WorksPhysics and Astronomy
Adsorbate-Free Screening of Dissociative Water Adsorption on Nickel Nanoparticles via Graph Neural Networks
Abstract
Dissociative water adsorption is a key step in heterogeneous and electrochemical catalysis, but on finite Ni nanosystems the OH and H fragments can occupy many combinations of surface sites, making exhaustive density functional theory (DFT) screening impractical. We introduce a graph neural network (GNN) framework with an adsorbate-free representation: candidate O- and H-binding sites are defined by sets of Ni atoms (anchors), and each site pair is encoded as a graph constructed solely from these anchors and their local Ni environment. Using only MACE-relaxed bare Ni geometries as input, separate GNNs predict the DFT-reference adsorption energies of OH+H configurations and whether the proposed binding sites would be retained upon DFT relaxation, eliminating the need to construct or relax adsorbates during screening. Trained across Ni nanoclusters and Ni(111)/Ni(100) surfaces, the energy model achieves a mean absolute error of 0.031 eV on the internal validation set and demonstrates transferability to unseen Ni, Ni, and Ni clusters and larger surface cells, particularly for relative-energy ranking, although quantitative accuracy varies across systems. These independent tests also support the retention classifier's use as a conservative filter for selecting configurations likely to preserve their proposed binding sites. Application to Ni nanoparticles containing up to 3000 atoms identifies favorable co-adsorption motifs and shows that lower O-site coordination generally stabilizes the dissociated state, whereas H-site coordination has a weaker influence. The framework thus offers an efficient route to screening adsorption energetics, with potential extensions to other metals and adsorbate combinations.