Bayesian reverse design of high-efficiency perovskite solar cells based on experimental knowledge constraints

To alleviate high costs and lengthy trial-and-error periods associated with traditional optimization methods for perovskite solar cells (PSCs), we developed a data-driven reverse design framework for high-efficiency PSCs. This framework integrates machine learning and Bayesian optimization (BO) to accelerate the optimization process of PSCs by intelligently recommending the most promising parameter configurations for PSCs, such as device structure and fabrication processes. To improve the robustness of the framework, we first designed a two-stage sampling strategy to alleviate the issue of imbalanced dataset classes. Subsequently, by integrating “experimental knowledge constraints” into the BO process, we achieved precise parameter configurations, thus avoiding discrepancies between predicted and actual results due to parameter mismatches. Finally, using SHapley Additive exPlanations, we unveiled key factors influencing the power conversion efficiency (PCE), such as the composition of perovskite solvents. Our framework not only precisely predicted the PCE of PSCs with an area under the curve of 0.861 but also identified the optimal parameter configurations, achieving a high probability of 0.981. This framework offers substantial support for minimizing redundant experiments and characterizations, effectively accelerating the optimization process of PSCs.