Machine Learning-Based Prediction of Activation Energies for Chemical Reactions on Metal Surfaces
Daniel J. Hutton, Kari E. Cordes, Carine Michel, Florian Göltl
Abstract
In computational surface catalysis, the calculation of activation energies of chemical reactions is expensive, which, in many cases, limits our ability to understand complex reaction networks. Here, we present a universal, machine learning-based approach for the prediction of activation energies for reactions of C-, O-, and H-containing molecules on transition metal surfaces. We rely on generalized Bronsted-Evans-Polanyi relationships in combination with machine learning-based multiparameter regression techniques to train our model for reactions included in the University of Arizona Reaction database. In our best approach, we find a mean absolute error for activation energies within our test set of 0.14 eV if the reaction energy is known and 0.19 eV if the reaction energy is unknown. We expect that this methodology will often replace the explicit calculation of activation energies within surface catalysis when exploring large reaction networks or screening catalysts for desirable properties in the future.