LoRA beats full fine-tuning for physics inference: same posterior, 1/5 the memory

bravo_abad · x · 2026-09-21

Bayesian inference for physical models is expensive because the solver must be re-adapted as the sampler explores nearby parameter values. In a Nature Communications (2026) paper, Ray Zirui Zhang and coauthors avoid retraining the whole neural PDE solver: they freeze the main network and apply only a low-rank LoRA correction to its weights.

Strikingly, LoRA ranks 4 and 8 recover essentially the same posterior as full fine-tuning. On the largest network tested, the low-rank update cuts peak memory to roughly one fifth of full fine-tuning and is faster per sample. The broader lesson: when successive scientific problems are nearby in parameter space, adapt the model locally instead of relearning it globally.

Original post →

More from Research

Research channel →