---
format: "aidr-story-markdown/v1"
id: "9bf6d686237ce909b72f47f00f4b5c941c115984f66e65f796ded52c7c7a0fe3"
canonical_url: "https://aidr.today/9bf6d686?lang=en"
title: "Frozen in Time: How a Physics Foundation Model Forgets Under Sim-to-Real Adaptation"
lang: "en"
requested_lang: "en"
available_langs: ["en","vi"]
translation_fallback: null
fallback_fields: []
published_at: "2026-11-02T05:00:00.000Z"
category: "Research"
topics: ["research"]
source_urls: ["https://arxiv.org/abs/2610.40857"]
summary: "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."
---

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

> [Open the canonical story](<https://aidr.today/9bf6d686?lang=en>)

**Published:** 2026-11-02T05:00:00.000Z
**Category:** Research
**Topics:** research

## Summary

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\.

## Sources

- [Story source](<https://arxiv.org/abs/2610.40857>)