Attention-Based Interpretable Multiscale Graph Neural Network for MOFs

Metal–organic frameworks (MOFs) hold great potential
in
gas separation and storage. Graph neural networks (GNNs) have proven
effective in exploring structure–property relationships and
discovering new MOF structures. Unlike molecular graphs, crystal graphs
must consider the periodicity and patterns. MOFs’ specific
features at different scales, such as covalent bonds, functional groups,
and global structures, influenced by interatomic interactions, exert
varying degrees of impact on gas adsorption or selectivity. Moreover,
redundant interatomic interactions hinder training accuracy, leading
to overfitting. This research introduces a construction method for
multiscale crystal graphs, which considers specific features at different
scales by decomposing the crystal graph into multiple subgraphs based
on interatomic interactions within varying distance ranges. Additionally,
it takes into account the global structure of the crystal by encoding
the periodic patterns of the unit cells. We propose MSAIGNN, a multiscale
atomic interaction graph neural network with self-attention-based
graph pooling mechanism, which incorporates three-body bond angle
information, accounts for structural features at different scales,
and minimizes interference from redundant interactions. Compared with
traditional methods, MSAIGNN demonstrates higher prediction accuracy
in assessing single-component adsorption, gas separation, and structural
features. Visualization of attention scores confirms effective learning
of structural features at different scales, highlighting MSAIGNN’s
interpretability. Overall, MSAIGNN offers a novel, efficient, multilayered,
and interpretable approach for property prediction of complex porous
crystal structures like MOFs using deep learning.