Physics-Audited AI Improves Scientific Machine Learning Reliability

In scientific machine learning (SciML), Large Language Model agents are increasingly used to discover surrogate models, often selecting the best one based solely on error metrics. However, a low error rate does not guarantee that the model's predictions adhere to fundamental physical laws, such as boundary conditions, causality, or superposition, which are crucial for mechanical systems. To address this, the Physics-Audited Agentic SciML (PA-SciML) workflow has been developed. This approach prioritizes verification, establishing a scoring evaluator and deriving machine-checkable physics requirements *before* the search for models begins. Each candidate model is then rigorously checked against these physical principles, not just its error performance. The workflow also includes advisory numerical probes before training and tests individual modeling changes to understand their impact on scores. In computational solid mechanics examples, PA-SciML successfully selected surrogate models that not only had low validation error but also passed essential linear-elastic checks. Crucially, in transient elastodynamics, an error-only baseline failed a causality check by responding to future events, while the PA-SciML selected model passed, demonstrating superior physical fidelity.