Evolutionary AI Optimizes Physics-Informed Neural Networks

Physics-Informed Neural Networks (PINNs) are powerful tools for solving scientific problems, but their performance is highly sensitive to numerous design choices, including architecture, activation functions, loss weighting, and optimization strategies. Manually configuring these parameters is complex and time-consuming. This research introduces an innovative approach where an evolutionary algorithm (EA) is used to guide large language models (LLMs) in the automated design of PINNs. The proposed system operates in a closed loop: the LLM generates complete PINN configurations, which are then executed and evaluated based on their training outcomes. This performance feedback informs the evolutionary algorithm, which then applies principles like mutation, crossover, and selection to refine subsequent generations of LLM-generated configurations. The algorithm maintains a population of evaluated solutions, preserving elite designs and promoting diversity. In experiments on a multiscale wave equation, this EA-guided LLM system successfully reduced the mean-squared error of the best PINN configuration by a substantial margin over multiple generations. The process autonomously discovered effective architectural elements, such as residual connections and increased depth, demonstrating the feasibility of using evolutionary algorithms to intelligently explore and optimize complex neural network designs for scientific applications.