Auditing Machine-Learning Models and Their Training Data with Explainability and First-Principles Verification: Application to Spin Hall Conductivity
Machine-learning models for materials properties rest on two assumptions that standard validation never tests: that a model's features reflect the physics of the property rather than accidents of the training distribution, and that the training labels are themselves correct. We introduce a model-agnostic audit protocol for both, combining SHAP attribution, counterfactual partial dependence analysis, and Rashomon-style cross-model verification, with every finding adjudicated by targeted density functional theory (DFT). Demonstrated on intrinsic spin Hall conductivity using a composition-only Random Forest, the model needs no relaxed crystal structure, reaching accuracy competitive with structure-aware graph networks while remaining applicable to the far larger space of compositions for which no structure has been computed. The model audit reveals that the average p-valence descriptor becomes statistically entangled with Pt content - a property of the learned representation rather than the physics; DFT confirms the consequence, a Pt-free compound (HgOsPb$_2$) whose true SHC is nearly four times the prediction. The data audit exposes a thirtyfold error in the HfC training label, inherited undetectably by every black-box model trained on the same data. The protocol audits a model and its training data for the cost of a few DFT calculations, wherever one element dominates the high-property regime.