Deciphering Experimental Reactivity of Metal Clusters Toward N(2) Activation Using Graph Neural Networks.

Machine learning (ML) analysis of gas-phase metal cluster reactivity has emerged as a pivotal approach in this field. However, existing ML studies relying on electronic properties have primarily focused on discrete features, with less consideration of continuous structural factors that also govern cluster reactivity. Here, we present the first graph neural network (GNN) framework to model N(2) activation reactivity across 245 metal clusters, combining DFT-optimized structures with experimental reaction rates collected from the literature and a public data set. Through encoding both topological connectivity and atomic-level features (e.g., natural charge, valence electron occupancy, and atomic number), the graph isomorphism network (GIN) achieves superior predictive performance on reaction rates of unseen clusters. Explainable analysis reveals that natural charge redistribution likely serves as the primary mechanism for ligand-mediated reactivity modulation. Furthermore, subgraph charge polarization shows potential as a reactivity descriptormetal-core subgraphs in highly active clusters exhibit significantly lower charge polarization than mixed metal-ligand subgraphs in less active clusters. This work establishes a graph-based interpretable framework for understanding structure-activity relationships of small-molecule activation by metal clusters.