Towards Atomistic Phase Diagrams for Surface Reconstructions and Nanoparticle Responses to Adsorbates using Machine-Learned Dynamics

Metal surfaces have long been known to reconstruct in both facile and activated fashions, exhibiting substantial influence in their resulting catalytic and mechanical performance. We provide an unbiased simulation method by which surface reconstructions can be studied with atomistic resolution under the effects of temperature and chemisorption of adsorbates. A machine-learned force field (MLFF) is trained from <i>ab initio</i> (first principles) calculations that can capture each of the low-index mesoscopic surface reconstructions of Au (e.g., the Au(111)-`Herringbone,' Au(110)-(1x2)-`Missing-Row,' and Au(100)-`Quasi-Hexagonal' reconstructions) using large scale molecular dynamics (MD) simulations. Additionally, we present another independent MLFF that can provide proper description of the Pt(100)-(1x1) to ‘hex’ reconstruction, as well as subsequent lifting via exposure to CO.<br/><br/>Analysis of these MD trajectories yields direct atomistic understanding of the dynamic evolution of these surfaces from their initial facets, providing previously inaccessible information such as nucleation timescales under the effects of strain, local deviations from the original stoichiometry, and adsorbate exposure. Both MLFFs can also be used to study surface reconstructions and responses to adsorbates on Au or Pt nanoparticles (NPs) under similar environmental stimuli. These methodological advancements set a new standard for determining surface reconstructions along the axes of surface stoichiometry, mechanical strain, and adsorbate exposure without bias and with atomic resolution, thus providing a comprehensive picture of the nucleation environments and time-scales for such phenomena and moving the surface science community towards surface phase diagrams for these complicated mesoscopic systems.