The Use of Deep Learning to Fast Evaluate Organic Photovoltaic Materials

Abstract

Abstract It is acknowledged that the structure of a material determines its activity or property. During the development of organic photovoltaic (OPV) materials, it is vitally important to build the relationship between chemical structures and photovoltaic properties. However, the conventional way based on trial‐and‐error experiments requires a significant amount of time and resources. Here, it is demonstrated that deep learning can be employed to quickly evaluate the performance of new OPV materials. The deep learning model allows direct use of pictures of chemical structures as input, possesses an excellent nonlinear analyzing capability, and has a low demand for computing power. After training the model with a database from the Harvard Clean Energy Project, it is able to predict the photovoltaic performance based on a given chemical structure of an OPV donor material. The prediction accuracy reaches 91.02% using a verification set of 5000 molecules. The codes are converted into visual pictures to understand how features are extracted by the model. In addition, the influence of database size on prediction accuracy is discussed. The model is further tested by using experimentally verified OPV materials and received positive results. Together, the results suggest that deep learning is promising for the quick evaluation of new OPV materials.