Improving Crystal Property Prediction from a Multiplex Graph Perspective.

Area of Science:

• Materials Science
• Computational Chemistry
• Data Science

Background:

• Graph Neural Networks (GNNs) are effective for crystal property prediction.
• Existing methods often overlook intrinsic crystal graph characteristics.
• A data science perspective is needed to fully leverage crystal graph information.

Purpose of the Study:

• To propose a novel Potential Multiplex Crystal Graph Neural Network (PMCGNN) for enhanced crystal property prediction.
• To explore and leverage the intrinsic information within crystal graphs from a data science viewpoint.
• To improve the accuracy and efficiency of predicting material properties.

Main Methods:

• Reconstructing crystal graphs into a multiplex graph with global and local views.
• Employing Graph Transformers (GTs) and Message Passing Neural Networks (MPNNs) for representation learning.
• Augmenting GTs with positional and structural encodings from local graphs for enhanced interaction.

Main Results:

• PMCGNN demonstrates superior performance in 9 crystal prediction tasks.
• The model effectively learns atomic representations by integrating global and local graph perspectives.
• Comprehensive experiments were conducted on JARVIS and Materials Project datasets.

Conclusions:

• The proposed PMCGNN significantly enhances crystal representation learning.
• The multiplex graph approach captures both infinite potentials and local atomic interactions.
• PMCGNN offers a powerful and computationally efficient tool for materials property prediction.