Deciphering Experimental Reactivity of Metal Clusters Toward N₂ Activation Using Graph Neural Networks

Area of Science:

• Computational Chemistry
• Materials Science
• Chemical Physics

Background:

• Machine learning (ML) in catalysis often focuses on discrete electronic properties, neglecting continuous structural factors influencing reactivity.
• Understanding metal cluster reactivity is crucial for catalysis and materials design.

Purpose of the Study:

• To develop the first graph neural network (GNN) framework for modeling nitrogen (N2) activation reactivity in metal clusters.
• To integrate structural and electronic features for improved reactivity prediction.
• To establish an interpretable model for understanding structure-activity relationships.

Main Methods:

• Utilized a graph isomorphism network (GIN) to encode topological and atomic features of 245 metal clusters.
• Combined Density Functional Theory (DFT)-optimized structures with experimental reaction rates.
• Applied explainable AI techniques to identify key reactivity descriptors.

Main Results:

• The GNN framework achieved superior predictive performance on reaction rates for unseen metal clusters.
• Identified natural charge redistribution as a primary mechanism for ligand-mediated reactivity modulation.
• Subgraph charge polarization emerged as a potential descriptor, with distinct patterns in highly active clusters.

Conclusions:

• Established a novel graph-based interpretable framework for analyzing metal cluster reactivity.
• Demonstrated the importance of incorporating structural information in ML models for catalysis.
• Provided insights into the fundamental mechanisms governing small-molecule activation by metal clusters.