Identifying correlation between the open-circuit voltage and the frontier orbital energies of non-fullerene organic solar cells based on interpretable machine-learning approaches

The open-circuit voltage (Voc) of the non-fullerene acceptors-based organic solar cells (NFAs-OSCs) under device operation conditions, as an essential photovoltaic parameter, is extensively studied for further power conversion efficiency (PCE) improvement. Generally, the Voc of binary bulk-heterojunction (BHJ) OSCs is roughly estimated by the energy level offset between the highest occupied molecular orbital of the donor (HOMO(D)) and the lowest unoccupied molecular orbital of the acceptor (LUMO(A)). Existing simulation and experimental approaches focus on studying the correlation between the Voc of PC61BM- and PC71BM-based OSCs with various donors. In solution-processed NFAs-OSCs, however, providing a numerical method for accurate estimates of Voc prediction is very difficult due to the extremely large pool of possible design donor-NFA combinations. It is also noted that using a conventional statistical model is challenging to accurately predict Voc for many different blends where the behavior between (|HOMO(D)|-|LUMO(A)|) offset and VOC is usually not a simple linear relationship. Herein, two tree-based ensemble machine-learning models of Random Forest and XGBoost are proposed to predict the Voc of NFAs-OSCs with a reasonable accuracy based on the intrinsic electronic parameters. In addition, the Shapley Additive Explanations (SHAP) analysis is applied not only for demonstrating the interpretability of the XGBoost model but also for visualizing the correlation between the Voc and the frontier orbital energies of NFAs-OSCs. This study demonstrates that the machine-learning approaches provide an empirical relation for accurately predicting the Voc of NFAs-OSCs, which might offer a strategy for efficiently designing new donor-NFA pairs to improve the Voc in devices.