Graph neural networks to predict atomic transition charges and exciton couplings in organic semiconductors

Exciton couplings between molecules in organic semiconductors are important parameters for simulating exciton diffusion, but they are time-consuming to compute from first-principles. Previous works have developed machine-learned models to predict exciton couplings, but most of these models are restricted to specific molecules and cannot generalize over databases of organic materials. In this paper, we present a graph neural network (GNN) that can predict exciton couplings between organic molecules by using atomic transition charges as an intermediary. Our GNN is shown to predict exciton couplings between important fused-ring electron acceptors (FREAs), as well as many other molecules found in the Cambridge Crystallographic Data Center crystal database, with good accuracy. We also show that the predicted couplings can be used for accurate simulations of exciton diffusion. This work, therefore, overcomes the key limitation of previous machine-learned models for exciton couplings and thereby brings us closer to the possibility of performing high-throughput virtual screening of organic materials for photovoltaic applications.