Table 1 Graph neural network models implemented in MatGL

From: Materials Graph Library (MatGL), an open-source graph deep learning library for materials science and chemistry

Name

Type

Brief description

Function

Ref

   

Prop. Pred.

MLIP

 

MEGNet

Invariant

GNN with global state vector.

Yes

No

5

M3GNet

Invariant

Extension of MEGNet with 3-body interactions. Used to implement the first FP as well as property models.

Yes

Yes

29

CHGNet

Invariant

GNN with regularization of node features using magnetic moments from DFT.

No

Yes

43

TensorNet

Equivariant

O(3)-equivariant GNN using Cartesian tensor representations, which is more computationally efficient compared to higher-rank spherical tensor models.

Yes

Yes

58

SO3Net

Equivariant

Minimalist SO(3)-equivariant GNN based on the spherical harmonics and Clebsch-Gordan tensor product.

Yes

Yes

49

  1. A brief summary of the architectures available in MatGL, including invariant models (MEGNet, M3GNet, CHGNet) and equivariant models (TensorNet and SO3Net), designed for property prediction (Prop. Pred.) and machine learning interatomic potentials (MLIP).
Back to article page