Property Prediction of Functional Organic Molecular Crystals with Graph Neural Networks

Predicting the properties of molecular crystals is imperative to the field of materials design. In lieu of alternative methods, advances in machine learning have made it possible to predict the properties of materials before synthesis. This is especially important for organic semiconductors (OSCs) that are prone to exhibit polymorphism, as this phenomenom can impact the properties of a system, including the bandgap in OSCs. While graph neural networks (GNNs) have shown promise in predicting the bandgap in OSCs, few studies have considered the impact of polymorphism on their performance. Using the MatDeepLearn framework, we examine five different graph convolution layers of ALIGNN, GATGNN, CGCNN, MEGNet, and SchNet, which all have graph convolutions implemented in torch geometric. A dataset of functional organic molecular crystals is extracted from the OCELOT database, which has calculated density functional theory (DFT) values for the bandgap as well as several sets of polymorphs. The trained models are then evaluated on several test cases including the polymorphs of ROY. In future work we plan to examine the impact of graph representations on the performance of these models in the case of predicting properties of polymorphic OSCs.