Deep Transfer Learning: A Fast and Accurate Tool to Predict the Energy Levels of Donor Molecules for Organic Photovoltaics

Abstract

Molecular engineering is driving the recent efficiency leaps in organicphotovoltaics (OPVs). A presynthetic determination of frontier energy levelsmakes the screening of potential molecules more efficient, exhaustive, andcost-effective. Here, a convolutional neural network is developed to predictthe highest occupied and lowest unoccupied molecular orbital(HOMO/LUMO) levels of donor molecules for OPV. The model takes a 2Dstructure image and returns a prediction of its HOMO/LUMO levelscomparable to experimental values. Insufficient experimental datasets areovercome with transfer learning where the model is initially trained on thelarge Harvard Clean Energy Project dataset and then fine-tuned usingexperimental data from the Harvard Organic Photovoltaic dataset. Errormargins on predicted HOMO/LUMO levels below 200 meV are achieved,without any chemical knowledge implemented. Noticeably, the model outputshave higher accuracy and precision than corresponding density functionaltheory (DFT) estimations. The model and its limitations are further tested ona home-built dataset of commercially available donor polymers reported inOPVs (e.g., P3HT, PTB7-Th, PM6, D18). The results demonstrate both thepractical utility of this model, to foster rational molecular engineering for OPVoptimization, and the potential for deep learning techniques, in general, torevolutionize the energy materials research and development sector.