Improving Crystal Property Prediction from a Multiplex\nGraph Perspective

Graph\nneural networks (GNNs) have proven to be effective tools\nfor the rapid and accurate prediction of crystal properties. While\nmost existing methods focus on enriching representations of crystal\nstructures, they do not deeply explore the characteristics of crystal\ngraphs and leverage their intrinsic information from a data science\nperspective. In this work, we propose the potential multiplex crystal\ngraph neural network (PMCGNN) for crystal property prediction. Based\non the characteristics of crystal graphs, we reconstruct the crystal\ngraph into a multiplex graph that includes two views: a global crystal\ngraph embodying infinite potentials and a local crystal graph capturing\nlocal atomic interactions. We employ graph transformers (GTs) and\nmessage passing neural networks (MPNNs) architectures to learn the\natomic representations of these two perspectives. Specifically, we\naugment the GT by incorporating positional encodings and structural\nencodings from the local crystal graph. This approach promotes interaction\nbetween the two perspectives, enabling the model to learn both node\npositional and graph structural information from different viewpoints\nthrough an attention mechanism. As a result, it enhances the model’s\nability to learn crystal representations. We conduct comprehensive\nexperiments on the JARVIS and the Materials Project data sets for\nevaluation. Results show that PMCGNN presents superior performance\nin 9 crystal prediction tasks while maintaining reasonable computational\nexpense.