Interpretable machine learning for predicting properties of multi-component oxide glasses and experimental validation

The precise prediction of multi-component oxide glass properties represents a critical frontier in advancing glass science and materials engineering. This study establishes an interpretable machine learning framework to predict physical and optical properties of glasses across a high-dimensional compositional space spanning 57 oxide components. By leveraging both composition-property and physics-informed composition-property datasets, we systematically evaluate six algorithms: Elastic Net Regression (ENR), Support Vector Machine (SVM), Random Forest (RF), Extreme Gradient Boosting (XGBoost), Multilayer Perceptron (MLP), and Convolutional Neural Network (CNN). Our analysis reveals that RF and XGBoost models achieve the highest predictive accuracy, with the physics-informed dataset effectively reducing overfitting compared to the sole composition-property dataset. SHapley Additive exPlanations (SHAP) analysis elucidates non-linear feature interactions, uncovering synergistic effects between components. For instance, we identify the strong coupling between rare earth oxides and network formers/modifiers, as well as a strong correlation between SiO 2 and B 2 O 3 . To validate the extrapolation capabilities, we apply the model to a thermal neutron-sensitive glass system (SiO 2 -MgO-Al 2 O 3 -Li 2 O-La 2 O 3 ), demonstrating the remarkable agreement between predicted and measured properties. This work underscores the potential of physics-informed machine learning to revolutionize the design and discovery of novel oxide glass materials with tailored properties, paving the way for innovations in various high-tech industries.