Novel structural feature-descriptor platform for machine learning to accelerate the development of organic photovoltaics

Abstract

Although the evolution of organic photovoltaics (OPVs) has been remarkable, the discovery of new combinations of materials for ideal bulk heterojunction (BHJ) blend systems remains challenging because of their numerous considerations. Aiming to eliminate the typically labor-intensive trial-and-error approaches, machine learning (ML) models coupled with quantitative structure-property relationships are constructed to efficiently predict the photovoltaic parameters. To this end, a unique structural-feature descriptor that fragments a molecular structure into functional group (FG) units is designed. These units are indexed and the indices are combined to translate the molecular structure into a concise matrix of integers. Our novel descriptor allows the ML models to trace the FGs with ease and assign larger weights to those FGs that significantly contribute to the photovoltaic performance. Furthermore, the descriptor enables the expansion of OPV ML to ternary or multicomponent BHJ systems because of its concise expression. The proposed ML model achieves a high correlation coefficient of 0.86 between experimental and predictive power conversion efficiency of OPVs, despite using only the molecular structure information without any additional input data. Combined with the superior prediction capabilities, our novel approach promises to perfectly screen numerous BHJ cadidates, resulting in accelerating the development of the OPV.