Graph Neural Networks for Molecules and Materials

A fundamental task of computational physics and chemistry is the calculation of quantum mechanical properties of molecules and materials. Unfortunately, these calculations require substantial computational resources and the involved electronic interactions are not amenable to manually designed approximations. Machine learning and graph neural networks (GNNs) in particular have recently emerged as an effective option to sidesteps manual approximations, enabling high accuracy while being several orders of magnitude faster than traditional methods.

In the first part of this talk, I will introduce the framework of message passing neural networks (MPNNs) and present ways of incorporating directional and geometric information in this framework. To alleviate the short-range nature of the MPNN framework, I will furthermore present a method to learn long-range interactions inspired by Ewald summation.

In the second part of this talk, I will dive into the evaluation of GNNs for energy and force prediction. I will first discuss evaluating GNNs for molecular dynamics simulations, highlighting possible pitfalls in this task. Finally, I will analyze the differences between models developed for small molecules and those optimized for large and diverse atomic systems such as the open catalyst (OC20) dataset.