Machine Learning Augmented Discovery of Chalcogenide Double Perovskites for Photovoltaics

Abstract

Hybrid organic inorganic perovskite solar cells based on CH3NH3PbI3 have drastically increased in efficiency over the past several years and are competitive with decades-old photovoltaic materials such as CdTe. Despite this impressive increase, significant issues still remain due to the intrinsic instability of CH3NH3PbI3 which degrades into carcinogenic PbI2. Recently, double halide perovskites which use a pair of 1+–3+ cations to replace Pb2+, such as Cs2InSbI6, and chalcogenide perovskites, such as BaZrS3, have been explored as potential replacements. In this work, double chalcogenide perovskites are explored to identify novel photovoltaic absorbers that can replace CH3NH3PbI3. Due to the large space of possible compounds, machine learning methods are used to classify materials as potential photovoltaic absorbers using data from the periodic table, eliminating wasteful computation. A random forest algorithm achieves a cross-validation accuracy of 86.4% on the constructed data set. Over 450 possible replacements are identified via traditional and statistical methods with Ba2AlNbS6, Ba2GaNbS6, Ca2GaNbS6, Sr2InNbS6, and Ba2SnHfS6 as the most promising alternative when thermodynamic stability, kinetic stability, and optical absorption are considered.