Machine learned defect level predictor for semiconductors

Electronic levels introduced by impurities and defects in the band gap are critically important in semiconductors for optoelectronic and photovoltaic (PV) applications. The energetics and energy levels of point defects can be reliably predicted using density functional theory (DFT) computations. However, the requirement of large supercells and inclusion of charged states make these computations very expensive, and trends and knowledge from previous calculations are not exploited in subsequent ones. In this work, we develop machine learned defect level predictors based on substantial DFT data for two classes of semiconductors: halide perovskites (MAPbX₃, MA = methylammonium, X = Cl/Br/I), and Cd-based chalcogenides (CdX, X=Te/Se/S). DFT data was generated for formation energies and transition levels of hundreds of vacancy, interstitial and substitutional defects, following which correlation analysis and random forest regression were used to map the properties to a set of unique numerical descriptors, resulting in cheap, accurate, quantitative prediction models. Such models can lead to accelerated prediction of defect states and allow efficient materials design of defect-tolerant semiconductors as well as semiconductors with suitably placed defect levels.