ML4Seismic Partners Meeting - 2023
Date:
Tuesday November 7, 2023 5:00 PM - 8:00 PM
Wednesday November 8, 2023 8:00 AM - 5:30 PM
Thursday November 9, 2023 8:30 AM - 5:00 PM
Venue:
The 2023 ML4Seismic Industry Partners Meeting will be held in person on the first floor (room 114) of the CODA Building at the Georgia Institute of Technology. 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
alregib@gatech.edu
Felix Herrmann, co-director
felix.herrmann@gatech.edu
The 2023 ML4Seismic Industry Partners Meeting will be held in person person on the first floor (room 114) of the CODA Building at the Georgia Institute of Technologyat the Georgia Institute of Technology. The meeting is scheduled for November 7—9, 2023.
Center for Machine Learning for Seismic (ML4Seismic)
A joint initiative at the Georgia Institute of Technology between Omni Lab for Intelligent Visual Engineering and Science (Olives) 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:
- Klaas Koster, Occidental Petroleum
- Chao Wang, Occidental Petroleum
- Reza Rasteger, Occidental Petroleum
- Daniel De Lilla, Occidental Petroleum
- Ahmad Mustafa, Occidental Petroleum
- John Washbourne, Chevron
- Tamas Nemeth, Chevron
- Anusha Sekar, Chevron
- Max Liu, Shell
- David Thanoon. Shell
Registered participants from Collaborators:
- Rishi Khan, Extreme Scale Solutions
- Peter Pirgov, Extreme Scale Solutions
- Mathias Louboutin, DevitoCodes
Registered participants from Center for Energy and Geo Processing (CeGP/OLIVES) and Seismic Laboratory for Imaging and Modeling (SLIM):
- Felix J. Herrmann (co-director, ML4Seismic, SLIM)
- Ghassan Al-Regib (co-director, ML4Seismic, OLIVES)
- Jinsol Lee (PhD student, OLIVES)
- Kiran Kokilepersaud (PhD student, OLIVES)
- Prithwijit Chowdhury (PhD student, OLIVES)
- Chen Zhou (PhD student, OLIVES)
- Yunlin Zeng (PhD student, SLIM)
- Ghazal Kaviani (PhD student, OLIVES)
- Yavuz Yarici (PhD student, OLIVES)
- Zoe Fowler (PhD student, OLIVES)
- Mohit Prabhushankar (Postdoctoral Fellow, OLIVES)
- Rafael Orozco (PhD student, SLIM)
- Ziyi (Francis) Yin (PhD student, SLIM)
- Abhinav P. Gahlot (PhD student, SLIM)
- Huseyin Tuna Erdinc (Master student, SLIM)
- Haoyun Li (PhD student, SLIM)
- Srikanth Avasarala (PhD student, SLIM)
- Yadhu Kartha (PhD student, SLIM)
- Richard Rex Arockiasamy (MSc student, SLIM)
- Grant Bruer (PhD student, SLIM)
- Seulgi Kim (PhD student, OLIVES)
- Mohammad Alotaibi (PhD student, OLIVES)
- Jorge Quesada (PhD student, OLIVES)
- Ryan Benkert (PhD student, OLIVES)
- Thales Souza (MSc student, SLIM)
Program 2023 ML4Seismic Partners Meeting
The 2023 ML4Seismic Industry Partners Meeting will be held in person on the first floor (room 114) of the CODA Building at the Georgia Institute of Technology. The meeting is scheduled for November 7-9, 2023, 8:30 AM - 5:00 PM EST. Address: 756 West Peachtree Street NW, Atlanta, GA 30308.
Tuesday November 7
05:00—08:00 PM | Industry-student mixer | El Burro Pollo in the CODA Collective Food Hall |
Wednesday November 8
Thursday November 9
Abstracts
Effective Data Selection for Seismic Interpretation through Disagreement
Ryan Benkert, Mohit Prabhushankar, and Ghassan AlRegib, OLIVES
Abstract. We discuss data selection for deep learning in seismic interpretation. Specifically, which samples are the most informative to generalize well to the target volume. For the interpretation of seismic volumes, selecting the training set from a target volume is a crucial component that determines the success of the deep learning algorithm. In this work, we advocate for the integration of interpretation disagreement as an intuitive and beneficial factor in the training set selection process. Drawing inspiration from established practices in seismic interpretation, we develop a novel data selection framework. Our framework leverages representation shifts to model interpretation disagreement within neural networks, and utilizes the disagreement metric to incorporate attention to geologically interesting regions during the data selection workflow. By combining this approach with active learning, a well-known machine learning paradigm for data selection, we present a comprehensive and innovative framework for training set selection in seismic interpretation. Furthermore, we provide a concrete implementation of our proposed framework, named ATLAS, as a method for data selection. Through extensive experimentation, we demonstrate that ATLAS consistently outperforms conventional active learning frameworks in the application of seismic interpretation, achieving improvements up to 10% in mean intersection-over-union.
Finding What You Want: Prompting Foundation Models for Zero Shot Segmentation
Prithwijit Chowdhury, Mohit Prabhushankar, Ahmad Mustafa and Ghassan AlRegib, OLIVES
Abstract. A key hindrance deep learning models face in their applications in geophysics is their inability to generalize on downstream tasks. This problem arises due to the inconsistencies in the seismic and geological data of different places as well as the interpretations deduced by groups of experts. Models trained on a particular dataset with a specific interpretation cannot generalize to a different set of data even if the task is the same. Foundation Models are multimillion parameter large scale models trained on huge corpuses of data which have proved to be useful and efficient in generalization tasks involving natural images. We aim to adapt their generalization capabilities to perform zero shot inferences, comparable to other pretrained models, on seismic images as well by using prompt interventions. We plan to showcase the SAM model’s behavior on seismic tasks.
A New Seismic Fault Label Uncertainty Dataset: Insights from Expertise, Certainty, and Consistency
Jorge Quesada, Chen Zhou, Mohit Prabhushankar, and Ghassan AlRegib, OLIVES
Abstract. We present the Seismic Fault Label Uncertainty Dataset, a comprehensive resource comprising 400 seismic images extracted from the Netherlands F3 block, which represents a valuable contribution to the seismic community, offering unique insights into fault labeling uncertainty through a multi-faceted approach. To create this dataset, we harnessed the collective efforts of 30 annotators across varying levels of expertise, including novice, mid-level, and expert annotators, facilitated by the Amazon Mechanical Turk platform. Each annotator evaluated the presence of seismic faults in these images and assigned labels with three distinct certainty levels: 0 (indicating absolute certainty of no fault), 1 (expressing moderate certainty of fault presence), and 2 (reflecting a high level of certainty regarding the presence of a fault). We have meticulously curated annotations to gain insights into labeling consistency, annotator confidence, labeling speed, and other relevant metrics. Researchers can leverage the variability in expertise and certainty levels to develop robust machine learning models, refine annotation guidelines, and explore strategies for handling uncertainty in seismic data interpretation. This dataset opens new avenues for improving the reliability and accuracy of fault detection systems, ultimately contributing to a deeper understanding of subsurface geological structures and enhancing decision-making in geological exploration and risk assessment.
On the Feedback between Expert and Machines in Seismic Active Learning
Kiran Kokilepersaud, Mohammed Alotaibi, Mohit Prabhushankar, and Ghassan AlRegib, OLIVES
Abstract. Active learning is the process of identifying optimal samples to achieve high performance while minimizing the necessary amount of annotation work. One issue with this setup is that the geophysicist cannot interact with the model during the sample selection process. Instead, they are forced to simply label what the model selects without integrating their expertise into the active learning workflow. This results in a discrepancy between an expert’s understanding of informative samples and that of the trained neural network. To address these issues, we analyze the effect of modifying active learning into a machine-expert system where the geophysicist and model act as feedback on each other’s selection decisions. We compare these approaches against the traditional seismic annotation and machine learning workflows.
What Uncertainties Do We Need in Deep Learning for Seismic Interpretation
Ryan Benkert, Mohit Prabhushankar, and Ghassan AlRegib, OLIVES
Abstract. In this talk, we discuss different sources of uncertainty when training deep neural networks for seismic interpretation. In particular, we provide a framework based on disagreement to develop theoretically grounded uncertainty estimators for seismic interpretation. We advocate for three uncertainty sources for deep learning models in seismic interpretation: The first source stems from the uncertainty in the input data; a type we refer to as data uncertainty. For seismic interpretation, we encounter data uncertainties in migration artifacts or lacking trace information under salt bodies. In machine learning literature, the uncertainty is frequently referred to as aleatoric. The second is of algorithmic nature (i.e. algorithmic uncertainty). Specifically, uncertainty arising from the stochasticity or lack of training data in deep neural networks. In practice, this arises from the scarce availability of training annotations and stochastic factors in the training setup. In literature, the source is referred to as epistemic. The third source of uncertainty arises from different interpretations of the data; a mechanism we refer to as interpretation uncertainty. Interpretation uncertainty refers to uncertainty in the labels and the stochasticity of different expert opinions with respect to the same data.
A Counterfactual Analysis of Interpretations for High Dimensional Tabular Data
Prithwijit Chowdhury, Mohit Prabhushankar, Ahmad Mustafa and Ghassan AlRegib, OLIVES
Abstract. In geophysics, hydrocarbon prospect risking involves assessing the risks associated with hydrocarbon exploration by integrating data from various sources. In recent days, ML classifiers trained on tabular data are used to make efficient and fast decisions for this kind of prospect risking tasks. The lack of transparency in the decision-making processes of these models has led to the emergence of explainable ai (XAI). However, discrepancies exist among XAI programs, raising concerns about their accuracy. The notion of what an “important” and “relevant” feature is, is different for different explanation strategies. Thus grounding them using causal probabilistic notions of necessity and sufficiency can prove to be a reliable way to increase their trustworthiness. We propose a novel approach to quantify these two concepts in order to explore which explanation method might be suitable for tasks involving the implementation of sparse high dimensional tabular datasets.
Exploiting Structures of Data for Application Specific Representations
Kiran Kokilepersaud, Yavuz Yarici, Mohit Prabhushankar, and Ghassan AlRegib, OLIVES
Abstract. Deep learning relies on producing representations of data that can generalize to downstream tasks of interest. Good representations should reflect the underlying structure of the dataset. Datasets are oftentimes organized in a hierarchical structure where a variety of component distributions reflect the relationship between instances. For example, within the medical domain the data distribution has component distributions that reflect both biomarkers and clinical data. Within autonomous vehicle settings, data reflects the interaction between semantic distributions and distributions related to sensor acquisition noise. Within the natural image domain, data is structured based on species taxonomies. Despite the existence of these structures, conventional strategies do not leverage these underlying distributions during the representation learning process. We show how to integrate this information within a variety of applications related to fisheye sensor data and clinical meta-data. We then discuss potential ways these concepts can also be applied within seismic deep learning problems.
Interventionist Uncertainty in Neural Networks
Mohit Prabhushankar, and Ghassan AlRegib, OLIVES
Abstract. Deep learning has shown a high degree of applicability in multiple fields that provide access to big data. As deep learning-based AI systems transition from academia to everyday life, their vulnerabilities must be understood before acceptance by the public. A key vulnerability is the lack of knowledge regarding a neural network’s operational limits. This vulnerability is viewed as a lack of robustness in neural networks. In the past, simple measures of robustness that have served the research community include input noise analysis and out-of-distribution recognition. However, large vision models with billions of parameters are vulnerable to other engineered and adversarial noise as well as being uncalibrated with regards to their prediction probabilities. Recently, prompt-based architectures that allow limited interventions from users during the inference stage have gained prominence. In this talk, we discuss the uncertainty associated with these interventions, for the application of facies classification. Specifically, we show that interventions are designed to only, partially, reduce predictive uncertainty. Additionally, we provide a quantification of trustworthiness as a function of interventionist uncertainty.
Attribute-based Model Training under Annotator Label Uncertainty
Chen Zhou, Mohit Prabhushankar, and Ghassan AlRegib, OLIVES
Abstract. Annotators exhibit disagreement during data labeling, which can manifests in variations of labeling quality. Training with low-quality annotations induces model reliability degradation. In this work, we aim to achieve the model’s performance as if annotations were expert-level while using novice-level annotations. Prior methods associate ambiguous samples with training annotations collected from different annotators to enhance generalizability. To avoid collecting massive annotations, we generate multiple plausible labels for ambiguous data in an objective manner. In the geophysics domain, seismic attributes such as texture and saliency, can be utilized to identify ambiguous samples. Hence, we integrate attributes within a multi-label training framework to enhance generalizability and prediction uncertainty. Through the experiments on fault prediction, we demonstrate that the proposed framework improves generalizability and prediction uncertainty of the model to approach the performance when using expert-level annotations.
Modeling Visual Attention for Seismic Interpretation with Incomplete Labels
Ahmad Mustafa, Reza Rastegar, Timothy Brown, Gregory Nunes, Daniel DeLilla, and Ghassan AlRegib, OLIVES
Abstract. Accurate interpretation of visual data for relevant information forms an important component of many real-world applications such as medical disease diagnosis, geological hazard assessment, hydrocarbon exploration, etc. Producing fine-grained annotations on images is an expensive, laborious, and time-consuming process. The human brain is wired to selectively focus its attention on certain aspects of the visual scene. This perception mechanism is driven both by low-level signal cues, such as changes in color, contrast, intensity, shapes etc., as well as high-level cognitive factors such as one’s prior knowledge, goals, expectations, and constraints with respect to the task at hand. These attentional factors, referred to as bottom-up and top-down attention respectively, play an important role in determining the final annotations that get produced for a given visual scene, often at the cost of leaving out a lot of visual information the brain deems to be unimportant with regard to the task of interest. Mapping geological faults on 3D seismic volumes is one such application where human attention selectivity results in highly incomplete fault annotations. Conventional supervised learning methods treat regions of missed fault labels as negatives, resulting in non-optimal learning for the machine learning model. We propose a method to model visual attention and incorporate it into data sampling and model training procedures. We demonstrate the utility of this approach for mapping faults on seismic volumes using pretrained 3D convolutional neural networks (CNNs). Using an annotated seismic dataset from NW Australia, we show quantitatively and qualitatively that modeling visual attention leads to significant performance gains even with limited, incompletely labeled seismic training data.
Non-linear joint recovery model for time-lapse seismic monitoring of CO2 plumes
Abhinav P. Gahlot and Felix J. Herrmann, SLIM
Abstract. During time-lapse seismic monitoring of CO2 plumes, a weak 4D signal below the level of inversion or migration artifacts poses challenges. To address these, low-cost randomized non-replicated acquisitions and a linear joint recovery model (JRM) have been introduced. It takes advantage of the shared information between different vintages in the time-lapse seismic data and subsurface structure undergoing localized changes. Since the relationship between seismic data and subsurface properties is seldom linear, we propose a more versatile nonlinear JRM (nJRM) to invert for the squared slowness of the vintages. The nJRM takes advantage of the full nonlinear relation between these squared slownesses and time-lapse data through the wave equation. Also, careful derivation of the gradients makes the computational cost of nJRM equivalent to the independent recovery. We present a synthetic study for geological carbon storage (GCS) which shows that the non-replication can be beneficial to time-lapse imaging, making seismic monitoring of GCS less costly for the long term sustainability of the technology.
Monitoring geological carbon storage with learned sequential Bayesian inference
Abhinav P. Gahlot, Ting-Ying (Rosen) Yu, Rafael Orozco, and Felix J. Herrmann, SLIM
Abstract. Reservoir engineers frequently employ two-phase flow simulations and history-matching to oversee and anticipate the behavior of CO2 plumes within geological carbon storage. These simulations, while valuable for gaining insights, face limitations due to several complex factors, such as uncertainties surrounding the plume’s dynamics. To investigate this phenomenon more comprehensively, we introduce the concept of stochasticity in the dynamics, accounting for uncertainties in the underlying permeability of the reservoir. To enhance the accuracy of CO2 plume predictions and quantify the uncertainties involved, we utilize machine learning techniques to condition these predictions on time-lapse seismic and well observations. This framework works on the principle of sequential Bayesian inference that continuously assimilates information from time-lapse observations, updates the CO2 plume predictions, and characterizes uncertainties about the plumes.
End-to-end permeability inversion from time-lapse seismic data: a case study on Compass model
Ziyi (Francis) Yin and Felix J. Herrmann, SLIM
Abstract. Effective geological carbon storage hinges on a deep understanding of CO2 plume behavior. The dynamics of these plumes can be modeled using multiphase flow equations, but their accuracy is tied to a precise permeability model. A significant challenge is that we often lack detailed permeability data, limiting our predictive capabilities. To bridge this gap, we’ve developed a multiphysics inversion method. This technique inverts for the permeability from observed time-lapse seismic data. Through a case study on the Compass model, we’ve compared this approach with traditional 4D FWI in forecasting CO2 plume movements. Additionally, our research delves into how different initial permeability models, acquisition setups, and survey frequencies affect the results. Across the board, the inversion method not only enhances our current estimations but also provides valuable insights into future plume dynamics, even without continuous monitoring.
Maximizing CO2 injectivity within fracture pressure via ML framework
Haoyun Li, Ziyi (Francis) Yin and Felix J. Herrmann, SLIM
Abstract. In geological carbon storage projects, optimizing CO2 injection strategies is paramount to enhance storage efficiency and prevent leakage. The objective is to maximize the CO2 injection volume without surpassing the fracture pressure. Traditional adjoint-based approaches necessitate extensive numerical simulations, leading to significant computational overhead. To circumvent this challenge, we introduce an optimization framework based on physics-informed deep convolutional neural networks. Trained on different permeability slices, our model can rapidly predict the maximal CO2 injection volume for new permeability fields in real-time.
Towards generative Earth model building from borehole wells and seismic data
Rafael Orozco and Felix J. Herrmann, SLIM
Abstract. Our goal is to build realistic parameterized (acoustic, velocity, permeability etc) earth models where the training and testing phase of our method uses only data that is available in the field. We first demonstrate the expressive power of normalizing flows to generate detailed realistic earth models by training on supervised pairs of full earth models and borehole wells. Our results are compared with traditional variogram kriging to show that our generated models can be used in parameterizations of various downstream tasks such as simulations of realistic acoustic waves and fluid flow for reservoir simulations. Then we introduce a novel unsupervised training objective that can train normalizing flows to generate full earth models without needing training pairs of the full earth models. By using a known proxy earth model as a testbed, we make preliminary prescriptions on how many wells our method needs to generate permissible earth models in a target area.
Large-Scale Parametric PDE Approximations with Model-Parallel Fourier Neural Operators
Richard Rex and Felix J. Herrmann, SLIM
Abstract. Solving PDEs to simulate two-phase flow is expensive since it involves the inversion of large ill-conditioned matrices. FNOs represent a special type of neural network capable of approximating solutions to two-phase flow equations. In order to speed up this computation, we develop a high-level software abstraction tool to exploit the linearly separable property of Fourier Transforms via Kronecker products. We perform a series of all-to-all operations where we partition the data to apply the operations in a distributed fashion. We apply these FNOs to predict the evolution of CO2-plumes in subsurface environments. Our model takes an input permeability model, and outputs time-varying CO2 saturations in a quick and cost-effective manner. Additionally, our research involves developing a distributed matrix-free abstraction library that can be used to represent any generic linear and nonlinear operator. This library is scalable and auto-differentiable thanks to the hand-written customized AD rules, allowing us to represent and train any network. We provide an evaluation on the accuracy of FNO simulations compared to traditional PDE simulations in solving various classes of PDEs.
Improved automatic seismic CO2 leakage detection via dataset augmentation
Huseyin (Tuna) Erdinc and Felix J. Herrmann, SLIM
Abstract. Previous works showed that neural classifiers can be trained to detect CO2 leakage from time-lapse seismic images. While this result is crucial to the global deployment of geological carbon storage (GCS), its success depends on relatively dense non-replicated time-lapse data acquisition. In this study, we present an approach to enhance the detection accuracy and robustness of CO2 leakage detection by augmenting the training dataset with a variety of coarsely sampled receiver data and their corresponding receiver numbers. This augmentation strategy is particularly beneficial for scenarios where low-cost coarse receiver samplings, such as with ocean bottom nodes (OBNs), are utilized. Furthermore, we explore interpretability of the classifier’s decisions by generating saliency maps for further analysis.
Extended full-waveform inference
Ziyi (Francis) Yin, Rafael Orozco, Mathias Louboutin, and Felix J. Herrmann, SLIM
Abstract. Quantifying uncertainty in full-waveform inversion is complex given the large sizes of both the model and data. A previous approach employed a variational inference framework, leveraging reverse-time migration to summarize observed data and approximate the posterior distribution through conditional normalizing flows. While reverse-time migration effectively summarizes the data when the background model is close to the true one, its accuracy diminishes with a less accurate background model. In our study, we suggest utilizing subsurface offset gathers as the summary statistics for the variational inference of full-waveform inversion. These gathers retain all the information in seismic data, even when the background model is cycle-skipped or fails to flatten the gathers. Through a case study on Compass model, we confirm our framework’s effectiveness and show that subsurface offset gathers offer a better summary statistic than just reverse-time migration.
CO2 reservoir monitoring through Bayesian data assimilation
Grant Bruer, Edmond Chow, and Felix J. Herrmann, SLIM
Abstract. Carbon capture and storage can be implemented by injecting supercritical carbon dioxide (CO2) into geological carbon reservoirs for long-term containment. Monitoring the pressure and saturation of the CO2 is necessary to optimize the injection amount without causing CO2 leakage or seismic activity. Directly measuring the CO2 at locations within the reservoir requires expensive drilling procedures that may damage the reservoir, so direct measurements are sparse and usually lie along injection and production wells. Indirect measurements such as seismic data are typically noisy, and inverting for the CO2 state is ill-posed. Bayesian data assimilation techniques allow us to integrate known physics for CO2 flow into this inversion process. The most well-established data assimilation algorithms are the family of Kalman filters. The ensemble Kalman filter is designed to efficiently work with large problem sizes and nonlinearity. In this work, we apply the ensemble Kalman filter to seismic measurements of a CO2 reservoir, yielding an estimate of the saturation and pressure fields with quantified uncertainties. This method models the CO2 plume state as a random field with a known distribution and assimilates information from seismic measurements with information from a physics model describing the CO2 flow. We show that the data assimilation strategy is a valuable contribution to advancing reservoir monitoring technology.
Uncertainty quantification…so what? Experimental design leveraged by probabilistic seismic inversion
Rafael Orozco and Felix J. Herrmann, SLIM
Abstract. Combining physics with recent developments of generative machine learning enables a scalable probabilistic framework for tackling seismic inversion problems including Full-Waveform Inversion. These probabilistic results can be proven to be from the Bayesian posterior but how exactly can we use them for practical downstream tasks? In this talk, we answer the question with a practical application of the probabilistic framework towards designing ocean bottom node placement of seismic imaging. With a simple adjustment to the original training objective, we show that jointly optimizing for an experimental design corresponds to maximizing the expected information gain used by the Bayesian community. After verifying this novel joint optimization with a stylized problem, we demonstrate its application for optimizing the placement of ocean bottom nodes in a synthetic seismic imaging experiment.