Machine learning-driven prediction of organic solar cell performance: a data-centric approach to molecular design

Organic solar cells (OSCs) offer a promising route toward flexible and sustainable energy technologies, yet predictive modeling of device parameters remains challenging due to the chemical diversity of donor-acceptor systems and morphology-dependent effects. In this work, we present the first systematic demonstration of using autoencoder-compressed molecular fingerprints with tree-based machine learning models to predict key OSC performance metrics-power conversion efficiency (PCE), open-circuit voltage (V_oc), short-circuit current (J_sc), and fill factor (FF)-from a broad experimental dataset of  2500 donor-acceptor pairs, including both fullerene and non-fullerene acceptors. These compact models, trained on compressed descriptors of only 32 dimensions, achieved strong predictive accuracy (Pearson $r>0.95$ , $MAE<0.4$ , $RMSE<0.95$ ) while remaining lightweight enough to run on standard computing hardware. As a complementary result, some k-nearest neighbor models achieved near-perfect correlations ( $r\sim 0.99$ ) and quite small errors ( $MAE<0.044$ and $RMSE<0.4$ ) in general, demonstrating the surprising strength of simple, instance-based learners when sufficient descriptive features are available. Supporting analyses reveal that fullerene datasets are more easily modeled than chemically diverse non-fullerene sets, that fingerprints encode substantial structural information, and that kernel density analyses identify critical ranges of molecular weight and energy offsets for high-efficiency devices. Collectively, this study establishes compressed fingerprint descriptors as a powerful, computationally inexpensive foundation for predictive modeling in OSCs, while also showcasing the unexpected efficacy of k-NN models trained on conventional descriptors. Together, these approaches provide a scalable path toward high-throughput prediction and guided molecular design of next-generation organic photovoltaic materials.

The dataset used in this work comprises approximately 2500 experimentally characterized donor-acceptor pairs from bulk heterojunction OSCs. These include both fullerene and non-fullerene acceptor systems. For each pair, the database provides electronic descriptors, polymerization-related metrics, and the SMILES representations of the donor and acceptor molecules. Molecular fingerprints were computed from SMILES codes using the RDKit and CDK cheminformatics toolkits. A variety of machine learning models were explored, including feedforward neural networks, autoencoders for feature compression, tree-based ensemble methods, and kernel-based regression algorithms. Hyperparameter tuning was carried out using the Optuna and BayesSearchCV libraries to ensure optimal model performance.