High-Throughput Prediction of Metal-Embedded Complex Properties with a New GNN-Based Metal Attention Framework

Area of Science:

• Computational Chemistry
• Materials Science
• Machine Learning

Background:

• Metal-embedded complexes (MECs), such as transition metal complexes (TMCs) and metal-organic frameworks (MOFs), are crucial in catalysis and materials science.
• Accurate modeling and property prediction for MECs are challenging due to their complex coordination environments.
• Traditional graph neural networks (GNNs) struggle to distinguish key metal-ligand interactions from standard covalent bonds.

Purpose of the Study:

• To introduce a novel Metal Attention (MA) framework for GNNs to improve the prediction of MEC properties.
• To address the limitations of existing GNNs in capturing the unique characteristics of metal coordination.

Main Methods:

• Developed a Metal Attention (MA) framework that converts heterogeneous graphs into homogeneous ones, emphasizing metal-feature coordination.
• Employed hierarchical pooling and metal cross-attention mechanisms within the MA framework.
• Compared the MA framework against 11 established GNN algorithms, including heterogeneous variants.

Main Results:

• Achieved an average accuracy improvement of 32.07% for predicting TMC properties.
• Demonstrated up to 23.01% improvement in predicting MOF CO2 absorption.
• Confirmed that enhanced performance is due to the MA architecture, not just increased model size, through robustness tests.

Conclusions:

• The MA framework offers a significant advancement in accurately predicting properties of metal-embedded complexes.
• Provides a powerful statistical tool for optimizing and designing novel materials, including catalysts and gas storage systems.
• Highlights the importance of specialized attention mechanisms for modeling complex chemical structures.