arXiv · 2611.12954
Physics2Weight: Direct Neural Network Architecture and Weight Generation from Physical Laws Using Large Language Models
Abstract
Recent advances in large language models (LLMs) have enabled weight-space programming (WSP), which directly generates neural network weights without gradient-based optimization. However, extending WSP to physical problems remains challenging, particularly in achieving both solution accuracy and physical consistency. In this paper, we propose Physics2Weight (PtW), a framework that extends WSP to solving partial differential equations by directly generating neural network weights and editing architectures based on governing equations and initial and boundary conditions. PtW introduces physics-constrained prompting and physics-guided architecture editing, combined with iterative weight refinement from WSP, to improve solution accuracy and physical consistency without gradient-based optimization. Experiments on four PDEs, namely the heat, advection, Burgers', and Allen--Cahn equations, demonstrate lower solution errors than conventional physics-informed neural networks, while revealing limitations in satisfying the governing equations for nonlinear PDEs.
Explore related subjects
Keep this discovery
Explore connections, maps & timelines
Shogo Sato, Tomoyasu Shimada, Kazuhiko Murasaki, Takashi Sano. 2026-11-07. Physics2Weight: Direct Neural Network Architecture and Weight Generation from Physical Laws Using Large Language Models. https://arxiv.science/abs/2611.12954
Citations retain the original authors and DOI and include this site’s record URL. If this discovery service contributed to your research, you can also cite arXiv Science by Indiciae: BibTeX · RIS.