Physics2Weight: Direct Neural Network Architecture and Weight Generation from Physical Laws Using Large Language Models

Researchers have developed Physics2Weight (PtW), a framework that uses large language models to directly generate neural network weights and edit architectures for solving partial differential equations. PtW employs physics-constrained prompting and architecture editing, along with iterative weight refinement, to improve solution accuracy and physical consistency without gradient-based optimization. Experiments show PtW achieves lower solution errors than conventional physics-informed neural networks for several PDEs, though it has limitations with nonlinear equations.