Program 2026 ML4Seismic Partners Meeting

The 2026 ML4Seismic Industry Partners Meeting will be held in person at the Georgia Institute of Technology. The meeting is scheduled for November 11-12, 2026. On both days, the meeting, including the tutorials on the afternoon of November 12, will be held on the second floor (room 230) of the CODA Building at the Georgia Institute of Technology. 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.

Wednesday November 11

Thursday November 12

Abstracts

3D Masked Autoencoders for Self-Supervised Pretraining of Seismic Foundation Models

Abdelrahman Musleh,
Araz Shafiyev, and
Ghassan AlRegib, OLIVES

Abstract. Seismic data is three-dimensional, so it is natural to learn representations directly from 3D volumes when building seismic foundation models. Faults, horizons, and salt bodies are volumetric; hence, a model that sees context along all three axes can use that structure. Learning from full volumes, however, runs into a practical problem: most field seismic volumes are unlabeled, and creating detailed interpretations is slow and expensive. Supervised learning therefore often relies heavily on synthetic data, where large amounts of labeled examples can be generated. Self-supervised pretraining offers an alternative, since models can learn representations directly from large collections of real field data before being adapted to downstream interpretation tasks. Masked autoencoders (MAEs), a self-supervised technique, fit this setting well, because they learn by reconstructing masked parts of a volume and need no labels. In this talk, we discuss why 3D self-supervised pretraining matters for seismic foundation models and how it differs from working on individual 2D slices. We then present our work using an MAE and show preliminary results. We also cover the practical challenges of 3D pretraining, such as compute cost, patch and masking design, and what work remains to be done.

Construction of a Foundation Model for Seismic Properties from Data in the UK National Data Repository

Ipsita Bhar,
Huseyin Tuna Erdinc, and
Felix J. Herrmann, SLIM

Abstract. Foundation models for subsurface properties call for large collections of representative Earth models, yet densely sampled velocity and density models are rarely available in the field. Building on SAGE (Subsurface foundational model for AI-driven Geostatistical Extraction) and on the curation pipeline we developed for the UK National Data Repository (UK NDR), we report on the construction of a generative foundation model for seismic properties that is trained on field data from the North Sea. Instead of relying on fully sampled 2D property models, score-based generative models are trained on a combination of sparse well logs and the depth-converted seismic lines to which these wells are tied, with a formulation that accounts for the difference in sampling density between these two complementary sources of information. Earlier experiments on synthetic models with diverse geological structures, including faults and salt bodies, showed that this approach captures long-range geological correlations and yields samples that align well with unseen ground-truth properties. Here, we move from fine-tuning on a small curated subset towards training on well-log and seismic data curated at scale from the UK NDR, and discuss the practical challenges this entails. The resulting model is intended to serve as a prior for full-waveform inference with WISE and ASPIRE, and as a source of training data for supervised learning and seismic-based subsurface modeling.

Probing Seismic Knowledge Learned by Vision Foundation Models: A Layer-Wise Analysis of DINOv2

Araz Shafiyev,
Mohammad Alotaibi, and
Ghassan AlRegib, OLIVES

Abstract. It remains unclear what geophysically meaningful information vision foundation models encode when adapted to seismic data, and where in the network that information lives. We adapt DINOv2 to seismic through self-supervised pretraining and, separately, supervised fault segmentation, then train linear probes on intermediate-layer features to predict conventional seismic attributes. Comparing probe performance across layers and training regimes shows whether adaptation yields attribute-aligned features, how the training objective redistributes them with depth, and whether downstream tasks benefit from using multiple layers rather than only the final one.

The Impact of Attention Collapse on Vision Transformer Performance in Seismic Data

Mohammad Alotaibi and
Ghassan AlRegib, OLIVES

Abstract. Vision transformers (ViTs) have become a prominent architecture in seismic workflows. However, their primary design for classification tasks poses challenges when applied to dense prediction problems like seismic segmentation. In these settings, models often exhibit reduced adaptability stemming from attention collapse, wherein multiple attention heads become redundant. In this work, we analyze the underlying causes of attention collapse and demonstrate that addressing it enhances the performance of ViTs in seismic segmentation tasks.

Causal Seismic Generation: Learning Actionable Geological Worlds

Prithwijit Chowdhury and
Ghassan AlRegib, OLIVES

Abstract. Synthetic seismic data are widely used to train and test interpretation models, but most existing diffusion-based methods focus on reproducing what seismic images look like rather than learning the geological causes that produced them. We build a generative model that learns processes such as faulting, folding, deposition, and erosion as reusable and controllable transformations across different geological settings. By separating the underlying geology, the geological event, and the appearance of a particular seismic survey, we can move from passive generation to actionable generation. Instead of only producing a realistic seismic image, the model can answer controlled questions such as what would change if a fault were added, removed, or made larger, while preserving parts of the geology that should remain unaffected. This causal structure allows us to generate matched versions of the same geological world before and after a specific intervention and directly measure the consequences of that change. We test whether these learned geological processes remain valid across new structures, event combinations, simulation methods, and real seismic surveys. Our goal is to build a controllable seismic world model in which generated data are not only realistic, but also causal, interpretable, and actionable for seismic interpretation and model development.

Scaling Wavelet-Whitened Patch Diffusion to 3D Seismic Velocity Models

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

Abstract. Three-dimensional generative modeling of subsurface velocity fields remains challenging because full-volume diffusion models are computationally expensive, while patch-based models can lose the long-range correlations that characterize geological structure. Building on our wavelet-whitened patch diffusion framework, we develop a scalable 3D score-based generative prior that combines local volumetric training with multiscale preconditioning. Instead of training directly on full 192³ volumes, the model is trained on position-aware 3D cubes with maximum edge length 64, substantially reducing memory and computational requirements while retaining coverage of the full volume. An invertible separable wavelet transform with subband-wise normalization whitens the strongly colored spectrum of velocity fields, improving the conditioning of local score estimation and reducing the burden of learning correlations dominated by low spatial frequencies. At sampling time, local denoiser evaluations are composed across complementary full-volume patch partitions, enabling information to propagate across patch boundaries while keeping the network input local. We evaluate the resulting model on full 3D velocity generation using spectral fidelity, directional long-range correlations, boundary artifacts, sample diversity, GPU memory, training throughput, and sampling cost. Preliminary experiments show promising local denoising accuracy and conditioning, while ongoing full-volume experiments test whether these gains translate into coherent 3D geological samples. This approach provides a practical route toward memory-efficient 3D diffusion priors for seismic inversion and uncertainty quantification.

Inverse Homogenization with Well-Log Generative Priors for Physical Model Downscaling

Huseyin Tuna Erdinc and
Felix J. Herrmann, SLIM

Abstract. Homogenization provides a framework for relating fine-scale variations in subsurface physical properties to the effective medium governing long-wavelength seismic propagation. Inverse homogenization seeks to downscale coarse-scale velocity and density models by inferring fine-scale structures consistent with their effective properties. In this study, we will formulate two-dimensional inverse homogenization as a super-resolution inverse problem, using homogenization as the forward operator and coarse-scale models as observations. To address the inherent nonuniqueness of this problem, we will adopt a Bayesian framework that combines consistency with these observations and prior information represented by generative models trained on one-dimensional well logs. A central contribution will be the use of these one-dimensional generative priors to constrain two-dimensional reconstructions, incorporating fine-scale geological information without requiring a training set of fully resolved two-dimensional models. By targeting an ensemble of plausible velocity and density models, this approach aims to characterize uncertainty in unresolved heterogeneity and provide a principled way to integrate well-log information into subsurface models beyond the resolution of seismic imaging.

Bridging the Simulation-to-Real Gap in Probabilistic Seismic Inversion

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

Abstract. Seismic inversion frameworks trained only on synthetic data often struggle to generalize to field observations because synthetic and real data differ in wave physics, acquisition, noise, processing, and geological complexity. These discrepancies create a substantial simulation-to-real distribution shift, while dense field velocity labels needed for supervised training are rarely available.

We introduce ElasNet, a probabilistic seismic inversion framework that combines labeled acoustic simulations with unpaired field observations. ElasNet learns the observation-to-velocity relationship from Compass simulations while aligning their learned representations with North Sea field data, encouraging the inversion model to preserve subsurface-relevant information while reducing sensitivity to simulation-specific features. Sparse well measurements are only used for evaluation rather than training. Experiments show that incorporating unlabeled field data improves agreement with well measurements compared with simulation-only training when the inversion and alignment objectives are appropriately balanced. ElasNet therefore provides a practical route toward uncertainty-aware seismic inversion that leverages large-scale simulation while adapting to real observations without requiring paired field velocity models.

Following the Faults: Layer-wise Gradient Suppression for Targeted Fine-Tuning Under Seismic Domain Shift

Jorge Quesada and
Ghassan AlRegib, OLIVES

Abstract. Our earlier fault segmentation benchmark found that transfer between seismic datasets often fails as a function of the distributional shift between datasets. The best way to adapt depends on how large the shift is, and biases from pretraining persist even after fine-tuning. Taken together, these results suggest that updating either every layer the same way or only the late ones during fine-tuning is the wrong approach, and the question becomes where in the network the mismatch actually sits. We address this issue with the Suppression Index, a layer-wise diagnostic that uses a small sample of target data to measure how much of each layer’s learning signal falls in directions the pretrained model barely represents. In a challenging, large-shift transfer scenario, the diagnostic consistently places the mismatch in intermediate layers rather than the deepest ones, and it does so for both convolutional and transformer architectures. Scaling each layer’s learning rate by its suppression level improves fault segmentation significantly compared to modern finetuning baselines. Our results show that diagnosing where transfer breaks can directly guide how seismic interpretation models are adapted.

Transformer-Based Reconstruction of Missing Density Logs

Ipsita Bhar,
Huseyin Tuna Erdinc, and
Felix J. Herrmann, SLIM

Abstract. Accurate reconstruction of missing well-log measurements is important for subsurface characterization, particularly when continuous measurements are unavailable because of acquisition limitations or poor data quality. In this work, we develop a Transformer-based framework for predicting missing density (DEN) logs from available petrophysical measurements, including gamma ray (GR), acoustic (AC), neutron porosity (CNL), and deep resistivity (RLLD) logs. After quality control and preprocessing, 126 wells are retained for model development. Each continuous well-log interval is resampled at a fixed 0.125 m spacing and divided into non-overlapping 720-sample windows, corresponding to 90 m depth intervals; incomplete windows are discarded. This process produces 1,307 training files and 386 validation files. The model learns the nonlinear relationships between the available log measurements and the target density response. To evaluate its ability to generalize to previously unseen wells, two blind wells are used exclusively for prediction and independent evaluation. The proposed framework provides a data-driven approach for reconstructing missing well-log intervals and assessing prediction performance on unseen well data.

GUEST talk: AthenaLabs–Scaling Seismic Inversion at the Speed of Compute and Tokens

Richard Rex, AthenaLabs

Abstract. Exploring seismic inverse problems requires repeated decisions about physical assumptions, numerical settings, and the experiments needed to distinguish competing explanations. We present AthenaLabs, a platform for human-guided, AI-assisted scientific experimentation, demonstrated through permeability inversion from time-lapse seismic observations of CO₂ injection. Building on existing differentiable multiphysics software, the workflow connects reservoir simulation, rock-physics mappings, and seismic modeling within reproducible experiments. Researchers define the scientific question and experimental constraints; AI agents assist with configuring experiments, launching parallel runs, diagnosing failures, and interpreting results. We illustrate this approach by varying rock-physics parameters and saturation models while reusing a fixed set of observations and initial conditions. Each inversion retains its sequential optimization loop, while independent hypotheses are explored concurrently on cloud infrastructure. Versioned code, configurations, and artifacts preserve the provenance of each result, and interactive visualizations support comparison of recovered permeability, predicted seismic observations, and convergence histories. Separate evaluation against synthetic ground truth distinguishes improvements in data fit from improvements in model recovery. The demonstration examines how access to simulation compute and language-model inference can expand the scope of scientific exploration while keeping experimental choices and numerical evidence available for researcher review.

Refining Digital-Twin Forecasts through Probabilistic Permeability Reconstruction

Abhinav Prakash Gahlot and
Felix J. Herrmann, SLIM

Abstract. Digital Twins for geological CO₂ storage infer evolving saturation and pressure fields from monitoring observations while accounting for uncertainty in reservoir permeability. Their predictive uncertainty, however, depends strongly on the permeability distribution used to generate the training ensemble. We propose a closed-loop workflow in which monitoring data are first used to narrow this distribution and the resulting permeability posterior is then fed back into the Digital Twin. Specifically, a conditional score-based generative model jointly reconstructs permeability, CO₂ saturation, and pressure from time-lapse seismic and sparse well observations, producing an observation-informed ensemble of plausible permeability fields. This ensemble replaces the broader initial permeability prior when regenerating flow and seismic simulations for the Digital Twin. Retraining or updating the amortized inference model on this refined ensemble is expected to concentrate the forecast distribution on reservoir scenarios consistent with collected observations, thereby reducing uncertainty in subsequent saturation and pressure estimates.

From Seismic Monitoring to Injection Control: A Digital Twin for Subsurface Operations

Haoyun Li, SLIM and
Felix J. Herrmann, SLIM

Abstract. Pressure buildup limits subsurface injection, but uncertain rock properties and incomplete observations make it difficult to determine how much fluid can be injected within prescribed pressure limits. We present a digital-twin framework that connects seismic-informed monitoring with pressure-constrained injection control. Building on a digital shadow based on conditional normalizing flows, we update posterior pressure and saturation samples from time-lapse seismic summaries and well measurements. These samples are paired with permeability realizations to predict reservoir evolution under candidate injection schedules. Model predictive control seeks to maximize injected fluid volume under alternative constraints: probability of failure limits the fraction of pressure exceedances across reservoir cells and times, while conditional value-at-risk limits the mean of the largest normalized pressure-exceedance losses. The empirical distribution of optimized injection-rate endpoints, together with bootstrap uncertainty estimates, guides the selection of one schedule, which is updated as new monitoring data arrive. Synthetic experiments based on the Compass reservoir model illustrate how the pressure constraints govern the trade-off between cumulative injection and pressure exceedance, linking uncertainty-aware monitoring to closed-loop operational decisions.

A Bayesian Perspective on the Joint Recovery Model

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

Abstract. The Joint Recovery Model (JRM) improves time-lapse seismic imaging and inversion by writing each vintage as the sum of a component common to all surveys and an innovation specific to that survey, so that the shared component benefits from all data without insisting on replicated acquisitions. While successful in practice, including in its recent probabilistic formulation (πJRM), where the vintages are represented by a shared generative model with per-vintage latent variables, the method has so far been motivated mainly on deterministic grounds. In this talk, we give a Bayesian interpretation of the JRM. We consider observations y_i = ℱ_i(z₀ + z_i) + ϵ_i for the vintages i = 1, 2, where z₀ is the common component and the z_i are the innovations, and identify the priors and the posterior implied by this hierarchical model. This perspective clarifies what the strong and weak forms of the shared-decoder formulation with per-vintage Gaussian-mixture latents approximate, and which role the Kullback-Leibler term, omitted in current implementations, plays in the resulting uncertainty estimates. We discuss the consequences for uncertainty quantification of time-lapse changes and for the monitoring of subsurface storage.

From Velocity to Interpretation: Separating Velocity-Model and Network Uncertainty in Machine-Learned Seismic Interpretation

Sahil Mithani,
Abdelrahman Musleh,
Huseyin Tuna Erdinc,
Araz Shafiyev,
Abhinav Prakash Gahlot,
Zijun (Venice) Deng,
Ghassan AlRegib, and
Felix J. Herrmann, OLIVES and SLIM

Abstract. A seismic interpretation depends on the velocity model, its migration, and the interpretation network; the uncertainty at each stage carries into the final map. A misplaced fault or mispicked horizon, for example, may change where a well is drilled, yet current uncertainty estimates neither account for all of these sources nor measure the contribution of each. In this talk, we discuss a framework that propagates uncertainty from velocity inference through migration to interpretation and attributes it to its source. Using WISE, a conditional normalizing flow, we generate a posterior of velocity samples that are all plausible given the same observed data, and image each sample with reverse time migration (RTM). We then run fault, horizon, and facies interpretation on these images and decompose the resulting uncertainty into data uncertainty, inherited from the velocity model, and model uncertainty, from the interpretation network. Because the experiments use synthetic velocity models with known geology and labels, we can test whether the regions flagged as uncertain are the regions where the interpretation is actually wrong. We present preliminary results showing how uncertainty propagates to the interpreted maps that drive decisions, and what remains to be tested on field data.

Two-stage Bayesian Recovery of Velocity and Impedance from Common-Image Gathers

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

Abstract. Common-image gathers migrated with a poor background velocity are kinematically wrong, so a network recovering velocity and impedance from them must correct the kinematics and estimate amplitudes at once. We present a two-stage simulation-based inference framework that separates these tasks. In the first stage, a score-based diffusion model recovers a velocity posterior from gathers migrated with a deliberately smoothed background. In the second stage, the posterior mean velocity becomes the migration background, and a second network recovers velocity and acoustic impedance jointly from gathers whose kinematics are close to correct. We compare this sequential recovery against recovering both parameters in a single network from the poorly migrated gathers. We also report a controlled comparison of which imaging condition carries velocity information. Networks are trained on Compass and SAGE models that are identical except for the conditioning gather, which is the standard adjoint-state gather, the inverse-scattering gather, the anti-ISIC gather, or an angle gather from a slant stack over subsurface offset. Source positions are verified identical across imaging conditions, so the gather type is the only difference.

From Focusing Diagnostics to Learned Uncertainty: A WEMVA-Gradient Extension of WISE

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

Abstract. Quality control of Full-Waveform Inversion typically relies on differential semblance focusing diagnostics, most commonly common-image gathers (CIGs). However, this diagnostic can become unreliable when the extension direction poorly captures the local scattering geometry: steep or near-vertical reflectors illuminated by diving waves may appear not focused even when the background velocity is correct. To address this limitation, we develop an interactive three-dimensional diagnostic that visualizes focusing directly around subsurface image points, using a preconditioned extended image volume computed with a computationally efficient photoacoustic wave-equation formulation. Building on this diagnostic intuition, we extend it into a quantitative training signal—the gradient of the focusing objective underlying wave-equation migration velocity analysis (WEMVA). The resulting gradient converts image-domain defocusing into a model-space sensitivity, providing an explicit physics-based representation of how the velocity model should change to improve focusing. We use this gradient to augment probabilistic velocity-model inference. Specifically, using the WISE formulation, we compare two conditioning strategies: CIGs alone, and CIGs jointly with the focusing gradient. We evaluate the approach on synthetic experiments and the Compass model, examining velocity reconstruction and posterior behavior. Because diving-wave illumination of steep structure is also strongly sensitive to anisotropy, extending this focusing-based framework to anisotropic velocity models is a natural direction for future work.

Joint Twist-Flow for Bayesian Inverse Problems

Shiqin Zeng,
Zijun (Venice) Deng, and
Felix J. Herrmann, SLIM

Abstract. Bayesian inverse problems are often ill posed, as a given observation can be compatible with multiple solutions. In high-dimensional settings, directly sampling the posterior is computationally expensive, motivating conditional generative models. Conditional flow-based models map latent noise to the target conditioned on the observation. For a given observation, however, the latent-to-target map must remain invertible. This can make posteriors with multiple separated modes difficult to represent, leading to mode dropping or artificial connections between modes. We propose joint twist-flow, an augmented flow-matching formulation that replaces

(z_x, y) ↦ x

with

(z_x, y) ↦ (x, z_y),

where z_x and z_y are Gaussian reference variables. Invertibility is required only for the full augmented map, so the map from z_x to x for a given observation does not need to be invertible by itself. The additional variable z_y retains the information needed to keep the full transport invertible, giving the target map more freedom to represent multiple plausible solutions.

We validate joint twist-flow on low-dimensional inverse problems with reference posterior samples, where it better preserves multimodal posterior support than a conditional flow-matching baseline. We further evaluate the method on image restoration and seismic subsurface velocity inversion, where it captures variability in weakly constrained components while maintaining observation consistency.

A Curvelet Neural Operator for WISE

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

Abstract. WISE (full-Waveform variational Inference via Subsurface Extensions) draws posterior samples of velocity models conditioned on common-image gathers computed for a single, possibly poor, background-velocity model. The quality of this inference hinges on how well these gathers preserve the information contained in the data. We propose a Curvelet Neural Operator (CNO) that acts on extended images in the curvelet domain. Because the normal operator is for good background-velocity models pseudodifferential, it is approximately diagonal with respect to curvelets, a beneficial property that may approximately extend to poor velocity models. We discuss the design of the operator, its training, and its use within WISE for amortized Bayesian velocity-model building.

Federated Learning and Privacy for Seismic Imaging

Carlos Marí and
Ghassan AlRegib, OLIVES

Abstract. Federated Learning enables multiple organizations to train a shared model without exchanging their data. Each participant trains locally and communicates only model updates, which makes the paradigm well suited to seismic interpretation, where survey data is costly to acquire and commercially sensitive. However, the participant’s data may differ substantially, and personalized methods that adapt to these differences rely on each participant’s data composition, which is itself sensitive. This work explores the boundary between personalization and privacy in federated learning: how much a participant’s model updates reveal about which categories its data contains and in what amounts, how personalization methods shift this exposure, and to what extent encrypted aggregation can limit it when only few organizations collaborate.

Underspecification in Seismic Interpretation: Decomposing Data, Model, and Interpretational Uncertainty

Sahil Mithani,
William Stevens,
Ryan Benkert,
Mohit Prabhushankar, and
Ghassan AlRegib, OLIVES

Abstract. Deep learning interpretations of seismic data are typically paired with uncertainty maps that treat the processed image and its labels as fixed, an assumption that rarely holds. A seismic image is one stochastic product of processing sparse measurements, and each voxel’s final label is assembled from several overlapping, often conflicting interpretations. Our previous work framed this as underspecification and derived a decomposition of total predictive uncertainty into data, model, and interpretational components, showing that they combine through nested expectations rather than as independent terms and that each is shaped by all three sources. Established methods such as Monte Carlo dropout and deep ensembles capture only the model term. We conduct controlled experiments on the F3 block that vary one source of underspecification at a time and measure whether its component responds while the others hold still, which shows where the terms separate cleanly and where they stay entangled. Because F3 and Parihaka both carry facies labels, every component is also checked against prediction error on two surveys, alongside dropout and ensemble baselines at matched compute, and we examine how much of a qualitative uncertainty map reflects the section it was drawn on rather than the estimator. We then extend the framework to fault detection, where disagreement among expert interpreters gives interpretational uncertainty a direct reference. Each component points to a different fix, whether reprocessing, retraining, or expert review. The goal is uncertainty quantification that tells an interpreter where to doubt a result and why.

Guest talk: Amortized Quadrature for Posterior Expectations in Inverse Problems

Ali Siahkoohi, University of Central Florida

Abstract. Uncertainty in the solution of an inverse problem and in the tasks performed on it is quantified by posterior expectations, each an average of an integrand over M posterior samples. While designed quadratures improve on the 𝒪(M^−1/2) error of Monte-Carlo estimation, they solve an optimization problem, often against the posterior density, for every new observation, which can be computationally costly. To address this limitation, we introduce the quadrature field, a set-equivariant network that maps an observation and its M posterior samples to an M-node signed-weight quadrature in one forward pass. Trained once on a family of posteriors to minimize the worst-case integration error over a class of functions, it serves any observation, any M and any integrand in that class with no further optimization. We show that, with high probability and up to a computable slack, the resulting quadrature is never worse than the Monte-Carlo estimate built from the same samples. We validate the quadrature field on closed-form and on learned posteriors, one constrained by a partial differential equation, where it improves on the Monte-Carlo estimate in median at every node count, often by orders of magnitude.

Instant Model Uncertainty in Seismic Interpretation

William Stevens,
Mohit Prabhushankar, and
Ghassan AlRegib, OLIVES

Abstract. Deep networks increasingly extend sparse expert labels across entire seismic surveys, but they rarely indicate where their outputs should not be trusted. Established uncertainty methods such as deep ensembles and Monte Carlo Dropout are reliable but costly, requiring many models or repeated inferences that scale poorly to full 3D volumes. We present a single-pass approach that estimates model uncertainty from the information flowing between a network’s internal layers, requiring no additional training, architectural changes, or repeated inference. We demonstrate this approach across a range of seismic tasks, from facies classification and fault detection to seismic inversion, and across several modern network architectures. We examine what this internal signal captures, how it relates to where models actually make errors, and where its limits lie. The result is a general, low-cost confidence measure that helps interpreters decide where automated interpretations can be trusted and where expert review is needed.

GeoBuddy: A Local AI Assistant for Geoscience

Ghazal Kaviani,
Araz Shafiyev,
Abdelrahman Musleh,
Neha Kunche,
Sihu Kim,
Kai Wen Khoo,
Leo Lin, and
Ghassan AlRegib, OLIVES

Abstract. Geoscience knowledge is spread across books, papers, online resources, and datasets, and it keeps changing. GeoBuddy is an ongoing effort to build a local AI assistant that ingests this kind of corpus and answers questions about it conversationally, for teaching and analysis. It builds on StudyBuddy, a course-grounded study assistant that answers from a student’s own materials with page-specific citations and runs its models locally. In this talk, we describe what carries over from StudyBuddy, what does not, and where GeoBuddy stands, including the work of assembling a geoscience corpus of books, papers, and datasets. We also discuss the open question of how to keep such an assistant current as new work appears.

Tutorials

Interactive, Uncertainty-Guided Geobody Interpretation in 3D Seismic: A Hands-on Tutorial

Araz Shafiyev,
Prithwijit Chowdhury, and
Ghassan AlRegib, OLIVES

Tutorial. Foundation models such as the Segment Anything Model can segment an image from just a few clicks, and they work best on seismic data when an interpreter guides them. This tutorial introduces a browser-based tool that brings this approach to 3D seismic interpretation. Its uncertainty estimation runs as a lightweight, modular head on top of a frozen SAM backbone. The backbone is never retrained or fine-tuned, and users need no labeled data to get started. Participants will load a SEG-Y volume and outline geobodies such as salt and gas-related bright spots with simple point clicks. The model’s uncertainty map then suggests where the next click will help most and highlights where the current interpretation needs a closer look. The tool offers a range of acquisition functions for building this map, including BALD with Laplace and MC-dropout approximations and several entropy-based alternatives. Users can also vary the number of posterior samples, k, to trade off speed against the reliability of the uncertainty estimate. We will also show how to fill gaps between interpreted inlines and export the result as a SEG-Y mask that can be loaded into any interpretation software. No installation is required.

Tutorial: StudyBuddy — A Local, Course-Grounded AI Study Assistant

Ghazal Kaviani,
Araz Shafiyev,
Abdelrahman Musleh,
Neha Kunche,
Sihu Kim,
Kai Wen Khoo,
Leo Lin, and
Ghassan AlRegib, OLIVES

Tutorial. StudyBuddy is a course-grounded study assistant that uses local AI agents to answer questions from the materials a student provides, such as lecture slides, textbooks, and local files, with page-specific citations back to those sources. Because the models run locally, coursework is not sent to third-party cloud services. It also includes features that support learning, such as quiz and flashcard generation, a weekly study planner, and a progress dashboard. In this tutorial, we walk through how StudyBuddy works, from ingesting course materials to retrieving relevant passages and synthesizing grounded answers. We then turn to GeoBuddy, an ongoing effort to apply the same local approach to geoscience, where the corpus is geoscience books, papers, and datasets and not course material. We close with an open discussion on features that would make GeoBuddy useful to geoscientists and industry.

Tutorial: Actionable Explainability for the Seismic Interpretation Workflow

Mohammad Alotaibi,
Prithwijit Chowdhury, and
Ghassan AlRegib, OLIVES

Tutorial. Seismic interpretation is an inherently iterative process in which domain experts repeatedly revise their analyses, and it often involves several experts deliberating over the same task. Recent efforts to integrate artificial intelligence into this process aim to bring greater consistency and efficiency to established workflows. However, current AI-assisted workflows offer little meaningful interaction between the domain expert and the model. Moreover, most AI models are trained in a supervised manner on labels drawn from a single expert source, labels with which other experts may disagree, leaving them no direct way to adjust how the model responds to the data. In this work, we introduce an actionable explainability framework that incorporates expert domain knowledge directly into the AI workflow. Rather than serving only as a post hoc justification, the model’s explanations actively guide the expert toward informed interventions, enabling an interactive loop in which human expertise and model behavior refine one another.