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

Area of Science:

• Computational Chemistry
• Materials Science
• Artificial Intelligence

Background:

• Superhalogens possess electron affinities (EAs) greater than halogen atoms, making them valuable for energy, catalysis, and electronics.
• Discovering new superhalogens is challenging due to the vast number of potential cluster compositions.
• Accurate and rapid prediction of cluster EAs is crucial for advancing superhalogen research.

Purpose of the Study:

• To develop a novel computational method for predicting the electron affinity (EA) of chemical clusters.
• To accelerate the discovery of new superhalogen materials with potential applications.

Main Methods:

• Proposed an Electron Affinity Graph Fusion Network (EAGFN) integrating graph-convolutional embeddings and many-body tensor representations.
• Trained the EAGFN model on the Cluster-AEA-2813 dataset.
• Employed an ensemble of five EAGFN models to screen 1.5 × 10^5 hypothetical clusters.
• Validated promising candidates using Density Functional Theory (DFT) calculations.

Main Results:

• The EAGFN model achieved a mean absolute error of 0.36 eV on the test set, surpassing conventional Graph Neural Networks (GNNs).
• Identified over 2 × 10^4 putative superhalogen candidates from screening.
• DFT validation confirmed high EAs for representative metal-ligand clusters.
• Fluorine ligands were identified as key contributors to enhanced electron-accepting ability.

Conclusions:

• The EAGFN model offers a rapid and accurate approach for predicting superhalogen EAs.
• This AI-driven method significantly accelerates the identification of novel superhalogen candidates.
• The findings provide insights into the structural factors governing superhalogen properties, particularly the role of fluorine.