Equivariant Graph Neural Networks for Prediction of Tensor Material Properties of Crystals
Abstract
Traditional machine learning methods applied to the material sciences have often predicted invariant, scalar properties of material systems to great effect. Newer, coordinate equivariant models promise to provide a coordinate system dependent output in a well defined manner, but recent applications often neglect a direct prediction of directional (i.e. coordinate system dependent) quantities and instead are used to predict still just invariant quantities. This component-wise prediction of tensorial properties is achieved by decomposing tensors into harmonic subspaces via a \textit{tensor spherical harmonic decomposition}, by which we may also associate arbitrary tensors with the irreducible representations of the rotation group. This essentially allows us to read off tensors component-wise from the output representations of these equivariant models. In this work, we present results for the prediction of various material property tensors directly from crystalline structures. Namely, given some material's crystalline structure, we may predict tensor components of dielectric, piezoelectric, and elasticity tensors directly from the output of a $SE(3)$ equivariant model.
- Publication:
-
arXiv e-prints
- Pub Date:
- June 2024
- DOI:
- 10.48550/arXiv.2406.03563
- arXiv:
- arXiv:2406.03563
- Bibcode:
- 2024arXiv240603563H
- Keywords:
-
- Physics - Computational Physics;
- Condensed Matter - Materials Science