Uncovering universal and system-specific mechanisms of glass formation across covalent, ionic, and metallic networks via SHAP-integrated machine learning

Establishing a unified yet system-specific physicochemical framework for glass formation across covalent, ionic, and metallic systems remains a fundamental challenge, as existing models are largely confined to a single glass family or heavily biased toward compositions that have been successfully vitrified. Here we bridge this gap using a success-bias-free dataset of 78,889 compositional entries drawn from glass-forming region maps. Encoding each composition as an elemental-property matrix derived from the Magpie database, we train a scalable XGBoost framework that attains 89.6% classification accuracy on the all-glass dataset, and generalizes robustly to extrapolative elemental substitutions and higher-dimensional spaces. Shapley analysis of all-glass and system-specific models shows that the four glass families share the same set of frustration mechanisms whose relative weights redistribute systematically with bonding character. Across all families, governing descriptors exhibit a bounded, non-monotonic dependence that we term frustration inversion, whereby compositional heterogeneity stabilizes the amorphous state only up to a family-specific threshold that narrows as network rigidity decreases, furnishing quantitative bounds on the confusion principle. The analysis further identifies BCC-polymorph-referenced elemental descriptors as compositional-level predictors that encode structural preferences toward vitrification. These results establish a cross-system foundation for amorphous stability and a data-driven route to functional glass design.