Large Language Model Based Agent for Automated Discovery in Computational Physics

Abstract

Scientific discovery in computational physics can often be framed as the optimization of quantitatively evaluable objectives subject to physical constraints. While researchers excel at formulating such problems, they frequently devote substantial effort to iterative refinement of methods and solution strategies. To accelerate this process, we introduce PhyNex, an autonomous agent that systematically explores the solution space of scorable scientific tasks by coupling large language model (LLM)-guided search with domain-specific computational tools that enforce physical consistency. PhyNex operates via progressive local search, accumulates reusable knowledge from both successful and failed attempts, and produces interpretable exploration trajectories that reveal which algorithmic components drive performance improvements. We validate PhyNex on three representative and scientifically important problems: predicting frequency-dependent dielectric spectra of semiconductors from crystal structure, designing probabilistic-circuit heuristics for Max-Cut on graphs, and optimizing charging protocols for Dicke quantum batteries in the chaotic coupling regime. Across the three tasks, PhyNex autonomously identifies solutions that match or exceed state-of-the-art approaches designed by human scientists, yielding search-averaged improvements of up to 3.8\% in spectral similarity, up to 15.0\% in normalized mean cut for Max-Cut, and 5.9\% in ergotropy at the $80\mathrm{k}$ training checkpoint in open exploration. These findings demonstrate that LLM-based agents with structured, feedback-driven exploration can substantially accelerate the path from problem specification to effective implementation, suggesting a practical division of labor in which scientists define objectives and constraints while automated systems navigate the methodological search space.