Full prediction of band potentials in semiconductor materials

Abstract

A machine learning (ML) framework to predict the physical band potentials for a range of semiconductor materials, from metal sulfide, oxide, and nitride, to oxysulfide and oxynitride, is hereby described. A valence band maximum (VBM) model was established via the transfer learning of a large dataset of 2D materials (1382 samples, but with incorrect VBM potentials) onto a much smaller dataset of physically measured VBM for bulk 3D materials (87 samples) on a crystal graph convolutional neural network. This resulted in predictions with experimental accuracy (RMSE = 0.27 eV), which is a 3-fold improvement compared with ML trained on the physical dataset without transfer learning (RMSE = 0.75 eV). When combined with the bandgap prediction model (RMSE = 0.29 eV), a full prediction of conduction and valence band potentials can be made, which to the best of our knowledge, is the first for any ML framework. The variation of band potentials across low-index surfaces was predicted correctly and verified with reported physical potentials. In fact, the framework is able to capture variation in band potentials associated with minor atomic position alterations. Based on this, a general trend between surface atomic displacement and VBM shift was observed across various semiconductor materials. The model is not yet able to cope with major rearrangement of atomic sequence on surface layers, i.e., surface reconstructions, since it was not trained with such data but can be easily done so with specifically designed dataset. As an example application, the ML framework was used for the screening of potential photocatalytic materials for visible light water splitting. A total of 824 materials was successfully identified, including those experimentally-verified in the literature.