Advances in Subsurface Data Analytics: Traditional and Physics-Based Machine Learning




Contributors
CONTENTS
About the Editors
Acknowledgments
Preface
PART 1 - Traditional machine learning approaches
Chapter 1 - User vs. machine-based seismic attribute selection for unsupervised machine learning techniques: Does human in ...
1.1 Introduction
1.2 Motivation
1.3 Dataset characteristics
1.4 Seismic attributes
1.5 Principal component analysis (PCA)
1.6 Self-organizing maps (SOMs): An unsupervised technique for seismic facies identification
1.7 Methodology
1.8 Results
1.8.1 Principal component analysis
1.8.2 From the eigenspectrum to principal components and attributes
1.9 Self-organizing maps analysis
1.10 SOM results and discussion from user-selected attributes vs. machine-derived inputs
1.11 Human vs. machine comparison (upsides and downsides of PCA vs multiattribute analysis)
1.11.1 Upsides of the approaches
1.11.2 Downsides and limitations
1.12 Recommendations
1.13 Conclusion
1.14 Acknowledgments
References
Chapter 2 - Relative performance of support vector machine, decision trees, and random forest classifiers for predicting p ...
2.1 Introduction
2.2 Methods
2.2.1 Dataset available for the North American shale plays
2.2.2 Machine learning approach
2.2.3 Data cleaning and preprocessing
2.2.3.1 Input feature cleaning
2.2.3.2 Dimensionality reduction
2.2.4 Training machine learning algorithms
2.2.4.1 Stochastic gradient descent kernel trained support vector machine classifier
2.2.4.2 Decision tree classifier
2.2.4.3 Random forest classifier
2.2.4.4 Sample splitting and hyperparameter tuning
2.2.4.5 Evaluating performance of trained algorithms
2.3 Results
2.3.1 Comparative performance of our algorithms
2.3.1.1 SVM-SGD classifier
2.3.1.2 Decision tree classifier
2.3.1.3 Random forest classifier
2.3.1.4 Error analysis
2.3.2 Further optimization for four production classes and application to shale plays
2.3.2.1 Further optimization of the random forest classifier
2.3.2.2 Random forest prediction result for different North American shale plays
2.4 Discussion
2.4.1 Algorithm overview and comparison
2.4.2 Geological features of importance for each North American shale play
2.4.3 Data limitations
2.5 Conclusion
2.6 Acknowledgments
References
PART 2 - Deep learning approaches
CHAPTER 3 - Recurrent neural network: application in facies classification
3.1 Introduction
3.2 Data types
3.2.1 Chronological data sets
3.2.2 Spatial data sets
3.3 Recurrent neural network (RNN) methods
3.3.1 Notation definition
3.3.2 Simple recurrent neural network (simple RNN)
3.3.3 Long short-term memory (LSTM)
3.3.4 Gated recurrent unit (GRU)
3.3.5 Convolutional recurrent neural network (ConvRNN)
3.3.6 Bidirectional recurrent neural network (BRNN)
3.4 Case study: Bi-LSTM-assisted facies classification based on well logging data
3.4.1 Geological setting
3.4.2 Problem definition and notation
3.4.3 Methods
3.4.4 Evaluation of the model performance
3.5 Results and discussion
3.6 Conclusion
References
Chapter 4 - Recurrent neural network for seismic reservoir characterization
4.1 Introduction
4.2 Methodolgy
4.3 Applications
4.3.1 Stanford VI-E dataset
4.3.2 Marmousi2 model
4.4 Conclusion
References
Chapter 5 - Convolutional neural networks: core interpretation with instance segmentation models
5.1 Introduction
5.2 Methods
5.2.1 Geological setting and data
5.2.2 Instance segmentation
5.3 Results
5.4 Discussion
5.5 Conclusion
5.6 Acknowledgments
References
Chapter 6 - Convolutional neural networks for fault interpretation – case study examples around the world
6.1 Introduction
6.2 Machine learning algorithms
6.3 Data and workflow
6.4 Case studies
6.4.1 Beagle Sub-Basin, Australia
6.4.2 East Shetland Basin, UK
6.4.3 Main Pass, Gulf of Mexico, USA
6.4.4 Cooper-Eromanga Basin, Australia
6.5 Discussions
6.6 Conclusions
6.7 Acknowledgments
References
PART 3 - Physics-based machine learning approaches
Chapter 7 - Applying scientific machine learning to improve seismic wave simulation and inversion
7.1 Introduction
7.1.1 Seismic imaging
7.1.2 Computational issues for wave propagation
7.1.3 Opportunity for data analytics
7.1.4 Scientific machine learning
7.2 Related work
7.2.1 Seismic wave simulation and inversion
7.2.2 Surrogate models using machine learning
7.2.3 Dimensionality reduction
7.2.4 Differentiable programming
7.3 Wave equations and RNN
7.3.1 Wave equations
7.3.2 Recurrent neural network
7.3.3 PyTorch RNN implementation
7.3.4 Seismic wave simulation
7.4 Differentiable programming
7.4.1 Automatic differentiation and adjoint-state method
7.4.2 Extended automatic differentiation
7.5 Seismic inversion
7.5.1 Seismic inversion using neural network
7.5.2 Autoencoder for dimensionality reduction
7.5.3 Results
7.6 Discussion
7.7 Conclusions
7.8 Acknowledgment
References
Chapter 8 - Prediction of acoustic velocities using machine learning and rock physics
8.1 Introduction
8.2 Conventional rock physics modeling
8.3 Machine learning methods
8.3.1 Support vector regression
8.3.2 Random forest
8.3.3 Multilayer perceptron
8.3.4 Metrics for model performance evaluation
8.4 Synthetic data test
8.5 Real data test
8.6 Discussion
8.7 Conclusions
8.8 Acknowledgments
References
Chapter 9 - Regularized elastic full-waveform inversion using deep learning
9.1 Introduction
9.2 Methodology
9.2.1 Correlation elastic FWI
9.2.2 Deep neural networks
9.2.3 Facies constraints
9.3 Numerical examples
9.3.1 A synethetic Marmousi example
9.3.2 The North Sea field data example
9.3.2.1 Facies extraction
9.3.2.2 Inversion results
9.4 Discussion
9.5 Conclusions
9.6 Acknowledgments
9.7 Appendix example training code
References
Chapter 10 - A holistic approach to computing first-arrival traveltimes using neural networks
10.1 Introduction
10.2 Theory
10.2.1 Eikonal equations
10.2.2 Approximation property of neural networks
10.2.3 Automatic differentiation
10.2.4 Solving eikonal equations
10.3 Numerical tests
10.4 Discussion and conclusions
References
PART 4 - New directions
Chapter 11 - Application of artificial intelligence to computational fluid dynamics
11.1 Introduction
11.1.1 Structure of the work
11.2 Background
11.2.1 NETL’s high-pressure combustor facility (B6 combustor)
11.2.2 Ansys fluent
11.2.2.1 Turbulence model
11.2.2.2 CFD reaction Eddy-dissipation model
11.2.2.3 CFD heat transfer model
11.2.2.3.1 Radiation model
11.2.3 Machine learning
11.2.3.1 Fuzzy clustering
11.2.3.2 Artificial neural networks
11.2.3.3 Artificial neural network performance evaluation metrics
11.2.3.4 Data batching
11.2.4 Previous work
11.3 B6 combustor model
11.3.1 B6 combustor problem definition
11.3.2 B6 combustor CFD simulation model
11.3.3 B6 smart proxy development overview
11.3.3.1 Data received from CFD simulation runs
11.3.3.2 Data visualization tool
11.3.3.3 Descriptive analytics
11.3.3.3.1 B6 combustor model sectioning
11.3.3.4 Predictive analytics
11.3.3.4.1 Data partitioning
11.3.3.4.2 Fuzzy clustering
11.3.3.4.3 Artificial neural network setup
11.3.3.4.4 Data batching
11.4 Smart proxy development steps
11.4.1 Model development step 1: Cell geometry and distances to wall boundaries
11.4.1.1 Model development step 1 – model training information
11.4.1.2 Model development step 1 – presentation of results
11.4.2 Model development step 2: Cell neighborhood, location, and Euclidian distances to wall boundaries
11.4.2.1 Model development step 2 – model training information
11.4.2.2 Model development step 2 – presentation of results
11.4.3 Model development step 3: Swirler distances
11.4.3.1 Model development step 3 – model training information
11.4.3.2 Model development step 3 – presentation of results
11.4.4 Model development step 4: Fuzzy clustering
11.4.4.1 Model development step 4 – model training information
11.4.4.2 Model development step 4 – presentation of results
11.5 Conclusions
11.5.1 Recommendations
11.6 Appendix
11.6.1 Model development step 3: Q1 combustor results – nitrogen
11.7 Acknowledgments
11.8 Disclaimer
References
Index