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

Accurate geological fault interpretation is critical for the safe construction and efficient operation of underground gas storage sites. However, traditional manual interpretation suffers from inefficiency and reliance on expert experience. Existing deep learning-based methods face three major challenges: limited generalization due to scarce seismic samples, unreliable annotations in low signal-to-noise ratio regions, and neglect of geophysical principles in generic models. To address these issues, a visual foundation model-driven framework with domain adaptation fine-tuning is proposed. First, a Fault-Aware Auto-Augmentation algorithm is adopted to generate diverse synthetic samples through reinforcement learning-based search for physically compliant augmentation strategies, overcoming data scarcity limitations. Second, an Uncertainty-Driven Self-Annotation Optimization mechanism is developed, establishing a high-reliability annotation loop through integration of prediction confidence with expert collaborative correction. Finally, Geophysics-Constrained Feature Alignment Fine-Tuning is introduced, incorporating prior knowledge such as structural tensors to enforce adherence to strata continuity principles. Experimental results demonstrate significant enhancement of fault identification robustness in complex structural zones, with segmentation outcomes strictly adhering to geological cognition. An efficient and interpretable intelligent interpretation paradigm is delivered for caprock integrity evaluation and fault sealing analysis in gas storage operations.