Dynamics of surfaces and interfaces: From first-principles modeling to machine-learning molecular dynamics

Abstract

Chemical production accounts for a substantial fraction of global energy use, and the majority of industrial processes rely heavily on precious metal heterogeneous catalysts. Fundamental knowledge of the surface structure of these metals is an absolute prerequisite for engineering of energy-efficient and sustainable catalysis. In this regard, first-principles atomistic modeling based on density functional theory (DFT) plays a crucial role in the advancement of rational catalyst design. However, the large computational cost of DFT precludes its use for dynamical simulations of realistic systems. Moreover, accurate identification of the active site remains a significant challenge, as surface composition and morphology are inherently dynamic at reaction conditions of interest, especially for soft metals such as Au, Ag, Pt, and Pd. Such restructuring has been shown to play a significant role in catalytic activity of many nanoparticle systems, but mechanistic understanding at the atomic level has remained scarce and challenging.

To investigate the dynamics of restructuring and reactivity in catalytic surfaces, we develop and apply novel simulation methods and analysis workflow to model systems of increasing complexity. We first demonstrate the utility and limitations of the simplest static method – coordinate dragging – using an example system with a chemically diverse interface: diffusion of a water molecule through periodically porous graphene derivatives. We then proceed to highlight the power of proper transition state modeling and microkinetic modeling, combined with catalysis and machine learning-enabled spectroscopic analysis, to uncover the dynamic changes in the active site for hydrogen-deuterium exchange reaction in dilute Pd-in-Au catalysts in response to different catalyst treatments.

Dynamic problems require dynamic solutions. As such, we employ molecular dynamics (MD) to elucidate the mechanism and timescale of surface restructuring. First, we perform MD simulation with an empirical potential followed by automated analysis to establish seven main classes of restructuring events by which a single Pd atom becomes incorporated into the Ag surface. Next, we develop and apply Bayesian active learning to train a fast and accurate machine-learning force field, which enables large-scale and long-timescale simulation of the restructuring of Pd deposited on Ag. Finally, the method is extended to reactive simulation and mechanistic analysis of H2 turnover on Pt at chemical accuracy. We close the thesis with a proposal of an automated workflow Catadynamics for discovery and characterization of surface dynamics and reactions directly from reactive MD simulations. The work presented in this thesis attests to the promise of machine-learning MD for further empowering computational catalysis and surface science.