Machine learning models for the discovery of direct band gap materials for light emission and photovoltaics

Abstract

Screening large materials spaces to find suitable candidates for optoelectronic applications requires a fast estimation of the material’s band gap and whether it is direct or indirect band gap. Obtaining accurate band structures with density functional theory (DFT) remains prohibitively computationally expensive. Machine learning is a promising approach to making such predictions faster. Multiple band gap prediction models have been demonstrated so far. Here, we expand such models to predict the direct–indirect band gap nature of the gap and apply them to discover new materials with desired properties. We explore two different models, a binary classifier and a regression-based model predicting the difference between the true band gap and the gamma point band gap. Our models achieved a true positive rate of up to 90 % in predicting the band gap type of materials. Starting from indirect band gap materials with band gaps relevant for PV and LED applications, we generate alloys that become direct-gap band gap materials, resulting in ∼30 candidates validated with DFT.