Graph neural networks for materials science
Title: Graph neural networks for materials science
SNIC Project: Berzelius-2022-164
Project Type: LiU Berzelius
Principal Investigator: Fredrik Lindsten <fredrik.lindsten@liu.se>
Affiliation: Linköpings universitet
Duration: 2022-08-22 – 2022-09-29
Classification: 10201
Keywords:

Abstract

Graph neural networks (GNNs) have recently shown very promising results in the materials science field. For example, they can be used to predict energies and forces of atomic systems, and large GNNs like GemNet [1] currently set the state of the art on various datasets. Although models like GemNet set the current state of the art, they are computationally expensive. This limits their use for long simulations and large systems. Therefore, an initial goal for this project is to improve the performance of computationally cheap GNNs like PaiNN [2], using so called knowledge distillation. Knowledge distillation for atomic systems is especially promising since we can leverage a pretrained (large and computationally expensive) teacher model to provide labels for an arbitrary amount of additional data, and we can choose any points in chemical space. In the end, this could pave the way for using models that are both cheap and making accurate predictions, and we plan on presenting this work at a top machine learning conference. [1] Gasteiger, Becker, Günnemann. GemNet: Universal Directional Graph Neural Networks for Molecules. NeurIPS 2021 [2] Schütt, Unke, Gastegger. Equivariant message passing for the prediction of tensorial properties and molecular spectra. ICML 2021