A machine learning workflow for large-scale discovery of direct bandgap double perovskites

Direct bandgap double perovskites are gaining great attention as a stable and eco-friendly alternative to single perovskites, exhibiting considerable potential in photovoltaics, luminescence, and catalysis. However, the high costs, lengthy experimental trial-and-error cycles, and the substantial computational demands of conventional first principles-based calculations have hindered the effective screening of direct bandgap double perovskites across the vast space of potential materials. In this study, we present a machine learning-based workflow capable of rapidly screening double perovskites with direct bandgap from the periodic table, achieving a remarkable 90 % accuracy. Leveraging this approach, we have identified 176 double perovskites that have a direct bandgap and exhibit excellent stability, 153 of which have not been reported. In addition, we find that the ionic radius of the B-site and B′-site of double perovskites are highly correlated with the nature of the bandgap by explainable machine learning. This framework provides an efficient and accurate method for identifying promising double perovskites for optoelectronic applications.