ML4Seismic Partners Meeting - 2022

Date:

Wednesday November 16, 2022 8:00 AM - 5:30 PM
Thursday November 17, 2022 8:00 AM - 5:30 PM
Friday November 18, 2022 8:30 AM - 5:00 PM

Venue:

The 2022 ML4Seismic Industry Partners Meeting will be held in person at the Atrium on the 9th floor of CODA Building at the Georgia Institute of Technology. The meeting is scheduled for November 16-18, 2022, 8:30 AM - 5:00 PM EST. Address: 756 West Peachtree Street NW, Atlanta, GA 30308.

Hotels:

Lodging options recommended by the institute (with Georgia Tech deals) can be found here.

More hotel choices can be found here.

Contact:

Ghassan AlRegib, co-director

Felix Herrmann, co-director

The 2022 ML4Seismic Industry Partners Meeting will be held in person at the Atrium on the 9th floor of CODA Building at the Georgia Institute of Technology. The meeting is scheduled for November 16-18, 2022, 8:30 AM - 5:00 PM EST. Address: 756 West Peachtree Street NW, Atlanta, GA 30308.

Center for Machine Learning for Seismic (ML4Seismic)

A joint initiative at the Georgia Institute of Technology between the Center for Energy & Geo Processing (CeGP) lead by professor Ghassan AlRegib (ECE) and the Seismic Laboratory for Imaging and Modeling (SLIM) lead by professor Felix J. Herrmann (EAS, CSE, ECE), innovators in the energy sector, and major Cloud providers.

Georgia Tech’s Center for Machine Learning for Seismic (ML4Seismic) is designed to foster research partnerships aimed to drive innovations in artificial-intelligence assisted seismic imaging, interpretation, analysis, and monitoring in the cloud.

Through training and collaboration, the ML4Seismic Industry Partners Program promotes novel approaches that balance new insights from machine learning with established approaches grounded in physics and geology. Areas of interest include, but are not limited to, low-environmental impact time-lapse acquisition, data-constrained image segmentation, classification, physics-constrained machine learning, and uncertainty quantification. These research areas are well aligned with Georgia Tech’s strengths in computational/data sciences and engineering.

ML4Seismic Industry Partners

Diamond Sponsors

Platinum Sponsors

Registered participants from Partner Companies:

  1. Chao Wang, Occidental Petroleum
  2. Daniel, Occidental Petroleum
  3. Klaas Koster, Occidental Petroleum
  4. Reza Rasteger, Occidental Petroleum
  5. Nikolaos Mitsakos, Occidental Petroleum
  6. Guangchi Xing, Chevron
  7. John Washbourne, Chevron
  8. Anatoly Baumstein, ExxonMobil
  9. Brent Wheelock, ExxonMobil
  10. Noah Vento, ExxonMobil
  11. Vanessa Goh, Shell
  12. Debanjan Datta, Shell
  13. Pandu Devarakota, Shell

Registered participants from Collaborators:

  1. Rishi Khan, Extreme Scale Solutions
  2. Petar Pirgov, Extreme Scale Solutions

Online participants:

  1. Hongbo Zhou, Equinor
  2. Mike Cogan, Equinor
  3. Sheng Xu, Equinor
  4. Meixia Wang, Equinor
  5. Bing Tang, Equinor
  6. Celine Ravaut, Equinor
  7. Philipp A. Witte, Microsoft (ex-SLIM)
  8. Yazeed Alaudah, Microsoft (ex-OLIVES)
  9. Antonio Serna, Occidental
  10. Bruce Power, Chevron
  11. Adam Halpert, Chevron
  12. Man Feng, ExxonMobil
  13. Valeriy Brytik, ExxonMobil
  14. Haiyang Wang, ExxonMobil
  15. Partha Routh, ExxonMobil
  16. Fuxian Song, ExxonMobil
  17. Jay Chen, Shell
  18. Xuefeng Shang, Shell
  19. Jan Stammeijer, Shell
  20. John Kimbro, Shell
  21. Side Jin, Shell
  22. Engin Alkan, Shell
  23. Satyakee Sen, Shell
  24. Ahmad Zamanian, Shell
  25. Farhad Bazargani, Shell
  26. Yang Xue, Shell
  27. Ying Ji, Shell
  28. Apurva Gala, Shell
  29. Tom Merrifield, Shell
  30. Hamish Macintyre, Shell
  31. Samuel Chambers, S & P Global
  32. Yong Ma, Aramco Americas
  33. Lei Fu, Aramco Americas
  34. Ali Almadan, Aramco
  35. Zvi Koren, AspenTech | Tel Aviv University
  36. Amir Shafiq, Apple
  37. Xiang Li, BGP (ex-SLIM)
  38. Wendel Lopes Moreira, Petrobras
  39. Sheng Dai, Georgia Tech
  40. Peng Chen, Georgia Tech
  41. Jinwoo Go, Georgia Tech
  42. Joyce Sim, Georgia Tech
  43. Motaz Alfarraj, King Fahd University of Petroleum and Minerals
  44. Konstantin Osypov, Halliburton
  45. Haibin Di, SLB
  46. Son Phan, SLB
  47. Wenyi Hu, SLB
  48. Shashin Sharan, SLB (ex-SLIM)
  49. Bingbing Sun, Saudi Aramco

Registered participants from Center for Energy and Geo Processing (CeGP/OLIVES) and Seismic Laboratory for Imaging and Modeling (SLIM):

  1. Felix J. Herrmann (co-director, ML4Seismic, SLIM)
  2. Ghassan Al-Regib (co-director, ML4Seismic, OLIVES)
  3. Ahmad Mustafa (PhD student, OLIVES)
  4. Jinsol Lee (PhD student, OLIVES)
  5. Kiran Kokilepersaud (PhD student, OLIVES)
  6. Ryan Benkert (PhD student, OLIVES)
  7. Yash-yee Logan (PhD student, OLIVES)
  8. Chen Zhou (PhD student, OLIVES)
  9. Rahul Jain (PhD student, OLIVES)
  10. Ghazal Kaviani (PhD student, OLIVES)
  11. Yavuz Yarici (PhD student, OLIVES) .
  12. Mohit Prabhushankar (Postdoctoral Fellow, OLIVES)
  13. Mathias Louboutin (Postdoctoral Fellow, SLIM)
  14. Ali Siahkoohi, Rice University (formerly SLIM)
  15. Yijun Zhang (PhD student, SLIM)
  16. Rafael Orozco (PhD student, SLIM)
  17. Ziyi (Francis) Yin (PhD student, SLIM)
  18. Thomas J. Grady II (PhD student, SLIM)
  19. Abhinav P. Gahlot (PhD student, SLIM)
  20. Huseyin Tuna Erdinc (Master student, SLIM)
  21. Haoyun Li (PhD student, SLIM)
  22. Srikanth Avasarala (PhD student, SLIM)
  23. Yadhu Kartha (PhD student, SLIM)
  24. Ting-ying (Rosen) Yu (Undergraduate student, SLIM)

2022 ML4Seismic Partners Meeting

The 2022 ML4Seismic Industry Partners Meeting will be held in person at the Atrium on the 9th floor of CODA Building at the Georgia Institute of Technology. The meeting is scheduled for November 16-18, 2022, 8:30 AM - 5:00 PM EST. Address: 756 West Peachtree Street NW, Atlanta, GA 30308.

Wednesday November 16

08:00—09:00 AM Everyone Breakfast (provided)
09:00—09:15 AM Felix J. Herrmann Introduction 1
09:15—09:25 AM Ghassan AlRegib Introduction 2
Theme: Explainability, Risk, and Trust in ML (Chair: Ahmad Mustafa)
09:25—09:45 AM Ahmad Mustafa Gathering Insights on DHI Attributes via Explainable Machine Learning
09:45—10:05 AM Huseyin (Tuna) Erdinc and Abhinav P. Gahlot Trustworthy CO2 Leakage Detection with Deep Neural Network Classifiers
10:05—10:25 AM Prithwijit Chowdhury A Causal Analysis of DHI Risk Assessment
10:25—10:45 AM Mohit Prabhushankar Manufacturing Trust in Seismic-Aware Neural Networks
10:45—11:00 AM Discussion
11:00—11:15 AM Break
Theme: Neural Operators & Seismic-Aware Networks (Chair: Thomas Grady)
11:15—11:35 AM Ziyi (Francis) Yin Uncertainty-aware Time-lapse CO2 Monitoring with Learned End-to-end Inversion
11:35—11:55 AM Thomas Grady Efficient Numerical Surrogates via Domain-Decomposed Fourier Neural Operators
11:55—12:15 PM Ziyi (Francis) Yin Amortized velocity continuation with Fourier neural operators
12:15—12:35 PM Discussion
12:35—02:05 PM Lunch (provided)
Theme: Human In the Loop Learning (HILL) (Chair: Kiran Kokilepersaud)
02:05—02:25 PM Kiran Kokilepersaud Contrastive Learning Opportunities in Seismic
02:25—02:45 PM Ryan Benkert Regression in Active Learning
02:45—03:05 PM Yash-yee Logan Deployable Active Learning
03:05—03:25 PM Ahmad Mustafa Robust Active Learning for Seismic Interpretation
03:25—03:40 PM Discussion
03:40—03:55 PM Break
03:55—05:25 PM Informal Sessions
5:30PM— Industry-Student Mixer @ “There on Fifth”
Table1Program for Wednesday November 16 of the ML4Seismic Partners Meeting

Thursday November 17

08:00—09:00 AM Everyone Breakfast (provided)
09:00—09:15 AM Felix J. Herrmann Introduction 1
09:15—09:25 AM Ghassan AlRegib Introduction 2
Theme: Seismic Methods & Software (Chair: Mathias Louboutin)
09:25—09:45 AM Yijun Zhang Time-lapse Seismic Survey Design by Maximizing the Spectral Gap
09:45—10:05 AM Mathias Louboutin Learned extensions for wave-based simulation and inversion
10:05—10:25 AM Mathias Louboutin ML4Seismic Open-Source Software: Updates & Developments
10:25—10:45 AM Rafael Orozco 3D Seismic Inverse Problem Regularization with Normalizing Flows
10:45—11:05 AM Discussion
11:05—11:20 AM Break
Theme: Uncertainty Quantification and Estimation (Chair: Ryan Benkert)
11:20—11:40 AM Chen Zhou Quantifying Human Label Uncertainty
11:40—12:00 PM Ryan Benkert Developing Reliable Uncertainty Estimates for Seismic Interpretation
12:00—12:20 PM Ali Siahkoohi Low-cost uncertainty quantification for large-scale inverse problems
12:20—12:40 PM Rafael Orozco Adjoint Operators as Summary Functions in Amortized Bayesian Inference Frameworks
12:40—01:00 PM Discussion
01:00—02:30 PM Lunch (provided)
Theme: Seismic Monitoring of Geological Carbon Storage (Chair: Ziyi (Francis) Yin)
02:30—02:50 PM Ziyi (Francis) Yin Simulation-based Framework for Geological Carbon Storage Monitoring
02:50—03:10 PM Felix J. Herrmann Meet our Digital Twin for Geological Carbon Storage
03:10—03:30 PM Ting-Ying (Rosen) Yu Monitoring with sequential Bayesian Inference
03:30—03:45 PM Discussion
03:45—04:00 PM Break
04:00—05:30 PM Informal Sessions
6:00PM— Dinner @ El Valle
Table2Program for Thursday November 17 of the ML4Seismic Partners Meeting

Abstracts

Introduction 1

Felix J. Herrmann

Abstract. During this presentation, an overview of the 2022 ML4Seismic Program will be given including organization of the meeting, setup of ML4Seismic, and the Informal Sessions in the afternoon.


Introduction 2

Ghassan AlRegib

Abstract. During this presentation, an overview of the 2022 ML4Seismic Program will be given including organization of the meeting, setup of ML4Seismic, and the Informal Sessions in the afternoon.


Theme: Explainability, Risk, and Trust in ML

Gathering Insights on DHI Attributes via Explainable Machine Learning

Ahmad Mustafa and Ghassan AlRegib, OLIVES

Abstract. Hydrocarbon risk assessment refers to the set of processes used by geophysicists to assign values for the probability of success to drillable prospects in a company’s portfolio after thoroughly examining pertinent data. Grading seismic amplitude anomalies termed as direct hydrocarbon indicators (DHIs) forms a major component of the work process. While significant research has been done to ascertain the significance of various DHIs to risk assessment, the works thus far are limited in terms of modeling higher order interrelationships. Whereas powerful AI models have demonstrated immense potential to extract latent representations from raw data to efficiently model various input-output relationships, their wider adoption has been hindered by their lack of explainability. Using the popular paradigm of explainable machine learning, we demonstrate the utility of using machine learning models to validate decisions made by machine learning models on a risk assessment dataset. In addition, we show such methods can be used to detect for bias in the data and reveal higher order correlations among DHI attributes.


Trustworthy CO2 Leakage Detection with Deep Neural Network Classifiers

Huseyin (Tuna) Erdinc and Abhinav P. Gahlot, SLIM

Abstract. With the global deployment of Carbon, capture and storage (CCS) technology to combat climate change, there is an associated risk of contamination with CO2 leaking back to the atmosphere. Thus, it requires continuous monitoring of CO2 after the injection stops at the storage site. In this work, we generated synthetic CO2 plume development data with both leakage and no leakage scenarios. We trained a convolutional neural network (CNN) discriminative classifier and also a generative classifier and compared their performances in CO2 leakage detection. The accuracy of our discriminative classifier on the test data is 85% and that of the generative classifier is 90%. The Class Activation Mapping (CAM) results of the discriminative classifier and the latent space representation of our dataset in the case of generative classifier strengthens our claims about trustworthy leakage classification.


A Causal Analysis of DHI Risk Assessment

Prithwijit Chowdhury, Ghassan AlRegib, Mohit Prabhushankar, and Ahmad Mustafa, OLIVES

Abstract. Finding association between a certain feature-set and the outcome in a high dimensional dataset for hydrocarbon detection, through straight-forward metric measurement of Neural Network (NN) performances caused by the observational data may not be a suitable approach due to the presence of unknown or unobserved correlation between the individual feature-set themselves. This may lead to unaccounted influences during the calculation of the association metrics. We put forward the idea of feature data for prospect risk assessment from a causal attribution perspective. In the interventionist definition of causality, we say that an event A causes another event B if we observe a difference in B’s value after changing A, keeping everything else constant. Causal analysis of a dataset allows us to find the exact effect a change in a desired feature-set might have on the outcome without having to consider the influence of the other features present, leading us towards better explainability and ultimately decision-making.


Manufacturing Trust in Seismic-Aware Neural Networks

Mohit Prabhushankar, Ahmad Mustafa, Ryan Benkert, and Prithwijit Chowdhury, OLIVES

Abstract. The nature of big-data in seismic interpretation is such that first-order principle based Artificial Intelligence solutions are insufficient to capture and analyze complex structures. Large-scale neural networks with millions of parameters have shown better task-based performance compared to interpretable decision and binary trees. Hence, interpretability and trust in neural networks must be manufactured after the training process. This is termed as post-hoc trust. In this talk, we present a taxonomy of post-hoc trust and identify three large categorizations. We detail each category and provide examples within them. The talk will be supported by a Tutorial session that details existing codes and software packages that quantifies individual categories of post-hoc trust.


Theme: Neural Operators & Seismic-Aware Networks

Uncertainty-aware Time-lapse CO2 Monitoring with Learned End-to-end Inversion

Ziyi (Francis) Yin, SLIM

Abstract. Seismic monitoring of CO2 sequestration is computationally expensive as it involves modeling of both fluid-flow physics modeling and wave physics and differentiation through the solvers with respect to the subsurface properties of interest. In this talk, we demonstrate the effectiveness of learned coupled inversion framework using a pre-trained Fourier neural operator as a learned surrogate for the fluid-flow simulator, which greatly reduces the cost associated with fluid-flow modeling and differentiation through the solver. We study the effectiveness and correctness of inversion based on Fourier neural operator surrogate and a normalizing flow prior. We also demonstrate the efficacy of this framework on monitoring the growth of CO2 plumes during sequestration, and on uncertainty quantification of the permeability and CO2 plumes with conditional normalizing flow. With this framework, we can further forecast the CO2 plume in the future without any acquired seismic data with uncertainty estimation.


Efficient Numerical Surrogates via Domain-Decomposed Fourier Neural Operators

Thomas Grady, SLIM

Abstract. Numerical surrogates are models which learn to mimic a complex physical process (such as the solution to a PDE produced by a solver) from a set of input/output pairs. Fourier neural operators (FNOs) are a specific type of numerical surrogate which use a learned matched filter to quickly approximate solutions to relatively smooth complex physical processes. In the case of carbon capture sequestration (CCS) technology, FNOs have been shown to well-approximate solutions to the two-phase flow equations, with speedups of 1,000 to 10,000 times at inference time versus a tradtitional solver. This speed combined with the fact that FNOs are differentiable with respect to their input parameters allows for inverse and uncertainty quantification problems to theoretically be solved on real 3D data, a previously intractible task. However, due to the size of the input data, network weights, and optimizer state, FNOs have thus far been limited to small to medium 2D and 3D problems, well below the size of an industry standard such as the Sleipner benchmark. Here we alleviate this problem by proposing a model-parallel FNO which makes use of domain decomposition of the input data and network weights, and exploits architectural features of FNOs to also include a natural form of asynchronous pipeline parallelism. Our network can scale to arbitrary problem sizes on CPU and GPU systems.


Amortized velocity continuation with Fourier neural operators

Ziyi (Francis) Yin, SLIM

Abstract. Velocity continuation aims to map the migration image using one background model to the image using another background model. It is of great importance to quantify the uncertainty in seismic imaging result from various background models. With Fourier neural operators as a learned surrogate, this continuation from a given background model to an unseen background model can be quite accurately estimated with near-zero cost. However, the limitation of the prior art is that the input background model and the survey area are assumed to be fixed. The main contribution of this work is to extend the Fourier neural operator surrogate to be amortized over different given background models and survey areas. We verify the effectiveness of our learned surrogates by a realistic example on different areas of Parihaka dataset against different background models.

Theme: Human In the Loop Learning (HILL)

Contrastive Learning Opportunities in Seismic

Kiran Kokilepersaud, Mohit Prabhushankar, and Ghassan AlRegib, OLIVES

Abstract. In seismic interpretation, pixel-level labels of various rock structures can be time-consuming and expensive to obtain. As a result, there oftentimes exists a non-trivial quantity of unlabeled data that is left unused simply because traditional deep learning methods rely on access to fully labeled volumes. To rectify this problem, contrastive learning approaches have been proposed that use a self-supervised methodology in order to learn useful representations from unlabeled data. However, traditional contrastive learning approaches are based on assumptions from the domain of natural images that do not make use of seismic context. In order to incorporate this context within contrastive learning, we propose a novel positive pair selection strategy based on the position of slices within a seismic volume. We show that the learned representations from our method out-perform a state of the art contrastive learning methodology in a semantic segmentation task.


Regression in Active Learning

Ryan Benkert, Mohit Prabhushankar, and Ghassan AlRegib, OLIVES

Abstract. Active learning has proven a valuable asset to cost-effective seismic interpretation. Instead of manually selecting singular sections for annotation, deep neural networks are leveraged for automating the labeling pipeline. Even though active learning has shown great promise, active learning algorithms fail when annotations are inaccurate or when only a few sections are available for training. In this work, we evaluate failure in active learning as performance degradation after each model update. We further discuss possible methodologies to limit degradation in active learning pipelines.


Deployable Active Learning

Yash-yee Logan, Mohit Prabhushankar, and Ghassan AlRegib, OLIVES

Abstract. Conventional machine learning systems that operate on natural images assume the presence of attributes within the images that lead to some decision. However, decisions in other domains, like medicine, are a result of attributes within medical diagnostic scans and electronic medical records (EMR). Hence, active learning techniques that are developed for natural images are insufficient for handling data from other domains. We focus on reducing this insufficiency by designing a deployable active learning framework within a bi-modal interface so as to add practicality to the paradigm. Our approach is a plug-in method that makes natural image based active learning algorithms generalize better and faster on imagery from other domains.


Robust Active Learning for Seismic Interpretation

Ahmad Mustafa and Ghassan AlRegib, OLIVES

Abstract. Active learning methods rely on an acquisition function to select unlabeled data samples to be added to the training dataset. However, noisy training data can hinder accurate sample characterization by the acquisition function, leading to selection of redundant data for labeling. We propose a framework whereby the manifold learnt by the network while training for active learning is made robust to small feature changes in the data, thereby improving sample characterization and selection. We demonstrate an application of the proposed method to seismic facies interpretation with promising initial findings.

Theme: Seismic Methods & Software

Time-lapse Seismic Survey Design by Maximizing the Spectral Gap

Yijun Zhang, SLIM

Abstract. While time-lapse seismic has been applied successfully to CO2 sequestration monitoring, it remains a challenging problem since replicated dense surveys come at a very high cost in the field. Wavefield reconstruction based on matrix completion (MC) from randomized subsampled data is an efficient way to reduce operational costs. This technique allows for accurate time-lapse reconstruction by employing the joint recovery model (JRM), which capitalizes on the fact that different vintages share a common component. However, combining JRM with optimal time-lapse acquisition survey design remains an unexplored area of research. In expander graph theory, spectral gap (SG) reveals the source-receiver layout connectivity and is related to reconstruction quality during MC. Building on these insights, we proposed a simulation free time-lapse survey design based on JRM that aims to get similar reconstructed quality without insisting on replicate surveys, which significantly reduces the cost in the field. This approach uses the simulated annealing algorithm to find subsampling masks for each vintage. Numerical experiments confirm a direct correlation between increased spectral gap and promising time-lapse reconstruction quality.


Learned extensions for wave-based simulation and inversion

Mathias Louboutin. SLIM

Abstract. We introduce a new method that explores velocities as an operator (extended velocities) for wave-equation based inversion. Through this extended formulation, we obtain the known benefits of working with subsurface offset volumes. The offset-dependence of these volumes has been studied in the linear case, i.e as part of extended Born scattering and extended least-squares reverse-time migration, but has been avoided for non-linear inversion due to computational concderns and challenges. By using techniques from randomized linear algebra, we will show that we can work with extended velocities for inversion while maintaining an acceptable computational cost much lower than solving one PDE per extended velocity model.


ML4Seismic Open-Source Software: Updates & Developments

Mathias Louboutin. SLIM

Abstract. Software is at the core of research and development in inverse problems. At SLIM, we have experience developing scalable and performant software, such as our legacy parallel MATLAB framework. With ML4Seismic, we are dedicated to build on this experience to develop HPC open source software (OSS) for the scientific community in collaboration with our partners. In this talk, we will describe our OSS Julia and Python environment, our high-level abstraction principles, and the range of solutions we offer for seismic processing and inversion and for machine learning. We will emphasize our aim to provide scalable software that can be easily applied to industrial problems.


3D Seismic Inverse Problem Regularization with Normalizing Flows

Rafael Orozco, SLIM

Abstract. We present the first known exploration of a normalizing flow (NF) for generative 3D volumes. First, we tackle computational issues surrounding the high dimensionality of our desired 3D volume output. This is of particular concern in normalizing flows since their invertibility constraint implies equal dimension of output and input. Our findings show that by “freezing” expensive layers we can efficiently train a normalizing flow on 3D volumes. Using this NF architecture, we train a generative model on volume sections of the 3D BG compass model. Our method produces visually plausible generative samples which are efficient to produce. We demonstrate its practical use by using our trained generative model as an implicit prior in a Maximum A Posteriori (MAP) framework. We evaluate this MAP framework by estimating the solution of a inverse problem in seismic imaging. Our method results in higher SNR estimates than the baseline and in less iterations, importantly saving the computational cost of evaluating the expensive 3D PDE solver during optimization. Finally, through scaling analysis of training cost, we show that NF convolutional layers allow this approach to scale favorably to larger volumes.

Theme: Uncertainty Quantification and Estimation

Quantifying Human Label Uncertainty

Chen Zhou, Mohit Prabhushankar, and Ghassan AlRegib, OLIVES

Abstract. In many scenarios, the raw data often comes with separate ambiguous labels collected from multiple human annotators. The lack of consensus between independent annotators leads to human label uncertainty. While there are numerous works to quantify uncertainty from a machine perspective, this human uncertainty is not always accurately captured. We find that not only existing machine uncertainty quantification methods do not deal with human label uncertainty, but also a few existing uncertainty metrics are unreliable. We quantify human label uncertainty from a data-centric perspective. The true labels for the training instances sharing similar features within a local manifold neighborhood are likely to be the same. Soft labels could then be obtained via the feature relation between the instances with uncertain labels and instances with label consensus. Such soft labels can be utilized as a metric to quantify the label uncertainty. In seismic interpretation, annotators might not achieve consensus due to subjectivity. The work can be extended to quantify the uncertainty in the seismic labeling processes.


Developing Reliable Uncertainty Estimates for Seismic Interpretation

Ryan Benkert, Mohit Prabhushankar, and Ghassan AlRegib, OLIVES

Abstract. In seismic interpretation, reliable uncertainty estimates are vital to ensure trust in the predictions of deep neural networks. However, conventional uncertainty estimation relies on the data representation directly which is inaccurate for seismic applications where annotated data is scarce. For this purpose, we reduce uncertainty estimation to representation shifts and decouple model calibration from the primary representation manifold. We empirically show that our uncertainty estimates are significantly more accurate and fine-grained than conventional techniques.


Low-cost uncertainty quantification for large-scale inverse problems

Ali Siahkoohi, SLIM

Abstract. Bayesian inference for large-scale inverse problems is challenged by the computationally costly forward operator evaluations during posterior distribution sampling. Recent advances in variational inference and deep learning reduce these costs by pretraining a neural network capable of sampling the posterior distribution for previously unseen observed data. In geophysical applications, however, the accuracy of these methods depends on sufficiently capturing subsurface variability through a training dataset, which is challenging given the heterogeneity of the Earth’s subsurface and our lack of access to it. Moreover, these methods may be unreliable in the presence of data distribution shifts, e.g., a change in the number of source experiments, noise distribution, or geological features to be imaged. As such, we present a solution that increases the robustness of deep-learning-based Bayesian inference approaches when faced with changes in data distribution. Our proposed method involves a physics-based adaptation to the latent distribution of a conditional normalizing flow that is pretrained to approximate the posterior distribution for previously unseen data. Instead of feeding standard Gaussian latent samples to the conditional normalizing flow, this method parameterizes the latent distribution by a Gaussian distribution with an unknown mean and diagonal covariance, estimated by minimizing the Kullback-Leibler divergence between predicted and true posterior distributions. This method is applicable to a wide range of inverse problems and has the potential to significantly reduce the costs of Bayesian variational inference. By means of a realistic seismic imaging example we demonstrate that the proposed latent distribution adaptation method mitigates the Bayesian inference errors induced by data distribution shifts, including shifts in the forward model and prior distribution.


Adjoint Operators as Summary Functions in Amortized Bayesian Inference Frameworks

Rafael Orozco, SLIM

Abstract. An important concern in seismic inverse problems is the large and varying size of observed data. The large size can cause computational cost concerns and its varying size (such as when changing receiver geometries) implies the need to rerun inference algorithms from scratch for each new observation. Motivated by these two problems, we take inspiration from the statistics literature which commonly relies on summary statistic of observed data. Summary statistic compress the observed data leaving only information needed for inference. In this work, we argue that the adjoint operator provides a natural candidate for a summary function in the context of physics-based inverse problems. We first mathematically show that for certain general assumptions transforming data under the adjoint operator defines a new conditional distribution which preserves the expectations of the original posterior. We validate our hypothesis by evaluating our framework in a learned amortized inference algorithm. Our seismic and medical synthetic experiments show computational gains and increased quality of point estimates using our framework. We discuss statistical metrics that show our learned posterior is well calibrated therefore justifying its use in uncertainty quantification.

Theme: Seismic Monitoring of Geological Carbon Storage

Simulation-based Framework for Geological Carbon Storage Monitoring

Ziyi (Francis) Yin, SLIM

Abstract. While various monitoring modalities exist to track the behavior of CO2 plumes to ensure safe operations and compliance with regulatory requirements, active 3D time-lapse seismic monitoring has proven superior but costly. At SLIM, we aim to reduce the operating costs by optimizing acquisition design, to help drive innovations in seismic monitoring acquisition design and imaging, and to test novel time-lapse acquisition and imaging technologies in silico at scale. In this talk, we will introduce our open-source software platform simulation-based monitoring design framework. We demonstrate how to make use of proxy models for seismic properties derived from real 3D imaged seismic and well data to conduct realistic synthetic geological carbon storage projects. Furthermore, we discuss our proposed sparse non-replicated seismic acquisition and cutting-edge methodology to recover the dense data or to directly image the sparse non-replicated via joint recovery model. This automatic workflow ends with deep neural classifiers to detect potential CO2 leakage over seal through pressure-induced fault openings. We envisage the development of an automatic workflow to handle the large number of continuously monitored CO2 injection sites needed to help combat climate change.


Meet our Digital Twin for Geological Carbon Storage

Felix J. Herrmann, SLIM

Abstract. By embracing recent developments in simulation-based Bayesian inference—i.e., the task of deriving statistical information from a system based on in silico simulations—we envisage the development of an uncertainty-aware Digital Twin for seismic monitoring of Geologic Carbon Storage (GCS). According to IBM, “A digital twin is a virtual representation of an object or system that spans its lifecycle, is updated from real-time data, and uses simulation, machine learning and reasoning to help decision-making”. This Digital Twin will be designed to maximally benefit from vastly improved abilities to simulate complex phenomena, including the development of CO2 plumes in saline aquifers, and from the ability of neural networks to learn by example as part of inference. For GCS, this means that systematic assessment of uncertainties now becomes possible when observing CO2 plumes from time-lapse geophysical data (e.g., seismic). Because the proposed Digital Twin’s neural networks are taught to produce samples from the probability distribution for the CO2 plume conditioned by the observed time-lapse data, this approach will provide access to this information on uncertainty. As part of ML4Seismic, we are working on various aspects regarding the development of the Digital Twin including: (i) capability to generate realistic time-lapse data in response to CO2 injection in large strongly heterogeneous reservoirs. This simulation framework will facilitate the design of high-fidelity monitoring systems and is unique since it uses proxy Earth models with realistic CO2 plumes and heterogeneity; (ii) An inversion framework capable of producing high-fidelity time-lapse images of CO2 plumes and reservoir properties from time-lapse data collected in response to CO2 injection; (iii) uncertainty-aware data-assimilation framework based on techniques from sequential Bayes and capable of rapidly producing high-fidelity CO2 plume forecasts that are consistent with observed time-lapse data; (iv) A scalable uncertainty-aware early warning system designed to safeguard CO2 injection operations built on the latest insights from interpretable and trustworthy (explainable and robust) machine learning. After describing how to build a Digital Twin for GSC, early results will be presented on the use of Fourier Neural Networks as surrogates for the two-phase flow equations, seismic monitoring with our joint recovery model, and the use of spectral ratio to design low-cost acquisitions for time-lapse seismic.}


Monitoring with sequential Bayesian Inference

Ting-Ying (Rosen) Yu, SLIM

Abstract. For this study, we apply sequential Bayesian inference to monitor the time evolution of subsurface flow of CO_2 from indirect acoustic measurements at the surface. Upon receiving new acoustic measurements, we infer the current state of the CO_2 by sampling from a learned posterior. Using the incoming data, we then perform online updates of the current posterior. This is accomplished by using the fluid flow model to advance the estimated state variable forward in time in order to update the learned posterior. With a synthetic experiment, we demonstrate this method can track the flow evolution accurately as measured by PSNR metrics. Since the posterior is a learned network, we can compute estimates faster than traditional least squares methods. This method can also quantify the uncertainty due to stochasticity in fluid flow model and the limited-azimuth imaging configuration.


Discussion

This time slot is designated to informal discussion and feedback. It is important for us to get input on the application and possible on the use of our work, and we would therefore greatly value your input. We hope that this format will be conducive to lively discussions.

Tutorials

Informal Sessions

Our research team will be available during this time for informal breakout sessions.

Industry-Student Mixer

This event provides opportunities to engage with both students and faculty for potential research projects, student recruitment, etc.

It will take place at “There on Fifth”, 22 5th St NW, Atlanta, GA 30308