No part of this publication may be reproduced or transmitted in any form or by any means, electronic or mechanical, including photocopying, recording, or any information storage and retrieval system, without permission in writing from the publisher. Details on how to seek permission, further information about the Publisher’s permissions policies and our arrangements with organizations such as the Copyright Clearance Center and the Copyright Licensing Agency, can be found at our website: www.elsevier.com/permissions
This book and the individual contributions contained in it are protected under copyright by the Publisher (other than as may be noted herein).
Notices
Knowledge and best practice in this field are constantly changing. As new research and experience broaden our understanding, changes in research methods, professional practices, or medical treatment may become necessary.
Practitioners and researchers must always rely on their own experience and knowledge in evaluating and using any information, methods, compounds, or experiments described herein. In using such information or methods they should be mindful of their own safety and the safety of others, including parties for whom they have a professional responsibility.
To the fullest extent of the law, neither the Publisher nor the authors, contributors, or editors, assume any liability for any injury and/or damage to persons or property as a matter of products liability, negligence or otherwise, or from any use or operation of any methods, products, instructions, or ideas contained in the material herein.
ISBN: 978-0-12-822295-9
For Information on all Elsevier publications visit our website at https://www.elsevier.com/books-and-journals
Publisher: Charlotte Kent
Acquisitions Editor: Amy Shapiro
Editorial Project Manager: Maria Elaine D Desamero
Production Project Manager: R.Vijay Bharath
Cover Designer: Mark Rogers
Typeset by Aptara, New Delhi, India
Contributors
Ayodeji Aboaba
West Virginia University, United States
Tariq Alkhalifah
Department of Physical Science and Engineering, King Abdullah University of Science and Technology, Thuwal, Saudi Arabia
Heather Bedle
School of Geosciences, The University of Oklahoma, Norman, OK, United States
Nasher BenHasan
EXPEC ARC, Saudi Aramco, Dhahran, Saudi Arabia
Alex Bromhead
Halliburton, United Kingdom
John P. Castagna
Department of Earth and Atmospheric Sciences, University of Houston, Houston, TX, United States
Edward Clee
Department of Computer Science, Prairie View A&M University, Prairie View, TX, United States
Kate Evans
Halliburton, United Kingdom
Hugo Garcia
Geoteric, United States
Dario Grana
Department of Geology and Geophysics, School of Energy Resources, University of Wyoming, Wyoming, United States
Chris Guenther
Department of Energy, National Energy Technology Laboratory, United States
Pablo Guillen
Department of Computer Science, University of Houston, Houston, TX, United States
Ehsan Haghighat
Department of Civil Engineering, Massachusetts Institute of Technology, MA, United States
Chris Han
Geoteric, United Kingdom
Lei Huang
Department of Computer Science, Prairie View A&M University, Prairie View, TX, United States
Lian Jiang
Department of Earth and Atmospheric Sciences, University of Houston, Houston, TX, United States
Cédric M. John
Department of Earth Science and Engineering, Imperial College, London, United Kingdom
Karelia La Marca
School of Geosciences, The University of Oklahoma, Norman, OK, United States
Mingliang Liu
Department of Geology and Geophysics, School of Energy Resources, University of Wyoming, Wyoming, United States
Yong Liu
Department of Energy, National Energy Technology Laboratory, United States; Leidos Research Support Team, United States
James Lowell Geoteric, United Kingdom
Yvon Martinez
West Virginia University, United States
Shahab D. Mohaghegh
West Virginia University and Intelligent Solutions, Inc., United States
Philippe Nivlet
EXPEC ARC, Saudi Aramco, Dhahran, Saudi Arabia
Rafael Pires de Lima
Geological Survey of Brazil, São Paulo, Brazil
Nishath Ranasinghe
Department of Computer Science, Prairie View A&M University, Prairie View, TX, United States
Mehrdad Shahnam
Department of Energy, National Energy Technology Laboratory, United States
Robert Smith
EXPEC ARC, Saudi Aramco, Dhahran, Saudi Arabia
Chao Song
Department of Physical Science and Engineering, King Abdullah University of Science and Technology, Thuwal, Saudi Arabia
Fnu Suriamin
Oklahoma Geological Survey, Norman, OK, United States
Peter Szafian
Geoteric, United Kingdom
Miao Tian
Department of Geosciences, The University of Texas Permian Basin, Odessa, TX, USA
Sumit Verma
Department of Geosciences, The University of Texas Permian Basin, Odessa, TX, USA
Umair bin Waheed
Department of Geosciences, King Fahd University of Petroleum and Minerals, Dhahran, Saudi Arabia
Jessica Wevill
Department of Earth Science and Engineering, Imperial College London, United Kingdom
Ryan Williams
Geoteric, United Kingdom
Jeffrey Yarus
Case Western Reserve University, United States
Zhendong Zhang
The Earth Resources Laboratory, Massachusetts Institute of Technology, MA, United States
Part
2.
machine-based
3.
5.
3.4
6. Convolutional neural networks for fault interpretation – case study examples around the world
Hugo Garcia, Peter Szafian, Chris Han, Ryan Williams, James Lowell
6.5
6.6
6.7
Part 3 Physics-based machine learning approaches
7.
Lei Huang, Edward Clee, Nishath Ranasinghe
8.
Lian Jiang, John P. Castagna, Pablo Guillen
8.7
9. Regularized elastic full-waveform inversion using deep learning 219
Zhendong Zhang, Tariq Alkhalifah
9.3
9.4
9.5
9.6
9.7
10. A holistic approach to computing first-arrival traveltimes using neural networks
Umair bin Waheed, Tariq Alkhalifah, Ehsan Haghighat, Chao Song
Part 4 New directions
11. Application of artificial intelligence to computational fluid dynamics 281 Shahab D. Mohaghegh, Ayodeji Aboaba, Yvon Martinez, Mehrdad Shahnam, Chris Guenther, Yong Liu
About the Editors
Dr. Shuvajit Bhattacharya is a Research Associate at the Bureau of Economic Geology, the University of Texas at Austin. He is an applied geophysicist/petrophysicist specializing in seismic interpretation, petrophysical modeling, machine learning, and integrated subsurface characterization. Prior to joining the Bureau of Economic Geology, Dr. Bhattacharya worked as an Assistant Professor at the University of Alaska Anchorage. He has completed several projects in the US, Norway, Netherlands, Australia, South Africa, and India. He has published and presented more than 70 technical articles in peer-reviewed journals, books, and conferences. His current research focuses on the pressing issues and frontier technologies in energy resources exploration, development, and subsurface storage of carbon and hydrogen. He completed his Ph.D. at West Virginia University in 2016 and an M.Sc. at the Indian Institute of Technology Bombay in 2010.
Dr. Haibin Di is a Senior Data Scientist in the Subsurface Data Intelligence team at Schlumberger. His research interest is in implementation of machine learning algorithms, particularly deep neural networks into multiple seismic applications, including stratigraphy interpretation, property estimation, denoising, and seismic-well tie. Dr. Di has published more than 70 papers in seismic interpretation and holds 7 patents on machine learning-assisted subsurface data analysis. Dr. Di received his Ph.D. degree in Geology from West Virginia University in 2016, worked as a postdoctoral researcher at Georgia Institute of Technology in 2016-18, and joined Schlumberger in 2018.
Acknowledgments
We would like to express heartfelt thanks to all 31 authors from industry and academia who contributed to this book. This book could not have been possible without their writing efforts and patience in tolerating editorial commands and reviews. These authors have spent years studying and experimenting with advanced machine learning concepts in geosciences and reservoir engineering. We thank them for that. We would also like to express thanks to numerous individuals and professionals with whom we had an opportunity to discuss artificial intelligence and learn from their experiences.
Special thanks go to Amy Shapiro, Elaine Desamero, Vijay Bharath, and Chris Hockaday at Elsevier for making this book happen. We also acknowledge the publishers and individuals who provided permission to use figures from their technical articles and websites. We deeply appreciate all your support and encouragement.
Shuvajit Bhattacharya Haibin Di
Preface
Advances in Subsurface Data Analytics: Traditional and Physics-Based Approaches is a compilation of selected case studies in the applications of traditional, emerging, and physicsbased machine learning (ML) algorithms and approaches in subsurface imaging and characterization of heterogeneous media. We edited this book to bring together the fundamentals of several ML algorithms with their detailed applications in subsurface analysis, including geology, geophysics, petrophysics, and reservoir engineering. While we focus on traditional and physics-based ML approaches, we also shed light on how ML technologies developed in the subsurface can be used in other disciplines.
The book is intended to be used as a detailed reference book primarily by professional geoscientists and reservoir engineers working on subsurface-related research problems (oil and gas, carbon sequestration, geothermal, and hydrogen storage). It can be used by faculty, researchers, and graduate students in the geosciences/energy resources engineering departments at universities.
The edited book comprises several chapters, each focusing on a particular application of ML algorithm(s) with detailed workflow and recommendations. In addition, some of the chapters also contain a comparison of an algorithm with respect to others to better equip the readers with different strategies to implement automated workflows for case-specific subsurface analysis.
The structure of the book is easy-to-follow for the readers, with mainly four parts. The first part focuses on traditional ML algorithms with case studies, which have been widely used. This is followed by emerging and complex deep learning algorithms. The third part focuses on physics-based ML. The last chapter is on future research directions in ML to solve problems in other disciplines.
Marca and Bedle study deep water seismic facies classification in the Taranaki Basin of New Zealand by investigating the impact of a user-controlled (interpreter-driven) selection of seismic attributes versus machine-selected results. The user-selected attribute allows the interpreter to inspect attributes individually and select the most geologically meaningful ones for their study; however, the reliability of the exercise depends on experience of the interpreters, including bias. Authors recommend detailed human inspection and approval of ML-based results in pivotal moments, including the validation and interpretation of the ML outputs regardless of the initial approach used.
Wevill et al. discuss various ML techniques to analyze the relationships between subsurface data and production success within seven North American shale plays. They use formation depth, thickness, porosity, resource concentration, pore pressure, geothermal gradient, and gas-to-oil ratio to identify the “heavy-hitters” influencing hydrocarbon production. They conclude that a large influence on production success from these
reservoirs is linked to geological factors such as depositional and tectonic history dictating play properties like mineralogy and pressure. The overall success of a hydraulically fractured well in these plays is also susceptible to the influence of completion methods.
Many geologic processes are gradational at spatial and temporal scales, which results in sequence-like features. The whole concept of sequence stratigraphy is based on sequence (and parasequence) pattern analysis. Tian and Verma demonstrate a deep learning workflow (i.e., convolutional neural network combined with long short-term memory) to classify and predict complex mudstone facies using electrical log data. Their workflow captures and employs the hidden information of data sequence and spatial dependencies, which results in a generalized and robust ML model for facies classification in multiple boreholes in the basin.
Liu et al. apply recurrent neural network (i.e., bi-directional long short-term memory) for seismic reservoir characterization. They estimate P-wave velocity, S-wave velocity, density, porosity, water saturation, and facies. Their results indicate the promising potential for the application of deep neural networks to seismic inversion. The proposed approach is a valid alternative to the conventional model-driven methods, and it avoids several data pre-processing steps, such as model calibration, and provides a robust and efficient way to efficiently automate the workflow of seismic reservoir characterization.
Core description and interpretation are vital to geologic studies but time-consuming, expensive, and subject to multiple interpretations. Lima and Suriamin show that computer vision instance segmentation models built with convolutional neural networks have the potential to greatly accelerate core interpretation in a consistent manner. They directly use standard core photographs from a 260 m (850 ft) core from a siliciclastic succession as input to their model with barely any data preprocessing for core facies classification.
Waheed et al. develop a workflow based on the emerging paradigm of a physicsinformed neural network to solve various forms of the Eikonal equation in computing first-arrival travel time. This has applications in seismic velocity modeling, microseismic source localization, and seismic migration. Their workflow obtains fast and accurate travel times, compared to the conventional workflow used over decades. This workflow is applicable to both isotropic and anisotropic media.
Scientific ML brings a new research dimension to solve complex problems in geosciences that traditional ML approaches cannot solve. Huang et al. propose a new method to improve the performance of seismic wave simulation and inversion by integrating deep learning (recurrent neural network and AutoEncoder) and differential programming with High-Performance Computing. Their workflow improved model accuracy and efficiency.
Jiang et al. discuss a novel workflow on integrating physics-based approaches with ML into building robust rock physics models (RPM) for acoustic velocity prediction.
Building RPM is a complex multi-step process with several assumptions. Jiang et al. demonstrate the feasibility of ML-assisted RPM using synthetic and real datasets. Their RPM establishes relationships between mineral composition, porosity, fluid properties, water saturation, temperature, and pressure with density and the P-wave and S-wave velocities.
Zeng and Alkhalifah present a deep learning-aided elastic full-waveform inversion (FWI) strategy using observed seismic data and available well logs using a synthetic dataset and field dataset in the North Sea. They link seismic facies to the inverted P- and S-wave velocities and anisotropy using trained neural networks. The estimated facies can be used as a physical constraint for conventional elastic FWI, which can be helpful to better resolve deep-buried reservoir targets.
Mohaghegh et al. discusses new research directions where ML-based technologies, specifically Smart Proxy Modeling developed in reservoir engineering, can be used in other disciplines, such as computational fluid dynamics (CFD). CFD is widely used in gas compressors, turbines, heating, and aerodynamics, etc. Gas supplies are expected to increase in North America due to shale gas and the upcoming hydrogen economy. The effect of gas composition on combustion behavior is of interest to allow end-user equipment to accommodate the widest possible gas composition. Mohaghegh demonstrates the feasibility of highly accurate and high-speed proxy modeling of such complex systems using artificial intelligence.
We hope this book will help popularize ML in the subsurface community, providing new research directions. We hope you will enjoy reading this book and solving your own problems.
Shuvajit Bhattacharya
Haibin Di
PART 1
Traditional machine learning approaches
1. User vs. machine-based seismic attribute selection for unsupervised machine learning techniques: Does human insight provide better results than statistically chosen attributes? 3
2. Relative performance of support vector machine, decision trees, and random forest classifiers for predicting production success in US unconventional shale plays 31
User vs. machine-based seismic attribute selection for unsupervised machine learning techniques: Does human insight provide better results than statistically chosen attributes?
Karelia La Marca, Heather Bedle
School of Geosciences, The University of Oklahoma, Norman, OK, United States
Abstract
In geosciences, a variety of machine learning (ML) algorithms are currently being employed for multiple purposes, for example, facies classification, fault prediction, and reservoir characterization. Among these are two clustering methods: principal component analysis (PCA) and self-organized maps (SOMs), which provide a fast organization of data into groups or clusters (with no geologic supervision) that aid in preliminary geological interpretation. With increasingly common usage of these techniques, the motivation of this chapter is to investigate the impact of a user-controlled selection of attributes to perform SOM for deepwater seismic facies classification versus a machine-selected result through PCA. Results reveal that whereas an appropriate combination of attributes with a clear interpretation objective enhances the SOM’s results and facilitates the interpreter understanding of the output classes, PCA provides insightful information regarding the contribution of seismic attributes that may not have been initially considered. While machine learning techniques are a powerful “tool” for geological interpretation, user control on initial input attributes and validation of output using an “in-context” interpretation is necessary for an optimal elucidation, at least in unsupervised machine learning methods.
Keywords
Machine learning techniques; Principal component analysis; Self-organized maps
1.1 Introduction
Conventional seismic interpretation techniques involve seismic attribute calculations and applications to determine geometries, lithologies, and reservoir properties. A seismic attribute is any measure of seismic data that helps us visually enhance or quantify features of interest for interpretation1. Seismic attributes are a response to the rocks’ physical properties. Therefore, the spatial distribution and relationships of these responses,
along with geological context, help make reliable seismic interpretation whether they are used individually or combined. Some examples of the use of seismic attributes can be found in (1) Partyka et al.,2 who used frequency attributes to study channels, (2) the use of textural attributes to recognize patterns in a turbidite system by Gao3, (3) the application of curvature attributes for mapping faults/fractures in Chopra and Marfurt1 and the list goes on.
In recent years, the use of machine learning (ML) has become a common practice in advanced seismic interpretation workflows. These methods are based on complex algorithms that allow quick and improved analysis of seismic data than traditional single attribute studies made by interpreters. Self-organized maps, geostatistics, and neural nets are used to extend the capability of recognizing patterns beyond the three dimensions that the interpreter is limited to4. With common usage of large datasets, the geophysical community has worked hard to replicate the human ability to cluster, but these methods need to be thoroughly validated before they can be relied upon.
ML applications within the geosciences extend from predicting seismic facies, to automatic fault detection to reservoir property prediction5–7,8–10. The choice of an unsupervised ML method over a supervised method can depend on the type of data available for the study. In exploration stages, it is often rare to find well logs or other data in the basin or the area of interest. In this scenario, unsupervised methods provide ways to find relationships among the variables in the data available. However, unsupervised methods require interpreter evaluation to approve or disapprove the output.
Methods such as Principal Component Analysis (PCA), introduced by Pearson11 and developed by Hotelling12, is a common technique to find patterns in data of high dimensions13. PCA seeks to reduce a multivariate space (dataset) down to a more manageable size of relatively independent variables. In seismic interpretation, PCA helps to determine the most meaningful seismic attributes14,15. Likewise, SOM is an unsupervised technique that can be used in multi-attribute analysis to extract additional information from the seismic response that is not possible with a single attribute16. This SOM method was first coined by Kohonen17, but became popular recently due to the increase in computational power. PCA and SOMs have been used for many purposes. For example, in the marketing arena, Das et al.18 applied PCA and SOM to find clusters in people’s responses regarding their preference for retail store personality. In seismic data, Coleou et al.,5 applied SOMs for seismic classification using a 1D latent space. More recently, Guo et al.,19 computed spectral attributes and applied PCA to determine the first three principal components of the spectral variation. Later, Matos et al.,20 Roy and Marfurt21, and Roy22 demonstrated the advantage of extending the SOM latent space to 2D and 3D, using 2D and 3D colorbars to delineate elements in depositional environments. More recent studies6,15,16 integrated PCA to select attributes, and SOMs to interpret facies in various depositional environments and others specifically in deepwater deposits8,23,24.
To better understand the effects of attribute choice, we investigated and compared a multi- attribute user-driven approach versus a machine-derived method (through PCA) to select suitable attribute combinations to use in the unsupervised self-organizing maps (SOMs) for facies prediction. We used the Pipeline 3D seismic dataset, in the southern Taranaki Basin of New Zealand, to interpret deepwater stratigraphic features based on SOMs results. We explored which attribute selection (human or machine driven) allows for a reliable classification to improve interpretation. Finally, we document the advantages and disadvantages of each method, recommending good practices to take advantage of ML tools in a suitable manner to obtain efficient and valuable geological results.
1.2 Motivation
This study’s main motivation was to explore, compare, and understand the differences that machine selected attributes (via PCA) and user-selected attributes (based on interpreter experience) produce in an unsupervised ML technique classification. To compare methods, we used a seismic volume that contains a deepwater channel section to evaluate channel architecture. We evaluated how effective each combination of seismic attributes was to make an “in context” interpretation that includes a detailed characterization of architectural elements’ geomorphologies. We sought to understand the best seismic attributes to characterize the deepwater setting and compare the generated clusters that helped define seismic facies and architectural elements. We then explored the advantages and disadvantages of a user-selected vs. machine-selected attribute approaches.
Geological setting
The Taranaki Basin is located in western-offshore New Zealand (Fig. 1.1). The basin deposits range from the Cretaceous to the Neogene. The basin formed and filled up due to the Tasman Sea spreading event25,26. The stratigraphic section studied corresponds to the Middle Miocene deposits characterized by the deposition of deepwater sequences controlled mainly by tectonic uplift in the hinterland. The high relief provided a south-east source of sediments carried and deposited in a north- northwest direction27 (Fig. 1.1). The Miocene succession comprises intercalation of fine-grained basin floor sandstones deposited by channels and fans, in addition to silty and mudstone dominated deposits. The Moki Formation corresponds to deepwater deposits characterized by: (1) channel complex width that ranges from 600 m up to 5,000 m and ranges between 10–30 m in thickness; (2) channel sinuosity varies from low to high; and (3) a system that becomes more incised and mud-dominated during the Late Miocene26.
1.3 Dataset characteristics
The area of study is located within the Pipeline 3D dataset (Fig. 1.1). The seismic volume data is zero phase and SEG negative polarity (a trough represents a positive change in acoustic impedance), with a sample interval of 4 ms, and bin size of 25 m by
