Interpretable attention framework for materials property prediction and mechanistic discovery

• Architecture-embedded attention weights ensure intrinsic interpretability. • Bayesian optimization drives framework universality for diverse systems. • High predictive accuracy and mechanistic discovery for Mo-based alloys. • DFT calculations validate the discovered W-induced embrittlement. Machine learning in materials science is often limited by black-box predictions and scarce experimental data, which hinder reliable discovery and mechanistic understanding. Here we present SEAM, a self-explaining attention network that embeds interpretability directly into model architecture rather than relying on post-hoc analysis. SEAM records physically meaningful information during prediction through attention mechanisms and integrates multi-method cross-validation using attention analysis, integrated gradients, and SHAP to ensure robust interpretation. Specifically, Bayesian optimization enables automatic adaptation, achieving a predictive accuracy of R 2 = 0.84 for the ductility of molybdenum-based refractory alloys. By distinguishing causal mechanisms from spurious correlations, SEAM reveals a critical “correlation inversion” for Tungsten: while statistically correlated with high ductility, decoupling analysis identifies W as an intrinsic embrittler. This is validated by complementary first-principles calculations showing Re-induced bond softening versus W-induced stiffness preservation. This integrated framework bridges data-driven prediction with physical understanding, enabling trustworthy learning and rational design in data-scarce and mechanistically complex materials systems.