ML4Seismic Industry Partners

Diamond Sponsors

• Occidental Petroleum

Platinum Sponsors

• Chevron Corporation
  
• ExxonMobil Corporation
  
• Shell plc
  
• Woodside

Guests

• Halliburton

Registered participants from Partner Companies:

1. Klaas Koster, Occidental Petroleum
2. Daniel De Lilla, Occidental Petroleum
3. Advait Balaji, Occidental Petroleum
4. Rafael Orozco, Occidental Petroleum
5. John Washbourne, Chevron Corporation
6. Anusha Sekar, Chevron Corporation
7. Guangchi Xing, Chevron Corporation
8. Brent Wheelock, ExxonMobil Corporation
9. Zhen Chen, ExxonMobil Corporation
10. Kun Wang, ExxonMobil Corporation
11. Arash Fathi, ExxonMobil Corporation
12. Ahmad Zamanian, Shell plc
13. Max Liu, Shell plc
14. Konstantin Osypov, Halliburton

Registered participants from Collaborators:

15. Mathias Louboutin, Devito Codes

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

16. Felix J. Herrmann (co-director, ML4Seismic, SLIM)
    
17. Ghassan Al-Regib (co-director, ML4Seismic, OLIVES)
    
18. Kiran Kokilepersaud, PhD student, OLIVES
    
19. Chen Zhou, PhD student, OLIVES
    
20. Mohit Prabhushankar, PhD student, OLIVES
21. Jorge Quesada, PhD student, OLIVES
22. Prithwijit Chowdhury, PhD student, OLIVES
23. Seulgi Kim, PhD student, OLIVES
24. Mohammad Alotaibi, PhD student, OLIVES
25. Zoe Fowler, PhD student, OLIVES
    
26. Ghazal Kaviani, PhD student, OLIVES
27. Sahil Mithani, PhD Student, OLIVES
28. William Stevens, PhD Student, OLIVES
29. Abdel Rahman Fuad Musleh, PhD Student, OLIVES
30. Zijun (Venice) Deng, PhD student, SLIM
31. Shiqin Zeng, PhD student, SLIM
32. Yunlin Zeng, PhD student, SLIM
33. Abhinav Prakash Gahlot, PhD student, SLIM
34. Huseyin Tuna Erdinc, PhD student, SLIM
35. Jayjay Jeongjin Park, MSc student, SLIM
36. Haoyun Li, PhD student, SLIM
37. Ipsita Bhar, PhD student, SLIM
38. Jorge Quesada, PhD Student, OLIVES

Program 2025 ML4Seismic Partners Meeting

The 2025 ML4Seismic Industry Partners Meeting will be held in person at the Georgia Institute of Technology. The meeting is scheduled for November 19-21, 2025. On November 19-20, the meeting will be held on the second floor (room 230) of the CODA Building at the Georgia Institute of Technology. On Friday, the tutorials will be held in rooms 232 (OLIVES) and 233 (SLIM) on the second floor as well. To go to the second floor, enter the Coda lobby and head to the escalators to head to room 230. The address of the CODA building is 756 W Peachtree St NW Atlanta, GA 30308.

For those of you wo can not make the meetings on November 19-20 in person, we provide remote access via this Teams meeting link.

Wednesday November 19

Thursday November 20

Friday November 21

Session 1 (Room 232)

Session 2 (Room 233)

Business meeting

Abstracts

Self-attention summary networks for velocity model building from common-image gathers

Shiqin Zeng, Abhinav Prakash Gahlot, Yunlin Zeng, Zijun (Venice) Deng and Felix J. Herrmann

Abstract. This study develops a self-attention-based summary network designed to efficiently encode and compress seismic information residing in subsurface-offset Common Image Gathers (CIGs) into low-dimensional latent representations suitable for conditional machine-learning enabled velocity-model building. The proposed network functions as a compact seismic encoder that captures global spatial dependencies through self-attention mechanisms while preserving essential local structural features critical for accurate reconstruction. When integrated within a flow matching generative network, the summary network learns to effectively summarize seismic information from CIGs via a new multi-scale attention mechanisms. This integration enables efficient training and inference, facilitating fast and physically consistent generation of seismic velocity models conditioned on CIGs.

Digital Shadow for CO₂-based Enhanced Oil Recovery (CO₂-EOR)

Haoyun Li,
Abhinav Prakash Gahlot,
Olav Møyner, and
Felix J. Herrmann, SLIM

Abstract. We present a Digital Shadow for CO₂-based Enhanced Oil Recovery (CO₂-EOR), a data-driven surrogate model that predicts reservoir dynamics by conditioning on time-lapse seismic observations. Built upon conditional normalizing flows, the Digital Shadow learns the posterior distribution of subsurface states—CO₂ and oil saturation—conditioned on seismic data generated through coupled multiphase flow and rock-physics modeling. The framework integrates high-fidelity simulations with JutulDarcy.jl, nonlinear seismic forward modeling and imaging with JUDI.jl, and nonlinear rock-physics mappings to relate saturation changes to changes in the acoustic properties. Trained on 128 stochastic realizations derived from a North Sea permeability distribution, the model achieves accurate posterior samples for the saturation fields with high structural similarity to the ground truth. This work demonstrates how integrating differentiable physics with generative modeling enables seismic-informed monitoring of CO₂ plume evolution and oil recovery, laying the foundation for a Digital Twin of CO₂-based EOR operations.

Risk-Aware Injectivity Control for Energy Storage

Haoyun Li,
Abhinav Prakash Gahlot, and
Felix J. Herrmann, SLIM

Abstract. We introduce a digital-twin–driven framework for real-time injectivity control that maximizes injected CO₂ while enforcing reservoir safety. The forward model is a multiphase flow simulator evolving reservoir states under injectivity controls q; time-lapse seismic acts as the observation operator to reduce uncertainty of the the state (reservoir’s saturation/pressure). Risk is handled via two complementary metrics: the Probability of Failure (POF) for exceedance likelihood and the less conservative CVaR (superquantile) for tail severity. For the optimization, we employ smooth surrogates—a logistic exceedance rate for POF and the Rockafellar–Uryasev formulation for CVaR—enabling gradient-based control with constraints. Within a finite-horizon Model Predictive Control (MPC) loop, we optimize a trajectory per posterior sample and aggregate the ensemble of optimal rates into an empirical (KDE-kernel) Cumulative Density Function (CDF) with binomial confidence bands; the final controlled injectivity q is chosen as the largest rate whose upper confidence bound satisfies the target risk level. Case studies demonstrate that POF-constrained optimization is more conservative than CVaR-based optimization, In addition, sensitivity of the policy to the CVaR tail level, α, is reduced compared to POF’s parameter sensitivity, allowing for adaptive, uncertainty-aware control that balances injectivity and fracture risk. The proposed framework also unifies time-lapse seismic-informed inference, risk metrics, and MPC into a practical approach for safe high-throughput underground energy storage.

SAGE – Subsurface modeling with AI-driven Geostatistical Extraction and evaluation on North Sea Data

Huseyin Tuna Erdinc, Ipsita Bhar, Thales Souza, Rafael Orozco, and Felix J. Herrmann, SLIM

Abstract. Traditional machine learning based approaches often depend on large, high-quality and complete datasets of 2D Earth propety models, which can be difficult to obtain in subsurface applications. In this study, we leverage score-based generative models to synthesize high-fidelity geophysical properties (velocity, density, etc.) trained on a combination of sparse but tied well logs and seismic lines without relying on densely sampled velocity models and adress the sampling-density differences by enabling coherent integration of these complementary sources of information. With experiments on multiple synthetic datasets from subsurface models featuring diverse geological structures (e.g., faults, salt bodies), we demonstrate that our approach can accurately capture long-range geological correlations and align well with unseen ground-truth properties. Finally, we highlight the potential of our method to generalize to field data by conducting fine-tuning on a curated dataset derived from real well-log and seismic data currated from the UK Data Repository for the North Sea. This study represents an important step toward the development of foundational models for inverting physical properties and provides inputs for applications such as full-waveform inference (WISE/ASPIRE), supervised learning algorithms, and enhanced seismic-based subsurface modeling.

Power-scaled Bayesian inference for velocity-model estimation

Huseyin Tuna Erdinc, Yunlin Zeng, Abhinav Prakash Gahlot, and Felix J. Herrmann, SLIM

Abstract. Score-based generative models have emerged as powerful tools for conditional generation in Bayesian inference. While promising, a common criticism of these methods is their reliance on strong, structured priors, which can dominate the inference process and limit the network’s dependence on observed data. To address this issue, we propose a modification to the generative sampling algorithm that incorporates power scaling of both priors and likelihoods. Our approach enables flexible control over the relative influence of the prior and likelihood without requiring retraining for different scaling configurations. We demonstrate this in the context of synthesizing seismic velocity models conditioned on imaged seismic data. By sampling from intermediate power posteriors, our method naturally supports sensitivity analysis and allows us to assess how varying prior and likelihood weights determines the behavior of the posterior. Through a comprehensive set of experiments, we examine the impact of the power parameters in three scenarios: scaling only the prior, scaling only the likelihood, and scaling both simultaneously. The results show that increasing the likelihood power up to a certain threshold improves the fidelity of posterior samples to the conditioning data, while reducing prior power increases structural diversity in the generated models. Furthermore, moderate likelihood scaling leads to reduced shot data residuals, demonstrating its effectiveness for posterior refinement.

Fisher-Informed Training of Neural Operators for Reliable PDE Inversion

Jeongjin (Jayjay) Park, Grant Bruer, Huseyin Tuna Erdinc, Richard Rex Arockiasamy, Nisha Chandramoorthy, and Felix J. Herrmann, SLIM

Abstract. Neural operators have shown strong performance for PDE-solution learning; however, their effectiveness in PDE-constrained optimization, such as inversion, remains limited, in part because inaccurately learned derivatives can mislead gradient-based updates. Standard training typically optimizes only the forward least-squares fit, so during iterative inversion the optimization variable can drift outside the training distribution, further degrading quality of derivative information and inversion performance. To address this limitation, we propose a training algorithm designed for inversion: the surrogate is trained not only to predict forward outputs but also to learn observation-relevant gradients, which leads to Fisher-information directed parameter updates. We evaluate forward and gradient accuracy, as well as the inversion trajectory, on problems governed by PDEs—Darcy flow and laminar (incompressible) Navier–Stokes—which are relevant to porous-media fluid flow and energy-storage applications. Our results suggest that training for gradient fidelity with forward accuracy unlocks a pathway to reliable, efficient, neural operator–based inversion.

Velocity Model Building with Jacobian-Informed Neural Operators

Jeongjin (Jayjay) Park, Huseyin Tuna Erdinc, and Felix J. Herrmann, SLIM

Abstract. Neural operators provide fast learned surrogates for forward PDE modeling. In seismic settings, given multiple representative ground-truth velocity models and associated background models, an amortized neural operator can produce seismic observations at near-zero marginal cost. Because inference is cheap, these surrogates can be embedded in inversion pipelines via automatic differentiation. In this work, we systematically analyze the accuracy of the learned gradients from the Neural operators and further improve inversion performance by explicitly incorporating a Jacobian-Informed objective during training. Our experiments based on synthetic velocity models highlight that Jacobian-Informed Neural Operator increase the accuracy of inversion trajectory and provide more scalable solution compared to traditional workflows such as full-waveform inversion.

Bridging the Acoustic–Elastic Gap in Seismic Inversion via Robust Summary Statistics

Zijun (Venice) Deng,
Abhinav Prakash Gahlot,
Shiqin Zeng, and
Felix J. Herrmann

Abstract. Simulation-based Bayesian inference (SBI) enables uncertainty quantification in seismic inversion but typically assumes that simulated and observed data follow the same physics. In reality, this assumption does not hold. The Earth is elastic, yet large-scale simulations are often acoustic due to the prohibitive cost of elastic modeling. This acoustic–elastic mismatch introduces model misspecification, leading to biased posterior estimates and reduced interpretability.

We propose a robust conditional generative framework that mitigates this mismatch by learning a shared representation between acoustic and elastic domains. The approach integrates a robust summary network within a conditional normalizing flow, trained primarily on acoustic simulations and guided by limited elastic supervision. It enables elastic-consistent posterior inference from inexpensive acoustic simulations, balancing computational efficiency with physical realism. This framework offers a scalable path toward uncertainty-aware subsurface characterization under model misspecification, bridging the gap between fast acoustic modeling and physically accurate elastic data.

Mitigating Rock-Physics Model Misspecification in Digital Shadows via Amortized Bayesian Inference

Abhinav Prakash Gahlot,
Ipsita Bhar, and
Felix J. Herrmann, SLIM

Abstract. Digital shadows for subsurface monitoring rely on simulation-based inference to map seismic observations to evolving reservoir states. However, when the simulations are generated using simplified or different rock physics models than those governing the observed data, model misspecification leads to biased posterior estimates and degraded uncertainty quantification. We address this challenge through Bayesian frameworks including rock-physics marginalization and a context-aware amortized inference using conditional normalizing flows. The framework learns a mapping from seismic responses to pressure and saturation while accounting for variability across different randomly drawn models for the rock-physics. To enhance sensitivity to fluid and stress effects, we also incorporate pressure-dependent rock physics into the seismic forward model and perform inference in the Radon domain of Common Image Gathers (CIGs) instead of conventional RTM images. Because fluids exhibit distinct angle-dependent signatures in AVA responses, Radon-transformed CIGs are expected to provide improved separation between pressure and saturation effects, leading to more accurate and physically consistent state estimates. Thus, it enables an uncertainty-aware, physics-informed digital shadows under rock-physics model mismatch.

Time-lapse Full-waveform Inversion with Uncertainty Quantification

Zijun (Venice) Deng, Rafael Orozco, Abhinav Prakash Gahlot, and
Felix J. Herrmann, SLIM

Abstract. Reliable management of underground energy storage systems, such as underground compressed Hydrogene or Air storage, rely on semi-continuous monitoring and rigorous quantification of subsurface uncertainty. Time-lapse seismic imaging provides valuable insights into fluid migration and storage integrity. However, traditional inversion techniques—particularly full-waveform inversion (FWI)—are inherently ill-posed and deterministic, limiting their applicability for uncertainty-aware decision-making. The Joint Recovery Method (JRM) improves consistency across multiple surveys by jointly reconstructing shared structures, but does not capture uncertainty. We introduce the Probabilistic Joint Recovery Method (πJRM), which embeds uncertainty estimation within a joint inversion framework. πJRM leverages a shared generative model to infer wave-equation–based posterior distributions for each monitoring survey, enabling probabilistic interpretation of time-lapse changes. Synthetic experiments demonstrate that πJRM accurately recovers dynamic reservoir evolution and provides uncertainty-aware reconstructions, offering a principled foundation for safe, efficient management of subsurface energy storage operations.

Reconstructing reservoir states from multimodal data via score-based generative models

Shiqin Zeng, Abhinav Prakash Gahlot, Haoyun Li, and Felix J. Herrmann, SLIM

Abstract. This study develops a score-based generative framework for reservoir simulation to reconstruct spatially varying permeability and saturation distributions in saline aquifers from sparse observations at two wells combined with time-lapse seismic data. We learn joint distributions of permeability, saturation, and pressure from high-fidelity multiphase-flow simulations, conditioned on observed imaged time-lapse seismic data. The framework provides a unified, any-subset inference interface that flexibly incorporates well logs and physics-based constraints to enforce mass balance and physically plausible behavior. An ablation without pressure on training yields markedly lower structural similarity within physically meaningful measurement ranges, underscoring the pressure’s importance for accurate reconstruction of the saturation. Our approach generalizes across varying geological settings and highlights the value of multimodal data fusion for practical reservoir management.

Seismic Dataset Curation from UK National Data Repository to Validate SAGE and WISE

Ipsita Bhar, Huseyin Tuna Erdinc, Thales Souza, Rafael Orozco, and Felix J. Herrmann, SLIM

Abstract. This work presents a curated seismic dataset pipeline developed for data residing in the UK National Data Repository (UK NDR) in support of advanced generative geophysical modeling and machine learning. The workflow integrates checkshots and 3D post-stack seismic data volumes to generate accurate depth-domain seismic sections. Custom interpolation algorithms are used to construct velocity models from checkshot data, which are then used for time-to-depth conversion with [OpendTect]https://dgbes.com/software/opendtect. The converted seismic cubes are subsequently processed in Python using [SegySAK]https://segysak.readthedocs.io/en/latest/, a python package, enabling efficient extraction and visualization of 2D lines. These datasets will be used to further validate the workflow of SAGE (Subsurface foundational model for AI-driven Geostatistical Extraction), which focuses on large-scale representation learning for the geosciences. The trained generative models will be used as input to WISE and ASPIRE that form foundation models for seismic inference designed to characterize the subsurface. This curated dataset pipeline streamlines seismic data preparation, enhances reproducibility, and bridges the gap between conventional geophysical workflows and emerging data-driven inference methods.

Predicting Subsurface CO₂ Saturation using Late Fusion of Seismic and Streaming Well Data

​Victor Henriksson, Abhinav Prakash Gahlot, and Felix J. Herrmann, SLIM

​Abstract. Determining subsurface flow of CO₂ in porous rock formations is a challenging task especially when it involves intergration of multimodal data that are collected at disparate timescales. In this work, we investigate how data fushion can be used to integrate multimodal data consisting on infrequently collected active-source seismic surveys and continuous streaming data of saturation measurements collected at monitoring wells. For the purpose, we propose a data fusion framework based on late fusion, which combines seismic images, computed from seismic surveys, with time-series data collected at the wells. By integrating the spatial information from seismic data with high-resolution but sparce spatial-temporal patterns from wells, we aim to better approximate the complex CO₂ flow patterns.

Multiscale Wavelet Score-based Posterior Approximations for Seismic Inversion

Ege Cirakman, Huseyin Tuna Erdinc, and Felix J. Herrmann, SLIM

Abstract. We present a cascaded conditional wavelet score–based posterior surrogate (CWSGM) for seismic inversion that generates samples of subsurface velocity fields conditioned on physics-derived summaries (e.g., RTM images). Through multi-scale wavelet-based whitening, we specifically address ill-conditioning of the velocity model’s covariance (due to its fractal-like 1/f power spectrum), whose long range correlations hamper training neural networks to approximate the score function. This strategy improves the conditioning of the score-learning problem and accelerates both training and sampling. Experiments on the North Sea BG Model demonstrate that CWSGM accurately captures long-range geological correlations while reducing GPU memory usage by approximately 50% and sampling time by 73% compared to a single-scale vanilla score model. Overall, our approach enables efficient multi-scale generation of subsurface velocities at lower computational cost and provides a foundation for exploring alternative multiscale transforms that are more optimally aligned with seismic observations.

Background-Conditioned Diffusion Inversion for Seismic Velocity Models

Yunlin Zeng Huseyin Tuna Erdinc Rafael Orozco Felix J. Herrmann, SLIM

Abstract. Accurate seismic imaging and velocity estimation are important for subsurface characterization, yet conventional full-waveform inversion remains computationally expensive and highly sensitive to initial velocity models. To address these challenges, we propose a simulation-based inference framework using conditional elucidated diffusion models for posterior velocity-model sampling. Our approach integrates both horizontal and vertical subsurface-offset common-image gathers to capture a wider range of reflector geometries, from gently dipping to steeply inclined layers. We further condition the model on the background-velocity field to improve generalization across diverse geological settings. Evaluations on the SEAM dataset, which includes complex salt geometries, demonstrate that this conditioning substantially improves performance, increasing SSIM from 0.717 to 0.733 and reducing RMSE from 0.381 km/s to 0.274 km/s. Uncertainty analysis further shows enhanced calibration, lowering uncertainty calibration error from 6.68 km/s to 3.91 km/s. These results confirm that our diffusion-based simulation framework enables robust seismic inversion with reliable uncertainty quantification.

Active Prompt Querying for Agentic AI in Seismic Tasks

Prithwijit Chowdhury, Mohit Prabhushankar, Ghassan AlRegib, OLIVES

Abstract. Transformer-based foundation models trained on large corpora now provide strong out of the box solutions for a wide range of tasks across multiple domains. However, their behavior and outputs become increasingly uncertain when the domain is difficult or the task is very specific, leading to fragile predictions. To address this, prompts have been added as an explicit control channel. Prompts not only let users interact with the system but also help guide these billion parameter models toward more appropriate responses without the need for fine tuning. In this presentation we i) explore how prompts interact with the data inside the model representation space while also informing one another about the task, and ii) provide a task and label agnostic robustness metric grounded in information theory and causality to tackle common problems of over prompting and mis-prompting in large vision models. Together, these contributions give users clearer levers to intervene, diagnose, and refine task specific outputs from the model while preserving the convenience of out of the box use.

Domain Shift & Model Generalization in Seismic Fault Segmentation

Prithwijit Chowdhury, Jorge Quesada, Mohit Prabhushankar, and Ghassan AlRegib, OLIVES

Abstract. We present a large-scale benchmark of over 200 fault segmentation models trained and evaluated across diverse seismic datasets to investigate how model capacity, training strategy, and data alignment affect generalization under domain shift. Our results show that fine-tuning remains effective when domains are similar but becomes unstable as distributional differences increase, while larger models exhibit greater adaptability than smaller ones. To complement traditional pixel and distance-based scores, we introduce a set of geometric and topological metrics that capture fault-level characteristics such as orientation consistency, tortuosity, and continuity. Our analysis reveals that models inherit structural biases from the datasets they are finetuned on, influencing the geometry of predicted faults beyond what conventional metrics capture. Taken together, the findings provide practical guidance for selecting pretraining and finetuning strategies, balancing model size and data similarity, and integrating structural metrics into the evaluation of deep-learning-assisted seismic interpretation workflows.

Stable Transfers: Domain Adaptation and Memory Preservation in Seismic Deep Learning

Jorge Quesada, Mohit Prabhushankar, and Ghassan AlRegib, OLIVES

Abstract. Deep-learning-based fault segmentation models often struggle to generalize across seismic datasets, suffering from domain shift and catastrophic forgetting. We study two complementary strategies to improve robustness in transfer settings: domain adaptation and regularization-based forgetting mitigation. Domain adaptation methods are shown to enhance model transferability when domains are strongly mismatched, but may reduce performance when the shift is moderate, underscoring the need for adaptive alignment strategies. Regularization approaches such as Elastic Weight Consolidation help preserve knowledge from prior domains, and are designed to stabilize performance during finetuning. Together, these case studies illustrate the trade-offs between adaptation and retention, and emphasize that optimal strategies depend jointly on dataset similarity and the degree of shift. Our analysis provides practical insights into how domain adaptation and forgetting mitigation can be systematically applied to improve the reliability of deep-learning-assisted fault delineation.

Gradient Features for Post-Training Model Selection and Label-Efficient Fault Detection in Seismic Volumes

Sahil Mithani, Mohit Prabhushankar, and Ghassan AlRegib, OLIVES

Abstract. In this talk, we propose a trained-model selection technique for zero-shot fault delineation in new seismic surveys. Deploying pretrained fault-segmentation networks are often hindered by label scarcity and data shifts. Fully supervised evaluation or fine-tuning on each new volume is costly and can be counterproductive, for example, fine-tuning on the Thebe field can degrade performance and trigger catastrophic forgetting under domain shift. We introduce a label-free, gradient-based evaluation pipeline that ranks pretrained models by their expected generalization. Without ground-truth labels, we compute a set of gradient-derived metrics from a forward/backward pass on unlabeled seismic slices. These metrics use a confounding target – a shifted, non-informative label – to probe how confidently and stably a model responds when it is not guided by true annotations. The resulting gradient signals serve as proxies for model quality. This enables post-training model selection,choosing the most promising fault detector for a new volume, with zero labels and no fine-tuning. We quantify the effectiveness of our approach across over 240 pretrained models evaluated on synthetic and real seismic datasets.

Revealing Scale Dependence in Self-Supervised Fault Segmentation

Jorge Quesada, Mohit Prabhushankar, and Ghassan AlRegib, OLIVES

Abstract. The high cost of manual fault interpretation in seismic analysis often leads to the label-constrained regime: abundant data but limited annotations. Self-supervised learning (SSL) has emerged as a promising solution in such settings, and while it has been shown to perform well for facies segmentation, it has not translated to fault delineation. In this work, we propose a scale-aware SSL strategy that embeds small-window extraction into the augmentation process, allowing models to “zoom in” on localized structural patterns. Across multiple real seismic datasets, this approach yields up to 13% improvements in segmentation accuracy under label constraints compared to standard SSL and supervised pipelines. By contrast, benefits are far less pronounced for larger-scale features such as facies. These findings highlight the importance of adapting SSL to the inherent scale of geological structures and demonstrate its potential to significantly enhance fault interpretation in real seismic data.

Understanding Attention: How Seismic Features are Attended in Transformer Models

Mohammad Alotaibi, and Ghassan AlRegib, OLIVES

Abstract. Transformers are increasingly used in seismic interpretation, but how seismic features attend to each other within attention maps is still not well understood. In this presentation, we analyze how different seismic structures interact through attention. We compare these behaviors with those in natural and medical images to understand differences in locality and globality. By studying which features attend to which, we aim to reveal how Transformers perceive the seismic subsurface and how this understanding can help adapt attention-based models to seismic data, even with limited labeled samples.

AI Robustness Certification in Subsurface Interpretation

Mohit Prabhushankar, and Ghassan AlRegib, OLIVES

Abstract. Foundation models are billion parameter large-scale neural networks that generalize across data domains and tasks. This generalizability allows the usage of foundation models as backbones for multifarious downstream applications and processing. Traditionally, model training involves minimizing some predefined empirical risk on a given task. By doing so, task-dependent minimal sufficient statistics (MSS) is extracted from the current data to make the inference. However, in foundation models, the task is unknown during training and hence, the notion of task-specific MSS at inference becomes degenerate. Hence, a fundamental aspect of optimization guarantees is unavailable. Paradoxically, the more a model trains, the less robust the outputs will be. In this talk, we show the following: (i) transitional information between layers must be preserved, (ii) task specific bias-variance tradeoffs must be replaced by dimensionality-mutual information tradeoffs.

Pseudo-labels as Signals: Retaining Informative Variability of Seismic Interpretation

Chen Zhou, and Ghassan AlRegib, OLIVES

Abstract. Interpretation disagreement provides informative variability in the understanding of data rather than mere noise. However, when modeling variable disagreement through latent representations, the representation space can suffer from dimensionality collapse, where only a subset of dimensions encodes meaningful disagreement. This phenomenon limits the model’s capability to capture diverse interpretations, especially underrepresented but critical interpretations. This talk explores the utilization of divergence- and information-based regularization to help retain the effective dimensionality of representation spaces, enabling models to better represent multiple plausible interpretations. These insights provide cross-domain implications. In the medical domain, preserving crucial yet rare inter-observer variability supports reliable diagnosis. In seismic interpretation, capturing subtle differences between plausible subsurface interpretations enriches understanding across distinct geophysical contexts.

Visualizing Uncertainty in Facies Segmentation by Tracking Epoch-wise Mutual Information

William Stevens, Mohit Prabhushankar, and Ghassan AlRegib, OLIVES

Abstract. Understanding how neural network representations evolve during training can offer a new lens into learning stability and model uncertainty that is not captured by traditional measures comparing representations with inputs or outputs. This work explores a layerwise, per-epoch, information theory approach using mutual information between sequential internal representations. By trying to understand learning as a dynamic flow through the representation space, this work has discovered trends such as early-layer stabilization and delayed convergence for deeper layers. These patterns give a promising baseline to help quantify uncertainty and effective capacity across training. In this talk, we visualize these representation dynamics and showcase model uncertainty. Looking forward, this framework will be applied to prompt-wise tracking in the Segment Anything Model (SAM) to augment uncertainty estimation and over-prompting indication in seismic imaging tasks such as salt dome identification.

Countering Multi-modal Representation Collapse through Rank-targeted Fusion

Seulgi Kim, Kiran Kokilepersaud, Mohit Prabhushankar, and Ghassan AlRegib, OLIVES

Abstract. Multi-modal fusion methods often suffer from two types of representation collapse: feature collapse where individual dimensions lose their discriminative power (as measured by eigenspectra), and modality collapse where one dominant modality overwhelms the other. Applications like human action anticipation that require fusing multifarious sensor data are hindered by both feature and modality collapse. However, existing methods attempt to counter feature collapse and modality collapse separately. This is because there is no unifying framework that efficiently addresses feature and modality collapse in conjunction. In this paper, we posit the utility of effective rank as an informative measure that can be utilized to quantify and counter both the representation collapses. We propose Rank-enhancing Token Fuser, a theoretically grounded fusion framework that selectively blends less informative features from one modality with complementary features from another modality. We show that our method increases the effective rank of the fused representation. To address modality collapse, we evaluate modality combinations that mutually increase each others’ effective rank. The talk presents lessons from fusing information-rich modalities like images and 3D point clouds against 1D measurements. The insights are applicable across domains, including seismic interpretation workflows.

Information Collapse in Deep Learning and Its Impact on Seismic Interpretation Workflows

Abdelrahman Musleh, Mohit Prabhushankar, and Ghassan AlRegib, OLIVES

Abstract. Information Collapse in Deep Learning appears in several forms, including Representation Collapse (where feature vectors of all samples map to an identical embedding), Dimensional Collapse (where feature vectors of all samples lie on a low-dimensional manifold), Model Collapse (where the model degrades by recursively training on its own generated or synthetic data, leading to overfitting or suboptimal solutions), and Neural Collapse (where features of a class converge to its class mean, class means are symmetrically arranged, and the classifier for each class aligns with its class mean). Each represents a distinct phenomenon where learned representations exhibit reduced variability or increased alignment, which may either enhance structure or degrade performance depending on context. This talk presents the literature and case studies of Information Collapse in seismic literature. Subsequently, it identifies suitable metrics for detecting and characterizing the occurrence and degree of each collapse type during training. Recognizing when and how these collapses emerge is crucial for understanding their implications on model generalization, interpretability, and stability. Looking forward, the insights will inform strategies for leveraging beneficial collapses while mitigating those that are detrimental for seismic interpretation workflows.

From Data-Centric to Human-Centric: Building Robust and Trustworthy Visual Intelligence

Ghassan AlRegib, and Mohit Prabhushankar, OLIVES

Abstract. Current paradigm in visual intelligence often prioritizes massive datasets and model scaling, leading to systems that are brittle, biased, and lack human-level understanding and transparency. This talk advocates for a crucial shift towards a human-centric approach. We will explore our ongoing research in visual AI that integrates explainability, robustness, uncertainty, privacy-preserving models, and multi-modal learning. Moving beyond optimizing for performance alone, we aim to build the next generation of visual intelligence systems that are inherently robust, responsible, trustworthy, and aligned with human expectations and real-world needs.

Tutorials

Understanding Attention: Comparing CNN and Transformer Locality and Globalness Across Natural, Seismic, and Medical Domains

Mohammad Alotaibi, andGhassan AlRegib,OLIVES Tutorial. In this tutorial, we compare Convolutional Neural Networks (CNNs) and Transformers to better understand how they process image information and optimize the balance between local and global features. We focus on the concept of attention in Transformers and how it differs from the local receptive fields in CNNs. By analyzing their behaviors across three domains: natural images, seismic data, and medical images, we highlight how attention helps the model capture both fine and large-scale structures. To gain deeper insight into the attention mechanism, we examine which features are attended to by each region of the image, allowing us to observe how the models balance local detail with global context. This understanding provides a clearer view of how attention operates across different data types and guides the adaptation of these models to new domains like seismic interpretation.

A Hands-on Tutorial on Prompt Interactions and Robustness in Vision Models

Prithwijit Chowdhury andGhassan AlRegib,OLIVES Tutorial. This tutorial treats prompting as an explicit control channel that lets users interact and steer billion parameter models without fine tuning. It examines how prompts engage the model representation space, including interactions between prompts, and introduces a task and label agnostic robustness metric grounded in information theory and causality to detect and mitigate over prompting and mis-prompting in large vision models.

The Return of Matrix Methods: Effective Rank to Evaluate Supervised, Self-supervised, and Multi-modal Learning

Mohit Prabhushankar, Seulgi Kim, Kiran Kokilepersaud, Jorge Quesada, andGhassan AlRegib,OLIVES Tutorial. A fundamental quantification of learning guarantees in machine learning involves bias-variance tradeoffs. However, task agnostic learning in foundation models does not allow this tradeoff due to the lack of labels. A recent strategy that evaluates learning looks at the matrix properties of the singular value projections of data on trained weights. In this tutorial, we elaborate on the effective rank metric for evaluating supervised, self-supervised and multi-modal learning. We provide case studies where effective rank showcases the: (i) differences between learning from synthetic data vs real data (F3 volume), (ii) learning from natural images vs computed biomedical images, and (iii) learning from multi-modal data when the data modalities have asymmetric information. –>

JUDIAgent: Intelligent Workflow Automation for Seismic Modeling

Haoyun Li, Abhinav Prakash Gahlot, and Felix J. Herrmann, SLIM

Tutorial. This tutorial introduces JUDIAgent, an intelligent agentic automation framework built on the JUDI platform for PDE-constrained seismic modeling and inversion. The first part of the session covers the core design of JUDI, showing how it enables scalable forward and adjoint simulations using high-level abstractions for sources, receivers, and operators. Participants will learn how to implement and differentiate seismic experiments on GPUs and CPUs.

The second part focuses on JUDIAgent, an agent-based system for automating acquisition setup and workflow orchestration. Through examples, participants will see how the agent configures survey geometries, boundary conditions, and experiment parameters, and manages modeling and inversion pipelines with minimal manual tuning.

The session concludes with a demonstration of adaptive experiment configuration, where the agent responds to simulation feedback to refine inputs and improve computational efficiency. Attendees will leave with hands-on experience using agentic tools to streamline and optimize seismic modeling workflows.

From Math to Agents: Hands-on Tutorial on Differentiable and Agentic Reservoir Simulation with JutulDarcy and JutulGPT

Jeongjin (Jayjay) Park, Huseyin Tuna Erdinc, and Felix J. Herrmann, SLIM

Tutorial. This tutorial provides a practical introduction to JutulDarcy, a fully differentiable reservoir simulator, and its LLM-based extension, JutulGPT, for building agent-driven reservoir simulations. The first part of the session covers the fundamental mathematical formulation behind JutulDarcy, showing how the governing equations of multiphase flow can be implemented and differentiated with JutulDarcy. Participants will learn how to compute gradients with respect to physical parameters, and how to run scalable GPU-based simulations, for example, on 3D CO₂ injection problems. The second part of the session explores how agents can be used to set up and interact with a reservoir simulator using JutulGPT. Participants will learn how to define the model itself, configure simulation scenarios, and analyze results such as CO₂ storage or energy recovery. Finally, the session demonstrates how agents can iteratively refine the setup, adjusting grid resolution, moving wells, or adding new features, based on analysis outcomes and goals. By the end of the tutorial, attendees will have a hands-on understanding of how to build differentiable and agent-driven reservoir simulations using JutulDarcy and JutulGPT.

Seismic Dataset Curation from UK National Data Repository to Validate SAGE and WISE

Ipsita Bhari, Huseyin Tuna Erdinc, Thales Souza, [Ziyi (Francis) Yin] (https://slim.gatech.edu/people/ziyi-yin) and Felix J. Herrmann, SLIM

Tutorial. This tutorial introduces Seismic Dataset Curation for ML. Because open-source, real-world seismic datasets are rare, we are collaborating with the UK National Data Repository (UK-NDR) to obtain raw data. These raw datasets are not directly trainable, so we curate them into ML-ready formats and plan to release the curated versions publicly for academic use. The curated datasets will be used to train SAGE (developed by Tuna) to validate his model’s performance. Outputs from SAGE can then be passed to WISE (developed by Francis) to validate his model. Successful training and inference across SAGE → WISE provide a practical test that the curated datasets are truly ML-ready.

The session concludes with a demonstration of Processing raw seismic data from UK-NDR into standardized, ML-ready format. The second half would focus on using the curated datasets to train SAGE, and then feeding SAGE outputs into WISE to evaluate end-to-end utility.