Physics Foundation Model Degrades in Sim‑to‑Real Transfer
| Source: ArXiv | Original article
A new arXiv paper shows that scientific foundation models, while excelling on PDEs and finetuning to new distributions, can lose previously learned knowledge when adapted from simulation to real‑world data.
A new arXiv pre‑print (arXiv:2610.40857v1) reveals that Walrus‑1.3B, a scientific foundation model (SciFM) for continuum dynamics, suffers catastrophic forgetting when it is adapted from simulation to real‑world fluid‑flow data. The authors evaluate the model in a two‑stage “sim‑to‑real” pipeline: first they fine‑tune Walrus‑1.3B on a simulated flow around a cylinder, then they continue training on experimental particle‑image‑velocimetry (PIV) measurements of the same setup. Across both phases the model’s performance on the original simulated task degrades sharply, confirming that the additional fine‑tuning erases knowledge acquired during pre‑training.
The finding matters because SciFMs have been celebrated for their ability to solve a broad spectrum of partial‑differential equations and to adapt quickly to new data distributions. If adaptation routinely leads to forgetting, the reliability of these models in safety‑critical or high‑cost experimental contexts could be compromised. The work also underscores a gap in current research: while many studies showcase the raw accuracy of SciFMs, fewer address their long‑term stability under continual learning scenarios.
The authors accompany the paper with a full code release that reproduces the two‑phase experiments and visualises the forgetting dynamics. Their NeurIPS 2026 workshop presentation is likely to spark follow‑up investigations into mitigation strategies such as replay buffers, regularisation techniques, or modular architectures. Watch for subsequent benchmarks that quantify forgetting across other SciFMs and for community‑driven toolkits aimed at preserving pre‑trained knowledge while still enabling rapid sim‑to‑real transfer.
Sources
Back to AIPULSEN