Speaker
Description
Computational modeling and simulation of catalytic processes, in particular of reactions like water electrolysis to produce hydrogen and CO or CO2 hydrogenation to produce alkanes or higher oxygenates, have become key ingredients in developing new catalytic materials for the sustainable generation of synthetic fuels. The calculation of relevant quantities such as adsorption enthalpies and reaction barriers is often carried out using density-functional theory (DFT). However, the computational cost of DFT remains a limiting factor, rendering the large-scale exploration of new catalysts difficult. In recent years, significant interest has been devoted to replacing the costly DFT calculations with computationally cheaper machine learning (ML) models. Advantages of this approach are that the models and training data sets can be flexibly tailored to arrive at the desired balance between cost and accuracy [1].
In recent works, we developed data-efficient and physically motivated ML models for the prediction of adsorption enthalpies [2]. For simple adsorbates, we focused on the compressed sensing method SISSO to explore representative multi-dimensional descriptors expressed as nonlinear functions of intrinsic properties of the clean surface. In the application to transition metal oxides (TMOs) for the oxygen evolution reaction our descriptors largely outperform previously highlighted descriptors in terms of accuracy and computational cost, and properties related to local charge transfer are identified to be crucial for the screening of promising TMO catalysts.
For complex adsorbates we considered species that occur in the reaction mechanisms to produce alkanes or higher oxygenates on metal and metal alloy catalysts. Here, the challenge is the combinatorial explosion of possible adsorption motifs arising from different ways and orientations in which the adsorbates can bind to the catalyst surface, including mono-, bi- or higher-dentate adsorption motifs. Relying on mathematical graph theory, we developed a ML method to predict adsorption motifs and associated adsorption enthalpies based on a customized Wasserstein Weisfeiler-Lehman (WWL) graph kernel and Gaussian process regression [3]. The model enables a prediction accuracy en par with DFT and has a built-in uncertainty estimation method, which makes it promising for active learning approaches to explore complex reaction networks comprehensively and efficiently.
[1] M. Andersen and K. Reuter, Acc. Chem. Res. 54, 2741 (2021).
[2] W. Xu, M. Andersen, and K. Reuter, ACS Catal. 11, 734 (2021).
[3] W. Xu, K. Reuter, and M. Andersen, Nat. Comp. Sci. 2, 443 (2022).
| Abstract Number (department-wise) | TH 13 |
|---|---|
| Department | TH (Reuter) |
