Graph neural network with self-attention for material discovery. (16th February 2023)

• Record Type: Journal Article
• Authors: Chen, Xuesi; Jiang, Hantong; Lin, Xuanjie; Ren, Yongsheng; Wu, Congzhong; Zhan, Shu; Ma, Wenhui
• Is Part Of: Molecular physics. Volume 121:Number 4(2023)
• Publication Date: 2023-02-16
• Subjects: Graph neural network -- machine learning -- material property prediction -- attention feature fusion; Molecules -- Periodicals; Chemistry, Physical and theoretical -- Periodicals; Molécules -- Périodiques; Chimie physique et théorique -- Périodiques; 539.6.05
• DOI: 10.1080/00268976.2023.2176701
• Languages: English
• ISSNs: 0026-8976

Abstract

Abstract : Technology has developed as a result of computerisation, and a wide range of other fields, such as physics and chemistry, have been involved in the application of machine learning. nodes and edges together form a crystal so that it is easy to represent as a graph. Some typical models such as MEGNET show good generalisation in material property prediction by using a graph neural network instead of the traditional density functional theory(DFT). The author proposes a fusion self-attention graph neural network (FSGN) model that incorporates a graph neural network with fusion and attentional mechanisms to predict material properties. The convolutional self-attention module is mainly used to extract the importance of autocorrelation and cross-correlation in node, edge and global information. Multi-head attention feature fusion is used after shallow additive fusion to get more discriminative features. Compared with other Machine Learning models like MEGNET and CGCNN, it can be demonstrated that the prediction accuracy(ACCU) of our model has been improved to some extent. GRAPHICAL ABSTRACT: UF0001