Obtaining accurate activation energies for reactions is the most challenging aspect in electronic structure calculations of metal surfaces1. This task requires experienced researchers to set up and run these calculations, and calculating the activation energy for a single reaction is associated with a computational cost of about 10000 CPUh. At the same time, an in depth understanding of the performance or the computational design of heterogeneous catalysts requires the description of a complex reaction network and the calculation of a large number of activation energies. Therefore, an acceleration of activation energy calculations would be highly desirable. In this contribution we develop a machine learning based method for the prediction of activation energies for reactions of C, O, and H containing molecules on metal surfaces. As a first step we generate a database of more than 7100 surface reactions based on literature data. For each datapoint we assign a set of descriptors, which include the reaction energy, the change in chemical identity2, molecular size, bonding of the molecules to the surface, the used methodology, and the surface metal and structure. We use this representation to apply Random Forest Regression, Support Vector Regression, and Gradient Boosted Regression, and our best approach leads to a Mean Absolute Error of 0.16 eV for the prediction of activation energies. Further analysis shows that we can predict activation energies without the knowledge of the activation energy or performing any other DFT calculations with a Mean Absolute Error of 0.18 eV. The machine learning based approach presented here significantly outperforms other methods that have previously been used to predict activation energies on metal surfaces, such as Brønsted-Evans Polanyi relationships or other machine learning based approaches. It is fast, easy to use, and we expect that in the future it will replace the explicit calculation of activation energies for many applications in theoretical surface catalysis.

Machine learning based prediction of activation energies for chemical reactions on metal surfaces

Where thousands of chemistry professionals meet to share ideas and advance scientific and technical knowledge. Spring 2023 abstracts, posters, and presentations will launch at the end of March.

ACS Spring 2023

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.