Table 4 Inference times of GNN models for property prediction

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

Model

QM9

Matbench

MEGNet

11.996

11.137

M3GNet

19.715

20.089

TensorNet

14.945

13.694

SO3Net

14.371

32.601

  1. The numbers represent the inference times (in seconds) for MEGNet, M3GNet, TensorNet, and SO3Net on the QM9 (free energy) and Matbench (formation energy) test sets, which contain 6541 and 6637 structures, respectively. All property predictions were performed using a single Nvidia RTX 3090 and A6000 GPU for QM9 and Matbench, respectively.
Back to article page