A bulk-surface-coupled neuroevolution potential reveals dynamic low coordinated Pd sites and interfacial water...

Dynamic low-coordinated metal sites and solvent organization are central descriptors of aqueous metal/semiconductor catalytic interfaces, yet their finite-temperature evolution is difficult to sample using density functional theory (DFT) alone. Here, we construct a bulk-surface-coupled neuroevolution potential (NEP) for Pd/GaN-water systems using a compact, descriptor-guided training strategy. An initial pool containing more than 7500 configurations was assembled from over 3300 surface-derived structures, including adsorption geometries, path-like high-energy configurations, and modified surfaces, together with over 4200 bulk-derived structures generated from doped, thermally sampled, and randomly perturbed GaN models. Principal component analysis combined with farthest point sampling was used to select 550 representative configurations for DFT single-point labeling. A direct transferability test using the general NEP89 potential gives large deviations from the DFT references, confirming the need for a system-specific Pd-Ga-N-H-O potential. The trained NEP reproduces DFT reference energies, forces, virials, and stresses with an energy R² of 0.9991, a mean absolute error of 9.49 meV per atom, and a root-mean-square error of 12.08 meV per atom. Reaction-path and adsorption energy benchmarks for H₂O₂- and OH-related configurations further show that the potential retains useful accuracy for reaction-relevant regions of the potential energy surface. Temperature-controlled GPUMD simulations of a 2304-atom GaN supercell and an explicit Pd/GaN-water model confirm numerical stability under long-time and elevated temperature conditions. Analysis of a 298 K production trajectory reveals an anchored but dynamically reconstructed Pd nanocluster, persistent low-coordinated Pd atoms, pronounced Pd-water correlations, and layered interfacial water. These dynamic structural descriptors provide a finite-temperature basis for understanding adsorption- and activation-relevant environments at Pd/GaN-water catalytic interfaces.