Gaussian Approximation Potentials for Functionalized AuRh Nanoparticles

Bimetallic Au–Rh nanoparticles are promising catalysts for oxidation and (de)hydrogenation reactions, however running first-principles simulations at realistic sizes and temperatures is too expensive. We build machine-learning based interatomic potentials for Au–Rh nanoparticles covered by O₂ and CO that keep density-functional accuracy while making longer and larger simulations possible. The training data comes from ab initio molecular dynamics at several temperatures which covers a wide range of particle sizes and common morphologies. We carefully filter and balance the snapshots so that the models contain diverse but physically meaningful structures. Validation tests show that the potentials match reference energies and forces and do not overfit to a single particle type. Finite-temperature analysis using radial distribution functions indicates that alloyed particles—especially core–shell and Janus—preserve short-range order better on heating than ordered alloys or monometallic particles, which supports their higher thermal robustness. With the acquired accuracy and stability, the fitted machine-learning based interatomic potential models open the door to study questions that are hard to access with direct electronic-structure methods such as coverage-dependent stability, adsorbate-driven restructuring, competitive adsorption and site blocking, and morphology changes under varying temperature and coverage.