LLM-Driven Loss Balancing for Physics-Informed Neural Networks

Denis Degtiarev, Anastassiya Ryabkova, Alexei Kornaev

Keywords: Physics-Informed Neural Networks (PINNs), Large language models(LLMs), Loss balancing, PDE benchmarks

Abstract:

We test whether LLMs can mitigate PINN loss imbalance by either directly controlling loss weights or ranking established balancers from PDE and method descriptions.
LLMs fail as weight controllers even on simple ODEs, but often identify strong candidates (e.g., GradNorm/LB-PINN/RAD/DB-PINN) as meta-evaluators on harder multiscale/stiff problems.

Journal Opt In: Yes, I want to participate in the IOP focus collection submission

Submission Number: 51