An Interpretable XGBoost Framework for Predicting Oxide Glass Density

Accurately predicting glass density is crucial for designing novel materials. This study aims to develop a robust predictive model for the density of oxide glasses and, more importantly, to investigate how physically informed feature engineering can create accurate and interpretable models that reveal underlying physical principles. Using a dataset of 76,593 oxide glasses from the SciGlass database, three machine learning (ML) models (ElasticNet, XGBoost, MLP) were trained and evaluated. Four distinct feature sets were constructed with increasing physical complexity, ranging from simple elemental composition to the advanced Magpie descriptors. The best model was further analyzed for interpretability using feature importance and SHapley Additive exPlanations (SHAP) analysis. A clear hierarchical improvement in predictive accuracy was observed with increasing feature sophistication across all models. The XGBoost model combined with the Magpie feature set provided the best performance, achieving a coefficient of determination (R2) of 0.97. Interpretability analysis revealed that the model’s predictions were overwhelmingly driven by physical attributes, with mean atomic weight being the most influential predictor. The model learns to approximate the fundamental density equation using mean atomic weight as a proxy for molar mass and electronic structure features to estimate molar volume. This demonstrates that a data-driven approach can function as a scientifically valid and interpretable tool, accelerating the discovery of new materials.