SEAM Framework Ensures Global Consistency in Scientific ML Explanations

Scientific machine learning (SciML) models are often validated based on local accuracy, such as performance within a subdomain or for individual predictions. However, ensuring that these local explanations cohere into a single, globally admissible scientific account remains a significant challenge. This paper introduces Scientific Explanation-Admissibility Machines (SEAM), a novel framework designed to make this local-to-global consistency computable. SEAM is generator-agnostic and represents explanations for each region as structured entities with state, closure, and observation channels, potentially including contract metadata. It compares neighboring explanations on their overlaps, converting any disagreement into a channel-resolved obstruction. This obstruction precisely locates inconsistencies and allows for testing competing accounts by restricting repairs to permissible revisions. The framework can refute or retain accounts based on exact feasibility, or provide empirical attribution when exact repair is not possible. SEAM also distinguishes inconsistency from non-identifiability and monitors learned generators under distribution shift. Through nineteen experiments on synthetic partial differential equation systems and out-of-distribution Fourier neural operator monitoring, SEAM successfully detected incompatible explanations even when local predictions were accurate, attributing failures to specific channels and overlaps.