casing and tubing can be imputed,
otherwise, the theoretical value of 0.0006 can be used if the roughness is
unknown. Table 11.5 shows downhole data.
Table 11.5. Downhole equipment data.

Equipment Measured Depth Tubing/Casing Inside Tubing/Casing Inside

(ft) Diameter (inches) Roughness (inches)
Tubing 12 300 4.778 0.0006
Casing 12 519 8.535 0.0006
 Artificial Lift Design for Future Inflow and Outflow Performance for Jubilee Oilfield 267

2.4.4 Average Heat Capacities Data Input

This section allows the user to input the average heat capacities for gas, oil,
and water. Table 11.6 shows the data imputed.
Table 11.6. Average heat capacities data.

Parameters Values
Cp oil 0.53 BTU/lb/F
Cp gas 0.51 BTU/lb/F

Step 4: Sensitivity Analysis

PROSPER helps users perform analysis by changing parameters such as GOR,
water cut, and reservoir pressure to know how changes in these parameters
will affect the well’s or reservoirs’ deliverability. For this work, water cut,
GOR, and reservoir pressure were varied.
Step 5: Vertical Lift Performance Correlations
Different correlations for VLP are provided. Test results showed that the
Petroleum Expert correlation was appropriate to the well circumstances and
therefore employed for VLP and sensitivity analysis. For surface equipment
horizontal piping, Beggs and Brill (1991) correlations were employed.
Step 6: Gas Lift Modelling
For this work, a continuous gas lift system was employed. The reason for
opting for the continuous gas lift was based on these criteria; wells in this
work are high productivity wells, lift gas is available for use and this warrants
the use of gas lift method and field under investigation produces under high
GOR conditions. PROSPER simulation software was used in the modelling
of the gas lift for all the wells under investigation. A gas injection rate was
10 MMscf/day was used. The casing pressure imputed was at 1,500 psi with a
kickoff pressure of about 1,900 psi. The pressure gradient of the kill fluid was
set to 0.465 psi/ft. All these datasets are shown in Table 11.7.

2.5 Artificial Neural Networks

ANN is a group of inputs and outputs of connected units where each
connection has a weight associated with its computer program (Pennel et al.,
2018; Akwensi et al., 2021; Brantson et al., 2022). This network helps build
predictive models from large databases. The model is built on the concept
of the human nervous system. For this work, two types of ANN (BPNN and
RBFNN) models will be considered.
268 Data Science and Machine Learning Applications in Subsurface Engineering

Table 11.7. Gas lift data.

Parameters Values
Minimum valve spacing 250 ft
Kill fluid gradient 0.465 ft/psi
Operating Injection Pressure 1300 psi
Kick off pressure 1900 psi
Gas lift valve type R-20 Monel
Port Size: 32
R-value: 0.25
Differential pressure dP across valve 250 psi
Vertical Lift Performance correlation Petroleum Experts 2
Surface Equipment correlation Beggs and Brill
Well Depth 12300 ft
Water Cut 14%
Maximum Liquid rate 12472 STB/D
Maximum Gas Available 10 MMScf/day

2.5.1 Back Propagation Neural Network

BPNN is a type of ANN model that involves fine-tuning the weights of the
neural network based on the error rate obtained in a previous iteration. Proper
tuning of the weights allows one to reduce the error rate which makes the
model more accurate thereby increasing the generalisation of the model.

2.5.2 Radial Basis Function Neural Network

RBFNN is a type of neural network which approximates multivariable
functions or multivariate by a linear combination of terms so that it can be
used in more than one dimension. It is applied to approximate functions or
data which is known at a finite number of points so that evaluations of the
approximate function can take place often and efficiently.

2.5.3 ANN Procedure

For this work, historical production data from the Jubilee field which is the
production data from the wells was grouped into training and testing data.
MATLAB was the software used in building the ANN models. The inputs for
the prediction were days, gas rate, cumulative gas produced, cumulative oil
production, and flowing bottomhole pressure. The output for the model was
the THP. To test the accuracy of the model, some statistical parameters were
measured as the MAPE and correlational coefficient (R). A value of R closer
to 1 means good accuracy while a value closer to 0 for MAPE indicates the
 Artificial Lift Design for Future Inflow and Outflow Performance for Jubilee Oilfield 269

Fig. 11.1. ANN procedure.

least amount of error. Computational time was also calculated to know which
model works faster. Figure 11.1 shows the process followed to obtain results
from the ANN models.

3. Results and Discussion

3.1 Production and Well Data of the Study Area
The Jubilee Field is located in Western Region, 60 km offshore, between
the Deepwater Tano and West Cape Three Points blocks, off the coast of
Ghana. Figure 11.2 shows the Jubilee concession and payout, respectively.
The field lies within the tropical climatic zone between latitude 4.492°N and
longitude 2.916°W. The Jubilee Field appraisal and development programme
began at the end of 2008 with the drilling of the Odum, Mahogany, and
Heydua wells. The field’s recoverable reserves are estimated to be more than
370 million barrels, with an upside potential of 1.8 billion barrels. It is located
at a water depth of 1100 m. The Jubilee Field was developed through an FPSO
(floating production storage and offloading system) with a target plateau oil
rate of 120,000 BOPD, water injection capacity of 230,000 BWPD, and gas
export and injection capacity of up to 160 MMscf/d (Schempf, 2011). Subsea
gathering receptacles are located below the riser base to receive production
from wells for onward transportation to an FPSO for processing. Compressors
are installed on the FPSO to process produced gas for delivery to an onshore
gas processing facility. Part of this processed gas is injected at the riser base
for intermittent subsea boosting.
The obtained production data used consisted of daily gas production
(MMSCF/D), BOPD, cumulative oil production (MMSTB), Tubing Head
Pressure (psig), Flowing Bottom Hole Pressure (psia), and number of days
of production.
 gas export and injection capacity of up to 160 MMscf/d (Schempf, 2011). Subsea gathering receptacles
are located below the riser base to receive production from wells for onward transportation to an FPSO
for processing. Compressors are installed on the FPSO to process produced gas for delivery to an onshore
gas processing facility. Part of this processed gas is injected at the riser base for intermittent subsea
270 Data Science and Machine Learning Applications in Subsurface Engineering
boosting.

Fig. 11.2. Jubilee field concession.

Fig. 11.2: Jubilee Field Concession.
Two random wells were chosen for evaluation. Obtained data conforms
to American Petroleum Institute Recommended Practice API RP 11V6 for gas
The obtained production data used consisted of daily gas production (MMSCF/D), BOPD, cumulative oil
lift design criteria (Agyemang, 2020). Well data obtained constitutes Pressure
production (MMSTB), Tubing
Volume and Temperature Head Pressure
(PVT) data, (psig),
well Flowing
inflowBottom Hole Pressure data,
characteristics (psia), and number
fluid
of days of production.
properties data, wellbore geometry, surface facilities, and fluid properties
data. For this work, the two randomly selected wells are named A and B. All
the wells are vertical wells. Tables 11.8–11.11 are the well data obtained from
Agyemang (2020).

Table 11.8. Production system data.

Parameters Well A Well B

Production Interval:
Top: (ft) 12300 11559
Bottom: (ft) 12519 11724
Casing
Inside Diameter: in Weight: Ib/ft 8.535 8.535
53.5 53.5
Tubing OD: in 5.5 5.5
Grade: ksi13Cr 80 80
Tubing ID: in 4.778 4.778
Burst Pressure: psi 9190 9190
Weight (Ib/ft) 20 20
PBTD: (ft) 12850 11832

(Source: Agyemang, 2020)

 Artificial Lift Design for Future Inflow and Outflow Performance for Jubilee Oilfield 271

Table 11.9. Well surface data.

Parameters Well A Well B

Separator/Well Head Pressure (Psi) 441 441
Temperature (°F) 110 110
Primary Power (V) 415 415
Frequency (Hz) 60 60
Oil specific gravity (°API) 36 36
Water specific gravity 1.03 1.03
Scaling nil nil
Gas specific Gravity 0.878 0.878
CO2 content nil nil
H2S content “ “
Paraffin “ “
Asphaltenes “ “
Sand production “ “
Water cut (%) 14 20
GOR (scf/bbl) 1243 1243
Bubble point pressure: (psia) 4419 4419

Table 11.10. PVT and viscosity data.

Well A B
Reservoir Pressure (psia) 6 014.7 6 014.7
Temperature (°F) 210 210
Oil Formation Volume factor (Bo) (bbl/stb) 1.595 1.595
Solution Gas Oil Ratio (GOR) (scf/stb) 1243 1 43
Oil Viscosity µo (cP) 0.7 0.7

Table 11.11. Well inflow characteristics data.

Well Test (production) A B

Static Pressure, psig 5516 4872
Test Rate (bbl/day) 14008 3026
Test Pressure, psig 3701 4635

3.1.1 Base Case Flow Rates

The results below graphically display the calculated average daily flow rates
of Well A and B. Table 11.12 shows the results from the calculation. This will
serve as the basis for checking if there has been an improvement in wellbore
deliverability after implementing the continuous gas lift method. Figures 11.3
and 11.4 show the daily production profile for well A and B with average base
case flow rates of 8,161 BOPD and 9,850 BOPD, respectively.
 272 Data Science and Machine Learning Applications in Subsurface Engineering

15
Table 11.12. Screening results.

Field Parameters
shows the results from the calculation. This will serve as theValues
basis for checking if there has been an
Depth after implementing the
improvement in wellbore deliverability 3657.6 m
continuous gas lift method. Fig. 11.3 and
Production
11.4 show the daily production profile Volume
for well 24average
A and B with 000 bbl/day
base case flow rates of 8,161 BOPD
and 9,850 BOPD, Temperature 98.89–106.11°C 15

Oil Gravity 36°API

respectively.
shows the results from the calculation. This will serve as the basis for checking if there has been an
Prime Mover Available Compressors
improvement in wellbore deliverability after implementing the continuous gas lift method. Fig. 11.3 and
Location Offshore
11.4 show the daily production profile for well A and B with average base case flow rates of 8,161 BOPD
and 9,850 BOPD,
respectively.

Fig. 11.3: Well A Daily Production Profile.

Fig. 11.3. Well A daily production profile.

Fig. 11.3: Well A Daily Production Profile.

Fig. 11.4. Well B daily production profile.

Fig. 11.4: Well B Daily Production Profile.
3.2 Artificial Lift Screening
Fig. 11.4: Well B Daily Production Profile.

Table 11.12 shows the field parameters compared to the parameters on the
chart by Takacs (2015) and the gas lift method was selected which satisfies
the chart. The following are the results obtained during screening for an
appropriate artificial lift method for the field under investigation in this work.
 Artificial Lift Design for Future Inflow and Outflow Performance for Jubilee Oilfield 273

Since both wells are high productivity index wells, a screening chart for
high productivity wells was used. Sucker rod and progressive cavity pumps
were not applicable in this field due to the depth restriction of not more than
3,050 m. Electrical Submersible Pump (ESP) and Gas Lift were the next in
line for the screening process. At the end of the screening process, the gas lift
was selected based on the following criteria. Gas lift can handle fields that
have wells with a fluid gravity greater than 15° API. Also, the temperature
for all the wells was in the acceptable range for gas lift 98.89–106.11°C. The
required operating volumes for the wells were within an acceptable range
for gas lift use. The required operating volumes for the gas lift range is
200–30,000 BOPD. Based on screening by advantages, gas lift is an excellent
choice for offshore applications due to its high efficiency rate when used
offshore. In addition to that, compressors are already available onsite for use.
For the above reasons, the gas lift was chosen as the artificial lift technique
for the wells under investigation for this research. Table 11.12 shows the field
data used for the screening.

3.3 PROSPER Simulation Results

3.3.1 IPR Curves
PROSPER simulation software was used to generate IPR family curves for
each well. PROSPER software generates the curves as well as calculates the
Absolute Open Flow (AOF) potential. Figures 11.5 and 11.7 show the current
IPR curves for well A and B, respectively. The intercept on the ordinate axis

Fig. 11.5. Present IPR curve Well A.

274 Data Science and Machine Learning Applications in Subsurface Engineering

Fig. 11.6. Future IPR curve Well A.

Fig. 11.7. Present IPR curve Well B.

indicates the shut-in reservoir pressure whereas the intercept on the abscissa
indicates the flowrate in stock tank barrels per day. The AOF for Well A is
28,486.5 STB/D and that of Well B is 37,826.4 STB/D. Future IPR curves
generated through sensitivity analysis can be seen in Fig. 11.6 and 11.8 for
well A and B, respectively.
 Artificial Lift Design for Future Inflow and Outflow Performance for Jubilee Oilfield 275

Fig. 11.8. Future IPR curve Well B.

3.3.2 Vertical Lift Performance Correlations

Figure 11.9 shows a vertical lift performance (VLP) correlation plot. For this
work, Dons and Ros modified the VLP and Petroleum Experts’ correlations
were applied. The purpose of performing this plot is to ascertain which
correlation best fits the wells under investigation. These results were obtained
by plotting a test point on a depth-pressure transverse curve as indicated as the
blue point at the bottom of Fig. 11.9.

Fig. 11.9. Shows a vertical lift performance correlation plot.

276 Data Science and Machine Learning Applications in Subsurface Engineering

3.3.3 Desired Flow Rates

The desired rate is the rate to achieve optimum production based on technical
and economic factors. This is usually a choice of management. In this work,
technical consideration was the basis to determine the desired flow rate. Desired
flow rate is required for gas lift. A sensitivity plot was obtained by varying
some parameters such as reservoir pressure, water cut and total gas-oil ratio
for each well. Figure 11.10 shows the sensitivity plot for Well A displaying
the test rate of 14,008 BOPD and pressure at 3,071 psig. The desired flow

Fig. 11.10. TPR vs IPR Well A.

Fig. 11.11. TPR vs IPR Well B.

 Artificial Lift Design for Future Inflow and Outflow Performance for Jubilee Oilfield 277

rate which is 17,000 BOPD (Fig. 11.10) still falls outside the intersection
indicating that the rate of production for that well must be changed since
that value is not optimum. It can be observed that the test point falls within
where the outflow performance curve and inflow performance curve intersect
meaning the desired test rate is feasible. It can be deduced that the well will
cease to flow and become a dead well if the bottom hole pressure falls below
1,400 psi at the prevailing reservoir conditions since the inflow performance
curves and the tubing performance curves cease to intersect. This observation
was made for Well A. The sensitivity plot obtained for Well B indicates a
maximum objective flowrate of 22,000 BOPD and a test rate of 3,026 STB/D
at a pressure of 4,872 psi. From the graph in Fig. 11.11, it can be deduced that
below 2,472 psi, Well B becomes a dead well since the inflow performance
curves and the tubing performance curves cease to intersect.

3.4 Gas Lift Results

The results were obtained for the various gas lift modelling for Wells A and
B and their respective gas sensitivity plots. The sensitivity plots show the
optimum light gas injection rate as well as the optimum production rate for
each well. For the parameters imputed into the simulation program, below are
the results obtained for each of the wells. Figures 11.12 and 11.13 show the
gas lift design for Wells A and B which were designed with four valves using
the spacing line procedure. Figures 11.14 and 11.15 show the gas sensitivity
plot for Wells A and B for the maximum economic gas to be injected.

Fig. 11.12. Gas lift design Well A.

278 Data Science and Machine Learning Applications in Subsurface Engineering

Fig. 11.13. Gas lift design Well B.

Fig. 11.14. Well A gas sensitivity.

3.4.1 Optimum Production Rates

Upon Simulation in PROSPER, the following production rates in Table 11.13
were obtained. Comparing the average daily base case flowrates calculated
using the production history of the field to the results obtained below after
applying gas lift modelling, there has been an increase in the deliverability of
both wells. There was an 83.63% increase for Well A and a 61.64% increase
for Well B.
 Artificial Lift Design for Future Inflow and Outflow Performance for Jubilee Oilfield 279

Fig. 11.15. Well B gas sensitivity.

Table 11.13. Gas lift results for Well A and B (Optimum Production Rates).

Well Liquid Rate Oil Rate Injected Rate Injection Pressure

(STB/D) (STB/D) (MMSCF) (Psig)
A 17432.2 14991.7 5.47332 1400
B 19903.1 15922.5 6.97933 1500

3.5 ANN Results

3.5.1 ANN Architecture
Table 11.14 shows the statistics of the historical production data input and
output for both wells used for the ANN models to predict THP. However,
production data of three wells were used to build the database for training the
model while one well data was used for testing the model.

Table 11.14. Input and output data statistics.

Parameters Minimum Maximum Average Standard Data

Deviation Type
Days 0 2455 1009.04 660.53 Input
Daily Oil Production (MMSTB/D) 0 27645 11308.43 5023.96 Input
Cumulative Oil Production 0 26.06 10.44 6.68 Input
(MMSTB)
Cumulative Gas Production 0 32.73 8.04 9.41 Input
(MMSCF)
Gas Rate (MMSCF/D) 0 23.66 5.94 6.51 Input
Flowing Bottomhole Pressure (PSIA) 2489.96 6212.41 4423.02 508.60 Input
Tubing Head Pressure (PSIG) 1095.03 4185.29 2564.75 430.48 Output
280 Data Science and Machine Learning Applications in Subsurface Engineering

3.5.2 Model Visualisation

Figure 11.16 shows the training plot whereas Fig. 11.17 shows the testing plot
obtained for the BPNN model. The results indicate that the prediction points
are closer to the ideal line with an architecture of five inputs, one neuron,
and one output. BPNN shows how best the prediction results are close to the
actual. It can also be observed that the testing well dataset generalises very
well on the training model.
Figure 11.18 shows the training plot whereas Fig. 11.19 shows the testing
plot obtained for the RBFNN model. The testing RBFNN shows how best
the prediction results are close to the actual with an architecture of six inputs,
17 neurons, and one output. The testing RBFNN shows how best the prediction
results are close to the actual. Also, it can also be observed that the testing well
dataset generalises very well on the training model.
RBFNN is a powerful AI technique used in well performance prediction,
and it demonstrated better prediction accuracy than the BPNN model in this
study. However, RBFNN has some limitations that could impact the accuracy
of the results. One of the primary limitations of RBFNN is its sensitivity to the
number and placement of basis functions, which are critical to the accuracy
of the model. The selection of the appropriate number and location of basis
functions can be challenging, and an improper choice can lead to overfitting
or underfitting of the data. Additionally, the choice of the radial basis function
itself could impact the accuracy of the model. The potential impact of these
limitations on the results of the study is that the accuracy of the RBFNN model

Fig. 11.16. Training plot for BPNN model.

 Artificial Lift Design for Future Inflow and Outflow Performance for Jubilee Oilfield 281

29

of the data used in the model training. Therefore, it is essential to carefully evaluate the quality of the data
used and the choice of basis functions to ensure the accuracy of the RBFNN model predictions.
Additionally, further research can explore the use of other AI techniques to complement RBFNN and
improve the accuracy of the predictions
Fig. 11.17. Testing plot for BPNN model.

Fig. 11.18. Training plot for RBFNN.

Fig. 11.18: Training plot for RBFNN.
could be influenced by the choice of basis functions and the quality of the
data used in the model training. Therefore, it is essential to carefully evaluate
the quality of the data used and the choice of basis functions to ensure the
accuracy of the RBFNN model predictions. Additionally, further research can
explore the use of other AI techniques to complement RBFNN and improve
the accuracy of the predictions
Table 11.15 shows the results compilation for the ANN with the accuracy
of each model and the computational time. RBFNN was the best model when
tested with unseen data while BPNN obtained the fastest computational time
due to the one neuron used than the 17 neurons by RBFNN.
 Fig. 11.18: Training plot for RBFNN.
282 Data Science and Machine Learning Applications in Subsurface Engineering

Fig. 11.19. Testing plot for RBFNN.

Fig. 11.1: Testing plot for RBFNN.
3.6 Discussion
The gas lift models developed in this study using PROSPER simulation
software have significant implications for the accuracy of well performance
predictions. The application of the gas lift model for both wells helps to
optimise the performance of the artificial lift system, resulting in improved
well productivity and efficiency. There was an 83.63% increase for Well A
and a 61.64% increase for Well B. By using historical production data, the
gas lift model can provide accurate predictions of future well performance
under different operating scenarios. This is important for optimising
production schedules, as well as for identifying potential production issues
and troubleshooting them in advance. The gas lift models also enable the
prediction of downhole and surface parameters, such as tubing head pressure
and gas injection rate, which are critical in determining the effectiveness of the
gas lift system. The development of an optimised gas lift model has significant
implications for the oil and gas industry. It can lead to more efficient and
cost-effective production, improved asset management, and increased profits.
Furthermore, the methodology developed in this study can be applied to other
oil fields with similar production characteristics, providing a standardised
approach to optimising artificial lift systems. Overall, this research highlights
the potential benefits of integrating AI techniques into gas lift modelling for
accurate well performance predictions and optimisation. While the findings
of this study are specific to the Jubilee Field, the methodology developed
can be applied to other oil fields with similar production characteristics. The
recommendation of implementing artificial lift at the start of production for
nearby fields, such as the Pecan Field, is based on the observed technical
problems faced in the Jubilee Field and the potential benefits of early
 Artificial Lift Design for Future Inflow and Outflow Performance for Jubilee Oilfield 283

Table 11.15. Statistical measures used for ANN models’ assessment.

MODELS (ANN) BPNN RBFNN

Training 0.96178 0.96328
R-Correlational Coefficient
Mean Absolute Percentage Error (MAPE) 0.01651 0.02055
Testing 0.96118 0.96290
R-Correlational Coefficient
MAPE 0.02764 0.02134
Computational Time (seconds) 6.83160 10.7950

intervention. However, the applicability of this recommendation to other fields

should be evaluated based on their unique characteristics and conditions.

4. Conclusions
In summary, this research has contributed to the development of an optimised
artificial lift system for the Jubilee Field, resulting in increased deliverability
of all wells. Additionally, the application of AI techniques, such as BPNN
and RBFNN, have demonstrated improved prediction accuracy for well
performance. These findings have important implications for the optimisation
of oil field production, and the methodology developed in this study can be
applied to other fields to improve efficiency and productivity. The following
conclusions can be drawn from the study.
• Successfully screen and design an optimum artificial lift system (gas lift),
for the field under investigation.
• Comparing the average daily base flow rates calculated using the
production history of the field to the results obtained after applying gas
lift modelling with an increase in the deliverability of all the wells. There
was an 83.63% increase for Well A and a 61.64% increase for Well B.
• RBFNN gave the best prediction for the tubing head pressure in the
Jubilee Field than the BPNN model used with an R of 96.290% and
MAPE of 0.02764.
• BPNN had less computational time due to the number of neurons used as
compared to the RBFNN model.
• Artificial lift method is recommended at the start of production for the
field nearby of Jubilee Field which is the Pecan Field operated by Aker
Energy to avoid technical problems faced in the Jubilee Field.
284 Data Science and Machine Learning Applications in Subsurface Engineering

Acknowledgment
We would like to thank anonymous reviewers for their contribution to this
research. We are also grateful to Ghana National Petroleum Corporation for
providing access to data for this research to be successful. We also acknowledge
PETEX for making the educational licence of the software available. Finally,
we thank the University of Mines and Technology, Tarkwa, Ghana for their
immense support.

References
Agyeman, B.K. 2020. Design of Gas Lift and Electric Submersible Pump: A Case Study at
Jubilee Field. Unpublished BSc Project Report, University of Mines and Technology,
Tarkwa, pp. 1–118.
Akwensi, P.H., Brantson, E.T., Niipele, J.N. and Ziggah, Y.Y. 2021. Performance evaluation
of artificial neural networks for natural terrain classification. Applied Geomatics
13(3): 453–465.
AL-Dogail, A.S., Baarimah, S.O. and Basfar, S.A. 2018. Prediction of inflow performance
relationship of a gas field using artificial intelligence techniques. In: SPE Kingdom of Saudi
Arabia Annual Technical Symposium and Exhibition. April. OnePetro.
Baudoin, C.R. 2016. Deploying the industrial Internet in oil & gas: Challenges and opportunities.
In: SPEntelligent Energy International Conference and Exhibition. OnePetro.
Bhatia, A. and McAllister, S. 2014 Artificial Lift: Focus on Hydraulic Submersible Pumps,
ClydeUnion Pumps, Tech 101(10): 3, 29–31.
Boyun, G., William, C.L. and Ali, G. 2007. Petroleum Production Engineering, A Computer
Assisted Approach. Oxford, UK: Elsevier Ltd., pp. 186–300.
Brantson, E.T., Ju, B., Omisore, B.O., Wu, D., Aphu, E.S. and Liu, N. 2018. Development of
machine learning predictive models for history matching tight gas carbonate reservoir
production profiles. Journal of Geophysics and Engineering 15(5): 22–35.
Brantson, E.T., Ju, B., Opoku Appau, P., Akwensi, H.P., Agyare Peprah G., Liu, N., Aphu, E.S.,
Annan Boah, E. and Aidoo Borsah, A. 2019a. Development of low salinity water polymer
flooding numerical reservoir simulator and smart proxy modeling for hybrid chemical
enhanced oil recovery (CEOR). Journal of Petroleum Science and Engineering 187: 1–23.
Brantson, E.T., Ju, B., Ziggah, Y.Y., Akwensi, P.H., Sun, Y., Wu, D. and Addo, B.J. 2019b.
Forecasting of horizontal gas well production decline in unconventional reservoirs using
productivity, soft computing and swarm intelligence models. Natural Resources Research
28(3): 717–756.
Brantson, E.T., Osei, H., Aidoo, M.S.K., Appau, P.O., Issaka, F.N., Liu, N., Ejeh, C.J. and
Kouamelan, K.S. 2022. Coconut oil and fermented palm wine biodiesel production for oil
spill cleanup: Experimental, numerical, and hybrid metaheuristic modelling approaches.
Environmental Science and Pollution Research, 1–19.
Brill, J.P. and Beggs, H.D. 1991. Two-Phase Flow in Pipes, Houston, 6th Edition, 1991,
pp. 1–15.
Brown, K.E. 1982. Overview of artificial lift systems. Journal of Petroleum Technology, SPE
9979-PA, 13 pp.
Chen, X., Wang, D.Y., Tang, J.B., Ma, W.C. and Liu, Y. 2021. Geotechnical stability analysis
considering strain softening using micro-polar continuum finite element method. Journal
of Central South University 28(1): 297–310.
 Artificial Lift Design for Future Inflow and Outflow Performance for Jubilee Oilfield 285

Flatern, R.V. 2015. Oilfield Review, Artificial Lift Systems, www.slb.com, Accessed November
2019.
Fleshman, R. and Lekic, H.O. 1999. Artificial lift for high production. Oilfield Review Spring
49–63.
Hill, T. and Remus, W. 1993. Neural network models for intelligent support of managerial
decision making. Hawaii University 11(5): 449–459.
Lea, J.F. and Nickens, H.V. 1999. Selection of Artificial Lift. SPE 52157, Oklahoma City,
Oklahoma 30 pp.
Mohammed, A.G.H. and Nasr, G.G. 2016. Gas lift optimisation to improve well performance.,
World Academy of Science, Engineering and Technology, International Journal of
Mechanical and Mechatronics Engineering 10(3): 512–520.
Neely, B., Gipson, F., Capps, B., Clegg, J. and Wilson, P. 1981. Selection of Artificial Lift
Method. SPE 10337 Dallas, Texas. 1 p.
Pennel, M., Hsiung, J. and Putcha, V.B. 2018. Detecting failures and optimising performance in
artificial lift using machine learning models. In: SPE Western Regional Meeting. OnePetro.
Renpu, W. 2011. Selection and Determination of Tubing and Production Casing Sizes. Advanced
Well Completions Engineering, 3rd Edn.. Houston,: Gulf Professional Publishing,
pp. 117–170.
Schempf, F.J. 2011. Jubilee brings Ghana to the fore among West African deepwater regions.
Offshore (Conroe, Tex.) 71(4).
Takacs, G. 2015. Sucker-rod Pumping Handbook: Production Engineering Fundamentals and
Long-stroke Rod Pumping. Houston: Gulf Professional Publishing.
Woods, J.D. and Lea, J.F. 2017. What Is New In Artificial Lift? World Oil Production Technology:
Artificial Lift, USA: Gulf Publishing, pp. 43–50.
Wu, D., Ju, B. and Brantson, E.T. 2016. Investigation of productivity decline in tight gas wells
due to formation damage and Non-Darcy effect: Laboratory, mathematical modelling and
application. Journal of Natural Gas Science and Engineering 34: 779–791.
 Chapter 12
Modelling Two-phase Flow
Parameters Utilizing
Machine-learning Methodology
Longtong Dafyak* and Buddhika Hewakandamby

1. Introduction
Two-phase flow is predominant in natural occurring and industrial processes;
from water droplets entrained in the air to domestic water distribution
lines, combustion engines, power generation plants, amongst others. In the
exploration industry, gas-liquid flow in pipes is a standard practice due to
the coexistence of these fluids in the subsurface (McCain, 1994; Soloveichik
et al., 2022). Furthermore, pipelines are still the safest and most economical
means of transporting fluids over long distances (Green and Jackson, 2015;
Canada Energy Report, 2020).
When gases and liquids flow simultaneously in pipes, unique flow
configurations are formed, which are referred to as flow patterns. These flow
patterns were classified as bubbly, dispersed bubbly, plug, slug, stratified, and
annular flows by Hewitt and Roberts (1969) and Mandhane et al. (1974) as
illustrated by Shoham (2006) in Fig. 12.1. The flow patterns formed depend on
the operating parameters, pipe geometric variables, and physical properties of
the fluids (Lu, 2015). Each flow pattern is associated with specific interaction
between the phases and the pipe walls, thus making one flow pattern entirely
different from another (Nie et al., 2022). The difference in momentum, phase
distribution, and velocity distribution, makes gas-liquid flow dynamics a lot
more complex than single-phase flow. However, some flow patterns share
similar intermittent behaviour and characteristic parameters.
University of Nottingham, UK.
* Corresponding author: [email protected]
 Modelling Two-phase Flow Parameters Utilizing Machine-learning Methodology 287

Fig. 12.1. Flow patterns in (a) horizontal and (b) vertical pipes (Shoham, 2006).

The wide application of gas-liquid flow and the curiosity of gas-liquid

enthusiasts has inspired a plethora of studies in this field leading to the
development of several gas-liquid models. Pioneering researchers (Nicklin,
1962; Dukler and Hubbard, 1975; Fernandes et al., 1983; Alves and
Shoham, 1991) proposed physics-based models to describe two-phase flow
dynamics; however, these models are either highly simplified or require high
computation time to derive a solution. For these reasons, several empirical
and semi-empirical models have been developed over the years to predict
two-phase flow parameters. These correlations predict the mean void
fraction and velocities for most flow patterns, the frequency, void fraction,
and length of flow structures for intermittent flow patterns amongst other
two-phase flow parameters. There are primarily two limitations of these
empirical/semi-empirical correlations; one of which is the constraint of its
applicability to flow conditions from which it is developed. Second, some
of these correlations have proven to be applicable over a wide range of
two-phase flow conditions; however, the simplistic curve fitting approach used
in developing most empirical correlations are often inadequate for optimally
modelling the complex relationships between the two-phase flow variables.
As Machine Learning (ML) application develops and gains popularity,
researchers have explored ML techniques in predicting two-phase flow
patterns and parameters. Flow pattern models developed by Cai et al. (1994),
Rosa et al. (2010), Al-Naser et al. (2016), Mask et al. (2019), Liu et al. (2020),
and dos Santos Ambrosio et al. (2022) using support vector machine, artificial
neural networks (ANNs) and tree-based algorithms show promising results
with an accuracy of approximately 96–100%. Yan et al. (2018) reviewed
soft computing techniques, ML, evolutionary computation, fuzzy logic, and
probabilistic reasoning, and proposed the trajectory of future development in
288 Data Science and Machine Learning Applications in Subsurface Engineering

multiphase flow research. Azizi et al. (2016) utilized ANN to predict liquid
fraction using 468 experimental data points for oil-water mixtures. Kim et al.
(2020) and Abdul-Majeed et al. (2022) proposed ML models to specifically
model slug flow parameters. More recently, the ANN model proposed by
Aliyu et al. (2023) for entrained liquid fraction in annular gas-liquid flows
show good performance with a Root Mean Squared Error (RMSE) and R2 of
0.005 and 0.97, respectively. These studies also compared existing correlations
and demonstrated the potential of ANN, random forest, and support vector
regression to outperform traditional two-phase flow correlations. Despite
the contributions of these studies to data-driven modelling, studies focused
on comparing the capabilities of varying ML algorithms in two-phase flow
modelling is sparse. Thus, this study is focused on evaluating the performance
of several ML algorithms in two-phase flow predictions and assessing the
applicability of these ML models in comparison to traditional empirical
correlations.
In this study, the interest is limited to the mean void fraction and structure
velocity of intermittent flow regimes. This is because intermittent flows
(slug, plug/elongated bubble, and churn) are the most persistent and often the
extremely challenging flow patterns in two-phase flow systems. These two
parameters are also generic characteristics of most two-phase flow patterns;
thus, the developed models are applicable across a wide breadth of two-phase
flow conditions. The two-phase flow variables considered in this study are
the operating conditions (the gas and liquid superficial velocities, the pipe
diameter and inclination) and the fluid properties (the gas and liquid viscosity,
the gas and liquid density).

2. Data Sources and Existing Correlations

The data utilized in this study was acquired from experiments conducted by
Escrig (2017) and Dafyak (2022) in the gas-liquid flow facility at University
of Nottingham. The experimental procedure is elaborately described in the
studies by Escrig et al. (2017) and Dafyak et al. (2021). Table 12.1 details the
data sources and the corresponding operating variables for the data utilized in
this study for mean void fraction and structure velocity.
Table 12.1. Data sources and summary of operating variables.

Author Number Fluid D θ USL USG ρL μL

(year) of Data System
Points
(mm) (°) (m/s) (m/s) (kg/m3) (cp)
Dafyak 985 Air-oil 67 0–90 0.03–2.82 0.08–0.3 928 93–230
(2022)
Escrig (2017) 493 Air-oil 67 0–90 0.06–2.92 0.02–0.47 917 5
 Modelling Two-phase Flow Parameters Utilizing Machine-learning Methodology 289

Both experiments were conducted in a 67 mm pipe for air and silicone

oil mixtures with a similar range of pipe inclination, gas and liquid velocities;
however, the fluid properties are different. A total of 1,478 experimental
runs were conducted at varying pipe inclination, superficial gas and liquid
velocities, and liquid viscosities. The data set covers upward inclined flows
between the horizontal and vertical axis (0°–90°) for low to medium liquid
viscosities; Escrig (2017) explored 7 pipe inclinations whereas Dafyak (2022)
covered 10 pipe inclinations within this range.
Table 12.2 shows a collation of some of the most widely used correlations
for predicting the mean void fraction and structure velocity. These empirical
correlations developed on the basics of slip-ratio and drift-flux model
are semi-empirical and have proven to be robust and easily scalable,
showing good performance over a wide range of fluid and flow parameters
(Woldesemayat and Ghajar, 2007; Godbole et al., 2011). For this reason, the
existing correlations considered in this study are limited to slip-ratio and
drift-flux correlations for both the mean void fraction and the structure
velocity. The performance of these correlations is evaluated using the test
data for this study in Section 5.

3. Methodology
The sequential procedure for data processing and model development in
this study are data acquisition and sourcing, data pre-processing, model
development, hyperparameter tuning and model evaluation, and validation.
All analysis was carried out in the Python environment using open-source
libraries; Pandas, Numpy, Matplotlib, and Seaborn for data manipulation and
visualization and Keras and Sklearn for predictive modelling.
The experimental data for this study was sourced from previous studies
as elaborated in Section 2. Data pre-processing and feature engineering
encompasses all the steps taken to prepare the data for optimal model
development and evaluation. The input parameters referred to as features and
output parameters referred to as target values in this study were identified
(Fig. 12.2). Correlation analysis is utilized to identify the magnitude of
the relationship between input variables and target variables. This helps in
selecting the most relevant variables for model development, eliminate
irrelevant variables to reduce noise, avoid multicollinearity, and improve the
overall performance of the model. Prior to model development, the data was
split into training and test data for model development and model evaluation,
respectively. The same number of data were used to train and test all the
ML algorithms considered in this study. The supervised learning algorithms
utilized in this study are Multiple Linear Regression (MLR), Polynomial
Regression (PR), Random Forest (RF), Support Vector Regression (SVR),
 290 Data Science and Machine Learning Applications in Subsurface Engineering

Table 12.2. Existing empirical correlations.

Authors Correlations
Mean void fraction
Lockhart, R.W., Martinelli −1
  1− x   ρG   µ L  
0.64 0.36 0.07
(1949) ε= 1 + 0.28       
  x   ρ L   µG  

Bendiksen (1984) U SG
ε=
1.2U m + 0.35sin θ gd + 0.54 cos θ gd

Rouhani and Axelsson −1

(1970)   x 1 − x  U GM 
ε x
=
ρG Co  ρG + ρ L  + G 
 

 1.18 
Where U GM 
=  ( gσ ( ρ L − ρG ))0.25 for P < 12.7
 ρ 
 L 

Premoli et al. (1970) −1

  1− x   ρG  
a b

ε= 1 + APRM     
  x   ρ L  

 y 
Where APRM =
1 + F1  − yF2  ,
1 + yF2 
0.22 −0.08
ρ  ρ 
=F1 1.578
= ReL−0.19  G  , F2 0.0273WeL ReL−0.51  G 
 ρL   ρL 
−1
 1 − x   ρ G   G2D GD
=y =
   , WeL = ,ReL
 x ρ
  L  σρ L µL

Woldesemayat and Ghajar U SG

(2007) ε=
B+C
 
0.1
 ρG 
 
  U SL  ρL  
=B U SG 1 +   
  U SG  
 
0.25 Patm
 gDσ (1 + cos θ )( ρ L − ρG )  Psystem
=C 2.9   (1.22 +1.22sin θ )
 ρ L2 
Structure velocity
Bendiksen (1984)
1.2U m + 0.35sin θ gd + 0.54 cos θ gd
U TB =

Hasan and Kabir (1988)

CoU M + U D (Vertical ) sin θ (1 + cos θ )1.2
U TB =

Woldesemayat and Ghajar UTB = B + C

(2007)
 Modelling Two-phase Flow Parameters Utilizing Machine-learning Methodology 291

Fig. 12.2. Workflow for data processing and model development.

and Deep Neural Network (DNN). The characteristics of these algorithms

are summarized in Table 12.3. The hyperparameters for each of the ML
algorithms were tuned to optimize model performance and improve accuracy
(Table 12.4). Selecting the optimal set of modelling parameters minimizes
overfitting, optimizes training speed, and overall model generalization.
In the final step, the performance of each ML model was evaluated using
the following statistical parameters: RMSE, Absolute Average Percentage
Error (AAPE), and the Average Percentage Error (APE), and correlation
coefficient (R2). The performance of the ML models and existing correlations
are also compared to evaluate the applicability of ML in modelling the two-
phase flow parameters.

4. Results and Discussions

4.1 Data Pre-processing
Figure 12.3 shows the Pearson standard correlation coefficients between the flow
variables and two-phase flow parameters. The pipe diameter, gas density, and
viscosity showed the least correlation; this can be linked to the constant values
of these parameters as air is used as the gas phase across all experimental points
and all experiments were conducted in the same pipe (diameter = 0.067 m).
Some existing models show a strong correlation between mixture velocity and
the two-phase flow parameters considered in this study.
For this reason, the mixture velocity (superficial gas + superficial liquid
velocities) is included as one of the variables in this study. The negative
correlations indicate (USL, ρL, θ, and μL) an inverse relationship, whereas
USG and Um show a positive relationship with the mean void fraction. The
USG shows the highest correlation for both the mean void fraction and the
structure velocity. Although the correlation matrix in Fig. 12.3 shows a
minimal correlation between the pipe inclination and the mean void fraction,
 292 Data Science and Machine Learning Applications in Subsurface Engineering

Table 12.3. Brief description of regression algorithms.

ML Algorithm Illustration
Linear regression assumes a linear relationship
between a dependent variable and independent
variable(s). Linear regression can be either simple
(one dependent variable) or multiple (2 or more
independent variables). Linear regression estimates
the coefficients (slopes and intercept) that minimize
ression modelsthenon-linear
loss functionrelationships
between the actualbetween
values and
predicted values
independent variables. (Montgomery
Polynomial et al., 2012).
models can fit
nships by adding
sion models higher-order
non-linear
Polynomial terms;,
relationships
regression however,
models between
nonlinear the
ynomial regression
ree of polynomial models non-linear
is anPolynomial relationships
important consideration
relationships between dependent and between
to
independent
dependent
Polynomial variables.
regression models non-linearmodels can fit
pendent and independent
variables. variables.
Polynomial modelsrelationships
Polynomial models
can fit between
complexcan fit
tting
hips (James
by adding
dependent et al., 2013).
andhigher-order
independent terms;,
variables.
relationships by adding however,
Polynomial
higher-order the can fit
models
terms;
mplex relationships by adding higher-order terms;, however, the
complex
eropriate
of relationships
polynomial anbyimportant
however, adding higher-order
the appropriate terms;,
degree however,
of polynomial
to isto the
degree of is polynomial consideration
is an important consideration
appropriate degreean important consideration
of polynomial to minimize
is an important overfittingto
consideration
ng (James
nimize
isminimize et al.,(James
overfitting
an ensemble 2013).
(James
overfitting
etal.,
et al.,2013).
2013).
learning
(James etmethod
al., 2013). developed as a
multiple decision trees. RF models are developed
ndom
equence: forestforest
is anisensemble learning method developed as a
anRandom
ensemble
mbination of multiple
an ensemble
learning
Random forest
decision
islearning
method
trees.
method
developed
an ensemble
RF models
developed
learning
are
as a as a
method
developed
and combination
ultiple tree building
decision
of multiple–- decision
developed
trees. randomly trees.sampling
as a combination
RF models
RF of
models are
multiple
are developed
developed
ofdecision
the
owing this
following sequence: trees.
this sequence: RF models are developed following this
to create multiple subsets of the data.
ence:
otstrapping
Bootstrapping and and sequence:
tree building
tree –- –randomly
building - randomly sampling
samplingofofthe
the
oving nodes or Bootstrapping
branches and tree
that dobuilding
not – randomly
improve the
ginal dataset
original to
dataset
d tree buildingsampling create
to –- randomly
multiple
create subsets
multiple of
subsets
of the original
the
of
sampling data.
the data.
of the
dataset to create multiple
mance.
ning –- removing
Pruning –- removing nodes
subsets oforthe
nodes branches
ordata.
branches thatthat
do donotnot
improve
improvethethe
create multiple subsets of the data.
–- aggregates the predictions of multiple decision
model's
del's performance.
performance. Pruning – removing nodes or branches that do not
ing nodes
Majority
jority voting –or- aggregates
voting branches
–improve thethat
- aggregates model’s
the the
do not improve
performance.
predictions
predictions
the
of multiple
of multiple decision
decision
final prediction (Breiman,
Majority voting2001).
– aggregates the predictions of
nce.trees to make a final
es to make a final prediction prediction
multiple decision
(Breiman,
(Breiman, 2001).
2001).
trees to make a final prediction
aggregates the(Breiman,
predictions
2001). of multiple decision
regression learns
Support vector the function
regression learns the that
Supportlearns maps
function
vector the that input
regressionmaps input
learns themaps
function
pport vector regression
alfeatures
prediction (Breiman, 2001). function that input
tinuous output
tures to continuous values
outputby
to continuous
that mapsoutput finding
input
values
values
featuresa tohyperplane
by
by finding
finding thatthat
a hyperplane
continuous output
a hyperplane
that
values
maximally separates thebyinput
finding a hyperplane
data while minimizingthat maximally
the error on
rates
ximallythe input the
separates datainput
whiledata minimizing
while
data minimizing
the error
the error on on
the training data.separates
SVR usesthe input
kernel while
functions tominimizing
transform input the error
data
egression
a.training
SVR uses learns
data.kernel
SVRon the
uses
the function
functions
kernel
training towherethat
transform
functions
data. SVRto maps input
transform inputdatadata
input
into a higher dimensional space, ituses kernel
is easier tofunctions
find a linear
ouous output
adecision
higher
mensional values
dimensional
space,
boundary by &
to transform
space,
where
(Chang finding
itinput
where
Lin, aithyperplane
data
is 2012).
easierinto toa higher
is easier findthat
findtoadimensional
a linear
linear
ision boundary space,
(Chang &where
Lin, it is
2012). easier to find a linear decision
es the input data while minimizing the error on
aryThe(Chang & Lin,of2012).
architectureboundary a deep (Chang
neural andnetwork
Lin, 2012).consists of multiple
VR uses kernel
architecture
layers of nodes. functions
of aThe deep
Each neural to
node of
architecture transform
network
performs a input
consists
a deep neural ofdata
different
network multiple
type
consistsof
e of a deep neural network consists of multiple
nsional
ers space,
of nodes.
computation and where
Each
of node
multiple
learns itlayers
complexisperforms
easier toa Each
of nodes.
relationshipsfind a linear
different
node
between type
performs of
a
input and
es.output
Eachandnode
mputation learns performs
different
firsttype
complex of ofa the
different
computation
relationships typeinput
andreceives
learns
between of and
complex
(Chang variables.
& Lin, The
2012). layer network
relationships between input and output variables.
the input
d variables,
putlearns complex
variables.andThe relationships
first
subsequent layerlayers
of the between
network
process input
receives
the data whilethe and
the input
output
The first layer of the network receives the input
of layer
s.ables,
The provides
first
and
a deep layer the
subsequent
neural offinal
the
layers
network
variables,
prediction.
network
and process The
the
consists
subsequent
weight
receives
data
ofwhile
layers
andthe
the
multiple
process
bias
input of the
theoutput
data
er nodes
providesin the
the network
final
while theare
prediction.iteratively
output The
layer adjusted
weight
provides and
the to minimize
bias
final of the
prediction.the
ubsequent
Each node layersperforms
process the data
abias while the
different type output of output
desdifference
in the networkbetween
Theare the
weight predicted
and
iteratively of and
the
adjusted the
nodes actual
toinminimize
the network the
the
earns final prediction.
complex
(Goodfellow, The adjusted
relationships
are
Bengio iteratively weight toand
between bias
minimizeinput ofdifference
the andthe
erence between the& Courville,
predicted 2016).
and the actual output
etwork
The are
first layer
odfellow, Bengioof
iteratively
between theadjusted
& the network
Courville,
predicted and
2016).
to minimize
the actual output
receives the input the
(Goodfellow et al., 2016).
ween the predicted
equent layers process the data while and the actual
the outputoutput
ngio & Courville,
4. Results The
e final prediction. 2016).
and dDiscussions
weight and bias of the
work4.are iteratively
Results adjusted to minimize the
and dDiscussions
Data- pre-processing
en the predicted and the actual output
Data- pre-processing
 Modelling Two-phase Flow Parameters Utilizing Machine-learning Methodology 293

Fig. 12.3. Correlation coefficient heatmap.

the empirical correlations by Bendiksen (1984) and Woldesemayat and

Ghajar (2007) suggest otherwise; thus, the pipe inclination is included in the
modelling data set for this study. The air properties and pipe diameter were
eliminated from the dataset as both variables show no correlation with the
output data. The pipe inclination, superficial gas and liquid velocities, liquid
viscosity and density showed some effect on the target variables; thus, all
these flow variables were utilized in the model development to capture their
impact on the mean void fraction and structure velocity. The entire data was
split in a train-test ratio of 4:1; that is, 1,146 train data and 287 test data. It
is worth noting that prior to model development, z-scores normalization was
utilized to transform all the variables of the data to a similar scale.

4.2 Model Development and Evaluation

For each of the models, the training parameters were tuned to ensure optimal
model performance while minimizing overfitting. For instance, several
294 Data Science and Machine Learning Applications in Subsurface Engineering

Table 12.4. Training hyperparameters.

ML Algorithms Parameters Values

PLR Polynomial degree 4
RF Number of decision trees 100
Depth of decision tree 20
SVR Kernel rbf
DNN Input layer activation function relu
Hidden layers activation function tanh
Kernal initializer for all layers normal
Optimizer adam
Batch size 20
Number of epochs 50

RF models were developed by varying the maximum number of trees and

maximum depth within the range of 50–300 and 20–50, respectively. The
model with the optimal performance and enough trees to ensure randomness
in the bootstrapping and tree-building stages of the RF model development.
The set maximum number of trees and tree depth in this case are 100 and
20 as shown in Table 12.4. Similar hyperparameter tuning techniques were
replicated for PR, SVR, and DNN and the optimal values are presented in
Table 12.4.
Table 12.5 shows the mean void fraction and structure velocity model
performances for the test data based on the predefined statistical error
parameters for each of the ML models. All the ML models, except multiple
linear regression, show an acceptable prediction for mean void fraction.
Although the performance of the nonlinear models for the mean void fraction
in all is above 90%, RF and DNN show superior predictive capabilities in
comparison to PR and SVR. As shown in Table 12.5, DNN shows the best R2
and RMSE while RF has the best APE and APPE; however, these values are
within similar range for both ML algorithms. For the structure velocity, all the
nonlinear models show similar capability with an RMSE of 0.1, APPE and R2
ranging between 3.5–3.7% and 99.0–99.2%, respectively. The MLR model
for UTB shows acceptable performance with an RMSE and R2 of 0.24 and 94.7;
however, it is not comparable to the other ML models considered in this study.
For this reason, the MLR models for both the mean void fraction and structure
velocity are considered unsuitable for further analysis due to their poor
performance in comparison to the other ML models. Figure 12.4 shows the
cross plots of the predicted parameters against the test data for each of the ML
models. For the void fraction, the predicted values all align with 20% of the
test data for RF and DNN. Figure 12.4, eludes the poor performance of MLR
 Modelling Two-phase Flow Parameters Utilizing Machine-learning Methodology 295

Table 12.5. Statistical error parameters for ML models.

ML Algorithms RMSE APE APPE R2

(-) (%) (%) (%)
Mean void fraction
Multilinear regression 0.08 24.92 9.49 83.81
Polynomial regression 0.02 3.98 –0.29 98.37
Random forest 0.02 2.21 0.27 99.32
Support vector regression 0.02 6.30 –0.61 98.60
Deep neural network 0.01 2.86 –0.60 99.55
Structure velocity
Multilinear regression 0.24 1.84 9.77 94.74
Polynomial regression 0.10 0.36 3.69 98.96
Random forest 0.10 0.61 3.52 99.01
Support vector regression 0.10 0.39 3.94 98.97
Deep neural network 0.10 –1.27 3.63 98.99

in prediction of the mean void fraction, sharing significant scatter especially

for Ɛ > 0.4. For the structure velocity, majority of the data falls within 20%
for all the ML models. The outcome of the MLR model supports the good
performance of linear correlations like Nicklin (1962) and Bendiksen (1984).
All the models developed in this study were training in less than 2 minutes;
thus, the training time was not considered as a performance criterion. However,
the ANN was more computationally challenging as it required tuning several
hyperparameters. For this reason, RF is selected as the best model for the data
set used in this study even though it shows similar statistical error parameters
with DNN.

5. Comparison between ML Algorithms and

Existing Correlations
This subsection evaluates the performance of existing correlations and
compares it to that of the ML algorithms evaluated in Section 4. The same
test data is used for evaluating the ML models and the existing correlations.
Table 12.6 shows that the Bendiksen (1984) and Woldesemayat and Ghajar
(2007) models perform better than the other correlations with RMSE and R2 of
0.04 and ~ 96%, respectively. These models cover all upward inclined flows. It
is worth noting that the Lockhart and Martinelli (1949), Rouhani and Axelsson
(1970), and Premoli et al. (1970) correlations were developed specifically for
vertical flows only. Although Bendiksen’s model does not consider the fluid
properties, it presents similar capabilities to the Woldesemayat and Ghajar
296 Data Science and Machine Learning Applications in Subsurface Engineering

Fig. 12.4. Cross plot of predicted and test data for Ɛ and UTB.
 Modelling Two-phase Flow Parameters Utilizing Machine-learning Methodology 297

Table 12.6. Statistical error parameters for empirical correlations.

Empirical Correlations RMSE APE APPE R2

(-) (%) (%) (%)
Mean void fraction
Lockhart and Martinelli (1949) 0.09 8.82 20.90 80.69
Bendiksen (1984) 0.04 –1.28 12.82 95.50
Rouhani and Axelsson (1970) 0.10 24.03 24.97 73.09
Premoli et al. (1970) 0.15 –0.81 29.35 46.97
Woldesemayat and Ghajar (2007) 0.04 7.54 8.77 96.24
Structure velocity
Bendiksen (1984) 0.31 5.98 13.17 91.22
Hassan and Kabir (1988) 0.37 –14.99 17.18 87.41
Woldesemayat and Ghajar (2007) 0.37 –16.40 18.00 87.46

(2007) correlation which takes into account the surface tension, gas, and
liquid densities.
Table 12.5 also shows the performance for the structure velocity
correlations. Across all statistical error parameters, the Bendiksen (1984)
correlation displays the best performance. For horizontal flows, drift flux
correlations developed based on the assumption of zero drift velocity tend to
predict the structure velocity with lower accuracy compared to correlations
that assume the opposite (Woldesemayat and Ghajar, 2007). Thus, it is not
surprising that the Bendiksen (1984) correlation performs better than the
Hasan and Kabir (1988) correlation. Although the Woldesemayat and Ghajar
(2007) correlation is developed based on the drift-flux model, assumes a
non-zero drift velocity in the horizontal pipe, and considers some fluid
properties, it fails to meet the predictive capabilities of the Bendiksen (1984)
correlation for the data set explored in this study.
To further validate the ML models developed, only the best performing
correlations are compared with the ML models development. For the
experimental conditions and two-phase flow parameters explored in this study,
the Bendiksen (1984) and Woldesemayat and Ghajar (2007) models show
the best performance for both the structure velocity and mean void fraction.
Comparing the statistical error parameters in Tables 12.4 and 12.5, all the ML
algorithms, except MLR, outperform the traditional empirical correlations for
predicting the mean void fraction. In the case of the structure velocity, all the
ML models surpass the correlations considered in this study by at least 8% of
the RMSE, APPE, APE, or R2. This comparative assessment demonstrates the
capabilities of ML algorithms in the development of data-driven models with
substantially higher accuracy compared to the existing empirical models.
298 Data Science and Machine Learning Applications in Subsurface Engineering

6. Conclusions and Recommendations

This study explores the capabilities of regression algorithms to predict two-
phase flow parameters. The mean void fractions and structure velocities of
intermittent flow patterns at varying gas and liquid superficial velocities,
liquid viscosities and densities, and pipe inclination were predicted using
multiple linear regression, polynomial regression, random forest, support
vector regression, and deep neural network. For both parameters, the MLR
predicted the target variables with the least accuracy, thus suggesting that the
relationship between the two-phase parameters and flow variables are complex
and cannot be adequately modelled using a linear relationship. Although
all other ML models show a similar performance, RF and DNN show the
best results with reference to the statistical error parameters considered in
this study. A cross plot of the predicted results shows that majority of the
data aligns within 20% of the corresponding test data. The ML predictive
models were validated by comparing the performance with well-established
correlations. All the ML models, except MLR, outperform the traditional
empirical correlations considered in this study by at least 8% of the RMSE,
APPE, APE, or R2.
In this study, the ML models were evaluated using test data having the
same configurations as the training data set. The authors suggest that future
work should explore validation data from other sources to adequately validate
the ML models, account for overfitting tendencies, and enable optimal
hyperparameters tuning to optimize the versatility and generalization of these
regression models. Furthermore, utilizing ML libraries to evaluate feature
importance should be considered. This technique quantifies the contribution
of each flow variable on the ML model and by extension, the magnitude of
its influence on the two-phase flow parameter which is challenging to capture
using the traditional correlation approach.

Nomenclature
A area of pipe x quality
Cd drift coefficient θ pipe inclination
Co distribution coefficient ε void fraction
D diameter of the pipe σ Surface tension
g acceleration due to gravity ρL liquid density
Re Reynolds number ρG gas density
Ud drift velocity μG gas viscosity
Um mixture velocity μL liquid viscosity
UTB translational velocity
 Modelling Two-phase Flow Parameters Utilizing Machine-learning Methodology 299

References
Abdul-Majeed, G.H., Kadhim, F.S., Almahdawi, F.H.M., Al-Dunainawi, Y., Arabi, A. and
Al-Azzawi, W.K. 2022. Application of artificial neural network to predict slug liquid
holdup. International Journal of Multiphase Flow 150(January): 104004. https://doi.
org/10.1016/j.ijmultiphaseflow.2022.104004.
Al-Naser, M., Elshafei, M. and Al-Sarkhi, A. 2016. Artificial neural network application for
multiphase flow patterns detection: A new approach. Journal of Petroleum Science and
Engineering 145: 548–564. https://doi.org/10.1016/j.petrol.2016.06.029.
Aliyu, A.M., Choudhury, R., Sohani, B., Atanbori, J., Ribeiro, J.X.F., Ahmed, S.K.B. and
Mishra, R. 2023. An artificial neural network model for the prediction of entrained droplet
fraction in annular gas-liquid two-phase flow in vertical pipes. International Journal of
Multiphase Flow 164: 104452. https://doi.org/10.1016/j.ijmultiphaseflow.2023.104452.
Alves, I.N. and Shoham, O. 1991. Slug flow phenomena in inclined pipes. Petroleum
Engineering, University of Tulsa, USA. Doctor of (July 2007).
Azizi, S., Awad, M.M. and Ahmadloo, E. 2016. Prediction of water holdup in vertical and
inclined oil-water two-phase flow using artificial neural network. International Journal
of Multiphase Flow 80: 181–187. https://doi.org/10.1016/j.ijmultiphaseflow.2015.12.010.
Bendiksen, K.H. 1984. An experimental investigation of the motion of long bubbles in
inclined tubes. International Journal of Multiphase Flow 10(4): 467–483. https://doi.
org/10.1016/0301-9322(84)90057-0.
Breiman, L. 2001. Random forests. Machine Learning 45: 5–32.
Cai, S., Toral, H., Qiu, J. and Archer, J.S. 1994. Neural network based objective flow regime
identification in air‐water two phase flow. The Canadian Journal of Chemical Engineering
72(3): 440–445. https://doi.org/10.1002/cjce.5450720308.
Canada. Energy Report. 2020. Oil Sands: Pipeline Safety. https://www.nrcan.gc.ca/energy/
publications/18754.
Chang, C.C. and Lin, C.J. 2012. LIBSVM: A library for support vector machines. ACM
Transactions on Intelligent Systems and Technology (TIST) 2(3): 27.
Dafyak, L. 2022. Hydrodynamics of Viscous Slug Flow in Inclined Pipes (Issue December).
University of Nottingham.
Dafyak, L.A., Hewakandamby, B., Fayyaz, A. and Hann, D. 2021. Taylor Bubbles of Viscous
Slug Flow in Inclined Pipes. Paper presented at the Offshore Technology Conference,
Virtual and Houston, Texas, August 2021. https://doi.org/10.4043/31238-MS.
dos Santos Ambrosio, J., Lazzaretti, A.E., Pipa, D.R. and da Silva, M.J. 2022. Two-phase
flow pattern classification based on void fraction time series and machine learning. Flow
Measurement and Instrumentation 83(November 2021): 102084. https://doi.org/10.1016/j.
flowmeasinst.2021.102084.
Dukler, A.E. and Hubbard, M.G. 1975. A model for gas-liquid slug flow in horizontal and near
horizontal tubes. Industrial & Engineering Chemistry Fundamentals 14(4): 337–347.
https://doi.org/10.1021/i160056a011.
Escrig, J. 2017. Influence of Geometrical Parameters on Gas-liquid Intermittent Flows. PhD
thesis, University of Nottingham.
Escrig, J., Hewakandamby, B. and Azzopardi, B. 2017. Influence of Diameter and Inclination
of the Pipes on the Velocity of Periodic Structures in Gas-Liquid Intermittent Flows.
November.
Fernandes, R.C., Semiat, R. and Dukler, A.E. 1983. Hydrodynamic model for gas‐liquid slug flow
in vertical tubes. AIChE Journal 29(6): 981–989. https://doi.org/10.1002/aic.690290617.
300 Data Science and Machine Learning Applications in Subsurface Engineering

Godbole, P.V., Tang, C.C. and Ghajar, A.J. 2011. Comparison of void fraction correlations for
different flow patterns in upward vertical two-phase flow. Heat Transfer Engineering
32(10): 843–860. https://doi.org/10.1080/01457632.2011.548285.
Goodfellow, I., Bengio, Y. and Courville, A. 2016. Deep Learning (Vol. 1). MIT Press.
Green, KP. and Jackson, T. 2015. Safety in the Transportation of Oil and Gas: Pipelines or Rail?
Fraser Institute: 1–14.
Hasan, A.R. and Kabir, C.S. 1988. Predicting multiphase flow behavior in a deviated well. SPE
Production Engineering 3(4): 474–482. https://doi.org/10.2118/15449-PA.
Hewitt, G.F. and Roberts, B.N. 1969. Studies of two-phase flow patterns by simultaneous
X-ray and flash photography. United Kingdom Atomic Energy Authority Research Group,
Berkshire.
James, G., Witten, D., Hastie, T. and Tibshirani, R. 2013. Introduction to Statistical Learning. An
Introduction to Statistical Learning with Applications in R: 176–178.
Kim, T.W., Kim, S. and Lim, J.T. 2020. Modeling and prediction of slug characteristics utilizing
data-driven machine-learning methodology. Journal of Petroleum Science and Engineering
195(June): 107712. https://doi.org/10.1016/j.petrol.2020.107712.
Liu, Y., Tong, T.A., Ozbayoglu, E., Yu, M. and Upchurch, E. 2020. An improved drift-flux
correlation for gas-liquid two-phase flow in horizontal and vertical upward inclined
wells. Journal of Petroleum Science and Engineering 195(June): 107881. https://doi.
org/10.1016/j.petrol.2020.107881.
Lockhart, R.W. and Martinelli, R.C. 1949. Proposed Correlation of Data for Isothermal
Two-phase, Two-component Flow in Pipes (pp. 45, 39–48). Chem. Eng. Progr. Univ. of
Berkeley California.
Lu, M. 2015. Experimental and Computational Study of Two-phase Slug Flow. PhD Thesis,
Imperial College of London, June.
Mandhane, J.M., Gregory, G.A. and Aziz, K. 1974. A flow pattern map for gas-iquid flow in
horizontal pipes. International Journal of Multiphase Flow 1(4): 537–553. https://doi.org/
https://doi.org/10.1016/0301-9322(74)90006-8.
Mask, G., Wu, X. and Ling, K. 2019. An improved model for gas-liquid flow pattern
prediction based on machine learning. Journal of Petroleum Science and Engineering
183(August): 106370. https://doi.org/10.1016/j.petrol.2019.106370.
McCain, Jr. W.D 1933. The Properties of Petroleum Fluids. PennWell Publishing Company,
Tulsa Oklahoma. ISBN 0-87814-335-1.
Montgomery, D.C., Peck, E.A. and Vining, G.G. 2012. Introduction to Linear Regression
Analysis. Wiley.
Nicklin, D.J. 1962. Two-phase bubble flow. Chemical Engineering Science 17(9): 693–702.
https://doi.org/https://doi.org/10.1016/0009-2509(62)85027-1.
Nie, F., Wang, H., Song, Q., Zhao, Y., Shen, J. and Gong, M. 2022. Image identification for
two-phase flow patterns based on CNN algorithms. International Journal of Multiphase
Flow 152(March): 104067. https://doi.org/10.1016/j.ijmultiphaseflow.2022.104067.
Premoli, A., Francesco, D. and Prima, A. 1970. An empirical correlation for evaluating
two-phase mixture density under adiabatic conditions. In: European Two-Phase Flow
Group Meeting, Milan, Italy.
Rosa, E.S., Salgado, R.M., Ohishi, T. and Mastelari, N. 2010. Performance comparison
of artificial neural networks and expert systems applied to flow pattern identification
in vertical ascendant gas-liquid flows. International Journal of Multiphase Flow
36(9): 738–754. https://doi.org/10.1016/j.ijmultiphaseflow.2010.05.001.
Rouhani, S.Z. and Axelsson, E. 1970. Calculation of void volume fraction in the subcooled and
quality boiling regions. International Journal of Heat and Mass Transfer 13(2): 383–393.
https://doi.org/10.1016/0017-9310(70)90114-6.
 Modelling Two-phase Flow Parameters Utilizing Machine-learning Methodology 301

Shoham, Ovadia. 2006. Mechanistic Modeling of Gas-liquid Two-phase Flow in Pipes. Society
of Petroleum Engineers. https://doi.org/10.2118/9781555631079.
Soloveichik, Y.G., Persova, M.G., Grif, A.M., Ovchinnikova, A.S., Patrushev, I.I., Vagin,
D.V. and Kiselev, D.S. 2022. A method of FE modeling multiphase compressible flow
in hydrocarbon reservoirs. Computer Methods in Applied Mechanics and Engineering
390: 114468. https://doi.org/10.1016/j.cma.2021.114468.
Woldesemayat, M.A. and Ghajar, A.J. 2007. Comparison of void fraction correlations for
different flow patterns in horizontal and upward inclined pipes. International Journal of
Multiphase Flow 33(4): 347–370. https://doi.org/10.1016/j.ijmultiphaseflow.2006.09.004.
Yan, Y., Wang, L., Wang, T., Wang, X., Hu, Y. and Duan, Q. 2018. Application of soft
computing techniques to multiphase flow measurement: A review. Flow Measurement and
Instrumentation 60(February): 30–43. https://doi.org/10.1016/j.flowmeasinst.2018.02.017.
 Index

A Deep Adaptation Neural Network (DaNN) 161,

162
Akaike Information Criterion (AIC) 75 Deep Convolutional Neural Network (DCNN)
Analytical models 88 3, 159, 164–166, 168, 169, 172, 173, 179,
Anthropogenic CO2 126, 127 181, 182, 189
Artificial Intelligence (AI) 1, 7, 35, 60, 106, Deep learning 39, 43, 160, 161, 163, 176, 178,
169, 182, 185, 208, 214, 261 179
Artificial lift methods 262, 272, 283 Deep learning-based algorithms 160
Artificial lift system 262, 263, 282, 283 Deep neural network 250, 252
Artificial Neural Networks (ANNs) 8, 208, 215, Distributed temperature sensing systems 105
220, 224, 261, 262, 267, 287 Downhole control systems 105
Downhole sensors 104
B Drilling 6–11, 17, 20, 23, 26–29
Backpropagation network 261 Drilling fluid 6–9, 11, 27, 28
Backpropagation neural network 227, 262 Drilling operations 7, 10, 11
Bayes theorem 75
Bayesian optimisation 1, 2 E
Bound fluid 33, 36–38, 40, 41 Electrical submersible pump 88, 90, 92, 100
Bound fluid volume 36, 37 Empirical correlation 58, 64, 75, 79–82, 105,
Brownfields 125 106, 111, 208, 248, 287–290, 293, 297, 298
Enhanced Oil Recovery (EOR) 3, 125
C Ensemble models 33, 43–45, 48, 49, 54
Capital and operating costs 105 Evolutionary computation 287
Carbon dioxide 3 Explainable AI 9, 15, 28, 29, 110, 114
Carbon Dioxide (CO2) injection 126
Climate change 126, 127 F
CO2 flooding 127, 128 Facies analysis 208, 234, 235, 250
CO2 storage 126, 127, 163 Faults 159, 160, 162, 169
Compressional sonic 58, 61–63, 82 Flow dynamics 286, 287
Computer vision 1 Flow metering 87, 88, 96
Flow pattern models 287
D Flow rate 87, 88, 90, 92–97, 99, 100
Data acquisition 105, 107, 110 Flow structures 287
Data analysis 11, 110, 163 Fluid distribution 105
Data mining framework 88 Fluid phase properties 87
Data science 1, 4, 5 Fluid sequestration 57, 58, 83
Data-driven models 88, 94–97, 100 Forecasting 261, 262
304 Data Science and Machine Learning Applications in Subsurface Engineering

Formation evaluation 57 M
Fossil fuel combustion 127
Fractal analysis 213 Machine Learning (ML) 1–5, 6–11, 13, 15, 16,
Free fluid 33–35, 37, 38, 40, 41, 47, 48, 54 18, 23, 26, 36, 38, 43, 47, 57, 58, 60, 61,
Free fluid index 33 64, 68, 73, 104, 106, 108–111, 113, 114,
Free fluid volume 33–35, 47, 48, 54 122, 160, 161, 178, 287–289, 291, 292,
Fuzzy logic 261, 287 294, 295, 297, 298
Machine learning algorithms 97
G Maximum likelihood 75
Mean void fraction 287–291, 293–295, 297,
Gas injection 125, 127, 129, 141, 142, 145 298
Gas lift 262, 264, 265, 267, 268, 270–273, Mercury Injection Capillary Pressure (MICP)
276–279, 282, 283 analysis 34
Gas oil ratio (GOR) 87 Metaheuristics algorithms 64
Gas reservoir 261 Miscible CO2 flooding 127, 128
Gas-liquid flow 286–288 Model agnostic 6, 9, 15, 16, 20, 22–24, 28
Geochemical modelling 132, 153 Mud loss 7, 9, 10, 12, 13, 18–20, 22–25, 27, 28
Geological formations 159 Multioutput supervised machine learning 58
Geomechanical properties 57 Multiphase flowrates 87
Geophysics 57 Multiphase physical flow meters (MPFMs) 87
Geothermal energy 58 Multivariate adaptive regression splines 132,
Ghana 261, 262, 269, 284 134
Global energy demand 125 Multi-zonal reservoirs 105, 106
Global optimisation 75
Global warming 127 N
Greenhouse gas emissions 126, 127
Greenhouse gases 126 NMR porosity 36, 39
Group method of data handling 132, 136 Nonconventional wells 104
Nuclear Magnetic Resonance (NMR) 34, 35
H
O
Heterogeneous formation 104
Hydrocarbon extraction 33 Objective functions 65, 66, 75
Hyperparameter tuning 75 Oil and gas industry 160, 161
Oil and gas production systems 87
I Oil and gas reservoir production 105
Oil recovery 261
Immiscible CO2 flooding 127 Oil recovery factor 125, 131–134, 136–138,
Inflow and outflow performance 261 142, 153, 154
Inflow Control Valves (ICVs) 105, 107, 108, 122 Oil reserves 125
Inflow performance relationship 263 Oil swelling 126
Intermittent flow 287, 288, 298 Original Oil in Place (OOIP) 126
Interval control devices 105
P
J
Permanent monitoring 105, 107
Jubilee field 262, 268–270, 282, 283 Permeability prediction 209, 223, 230, 231,
235, 240, 247
L Petroleum reservoir 125
Phase fractions 87, 88
Liquid rates 87
Physics-based models 287
Lost circulation 6–9, 11, 26–28
Pore fluids 33, 34, 37, 38, 57
Low Salinity Water (LSW) 126
Pore pressure 57
 Index 305

Pore volume fraction 33, 35, 36, 54 Smart wells 104, 105, 107–109, 114, 122
Porosity 33–42, 47–50, 52–54 Soft computing techniques 287
Porosity prediction 230 Sonic waves 57, 63
Predictive models 93–95 Stratigraphic features 160
Pressure 104–108, 110, 111, 115, 116 Structure velocity 288–291, 293–295, 297
Pressure control valves 104 Subsurface characterisation 160
Primary recovery 125, 126, 130 Subsurface engineering 1, 2, 4, 5
Probabilistic model 75 Subsurface structures 159, 160, 164
Probabilistic reasoning 287 Support Vector Machine (SVM) 7, 8, 10, 18,
Production rate 261–263, 277–279 208, 220–223, 225, 228, 230–233, 236,
Production systems 87–91, 99 238, 240, 243, 245, 287
Production testing 87 Surface control and monitoring systems 105
Production tubing 104
Proxy models 132, 153 T
Temperature 104, 105, 107, 110, 115
R
Tertiary oil recovery 126
Radial basis function neural network 262, 268 Test separator 87, 91, 92
Real-time tracking 87 Transfer learning 159, 161–164, 168, 169, 173,
Recovery factor 125, 126, 129, 131–134, 174, 177–179
136–138, 142, 145, 148, 153, 154 Tree-based algorithms 287
Reserves estimation 261 Tubing head pressure 262, 269, 279, 282, 283
Reservoir characterisation 1, 2, 4, 57, 207–209, Tubing performance relationship 261
213–215, 247–252 Two-phase flow 286–288, 291, 297, 298
Reservoir conditions 104
Reservoir exploration 57 U
Reservoir management 34, 35, 261
Unconventional resources 58
Reservoir pressure 125, 142
Reservoirs 159, 160, 163
Residual U-net architecture 168 V
Rising sea levels 126 Velocity 87, 96
Rock physics 235, 238, 248, 250–252 Velocity model 159, 162
Virtual flow metering (VFM) 2, 87, 88, 96
S Viscosity reduction 126
Salt mapping 159, 163–165, 168
Salt segmentation 161, 163, 164, 178 W
Salt tectonics 160 Water cuts 87, 88, 90–92
Sand contamination 262 Water injection 104–108, 120, 122, 130, 132,
Secondary recovery 125, 126 141, 142, 145
Seismic characterisation 252 Water injection management 122
Seismic edge-detection algorithms 160 Water saturation prediction 247
Seismic image resolution 181–184, 186, 189 Water-cut prediction 88
Seismic imaging 181–183, 186, 201, 203 Wellhead 87, 90–93
Seismic interpretation 159, 160, 162, 178, 179 Well-to-Seismic inversion 58
Seismic salt imaging 160 Wireline logs 35, 36, 40, 42, 43, 47, 54, 58, 60,
Semantic segmentation 163, 169, 174 75, 81–83
Sequential model-based optimisation 75
Shapley values 15, 16, 23, 24, 26–29 Z
Shear sonic 57, 58, 61–63, 83
Smart well completion 104, 105, 107, 122 Zonal isolation 105, 107