Accurate bandgap predictions of solids assisted by machine learning

The bandgap of the material is a primary property, which affects their performance and applications. Recently, with the emergence of high-throughput simulations, various materials databases are developed based on the density functional theory (DFT). However, for existing databases, the bandgaps are often underestimated since the exchange-correlation functional is treated by the generalized gradient approximation (GGA) with Perdew-Burke-Ernzerh (PBE) approach during the DFT calculations. To better describe the bandgaps, more accurate approach should be employed, such as Heyd-Scuseria-Ernzerh (HSE) hybrid functional. However, this method is extremely time-consuming, which limits its applications. In this work, we employ the machine learning (ML) approach to predict the bandgaps of solids at the HSE level. We first develop a classifier model to identify nonmetals from the database, which shows excellent performance with the area under curve (AUC) up to 0.99. To predict the bandgaps of nonmetals, three ML models are trained and tested based on the selection of different features. These models can accurately predict the HSE bandgaps of solids, with the cross-validation score of 96% and root mean square error (RMSE) of 0.28 eV. Moreover, we apply these ML models to predict the bandgaps from Materials Project database at the HSE level, which contain 126324 inorganic compounds. These data are fully accessible from our newly released code for further study. Thus, our work not only provides an efficient approach to accurately predict the bandgaps of solids, but also accelerates the discovery and development of functional materials.