(360as) Bayesian Forcefield Driven Monte Carlo and Molecular Dynamics Simulations of O and Cl Promoted Silver Surface Reconstructions

Silver nanoparticles (NPs) are commercial catalysts used for selective oxidation of ethylene to ethylene oxide (EO). Higher selectivity towards EO formation is achieved by incorporation of chlorine promoter. Ultra-high vacuum studies indicate that the Ag surface undergoes significant reconstruction upon exposure to O₂, or mixtures of O₂ and Cl. However, the surface structure formed by contact between Ag NPs and O₂ at relevant EO conditions remains debated, in both the presence and absence of Cl. Atomistic modeling of surface reconstructions with density functional theory (DFT), is challenging due to spatial and temporal limitations associated with electronic structure calculations. Here, we use DFT sampling of potential energy surfaces implemented in the Fast Learning of Atomistic Rare Events (FLARE) code [1] to generate a Gaussian Process (GP) interatomic potential descriptive of O and Cl promoted Ag surfaces.

To sample a diverse distribution of Ag-O and Ag-O-Cl interactions we performed: 1) on the fly active learning with ab initio molecular dynamics (MD) simulations of O and Cl single and mixed coverages on Ag(111) and Ag₂O(001) slabs, in the NPT ensemble, at 1500 K for 5 ps; 2) offline learning of DFT optimized reconstructed Ag surfaces proposed in literature. Trained models were validated with DFT calculations of structures sampled from the GP driven MD and Monte Carlo (MC) simulations, which were not included in the training data. The Ag-O trained model resulted in a RMSE of ~10 meV/atom for energy and 0.2 eV/Angstrom for forces comparable to machine learning forcefield errors reported in the literature.

To determine how Ag surfaces dynamically evolve under different conditions, we performed GP driven NVT MD simulations on the Ag(111), (110), (100) and reconstructed c(4x8) surfaces at different temperatures and O coverages in LAMMPS. We observed significant surface reconstruction and O subsurface diffusion for all surfaces. We analyzed all simulated surfaces by computing the time-averaged radial distribution functions and found that regardless of the initial state, all reconstructions reach similar final environments. Next, to overcome the activation barriers inherent to MD, we performed NVT MC simulations using an inhouse code. The code implements swap moves, which allow for subsurface diffusion, followed by geometry optimization, which minimizes atomic forces and increases probability of moves being accepted. In NVT MC simulations we observed surface reconstruction, O subsurface diffusion and formation of surface dioxygen species. The dioxygen species formation on Ag surfaces containing subsurface O, which we observed in MC simulations only, is consistent with recently reported experiments and DFT calculations [2,3]. We also performed grand canonical ensemble MC to simulate equilibrium surface coverages of atomic O and dioxygen species at varying temperatures and chemical potentials of oxygen, and related these to oxygen pressures used experimentally by our collaborators. Finally, for the Ag-O-Cl system we trained a GP forcefield and used it for both MD and MC. Our preliminary results suggest that the Ag surface reconstructs and both O and Cl species diffuse subsurface. This work improves our understanding of how both separately and together O and Cl affect Ag surface dynamics under EO reaction conditions.

References

1. Vandermause, J., Torrisi, S. B., Batzner, S., Xie, Y., Sun, L., Kolpak, A. M., Kozinsky, B,. npj Comput Mater 6, 20 (2020).

2. Liu, C., Wijewardena, D. P., Sviripa, A., Sampath, A., Flaherty, D. W., Paolucci, C., J. Catal, 405 (2022).

3. Pu, T., Setiawan, A., Mosevitzky Lis, B., Zhu, M., Ford, M. E., Rangarajan, S., Wachs, I. E., ACS Catal., 12 (2022).