Benchmarking PINNs Across Precision and Optimizers

Benchmarked physics-informed neural networks (PINNs) on reaction and wave PDEs across every dtype and optimizer combination. Found that precision matters far more under L-BFGS than under Adam, and that every low-error run finished in L-BFGS. Accepted to the NeurIPS OPT workshop.

See the paper ↗︎