Accelerating the Discovery of Superhalogens via Physics-Informed Graph Neural Networks

Superhalogens are clusters with electron affinities (EAs) exceeding those of halogen atoms and are promising for applications in energy, catalysis, and electronics. However, their discovery remains difficult due to the enormous number of possible compositions. Rapid, quantitatively reliable prediction of cluster electron affinity (EA) is therefore essential. This work proposed an Electron Affinity Graph Fusion Network (EAGFN) that combines local graph-convolutional embeddings of SOAP-augmented atomic features with a global many-body tensor representation (MBTR) within a lightweight residual architecture. Trained on the Cluster-AEA-2813 data set compiled from databases, EAGFN achieves a mean absolute error of 0.36 eV on the test set, outperforming conventional Graph Neural Networks (GNNs). An ensemble of five independently trained EAGFN models was then used to prioritize 1.5 × 10⁵ hypothetical cluster formulas, yielding over 2 × 10⁴ putative candidates with ensemble-mean EA values above the superhalogen threshold. Subsequent DFT validation of representative metal-ligand clusters, including MM'X_n-type systems, supported their high EAs, while ligand-substitution analysis provided a qualitative explanation for the role of fluorine ligands in enhancing electron-accepting ability.