Researchers at Shanghai Jiao Tong University, Shanghai University of Electric Power and Shandong Normal University have addressed the traditional trial-and-error method for preparing high-efficiency perovskite solar cells (PSCs) by introducing a goal-driven approach that integrates machine learning and data mining techniques to rapidly screen high-efficiency PSCs based on key features.
By predicting high-efficiency PSCs and identifying the dominant factors affecting their performance, namely the perovskite bandgap and the total thickness of the electron transport layer (ETL), this research aims to provide valuable insights for optimizing preparation processes and advancing the development of high-efficiency PSCs, thus significantly contributing to the renewable energy sector.
The team compiled an extensive dataset based on the 43,612 PSCs composition dataset of The Perovskite Database Project and generated elemental features with the Magpie tool. The scientists generated categorization features HPCE(high PCE) based on PCE. They investigated various parameters and explored the importance of the selected features on the prediction of the HPCE by means of an interpretable ML approach.
Finally, the team provides a prediction of the bandgap of perovskite materials and analyzed the impact of a range of perovskite-related descriptors on bandgap prediction, which helps to identify the indirect impact of perovskite layer-related features on the prediction of high-efficiency PSCs, and provides a reference to optimize the impact of parameters on PCE.
In synergy with cutting-edge ML techniques, the team introduced a rapid and objective-driven framework, specifically tailored to forecast the potential of high-efficiency PSCs. Simultaneously, this framework stands poised to illuminate the intricate interplay of optimization parameters on the PCE. Within this methodology, the team constructed an intricate map delineating the individual contributions of features to the realization of high-efficiency PSCs, harnessing the interpretability of ML.