Machine Learning Accelerated Study of Defect Energy Levels in Perovskites

Manipulating the defect tolerance is one of the effective ways to maintain the high power conversion efficiency and keep the stability of perovskite semiconductor materials. So, rapid screening for defects and trap states in the perovskite semiconductor candidates is urgently needed. Theoretical investigations of defects based on density functional theory (DFT) are still limited by their extremely high consumption of computational resources and time. We implement an accelerated material discovery approach using artificial intelligence and DFT, which can predict the defect transition levels in the candidate perovskite semiconductor materials. To verify the accuracy of our models, Cs3Sb2Br9 and Cs2SnBr6, which are out of the dataset that we used in machine learning (ML) model construction, are taken as examples. The extrapolation of ML prediction models and the results given by DFT calculations are compared for defect transition energy levels. The two methods are consistent with each other with very small errors. Our strategy avoids complex and time-consuming computational work based on DFT and provides quick and efficient screening of physical properties with low cost.