DeGAT: A Dual-Expert Graph Attention Network for Partial Charge Prediction in Metal-Organic Frameworks.
TL;DR
DeGAT enables efficient and accurate prediction of partial atomic charges in MOFs while maintaining charge neutrality and physical consistency, providing a scalable parametrization scheme for high-throughput screening and molecular simulations of porous materials.
Abstract
The high computational cost of deriving REPEAT charges via periodic density functional theory (DFT) limits the large-scale screening of metal-organic frameworks (MOFs). To address this, we developed DeGAT, a dual-expert graph attention network for the rapid prediction of partial atomic charges. By incorporating an uncertainty-driven active learning strategy on the ARC-MOF database, the model achieves a test-set R2 of 0.985, with a mean absolute error (MAE) of 0.0314 e and a species-averaged MAE (SMAE) of 0.0527 e. Subsequent grand canonical Monte Carlo and Widom insertion simulations demonstrate that CO2, N2, and water adsorption properties calculated using DeGAT charges closely match those derived from standard REPEAT charges. These results demonstrate that DeGAT enables efficient and accurate prediction of partial atomic charges in MOFs while maintaining charge neutrality and physical consistency, providing a scalable parametrization scheme for high-throughput screening and molecular simulations of porous materials.