Predicting Electronic Properties of Organic π-Conjugated Systems with Graph Neural Networks

Organic π-conjugated molecules have widespread applications namely in field-effect transistors, solar cells, (bio)sensors, and dyes. Accurate estimations of molecular electronic and optical properties using state-of-the-art density functional theory are time-consuming for large systems. Using graph neural network (GNN) and data from the OCELOT database, we trained machine learning (ML) models to predict properties, namely frontier molecular orbital energies, reduction and oxidation energies, relaxation energies, and low-lying singlet and triplet excitation energies. The trained GNNs can provide explanations for the predictions, thereby providing insights into the structure-property relationships. Further, we trained ML models to predict material properties, including the intermolecular electronic couplings between molecular dimers and crystal properties including bandgaps and cohesive energies. These models provide distinct connections between molecular and crystal properties, which we aim to build on as we develop and deploy machine-informed materials design for organic semiconductors.