Toward Predicting Efficiency of Organic Solar Cells via Machine Learning and Improved Descriptors

Abstract To design efficient materials for organic photovoltaics (OPVs), it is essential to identify the largest number of parameters that control their properties and build a model using these parameters (known as descriptors) for the prediction of the power conversion efficiency (PCE). By constructing a dataset for 280 small molecule OPV systems, it is found that for all high‐performing devices, frontier molecular orbitals of donor molecules are nearly degenerated and in such cases, orbitals other than just highest occupied molecular orbital (HOMO) and lowest unoccupied molecular orbital (LUMO) are involved in exciton formation, exciton dissociation, and hole transport processes influencing the macroscopic properties of OPVs. Machine learning approaches, including random forest, gradient boosting, deep neural network are used to build models for the prediction of PCE using 13 important microscopic properties of organic materials as descriptors. Quite impressive performance of the gradient boosting model (Pearson's coefficient = 0.79) indicates that it can certainly be applied to high‐throughput virtual screening of promising new donor molecules for high‐efficiency OPVs.