How Machine Learning Predicts and Explains the Performance of Perovskite Solar Cells. Issue 6 (7th March 2022)

• Record Type: Journal Article
• Authors: Liu, Yiming; Yan, Wensheng; Han, Shichuang; Zhu, Heng; Tu, Yiteng; Guan, Li; Tan, Xinyu
• Is Part Of: Solar RRL. Volume 6:Issue 6(2022)
• Page Start: n/a
• Page End: n/a
• Publication Date: 2022-03-07
• Subjects: machine learning -- perovskites -- power conversion efficiencies -- SHAP; Solar energy -- Periodicals; Photovoltaic power generation -- Periodicals; Solar energy -- Research -- Periodicals; Photovoltaic power generation -- Research -- Periodicals; Periodicals; 333.7923
• DOI: 10.1002/solr.202101100
• Languages: English
• ISSNs: 2367-198X

Abstract

Abstract : Characterizing the electrical parameters of perovskite solar cells (PSCs) usually requires a lot of time to fabricate complete devices. Here, machine learning (ML) is used to reduce the device fabrication process and predict the electrical performance of PSCs. Using ML algorithms and 814 valid data cleaned from 2735 peer‐reviewed publications, ML prediction models are built for bandgap, conduction band minimum, valence band maximum of perovskites, and electrical parameters of PSCs. These prediction models have excellent accuracy, and the root mean square error of the prediction models for bandgap and power conversion efficiency (PCE) reaches 0.064 eV and 1.58%, respectively. Among the many factors that affect the performance of PSCs, those factors play a major role in the lack of comprehensive explanation. Through the prediction model of electrical parameters and Shapley Additive explanations theory, the factors affecting the PCE of PSCs are explained and analyzed. It can not only verify the objective physical laws from the perspective of ML, but also conclude that among the 13 features, the content of formamidinium/NH2 CHNH2 + plays the most important role in improving the PCE of PSCs. These results show that ML has great application possibilities in the PSC field. Abstract : Seven prediction models built by machine learning algorithms demonstrate incredible prediction performance. A series of Shapley Additive Explanations explanation strategies for power conversion efficiency (PCE) were proposed, which not only comprehensively analyzed the 13 main features affecting PCE, but also ranked the most influential feature as FA (formamidinium/NH2 CHNH2 + ).