Prediction of non-radiative voltage losses in organic solar cells using machine learning

One of the major hurdles that are preventing organic solar cells (OSCs) from leading the efficiency chart is non-radiative voltage loss (ΔVNR). So far, however, not much effort is made to predict voltage losses and unravel the correlation of losses with electronic and structural descriptors. From the literature, we create a dataset consisting of 154 unique donor:acceptor combinations with reported ΔVNR. The dataset includes information about frontier molecular orbitals (FMO), optical bandgap (Eg), molecular descriptors, and molecular fingerprints. Four machine learning (ML) algorithms (random forest regressor, gradient boosting regressor, support vector regressor, and artificial neural network) are used to predict non-radiative voltage loss and the results obtained are compared on the basis of Pearson rs, root mean squared errors, and mean absolute percentage errors. Best results are obtained with gradient boosting regressor by using FMO + Eg + RDKit descriptors (Pearson r = 0.859) and FMO + Eg + MACCS fingerprints (Pearson r = 0.857). We have also applied these ML models by using only molecular descriptors and only molecular fingerprints and got impressive results (Pearson r = 0.78 and 0.726). These results indicate that ML models can be effectively used for the prediction of ΔVNR and virtual screening of promising donor:acceptor combinations with reduced ΔVNR.