PHYSICS-AUDITED AGENTIC DISCOVERY IN SCIENTIFIC MACHINE LEARNING

In agentic scientific machine learning (SciML), Large Language Model (LLM) agents are increasingly used to automate the discovery and selection of surrogate models. A recurring problem is that these agents typically select models based on aggregate error metrics (like validation loss). However, a model with low error may still produce physically impossible results, such as violating boundary conditions, symmetry, or causality. This paper asks: how can we integrate rigorous, machine-checkable physics audits into the automated agentic discovery process to ensure selected surrogates are physically trustworthy?

The authors propose Physics-Audited Agentic SciML (PA-SciML), a verification-first framework. Unlike standard agentic workflows that prioritize error metrics, PA-SciML enforces a strict separation between predictive accuracy and physical consistency. The workflow includes:

• Fixed Evaluators: A standardized scoring routine defined before the search begins to ensure consistent comparison.
• Physics Auditor: A module that applies machine-checkable physics requirements to the outputs of every trained candidate.
• Adversarial Search: A separate process that searches for high-violation cases within the input domain to stress-test the surrogate without requiring reference solution fields.
• Advisory Probes: Numerical tests performed before full training to evaluate the potential of proposed model changes.

The authors demonstrate the framework using computational solid mechanics examples. In a static elasticity problem, the PA-SciML workflow successfully identified a surrogate that maintained low error while satisfying all linear-elastic checks. In a more complex transient elastodynamics problem, the framework proved decisive: an error-only baseline achieved a similar mean error to the PA-SciML-selected model but failed a critical causality check by responding to future loading history. The study concludes that per-candidate physics evidence is essential for model selection, as aggregate error metrics are insufficient to guarantee physical validity.