Reproduce “Frozen in Time: How a Physics Foundation Model Forgets Under Sim-to-Real Adaptation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Haaris Mian, Miles Cranmer, Alessandro Favero, Payel Mukhopadhyay
The authors study catastrophic forgetting when adapting Walrus-1.3B from simulated cylinder flow to particle-image-velocimetry measurements. After 20 epochs, Helmholtz Staircase’s one-step NRMSE rises to 34.89 times its pretrained value. During real-data adaptation, 1% simulated-data replay recovers over 65% of lost cylinder-simulation performance; replay metrics are labeled inconsistently as NRMSE and VRMSE, and evidence covers one model, scenario, and seed.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.