Application of Machine Learning Methods to Approximate the Binding Energy of CO Molecules on the Surface of Pd Nanoparticles

Abstract

This manuscript presents a study of the applicability of machine learning methods for the problem of binding energy approximation of carbon monoxide adsorbates on the surface of a palladium nanoparticle. Machine learning algorithms were trained using a set of structures that represent models of CO interaction with different parts of the Pd₅₅ nanocluster with a variable distance from the molecule to the surface, for which the energy was calculated using the density functional theory methods. For structures that make up a training set, the radial distribution functions were calculated. Using these functions and their segments as descriptors, the effectiveness of various machine learning algorithms, such as “gradient boosting,” “ridge regression,” “extra trees,” and “support vector machines” for calculating the binding energy, was tested. Based on three different metrics, it was found that the error in determining the binding energy was the smallest when using the “support vector machines”: the mean absolute error was 0.093 eV. The efficiency of various individual sections of the distribution function used as descriptors was compared. It was found that it is crucial to take into account the part of the radial distribution function from 1.5 to 2.5 Å for a correct approximation of the energy.