Deep Mining Stable and Nontoxic Hybrid Organic–Inorganic Perovskites for Photovoltaics via Progressive Machine Learning

Abstract

As promising new-generation sunlight-harvesting materials, hybrid organic-inorganic perovskites (HOIPs) have attracted a great deal of attention because of their outstanding advantage of high-power conversion efficiency and low-cost experimental synthesis. Tremendous chemical space and complexity of HOIPs, however, seriously hinder the applications of traditional trial-and-error and high-throughput density functional theory (HT-DFT) methods. Although the machine learning methods successfully accelerate the discovery of new stable and nontoxic HOIPs for photovoltaics, the performance of the current machine learning strategy is still severely limited by the quality of training input database, resulting in a large chemical space for further exploration. A progressive machine learning strategy is therefore introduced in the current study to investigate the impact of an input database enriched by a previous machine learning study, aiming to provide a more reliable and accurate approach to deep mining of the hidden HOIPs for sunlight harvesting. Enhancement in the performance indicators of a progressive machine learning strategy indicates that the data set generated by the previous round of machine learning study could dramatically enrich the training input database and improve its quality. Further DFT validations confirm that 96 out of 209 machine learning selected candidates have promising band gaps for light harvesting, so the prediction success rate of the current work is significantly enhanced compared to that of the previous work. Current study thence successfully verifies the feasibility of a progressive machine learning strategy for accurate and deep mining of hidden novel functional materials.