Search for oxide glass compositions using Bayesian optimization with elemental-property-based descriptors

Abstract

Our study shows that machine learning technique, Bayesian optimization (BO) can efficiently find high-refractive-index glasses from a large number of candidate compositions using data from the INTERGLAD database. The effect of the parameters (i.e., descriptors) input to the BO algorithm on search performance is described. The results show that elemental-property-based (EPB) descriptors, recently applied in materials science, are more effective than the component-amount-based ones traditionally used in the study of glass. The results suggest that BO with EPB descriptors can accelerate the search for glass compositions with desirable properties.