Predictive\nModeling\nand Design of Organic Solar Cells:\nA Data-Driven Approach for Material Innovation

We present a robust machine learning\nmethodology to accurately\npredict key photovoltaic parameters in organic solar cells (OSCs).\nOur approach involves curating a comprehensive quantum mechanical\ndatabase of 300 experimentally validated OSC devices with distinct\ndonor/acceptor combinations. Through a two-step screening process,\nwe identify descriptors correlated with crucial properties such as\nshort-circuit current (<i>J</i><sub>SC</sub>), open-circuit\nvoltage (<i>V</i><sub>OC</sub>), fill-factor (FF), and power\nconversion efficiency (PCE<sub>max</sub>). Utilizing a LASSO model\nfor feature selection and four different supervised machine learning\ntechniques for prediction, our model achieves high accuracy, with\ngradient boosting showing exceptional performance for <i>J</i><sub>SC</sub>, <i>V</i><sub>OC</sub>, and PCE<sub>max</sub>. Shapley additive explanations (SHAP) analysis reveals the influential\nfeatures and the intricate nonlinear relationships governing OSC performance.\nAdditionally, we extend our model’s utility by predicting photovoltaic\nparameters for a larger data set of 4680 donor–acceptor combinations,\nincluding 120 newly designed nonfullerene acceptors and 39 experimentally\nknown donor polymers. Our results highlight 18 donor–acceptor\ncombinations with a power conversion efficiency exceeding 15%, emphasizing\nthe efficacy of our approach in evaluating OSC materials. This work\nprovides valuable insights for advancing photovoltaic research and\nserves as a powerful tool for the virtual screening of promising donor/acceptor\npairs, accelerating the development of high-performance OSC materials\nand devices.