Frozen in Time: How a Physics Foundation Model Forgets Under Sim-to-Real Adaptation

Scientific foundation models achieve performance on PDEs and can adapt to unseen data distributions with finetuning. Finetuning Walrus-1.3B on simulated flow rapidly degrades rollout performance on original pretraining datasets, with one-step errors increasing up to 35x after 20 epochs. Subsequent adaptation to real data further erodes performance on the simulated task.