Graph Neural Networks for Materials Chemistry

Graph neural networks (GNNs) are a rapidly expanding class of machine learning models with exceptional potential for materials chemistry problems across a wide range of spatial and time scales. A key strength in GNNs lies in their high expressivity and capacity for learning from large datasets over traditional feature engineering approaches in the materials sciences. Here, we develop a GNN-based software platform, MatDeepLearn, which offers broad customizability in data processing, model construction, hyperparameter selection and post-processing for materials informatics applications. In particular, we show the importance of the large hyperparameter space offered by the MatDeepLearn framework towards obtaining high performance in materials property predictions. We also demonstrate the viability of GNNs for the prediction of atomistic to macroscopic quantities, and across a large range of applications, including in catalysis, gas capture, and for molecular dynamics.