An end-to-end fault interpretation method driven by visual foundation model with domain adaptation fine-tuning

Area of Science:

• Geophysics
• Artificial Intelligence
• Geological Engineering

Background:

• Accurate geological fault interpretation is crucial for underground gas storage safety and efficiency.
• Manual interpretation is inefficient and relies heavily on expert experience.
• Current deep learning methods struggle with limited data, unreliable annotations, and lack of geophysical principles.

Purpose of the Study:

• To develop an advanced AI framework for robust geological fault interpretation.
• To overcome limitations of existing deep learning models in seismic data analysis.
• To enhance the safety and efficiency of underground gas storage operations through improved fault analysis.

Main Methods:

• A visual foundation model-driven framework with domain adaptation fine-tuning.
• Fault-Aware Auto-Augmentation using reinforcement learning for synthetic data generation.
• Uncertainty-Driven Self-Annotation Optimization for reliable annotation loops.
• Geophysics-Constrained Feature Alignment Fine-Tuning incorporating prior geological knowledge.

Main Results:

• Significant enhancement in fault identification robustness, especially in complex geological zones.
• Segmentation outcomes demonstrated strict adherence to geological principles and cognition.
• The framework provides an efficient and interpretable intelligent interpretation paradigm.

Conclusions:

• The proposed framework effectively addresses data scarcity, annotation unreliability, and geophysical principle integration.
• It offers a superior approach to geological fault interpretation for underground gas storage.
• This advancement supports better caprock integrity evaluation and fault sealing analysis.