Evolutionary Algorithm-Guided LLMs for Physics-Informed Neural Network Design

2026-07-16

AI Summary

This work addresses the sensitivity of physics-informed neural networks (PINNs) to design choices such as architecture, loss weighting, and collocation point placement, and the inability of current large language models (LLMs) to accumulate training experience for continuous optimization. The authors propose a novel closed-loop automated design framework that integrates evolutionary algorithms with LLMs, generating executable PINN configurations over multiple generations. The search is dynamically guided by actual training outcomes, with prompts enriched by feedback on the successes and failures of parent configurations. This approach enables experience accumulation, elite preservation, diversity maintenance, and effective deduplication. In experiments spanning ten generations with 60 PINNs total, the best configuration in the final generation reduced solution errors by 2.97% and 95.38% compared to the initial population, while also revealing that low solution error can coexist with high PDE residual.

Abstract

Physics-informed neural networks (PINNs) are unusually sensitive to interacting choices of architecture, activation, loss weighting, collocation, optimization, and constraint enforcement. Large language models (LLMs) can propose these choices, but independent recommendations do not accumulate experience from previously trained PINNs. We propose a closed-loop evolutionary algorithm that guides an LLM to generate complete, executable PINN configurations across generations, using measured training outcomes to determine subsequent search decisions. The algorithm maintains an evaluated population and lineage, applies parent-conditioned mutation and crossover, preserves elite and diverse solutions, rejects effective duplicates, and converts parent-relative successes and failures into the next-generation context supplied to the LLM. Every proposed configuration is executed directly under an exact optimizer-step budget. On a one-dimensional multiscale wave equation, two independent ten-generation runs trained 60 PINNs for 600,000 optimizer steps. In both runs, the best configuration appeared in the final generation, with best mean-squared error reduced by 2.97\% and 95.38\% relative to the initial population. The stronger run validated residual connections and increased depth on separate branches, combined them in a later generation, and then refined width and collocation density. It also revealed that low solution error can coexist with a high PDE residual. These results demonstrate the feasibility of evolutionary-algorithm-guided LLMs for PINN design on a controlled PDE while motivating broader, physics-aware evaluation.

Problem

• Physics-informed neural networks
• Neural network design
• Design sensitivity
• Configuration optimization
• PDE-constrained learning

Innovation

• evolutionary algorithm
• large language models
• physics-informed neural networks
• automated design
• closed-loop optimization