Autonomous LLM Agent for Materials Science Theory Development

An innovative autonomous large language model (LLM) agent is capable of conducting comprehensive, data-driven development of materials theory independently. This agent autonomously chooses equation forms, creates and runs code, and evaluates the fit between theory and data. It integrates methodical reasoning with specialized tools, adapting its strategy while documenting its choices. For well-established equations like the Hall-Petch equation and Paris law, it accurately identifies the governing equations and provides dependable predictions. In contrast, for more specialized equations such as Kuhn's equation regarding the HOMO-LUMO gap in conjugated molecules, its effectiveness varies based on the model used, with GPT-5 demonstrating superior equation recovery. Additionally, the agent can suggest new predictive relationships that extend beyond existing theories.