(688c) Molecular Dynamics Simulations of the Dissociation of Hydrogen on Single Crystal Surfaces Using Neural Networks

Ludwig, J. J. - Presenter, University of Delaware

Microkinetic models have become an indispensable aid in advancing the understanding of elementary surface reactions occurring in ultra-high vacuum (UHV) and on supported catalysts. With a detailed description of the underlying surface physics, they allow simulations of reaction networks under a much larger parameter space than simple power law rate expressions. However, this additional sophistication comes at a cost ? the large number of parameters used in microkinetic models must be based on accurate experiments or first principles calculations to ensure uniqueness of the underlying reaction network.

For activated surface reactions, the use of density functional theory (DFT) with a saddle-point seeking algorithm allows the identification of the structure of transition states, and an estimation of the electronic energy and vibrational frequencies. This information can in turn be used with transition state theory (TST) to estimate the activation energy and pre-exponential factors used as parameters in microkinetic models. However, if a process is non-activated, such as the adsorption of some diatomic molecules on a metal surface, or the reaction coordinate is complex and multiple saddle points exist, molecular dynamics (MD) simulations performed on a dynamically relevant region of the potential energy surface (PES) should be used to estimate reactivity.

Neural networks have recently been employed as a tool to obtain highly accurate ab-initio based PES for performing six dimensional MD simulations aimed at estimating the reactivity of hydrogen on Pd(100) surfaces [1,2]. In this talk we explore a hybrid algorithm employing a corrugation reduction procedure [3], novelty sampling, and neural networks [1,2] to obtain a dynamically accurate PES for simulating the dissociative chemisorption of hydrogen on single crystal surfaces. We address PES accuracy requirements, neural network topology and training method, and identification of the most relevant region of phase space with a focus on the impact on subsequent dynamics simulations. By selective sampling of phase space guided by dynamics simulations, we show that we can reduce the number of DFT calculations necessary for accurate estimate of reactivity by an order of magnitude without sacrificing accuracy.

(1) Lorenz, S.; Gross, A.; Scheffler, M. Chem. Phys. Lett. 2004, 395, 210.

(2) Lorenz, S.; Scheffler, M.; Gross, A. Phys. Rev. B 2006, 73, 115431.

(3) Busnengo, H. F.; Salin, A.; Dong, W. J. Chem. Phys. 2000, 112, 7641.