Predicting glass properties by using physics- and chemistry-informed machine learning models

Physics-and chemistry-informed machine learning (ML) models were trained by using descriptors in the element physical and chemical properties domain, which include stoichiometric, elemental-property-based, valance orbital occupation, and ionicity features. Young's modulus, shear modulus and electrical resistivity (rho) data for a group of oxide glasses were used to train artificial neural network (ANN), support vector machine (SVM), and random forest (RF) models. In comparison with experimental values, the ANN performs the best in predicting elastic moduli, whereas the RF is the best in predicting the temperature dependence of rho in terms of the coefficient of determination (R-2) value. The benefits of the ML models using descriptors in the element physical and chemical properties domain were demonstrated by revealing the relationships between the predicted glass properties and their first and second important features through a grid search.

Automated summary

Physics and chemistry-informed machine learning models were trained using descriptors in the element physical and chemical properties domain. These models successfully predicted the elastic moduli and temperature dependence of electrical resistivity for oxide glasses, and revealed the relationships between predicted glass properties and elemental features.