Machine-learning-driven prediction and chemistry-informed screening of mechanical and thermal properties of silicate-based oxynitride glasses

Silicate-based oxynitride glasses combine high hardness, elastic stiffness, and thermal stability, but systematic exploration of their multicomponent composition space remains experimentally demanding because synthesis requires high temperatures, controlled atmospheres, and careful control of nitrogen incorporation and glass formation. Here, we develop a chemistry-informed machine-learning framework for predicting and screening the mechanical and thermal properties of oxynitride glasses. The broader literature compilation assembled for this project contained 482 records; the final harmonized working table used for the present descriptor analysis contained 127 compositions with the elemental fields required by the revised descriptor definitions. After target-specific data-quality checks, 98 hardness, 72 Young’s-modulus, and 122 glass-transition-temperature measurements remained available. Descriptor-complete common subsets of 98, 72, and 120 compositions, respectively, were used for fair comparison of baseline, chemistry-informed, and combined representations. Three aggregate composition descriptors were evaluated together with chemistry-informed quantities describing nitrogen substitution, modifier/former balance, formal cation and anion charge relations, and nominal cation field strength. Ridge regression, decision trees, random forests, XGBoost, and multilayer perceptrons were evaluated using identical five-fold cross-validation splits. The combined representation provided the highest mean cross-validated performance, with R 2 = 0.766 for Vickers hardness, 0.752 for Young’s modulus, and 0.739 for glass-transition temperature, corresponding to RMSE values of 0.452 GPa, 4.495 GPa, and 36.24 °C, respectively. Permutation analysis identified the N/O ratio as the strongest predictive coordinate for hardness and T g , whereas Young’s modulus depended jointly on N/O, nominal cation field strength, and modifier chemistry. Descriptor-correlation analysis revealed substantial algebraic redundancy among several engineered variables, emphasizing that permutation sensitivity should be interpreted as predictive association rather than mechanistic causality. Finally, composition-first charge-neutral screening was used to investigate three target-property scenarios. The screening identified composition regions approaching supported hardness and thermal targets while failing to reproduce strongly extrapolative targets such as 15 GPa hardness and 165 GPa Young’s modulus. The framework is therefore positioned as an interpretable decision-support approach for experimental prioritization rather than unconstrained inverse glass design.