Equivariant graph neural network surrogates for predicting the properties of relaxed atomic configurations

TL;DR

This paper introduces an equivariant graph neural network model that predicts DFT-calculated properties of atomic structures, offering greater flexibility and accuracy over traditional methods.

Contribution

The authors develop an EGNN framework that inherently respects system symmetries, enabling accurate predictions of atomic configurations and energies beyond training data.

Findings

EGNN accurately predicts atomic displacements and strain tensors

The model generalizes well to unseen configurations

It provides insights comparable to DFT calculations

Abstract

Density functional theory (DFT) calculations determine the relaxed atomic positions and lattice parameters that minimize the formation energy of a structure. We present an equivariant graph neural network (EGNN) model to predict the outcome of DFT calculations for structures of interest. Cluster expansions are a well established approach for representing the formation energies. However, traditional cluster expansions are limited in their ability to handle variations from a fixed lattice, including interstitial atoms, amorphous materials, and materials with multiple structures. EGNNs offer a more flexible framework that inherently respects the symmetry of the system without being reliant on a particular lattice. In this work, we present the mathematical framework and the results of training for lithium cobalt oxide (LCO) at various compositions of lithium and arrangements of the lithium atoms. Our results demonstrate that the EGNN can accurately predict quantities outside the training set including the largest atomic displacements, the strain tensor and energy, and the formation energy providing greater insight into the system being studied without the need for more DFT calculations.