# NERSC Initiative for Scientific Exploration (NISE) 2010 Awards

## Modeling the dynamics of catalysts with the adaptive kinetic Monte Carlo method

**Graeme Henkelman, University of Texas at Austin**

#### Associated NERSC Project: Correlation of Theory and Function in Well-defined Bimetallic Electrocatalysts (m1069), Principal Investigator: Graeme Henkelman

NISE Award: | 1,700,000 Hours |
---|---|

Award Date: | February 2010 |

A mission of the Department of Energy is to develop alternative energy sources. In order to store energy in a transportable chemical form and most efficiently extract the energy again, better catalysts are needed. The proposed research is a computational effort to identify new nanoparticle catalysts and correlate their electronic structure to measured catalytic function. In this supplementary request, a new computational methodology will be used to model the reaction dynamics at the catalysts directly using density functional theory. In these adaptive kinetic Monte Carlo calculations, reaction pathways are determined without bias from the modeler so that unexpected mechanisms can be discovered. It is likely that catalytic nanoparticles will be more dynamic than previously assumed and that the dynamics of the nanoparticle surfaces will play an important role in determining their catalytic activity and stability.

Our current DOE sponsored research at NERSC uses density functional theory (DFT) to model nanoparticle catalysts and to understand correlations between their structure and catalytic function. The computational approach for this is to identify reactivity descriptors, such as the binding of reactant molecules to the surface or the average energy of the d-electrons in the metal surface. These descriptors have been shown to correlate well with experimental reactivities for some reactions, such as the oxygen reduction reaction on nanoparticles made from alloys of Pt-group metals.

What is missing from this approach, and what a new computational methodology can provide, is the possibility of discovering reaction mechanisms that we do not anticipate. One strategy for modeling reaction dynamics using DFT without having to anticipate the reaction mechanisms is called adaptive kinetic Monte Carlo (AKMC)^{1}. Here, reaction mechanisms are determined by searching for saddle points on the potential energy surface leading from an initial state to any final state. The rate of each mechanism is calculated using a harmonic approximation to transition state theory, and the state-to-state dynamics is calculated using the kinetic Monte Carlo algorithm. This approach avoids several major limitations of standard kinetic Monte Carlo: reaction mechanisms are a result of the simulation instead of an input; these mechanisms can be complex, unexpected, and involve collective motions of many atoms; and there is no need to base the simulation on a lattice.

What makes this new methodology particularly relevant for nanoparticle catalysts is the mounting experimental evidence that the particles are not static and can dramatically rearrange under reaction conditions. One recent example of this is the inversion of core-shell nanoparticles under oxidizing/reducing conditions; another is the local oxidation of the Au(111) surface in the presence of atomic oxygen. In both cases, a theoretical model in which reactions are assumed to take place on a static metal surface will be qualitatively wrong. As we investigate and try to discover new nanoparticle catalysts, it will be important to model the actual dynamics and not just the dynamics that we assume should happen.

The AKMC method is, however, more computationally expensive than modeling assumed reaction mechanisms. Massively parallel resources are just now able to make these calculations tractable since the searches for reaction mechanisms are independent of one another. In this supplementary request for computer time, I am asking for a million node hours, which will allow for several AKMC simulation of the oxygen reduction reaction at mono- and bi-metallic nanoparticles to investigate the mechanism(s) of reaction directly from DFT. Modeling the (possibly) dynamic role of the nanoparticle structure will be a significant step forward in predicting the activity of new nanoparticle catalysts.

^{1}L. Xu, D. Mei, and G. Henkelman. Adaptive kinetic Monte Carlo simulation of methanol decomposition on Cu(100). J. Chem. Phys. 131, 244520 (2009).