Adapting general representations of pretrained vision foundation models to seismic understanding

Interpreting seismic data remains challenging because most learning-based approaches rely on specific architectures that require retraining for each task and generalize poorly across subsurface regimes when training data are limited. Here, we introduce a cross-domain transfer learning framework that adapts a vision foundation model for seismic understanding. A lightweight seismic-to-vision bridge maps seismic data into the representational space of pretrained backbone, while low-rank adaptation and prefix tuning enable efficient parameter updates that preserve general visual priors. Geological constraints are incorporated via stratigraphic prompting that injects stratigraphic order into the latent space, guiding predictions to respect structural consistency. A task-adaptive decoder enables flexible downstream applications, including facies segmentation, structural interpretation, and property inversion. The framework outperforms baseline models across diverse tasks and datasets with fewer parameters, demonstrating that adapting pretrained vision foundation model via geologically informed prompting and lightweight tuning enables transferable seismic interpretation and broader data-driven geoscientific applications.