Uncovering Structure–Performance Relationships in Organic Photovoltaics: Interpretable Machine Learning Model for Predicting the Power Conversion Efficiency

Abstract

Organic photovoltaics (OPVs) represent a promising photovoltaic technology, but the design of candidate molecules has traditionally followed a trial-and-error approach, which is inefficient. However, machine learning provides a data-informed strategy by learning from large OPV material datasets, supporting the accelerated discovery and optimization of high-performance OPV materials. In this study, we use an extreme gradient boosting (XGBoost) model to predict the density functional theory-calculated power conversion efficiency (PCE) of OPV donor materials using structural features extracted from the Harvard Photovoltaic Dataset (HOPV15) dataset. To enhance predictive performance, we select the most informative molecular fingerprints based on the averaged feature importance scores from both random forest and XGBoost. Our XGBoost model achieves state-of-the-art predictive accuracy on HOPV15 with $R^{2}$ = 0.918 and RMSE = 0.302%, outperforming prior methods. Using SHapley Additive exPlanations, we identify key Morgan and PubChem substructures that influence PCE, offering interpretable insights. This framework supports accurate, explainable OPV prediction and holds promise for high-throughput screening.