Anomalous Diffusion of Metal Atoms on Oxide Surfaces: A Machine Learning Molecular Dynamics Study of Pt₁/TiO₂

We investigate the high-temperature evolution of the metal-support interface in atomically dispersed catalysts to probe the dynamical stability of the metal atom. Since high computational costs limit ab initio molecular dynamics (AIMD) trajectories to a few picoseconds, we deploy a machine learned interatomic potential (MLIP) trained using the FLARE program. We generate multiple 100 ps trajectories for atomically dispersed Pt on rutile TiO₂ (110) initiated at different positions on the stoichiometric surface. We find that Pt diffusion is sub-diffusive (not Brownian) in most cases, quantified by the parameters constituting the fractional Fokker-Planck equation. The sub-diffusive behavior originates in strong interactions of Pt with bridging oxygen atoms of the support. Diffusion is completely quenched when Pt coordinates with two bridging oxygens to form a near-linear O-Pt-O complex, observed in several trajectories across all starting positions of the metal atom. Diffusion that resembles Brownian motion is only observed at the highest simulation temperature examined (1000 K) for sites at which Pt is far from a bridging oxygen. The study therefore shows that operando stability of the metal atom, which is necessary for consistent catalytic activity, requires coordination with bridging oxygens when the metal atom is situated far from any surface vacancies.