Prediction of High-Performance Donor–Acceptor Pairs for Organic Photovoltaics with Machine Learning

Organic solar cells are widely recognized for their flexibility, light weight and semitransparency, all relevant for niche applications. A significant challenge in their development lies in the accurate prediction of their power conversion efficiency depending on the combination of the donor and acceptor selected in the bulk heterojunction. To address this issue, we developed a robust machine learning (ML) framework designed to establish correlations between molecular structure and device performance. A feature selection strategy, incorporating SHapley Additive exPlanations and Boruta algorithms, was employed to extract the most informative descriptors. Among the regression models that were systematically evaluated on a curated dataset comprising 1575 experimentally characterized donor–acceptor pairs, histogram-based gradient boosting demonstrated superior predictive performance, giving an R2 score of 0.79, with a low root mean square error of 2.16. Subsequently, the optimized model was used to predict new donor–acceptor pairs with PCEs above 20% and identify prospective candidates for further experimental validation.