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

2026-07-17 · Source: Artificial Intelligence · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Physical Sciences & Chemistry, Mathematics & Computational Sciences · Depth: Expert, quick

Summary

A closed-loop evolutionary algorithm guides large language models (LLMs) to generate complete, executable Physics-informed neural network (PINN) configurations. This approach addresses PINNs' sensitivity to architectural and training choices, which LLMs alone struggle with due to a lack of cumulative experience. The algorithm uses measured training outcomes to inform subsequent search decisions, maintaining an evaluated population and lineage, applying parent-conditioned mutation and crossover, and preserving elite solutions. Each proposed configuration is executed 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. The best configurations emerged in the final generation, reducing mean-squared error by 2.97% and 95.38% relative to the initial population. One run specifically validated residual connections and increased depth.

Key takeaway

For AI Scientists and Machine Learning Engineers designing Physics-informed neural networks, you should consider integrating evolutionary algorithms with large language models for architecture search. This approach overcomes the limitations of static LLM recommendations by iteratively refining configurations based on training outcomes, leading to substantial performance gains, such as a 95.38% reduction in mean-squared error. Furthermore, ensure you monitor both solution error and PDE residual, as low solution error does not guarantee a low PDE residual.

Key insights

Evolutionary algorithms can guide LLMs to iteratively optimize Physics-informed neural network designs.

Principles

• PINN design benefits from cumulative experience.
• Evolutionary search can refine LLM-generated configurations.
• Low solution error can coexist with high PDE residual.

Method

A closed-loop evolutionary algorithm guides an LLM to generate PINN configurations. It uses training outcomes to determine search decisions, applies mutation/crossover, and converts results into next-generation context for the LLM.

In practice

• Use evolutionary algorithms to optimize PINN architectures.
• Explore residual connections and increased depth in PINN design.
• Monitor both solution error and PDE residual during PINN training.

Topics

• Physics-informed Neural Networks
• Evolutionary Algorithms
• Large Language Models
• Neural Architecture Search
• Multiscale Wave Equation
• PINN Optimization

Best for: AI Scientist, Machine Learning Engineer, Research Scientist

Related on AIssential

• Evolutionary Two-Stage Hyperparameter Optimization Strategies for Physics-Informed Neural Networks — A two-stage evolutionary approach effectively optimizes PINN hyperparameters, improving accuracy and robustness under computational constraints.
• Neural Operator-enabled Topology-informed Evolutionary Strategy for PDE-Constrained Optimization — NOTES efficiently optimizes PDE-constrained inverse design by combining neural operators with evolutionary strategies in a compact latent space.
• Algorithmic algorithm development with LLMs: A Case Study on LLM-Usage for Contraction Order Optimization in Tensor Networks — LLMs can develop and improve algorithms for complex problems like tensor network optimization, guided by verifiers, but human oversight remains crucial.
• Deep Neural Networks as Discrete Dynamical Systems: Implications for Physics-Informed Learning — PINNs and finite-difference methods acquire similar knowledge via distinct matrix structures, revealing a robustness-explainability tradeoff.
• A Unified Physics-Informed Neural Network for Modeling Coupled Electro- and Elastodynamic Wave Propagation Using Three-Stage Loss Optimization — PINNs can model coupled electro-elastodynamic systems, but face challenges with error propagation and time-dependent accuracy in multiphysics.
• Large Language Model-Driven Full-Component Evolution of Adaptive Large Neighborhood Search — LLM-driven evolution can automate and enhance complex metaheuristic algorithm design by decoupling and optimizing core components.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Artificial Intelligence.