Glass‐Forming‐Region‐Informed Machine Learning Reveals Quantitative Criteria for Network Glass Formation

The accurate prediction of glass formation in inorganic materials remains a formidable challenge in condensed matter science. Current machine learning approaches often rely on glass‐forming ability datasets biased toward successful glass formers, compromising their generalizability and interpretability. Here, we present a comprehensive glass‑forming‐region‐informed machine learning (GFRI‐ML) framework, integrating both composition and physical‐property descriptors to quantitatively predict glass formation across diverse oxide systems. Our curated dataset comprises 44 753 oxide glass samples from 2771 ternary systems, specifically incorporating negative (non‐glass‐forming) instances to eliminate selection bias. We developed two hierarchical predictive models: (i) a Composition‐GFR model using only compositional features, achieving a cross‑validation accuracy of 90.6%, and (ii) a Property‑GFR model incorporating 51 weighted physical‐chemical descriptors, reaching 94.6% accuracy. Both models demonstrated superior generalization performance on independent, held‐out experimental systems. Furthermore, Shapley Additive Explanations analyses were employed to elucidate generalizable physical laws governing vitrification, revealing critical thresholds within the kinetic, thermodynamic, and topological dimensions. This work provides an interpretable, high‐fidelity tool for glass prediction and establishes a quantitative, physics‐grounded foundation for the accelerated discovery of novel inorganic glasses.