Machine Learning in Non-fullerene Organic Solar Cells: Accelerating Discovery, Design, and Understanding.

The emergence of nonfullerene acceptors (NFAs) has transformed the performance of organic solar cells (OSCs), driving laboratory power conversion efficiencies beyond 19%. Yet the immense combinatorial space of possible donor-acceptor materials renders exhaustive experimental exploration impractical. Machine learning (ML) and artificial intelligence have therefore become powerful tools for accelerating materials discovery, performance prediction, and molecular design in organic photovoltaics. This review provides a critical and systematic assessment of predictive and generative ML approaches applied to OSCs, encompassing methods ranging from random forests and gradient boosting to deep learning architectures such as graph neural networks, transformers, variational autoencoders, generative adversarial networks, and genetic algorithms. We examine advances in molecular representation and feature engineering, evaluate model performance in predicting key device metricsincluding power conversion efficiency, open-circuit voltage, short-circuit current density, and fill factorand assess the extent to which computational predictions translate into experimentally validated devices. We further discuss key limitations, including data set bias, distribution shift, morphology-related effects, chemical validity, and the synthetic accessibility of generated candidates, as well as the persistent gap between predicted and realized performance. Finally, we highlight emerging directions poised to shape the next phase of ML-guided OSC research, including physics-informed learning, multiobjective optimization, interpretable models, active learning coupled with first-principles calculations, and human-in-the-loop discovery pipelines. Collectively, these advances position machine learning as a central component in the rational design and accelerated development of next-generation organic photovoltaic materials.