Toward Realistic Modeling Of Catalytic Surfaces: From First Principles To Machine Learning

Computational catalyst design has the potential to revolutionize the energy and chemical industries by alleviating their reliance on fossil fuels, precious metals, and toxic elements. Despite recent advances in understanding catalytic trends, e.g. the chemisorption scaling relations and the d-band model, the description of catalytic surfaces has, for the most part, been far from realistic. It is well known that surfaces can undergo reconstruction where the structure and composition of the surface differs from that of the bulk and the nature of this reconstruction depends on the temperature, pressure, and chemical potentials of the elements in the system. Since catalytic transformations, i.e. bond breaking and formation, occur at the surface, an accurate picture of surface structure and composition is vital. In this thesis, we apply and develop state-of-the-art computational methods for studying the reconstruction of catalytic surfaces and investigate the effect of surface reconstruction on catalysis. First, we show using ab initio thermodynamics that the surfaces of nickel phosphide catalysts for the hydrogen evolution reaction (HER) are P-enriched, which was not previously considered in computational studies. Building on this discovery, we reevaluate the HER mechanism on Ni2P and Ni5P4 and find that P sites are the key to their catalytic activities. While the P sites on Ni2P are highly active toward the HER, they are not stable at conditions suitable for commercial electrolyzers. Under these conditions, the stable surface of Ni2P binds hydrogen too strongly at Ni sites. We demonstrate that these Ni sites can be activated by doping the surface of Ni2P with S, Se, and Te. Additionally, using tree-based machine learning methods, we reveal that nonmetal dopants induce a chemical pressure-like effect on the Ni sites, changing their reactivity through compression and expansion. Finally, we develop a software package for ab initio grand canonical Monte Carlo that automatically predicts surface phase diagrams. The results presented herein provide strong motivation and a methodological foundation for moving toward more realistic modeling of heterogeneous catalysts