Graph Transformer with Convolution Parallel Networks for Predicting Single and Binary Component Adsorption...

Area of Science:

• Materials Science
• Computational Chemistry
• Artificial Intelligence

Background:

• Metal-organic frameworks (MOFs) are crucial for carbon capture and storage (CCS) due to their tunable properties.
• Machine learning (ML) accelerates the discovery of high-performance MOFs for CCS applications.
• Existing ML models often require non-graph data, limiting their direct application to crystal structures.

Purpose of the Study:

• To develop a novel graph-based ML model for predicting MOF adsorption properties.
• To enable accurate prediction of both single- and binary-component adsorption using only crystal structure information.
• To screen large MOF databases and identify promising candidates for carbon capture.

Main Methods:

• Developed GC-Trans, a graph transformer model incorporating parallel convolution networks.
• Utilized graph-structured data from crystal diagrams as input features.
• Extracted and fused local and global feature information for enhanced model expression.
• Applied transfer learning to assess model generalizability for CO2/CH4 separations and CH4 uptake.

Main Results:

• GC-Trans accurately and efficiently predicts MOF adsorption performance for single- and binary-component systems.
• The model demonstrates strong expression and generalization capabilities by integrating local and global features.
• Screening of the ARC-MOF database identified MOF structures meeting target adsorption requirements.
• Transfer learning applications for CO2/CH4 separation and CH4 uptake prediction showed good performance.

Conclusions:

• GC-Trans offers a powerful, graph-only approach for predicting MOF adsorption properties.
• The model facilitates efficient screening and discovery of advanced MOFs for carbon capture.
• The demonstrated transferability highlights the model's potential for diverse gas separation and storage applications.