Adaptive Kinetic Monte Carlo Simulation on Platinum Oxidation at Electrochemical Conditions using Machine-Learned Potentials

Platinum is the most commonly used electrochemical catalyst. However, platinum suffers from oxidation in the operation conditions, which gradually degrades activity and lifetime of catalysts in applications like fuel cells and CO oxidation. Although many experiments have investigated oxidation kinetics of Pt, a comprehensive understanding has not yet been achieved. For instance, it is known that platinum oxidation occurs through a place-exchange mechanism, but the detailed atomic pathway remains elusive. While computational study could supplement the experiment, the spatial and temporal scales of platinum oxidation reactions are beyond the capabilities of computational approaches such as density functional theory (DFT) and classical molecular dynamics (MD) simulations. Kinetic Monte Carlo (kMC) simulations, which use an event table to transition between local minima, can investigate long-term evolution but require an efficient algorithm for identifying saddle points on-the-fly, in the open-ended style.<br/>In this presentation, we will discuss our adaptive kMC simulation for platinum oxidation behavior on a 2 nm × 2 nm Pt(111) surface. This is achieved by combining two computational advancements. Firstly, we employ Behler-Parinello type machine-learned potentials (MLPs) as surrogate models of DFT for increased computational speed. The present MLP achieved energy and force root-mean-square errors of 7 meV/atom and 0.22 eV/Å, respectively, on the DFT validation set consisting of platinum oxide crystal and platinum surface with randomly distributed oxygen atoms at various coverages. Secondly, we found that the state-of-the-art algorithm for identifying saddle points, such as ARTn, is inefficient in the presence of soft modes prevalent on the surface. To address this, we developed the SHERPA (Saddle point Hunting based on Energy surface for Reaction PAthways) package, incorporating two key features: constrained-orthogonal relaxation within ARTn and dynamic-active volume. These improvements resulted in a 50% higher success ratio and halved computational time compared to the conventional ARTn approach.<br/>To emulate oxygen flux from the electrolytes, we combine kMC simulation with a grand-canonical Monte Carlo (GCMC) simulation, controlling the oxygen chemical potential based on voltage and pH. Our simulation provides detailed pathways for the platinum surface oxidation via a place exchange mechanism, with oxygen atoms penetrating the platinum subsurface and displacing platinum atoms. For comparison, we also perform a hybrid GCMC-MD simulation at elevated temperatures. We believe this study deepens the understanding of platinum oxidation and paves the way to long-term atomistic simulation.