Domain-Validity-Gated Metamorphic Testing of Scientific ML Surrogates
Researchers have developed a new method for testing scientific machine-learning (SciML) surrogates, which approximate complex simulations. The proposed approach, called Domain-Validity-Gated Metamorphic Testing, addresses the challenge of verifying these surrogates when exact outputs are unavailable. It introduces a rubric to screen candidate metamorphic relations for domain validity and an executable asset format to record test details and verdicts. Case studies on MeshGraphNets and PhysicsNeMo demonstrated the method's ability to distinguish between model-level violations and out-of-domain applications. AI
IMPACT Enhances the reliability and trustworthiness of scientific machine learning models used in complex simulations.