Bandgap prediction of hybrid metal halide perovskites by transfer machine learning: From three to two dimensionality

In the past decade, three-dimensional (3D) and two-dimensional (2D) hybrid organic–inorganic perovskites (HOIPs) have attracted much attention due to their excellent power conversion efficiency, becoming candidate materials in the fields of photovoltaics, thermoelectric and optoelectronics. Although machine learning could be utilized to predict the bandgaps of HOIPs, conveniently improving their prediction accuracy presents a significant challenge due to the limitations on the size and homogeneity of datasets. In order to increase the prediction accuracy of 2D HOIPs, we have theoretically investigated the bandgaps of 3D and 2D HOIPs by using transfer machine learning methods, built the database including 26,920 and 15,060 compounds. We found that the features of chemical elements and B-sites exert a more significant influence on the bandgaps of 2D HOIPs than they do on 3D. Our findings suggest that the prediction accuracy of 2D HOIPs is markedly inferior to that of 3D, therefore more features in the former are necessary than those in the latter to achieve the same prediction accuracy, which can be attributed to the complex crystal structures and irregular dependency on halide ions in the former. The calculation results demonstrate that the mixed models based on both of 3D and 2D HOIPs datasets have significantly enhanced the prediction accuracy for HOIPs, especially for 2D HOIPs. Ultimately, we found the machine learning models can effectively discover the hidden nonlinear patterns and complex relationships that exist between crystal lattice, chemical element features and target bandgaps to provide valuable insights into the optoelectronic properties of HOIPs.