Study on bandgap predications of ABX3-type perovskites by machine learning

Abstract

The accurate bandgap prediction of perovskites has great significance for solar cell devices. Although density functional theory can be used for the calculations of material bandgaps, this method requires rich material calculation knowledge, and there are often some questions about whether the results are consistent with actual experimental results or not. To address this, the present work adopts machine learning (ML) to predict bandgaps of perovskites, where we collect 227 sets of experimental bandgap data of perovskites from the latest 1254 publications, to establish and identify 4 models from 24 kinds of ML models. The results of the models achieve high accuracy with root mean square error (RMSE) of down to 0.55 and meanwhile, the pearson correlation coefficient of up to 99%. In addition, our ML models give the effect of each chemical composition constituting the ABX3-type perovskites on the bandgaps by using the SHAP value, and they can be well explained in physics. These results all show the powerful potential of machine learning to fast and accurately predicate the bandgaps of perovskite for solar cell devices.