Theoretical Prediction of Vickers Hardness for Oxide Glasses: Machine Learning Model, Interpretability Analysis, and Experimental Validation

Abstract

Accurate prediction of the Vickers hardness for multi-composition oxide glasses is critical in glass science and technology. Here, we developed machine learning models to efficiently learn and predict Vickers hardness of oxide glasses across a high-dimensional composition space including 56 oxides. Three algorithms, LASSO, support vector machine (SVM), and random forest (RF), were employed, and their predictive performances were evaluated. The RF model exhibits the highest accuracy on the hardness prediction, especially under high loads, and can capture the nonlinear relations between Vickers hardness and compositions. Shapley additive explanation analysis implemented on the RF model provides valuable insights into the design of high-hardness glasses. To experimentally validate the RF model, systematic prediction in a typical SiO2-Al2O3-B2O3-CaO-MgO-Li2O-Na2O-K2O glass system has been carried out, and a glass comparable to Corning® Gorilla® Glass Victus® 2 was prepared, showing that the application values of the machine learning method for finding novel glass materials.