Adapting General Representations of Pretrained Vision Foundation Models to Seismic Understanding via Geologically Informed Prompting and Lightweight Tuning

Abstract Interpreting seismic responses to subsurface structures is essential for reservoir characterization and energy exploration, yet remains challenging due to the spatial complexity and heterogeneity of geological formations. While deep learning has accelerated progress, most existing models depend on task-specific architectures trained on limited labeled data, which hinders their generalizability. Here, we introduce a transfer learning framework that repurposes vision foundation model for seismic interpretation through cross-domain adaptation. A lightweight bridge model maps seismic amplitudes into the representation space of vision backbones, and efficient adaptation through low-rank updates and prefix tuning balances flexibility with preservation of the foundation model’s representational priors. By embedding geological priors into the latent space, the model learns to internalize stratigraphic order and to deliver predictions that preserve structural consistency. A task-adaptive decoder further supports various objectives in seismic interpretation. Benchmark experiments across various geological settings demonstrate that our framework consistently surpasses baseline architectures, delivering superior accuracy with fewer trainable parameters. These results underscore the promise of foundation models not only as scalable backbones for seismic interpretation, but also as a basis for broader data-driven advances across geoscientific disciplines under supervision-limited conditions.