Intuitive tests to validate machine learning models against physics and domain knowledge

Abstract

The integration of machine learning (ML) faces a fundamental challenge because current explainability methods provide statistical attribution and lack systematic frameworks for validating explanations against physical laws, constitutive relationships, and engineering principles. From this lens, this work aims to address the following two questions: when physics-informed models are inapplicable/infeasible, how can we relate ML model predictions to established physics principles and engineering domain knowledge, and when are such relationships reliable enough? This review introduces an intuitive hierarchical framework comprising ten validation levels, from fundamental physics compliance to operational context alignment. Our framework addresses critical gaps, including conservation law validation, multiscale physics consistency, temporal dependency verification, and uncertainty quantification in physics-constrained explanations. We also provide comprehensive implementation guidelines for both engineers and ML developers, including decision matrices, risk assessment protocols, and domain-specific validation procedures. This work concludes by identifying fundamental research questions and proposes standardized benchmarks for evaluating physics-informed explainability.