Explainable machine learning for predicting the glass transition temperature of fluoride glasses

Abstract

The glass transition temperature (Tg) is a fundamental physical parameter for evaluating the thermal stability and processing window of fluoride glasses. Traditional empirical formulas and experimental trial-and-error methods suffer from low efficiency and limited accuracy when dealing with multi-component complex systems. In this study, based on the SCIGlass Next database, 1,474 data records of fluoride glass compositions and their corresponding Tg values were utilized to construct and compare four machine learning models: Extreme Gradient Boosting Regressor (XGBR), Random Forest Regressor (RFR), Back-Propagation Neural Network (BPNN), and Support Vector Regression (SVR). The results demonstrate that the Bayesian-optimized XGBR model exhibits superior performance, achieving a coefficient of determination R2 of 0.9700 and a mean absolute error (MAE) of only 6.25 °C on the test set, significantly outperforming BPNN and SVR. To improve model interpretability and address the black-box nature of machine learning, the SHAP (SHapley Additive exPlanations) interpretability framework was introduced to quantify the contribution of each chemical component to Tg from both global and local perspectives. The analysis reveals that AlF3 and GeO2, acting as networkenhancing components, exert a significant positive influence on Tg, whereas alkali metal fluorides such as NaF and LiF, serving as network modifiers, markedly decrease Tg. This work not only provides an accurate data-driven tool for predicting the properties of fluoride glasses but also elucidates the physicochemical links between composition and structural stability through a data-driven approach, offering valuable guidance for the compositional design of novel fluoride glasses.