Examining the Influence of Graph Representation on Property Prediction of Polymorphic Organic Molecular Crystals

Organic semiconductors (OSCs) can exhibit polymorphism, which may impact the physichochemical properties of the material, including the bandgap. With the advent of machine learning, graph neural networks have been used to predict the bandgap of OSCs. However, these frameworks struggle with differentiating between polymorphs. Here, we examine the performance of two graph representations, one including incorporation of angular information and one including only edge and node information, on two datasets, the Organic Materials Database (OMDB) and Organic Crystals in Electronic and Light-Oriented Technologies (OCELOT). We find that incorporating angle featurization improves the performance of the model on both databases. We also test the models on two cases: the polymorphs of 5-methyl-2-[(2-nitrophenyl)amino]-3-thiophenecarbonitrile (ROY) and an augmented version of the PAH101 dataset. We find that angular featurization significantly improves the ability of the model to differentiate between polymorphs in the prediction of the bandgap.