Program Overview
Seismic data presents unique challenges to foundation models due to the complexity of its 3D volumetric structure and geological features. This research program explores cutting-edge foundation models, such as Segment Anything Model 2 (SAM2), and in-context learning-based frameworks to meet the distinct requirements of seismic data interpretation. The program proposes novel in-context learning paradigms that overcome the limitations of traditional point-based human-foundation model interaction and support multi-instance segmentation across discontinuous volumes. The program develops and validates models that treat seismic volumes like videos, utilize rich seismic embeddings, and exploit structural priors such as linearity and continuity commonly presented in geophysical data. These innovations aim to reduce manual annotation, improve generalization, and scale up interpretation workflows for large seismic datasets.
Objectives
- Embedding-based Prompt Transfer: Adapt SAM2 to seismic data through embedding-based prompt transfer, enabling non-continuous and multi-instance segmentation across seismic sections.
- Geophysical In-context Learning: Develop line-based in-context learning strategies to overcome the limitations of point-based prompts, improving boundary adherence and geobody representation.
- Time-series Analyses of In-context Learning. Develop an interactive seismic interpretation toolkit leveraging foundation models with flexible prompting interfaces. Analyze the model behavior in multiple cycles of interaction through in-context learning.
- Leverage Model Disagreements to Guide the Labeling of Geophysicists. Utilize representation shifts between neural networks to characterize attention towards geologically interesting regions.
- Leverage Label Disagreements to Guide Labeling. Modeling differences between delineations of multiple interpreters through generative distribution to characterize geologically difficult and ambiguous data.
- Develop Explanatory Frameworks that Generate Trustworthy Explanations for Subsurface Activities. Utilize human expertise-assisted explanation techniques, including causal metrics such as sufficiency/necessity scores and model statistics from back-propagated gradients to enhance the inferential behavior of foundation models.
- Interpreter-in-the-loop Frameworks for Training Models. Develop frameworks such as active learning paradigms that utilize representation manifold statistics to reduce the demand for expert annotations.
Key themes
Foundation models.
- Geophysical in-context learning: Investigating how in-context learning strategies (e.g., point-, line-, and embedding-based prompting) can efficiently guide foundation models, such as SAM2, in segmenting complex seismic structures.
- Adapting foundation models for seismic data. Extend general-purpose models to seismic data through architectural adaptation, domain-specific in-context learning, and zero-shot inference strategies.
- Multi-instance and discontinuous segmentation. Develop methods to detect and segment repeating but sparse or discontinuous geological structures across sections.
- Human expertise-assisted explanatory frameworks.
- Human interpretation disagreement. Differences between multiple interpreters are modeled as interpretation uncertainty.
- Model interpretation disagreement. Disagreements between two networks are modeled as interpretation uncertainty.
Participation and Governance
This CRP operates within the ML4Seismic philanthropic partnership framework. Participation requires enrollment as an Executive Member with at least one CRP selected. Partners are
invited to:
- Participate in CRP-specific workshops and benchmarking initiatives
- Collaborate on shared scientific objectives and data challenges
- Engage with students and researchers in community events
All research outcomes are shared via open-source repositories and peer-reviewed publications in accordance with ML4Seismic operating guidelines.
Contact
Prof. Ghassan AlRegib alregib@gatech.edu
