Understanding Structure of Supported Ag Nanoparticles by Machine Learning Interatomic Potentials
Tristan Maxson, University of Alabama
Small nanoparticles (1-5 nm in diameter) are effective catalysts for many applications due to their higher activity/selectivity than larger nanoparticles or thin film catalysts. Small nanoparticles also maximize the atom economy as more atoms are part of the reactive surface. Winterbottom constructions are simple computational models for small nanoparticles; they are based on surface energies of ideal surfaces and as such they are an approximation for nanoparticles with large finite-size effects (corner energies and edge energies). It is still unknown experimentally what the exact surface structure of common catalysts is like at ambient or reaction conditions (where rearrangements are possible) and computational studies have been limited previously. In this work, machine learning interatomic potentials (MLIPs) based on data from Density Functional Theory (DFT) calculations are developed to accelerate simulations and utilize GPU resources. Supported Ag nanoparticles simulated connect to benchmark microcalorimetric experiments of Prof. Charlie Campbell. MLIPs reproduce the heat of adsorption and the chemical potential of metal nanoparticles as a function of the nanoparticle size within the error of the underlying DFT method (~0.2 eV). Simulated annealing of the ideal Winterbottom constructions changes the shape and radius of nanoparticles to a more stable structure (confirmed using DFT single-point calculations). The change in structure implies that the Winterbottom construction is not an appropriate model for nanoparticle studies of catalysis. The distribution of the coordination number of surface metal atoms of the most stable nanoparticles is investigated as a function of nanoparticle size. The most stable (111) surface facet is less proportionally represented on the surface when compared with Winterbottom constructions up to ~600 atoms (~3.5 nm in diameter). This indicates common DFT models of catalyst metal surfaces that use the (111) facet as a periodic model may represent significantly fewer active sites than typically assumed.