A new framework lets a large language model watch its physics-simulating neural network fail in real time, then rewrite the design instead of guessing again from scratch.
The system, called PINNsForge, targets physics-informed neural networks (PINNs), which solve partial differential equations but require careful tuning of network architecture, sampling strategy, and loss construction that varies wildly from one equation to the next. Instead of testing configurations blind, PINNsForge trains each candidate, feeds the LLM the actual optimization behavior it observed, and uses that evidence to refine, recombine, and generate the next round of designs. The researchers ran it across 25 PDE benchmarks against three existing automated design methods - RoPINN, PINNsFormer, and PINNsAgent - and it produced the lowest mean error on 24 of them. Stripping out the PDE knowledge base, the execution feedback loop, or the evolutionary search each made results dramatically worse, with mean error rising as much as twelvefold.
PINNs have a reputation for being finicky: a setup that solves one equation cleanly can fail entirely on a structurally similar one, and researchers often hand-tune for weeks. Turning the training process itself into feedback for design search, rather than just a pass or fail evaluator, is a meaningful shift from one-shot automated tuning toward something closer to iterative engineering judgment.
The benchmarks are curated academic PDEs, though - whether this feedback loop holds up on the messier equations engineers actually get paid to solve is still untested.