Physics-Audited Agentic SciML: verification-first workflow

Physics-Audited Agentic Discovery in Scientific Machine Learning, an arXiv paper submitted 8 Jul 2026 by Diab W. Abueidda, Bilal Ahmed, Panos Pantidis and Mostafa E. Mobasher (arXiv:2607.07379), introduces Physics-Audited Agentic SciML, or PA-SciML, a verification-first workflow for agentic discovery of surrogate models.

PA-SciML fixes the scoring evaluator before search, derives reviewable machine-checkable physics requirements, checks each trained candidate on its outputs, and separately searches prescribed input ranges or measured load-history spans for high-violation cases without reference solution fields. The workflow reports a surrogate as verified only under the stated checks.

What is Physics-Audited Agentic SciML?

PA-SciML is a verification-first agentic SciML workflow that enforces per-candidate physics evidence on predicted fields rather than relying solely on aggregate error scores. It fixes the scoring evaluator prior to search, produces reviewable machine-checkable physics requirements, and runs those checks on each candidate before reporting verification.

The authors present PA-SciML as a procedural shift: candidates are not simply ranked by an error metric; each model is subjected to explicit physics checks such as boundary-condition satisfaction, superposition, stiffness scaling and causality checks. The workflow also searches input or load-history spans to find high-violation cases even when no reference solution is available.

How does PA-SciML change agentic surrogate selection in the reported experiments?

In computational-solid-mechanics examples reported in the paper, PA-SciML selected surrogates that met physics checks while matching or improving validation error compared with error-only baselines. In a static elasticity run the selected surrogate had lower validation error than the error-only baseline while both models passed common linear-elastic checks.

In a transient elastodynamics run the authors report that an error-only baseline with similar mean error failed a stricter causality check by responding to future parts of the loading history, while the PA-SciML-selected surrogate passed the stated checks. The paper emphasizes that the main distinction is per-candidate physics evidence on predicted fields, not a richer aggregate score.

How does the workflow operate in practice?

PA-SciML first fixes an evaluator and then derives a set of machine-checkable physics requirements; it runs those checks on candidate outputs and separately probes input ranges or measured load histories for violations. When enabled, it adds advisory numerical probes before training and tests single modeling edits to record which isolated changes produce score gains.

Those advisory probes and single-edit tests are used to document which modeling changes correlate with score improvements before reusing them. The workflow therefore combines pre-search probes, per-candidate checks, and targeted post-training violation searches that do not require reference fields.

Why it matters

PA-SciML pushes agentic SciML from purely error-driven selection to verification by physics constraints, shifting accountability from aggregate scores to explicit, reviewable evidence on each surrogate. That matters for mechanics domains where boundary conditions, scaling laws and causality are essential and where low error alone can hide physics-violating behavior.

The paper shows concrete failure modes: models with similar mean error can still fail causality checks. PA-SciML’s requirement that a surrogate be "verified only under the stated checks" forces model choices that respect domain constraints rather than just minimizing validation error.

What to watch

Look for implementations or code and datasets linked from the arXiv entry for arXiv:2607.07379 and for follow-up studies that apply PA-SciML to larger, multi-physics problems. The paper was submitted on 8 Jul 2026 and the arXiv PDF metadata lists a file size of 3,438 KB; those entries may carry code or data links in future revisions.

References: Diab W. Abueidda; Bilal Ahmed; Panos Pantidis; Mostafa E. Mobasher, "Physics-Audited Agentic Discovery in Scientific Machine Learning," arXiv:2607.07379 (submitted 8 Jul 2026).