7–10 Nov 2022
Europe/Berlin timezone

Understanding Elementary Steps in Surface Catalysis with Machine-Learned Surrogate Models

Not scheduled
20m

Speaker

Hyunwook Jung (Fritz-Haber-Institut)

Description

Predictive-quality first-principles based microkinetic models are increasingly used to analyze (and subsequently optimize) reaction mechanisms in heterogeneous catalysis. In full rigor, such models require the knowledge of all possible elementary reaction steps and their corresponding reaction barriers. Unfortunately, for complex catalytic processes (such as the generation of ethanol from syngas) the number of possible steps is so large that an exhaustive first-principles calculation of all barriers becomes prohibitively expensive. Furthermore, even the barriers of individual reaction steps can have non-trivial dependencies on the adsorbate geometry, active-site structure and temperature.

In computational catalysis research, these complexities are currently mostly ignored in favor of simple idealized reaction networks and catalyst models. This is mainly due to the prohibitive computational cost of a more realistic description, which is in turn related to the complexity of the underlying first-principles electronic structure calculations. To overcome this limitation, we are currently exploring the potential of using machine learned (ML) interatomic potentials as computationally efficient surrogates. Here, the main challenge is developing data-efficient strategies for training set generation, in order to keep the overall computational costs small. This can be achieved via iterative workflows, which we are applying to various simulation tasks.

One highly promising direction in this context is the global optimization of adsorbate/surface geometries. This is essential for accurately predicting reaction energies and barriers of chemical reactions on a catalyst, since the adsorption energies of different locally relaxed geometries of the same adsorbate can vary substantially. To tackle this, we develop an approach using Gaussian Approximation Potentials (GAPs) that are fitted on-the-fly during minima-hopping simulations. These simulations automatically explore different adsorbate binding sites and motifs, thus combining the training set generation with the global optimization itself in a highly efficient manner.

In a related project, a GAP potential is fitted to study an individual elementary reaction step (the hydrogenation of CO during ethanol synthesis on rhodium) in detail. In particular, the temperature dependence of the corresponding free energy barrier is studied, using extensive umbrella sampling simulations. In this context, the computational efficiency of the GAP potential is essential to go beyond the static picture typically obtained at the first-principles level. This leads to the surprising result that the reverse barrier of this reaction vanishes at operating temperatures, indicating that the corresponding intermediate is not metastable under realistic conditions.

Abstract Number (department-wise) TH 05
Department TH (Reuter)

Primary authors

Hyunwook Jung (Fritz-Haber-Institut) Sina Stocker (Fritz-Haber-Institut) Karsten Reuter (FHI Berlin) Johannes T. Margraf (FHI Theory Department)

Presentation materials

There are no materials yet.