EvoPINN: LLM Agent Automates Physics-Informed Neural Network Design

Researchers have introduced EvoPINN, a novel framework that leverages a Large Language Model (LLM) to autonomously create algorithms for physics-informed neural networks (PINNs). While PINNs are effective in addressing partial differential equations (PDEs), they typically involve extensive manual engineering of neural representations, loss functions, and optimization processes. EvoPINN transforms this challenge into an algorithm discovery task grounded in execution, separating neural representations from the training routines. An LLM agent iteratively suggests modifications based on memory within a modular search framework, ensuring the scientific integrity by dismissing mathematically unsound or numerically unstable solutions. This approach seeks to automate the previously labor-intensive design process. The research paper can be found on arXiv with the identifier 2607.26490.