Attention Mechanism in Line Graph Neural Network for Material Property Prediction

ABSTRACT Graph Neural Networks (GNNs) models are intensively used for predicting the properties of molecular materials. To improve the extraction of structural features from molecular representations, the line graph representation introduced in Atomistic Line Graph Neural Network (ALIGNN) can explicitly incorporate bond angle information. However, in terms of the convolution process, ALIGNN still employs the edge‐gated graph convolution from previous GNNs, making it challenging to extract more effective information. In this work, combined with the line graph representation, we introduce multiple attention mechanisms into GNNs and propose two GNN models, N‐ALIGNN and T‐ALIGNN, which include the Node‐Attention Layer (NAL) and Self‐Attention Layer (SAL) respectively. These models can extract more structural features from both bond graph and line graph, thereby improving prediction accuracy. We conducted tests on three public datasets: JARVIS‐DFT, Materials Project, and QM9. The results demonstrate that the performance of the two models is improved in different prediction tasks, proving that the introduction of appropriate attention mechanisms can enhance the prediction accuracy of the models.