Physics-Audited Agentic Discovery in Scientific Machine Learning

This paper introduces Physics-Audited Agentic SciML (PA-SciML), a verification-first workflow for LLM-driven surrogate-model discovery. It fixes the evaluator before search, derives reviewable and machine-checkable physics requirements, audits every candidate on predicted fields, and searches prescribed input ranges or measured load histories for violations without reference solution fields. In computational solid mechanics examples, the workflow selects a lower-validation-error surrogate for static elasticity, while in transient elastodynamics it rejects an error-only baseline with similar mean error because it responds to future loading information. The central contribution is candidate-level physics evidence, rather than adding physics terms to an aggregate score.