A transferable prediction model of molecular adsorption on metals based on adsorbate and substrate properties

TL;DR

This paper presents a machine learning-based model that predicts molecular adsorption energies on metal surfaces quickly and accurately using intrinsic properties, aiding rapid screening in catalysis and related fields.

Contribution

The authors develop a transferable, fast prediction model for adsorption energies based on substrate and adsorbate properties, trained on first-principles calculations, with high accuracy and broad applicability.

Findings

Correlation coefficient of 0.93 between predicted and computed energies

Mean absolute error of 0.77 eV in predictions

Eliminates around 90% of candidates in screening studies

Abstract

Surface adsorption is one of the fundamental processes in numerous fields, including catalysis, environment, energy and medicine. The development of an adsorption model which provides an effective prediction of binding energy in minutes has been a long term goal in surface and interface science. The solution has been elusive as identifying the intrinsic determinants of the adsorption energy for various compositions, structures and environments is non-trivial. We introduce a new and flexible model for predicting adsorption energies to metal substrates. The model is based on easily computed, intrinsic properties of the substrate and adsorbate. It is parameterised using machine learning based on first-principles calculations of probe molecules (e.g., H$_2$O, CO$_2$, O$_2$, N$_2$) adsorbed to a range of pure metal substrates. The model predicts the computed dissociative adsorption energy to metal surfaces with a correlation coefficient of 0.93 and a mean absolute error of 0.77 eV for the large database of molecular adsorption energies provided by \textit{Catalysis-Hub.org} which have a range of 15 eV. As the model is based on pre-computed quantities it provides near-instantaneous estimates of adsorption energies and it is sufficiently accurate to eliminate around 90\% of candidates in screening study of new adsorbates. The model, therefore, significantly enhances current efforts to identify new molecular coatings in many applied research fields.