PIAL

In this study, we introduce a novel physics-informed adaptation learning framework, termed PIAL, for efficient and generalizable modeling of dynamic subsurface flow. By leveraging recent advances in scientific foundation models (SciFMs) and parameter-efficient transfer learning, our approach enables effective adaptation of a large pretrained operator model to subsurface flow problems under data-scarce conditions. Specifically, PIAL combines lightweight adapter-based fine-tuning with a physics-informed post-hoc distillation strategy, in which governing equations, initial conditions, and boundary conditions are embedded into the learning objective. This dual-stage design ensures that the adapted model preserves pretrained knowledge while achieving strong physical consistency, robustness to noise, and high computational efficiency.