Graph neural networks (GNNs) have received intense interest as a quickly expanding class of machine learning models with exceptional potential for materials chemistry. A primary strength in GNNs is their flexible and learnable representations which make them adaptable to wide variety of chemistry-based applications. Here, we present a workflow and testing platform, MatDeepLearn, for reproducibly assessing and comparing GNNs and other machine learning models, implemented as a python package. Using this package, we demonstrate the effectiveness of GNNs for a broad range of chemical and materials systems and prediction tasks, and benchmark the comparative performance of several state-of-the-art models. We show how MatDeepLearn can create top-performing models with minimal human effort or technical expertise, and also outline the potential GNN model space which can be further tuned to achieve an even higher performance for materials predictions. Using this approach, it is now feasible to effectively harness the rapidly growing wealth of high-throughput materials databases for rapid, data-driven materials discovery.

Graph neural networks for materials chemistry

ACS Fall 2021

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.