Machine-learning-assisted exploration of new non-fullerene acceptors for high-efficiency organic solar cells

Highlights•Comprehensive database of training models•Explainable models are used to predict the PCE of organic solar cells•Three non-fullerene acceptor molecules were designed based on SHAP analysis•The error between model prediction and experimental verification is less than 3%SummaryThe power conversion efficiency (PCE) of organic solar cells (OSCs) has exceeded 19% with the development of non-fullerene acceptors (NFAs). Here, machine learning (ML) models based on the inputs of both molecular descriptors and fingerprints with different algorithms are investigated to assist the exploration of NFAs. Although the model based on the fingerprints exhibits slightly inferior performance parameters, it can deliver faster and high-throughput computation and much stronger generalization ability due to the decreased model complexity to avoid overfitting. Moreover, Shapley additive explanations (SHAP) techniques are used to explain the models for the design and synthesis of three NFAs. An excellent agreement between the experimental and predicted PCEs is achieved, with a relative error of less than 3%. Therefore, our study can offer a strategy for rapid PCE prediction and detailed analysis of molecular fingerprints and descriptors for high-throughput screening of NFAs for high-efficiency OSCs.Graphical abstract