Attention-Based Interpretable Multiscale Graph Neural Network for MOFs.

Area of Science:

• Materials Science
• Computational Chemistry
• Machine Learning

Background:

• Metal-organic frameworks (MOFs) are promising for gas separation and storage.
• Graph neural networks (GNNs) are useful for MOF structure-property prediction.
• Crystal graphs require handling periodicity and multiscale features, unlike molecular graphs.

Purpose of the Study:

• To develop a novel method for constructing multiscale crystal graphs for MOFs.
• To propose an advanced GNN (MSAIGNN) that incorporates multiscale structural information and attention mechanisms.
• To improve the accuracy and interpretability of deep learning models for MOF property prediction.

Main Methods:

• Decomposition of crystal graphs into subgraphs based on interatomic interaction distances.
• Encoding of unit cell periodicity to capture global crystal structure.
• Development of MSAIGNN incorporating three-body bond angles and self-attention graph pooling.

Main Results:

• MSAIGNN achieved higher prediction accuracy for single-component adsorption and gas separation compared to traditional methods.
• The model effectively learned and utilized structural features across different scales, as confirmed by attention score visualization.
• The proposed method demonstrated reduced overfitting by minimizing interference from redundant interatomic interactions.

Conclusions:

• The multiscale crystal graph construction and MSAIGNN provide an efficient, multilayered, and interpretable deep learning approach for MOF property prediction.
• This method addresses the complexities of crystal structures and interatomic interactions at various scales.
• MSAIGNN offers a significant advancement in computational materials science for designing and discovering new MOFs.