Descriptions of surface chemical reactions using a neural network representation of the potential-energy surface

Abstract

A neural network (NN) approach is proposed for the representation of six-dimensional ab initio potential-energy surfaces (PES) for the dissociation of a diatomic molecule at surfaces. We report tests of NN representations that are fitted to six-dimensional analytical PESs for ${\mathrm{H}}_{2}$ dissociation on the clean and the sulfur covered $\mathrm{Pd}(100)$ surfaces. For the present study we use high-dimensional analytical PESs as the basis for the NN training, as this enables us to investigate the influence of phase space sampling on adsorption rates in great detail. We note, however, that these analytical PESs were obtained from detailed density functional theory calculations. When information about the PES is collected only from a few high-symmetric adsorption sites, we find that the obtained adsorption probabilities are not reliable. Thus, intermediate configurations need to be considered as well. However, it is not necessary to map out complete elbow plots above nonsymmetric sites. Our study suggests that only a few additional energies need to be considered in the region of activated systems where the molecular bond breaks. With this understanding, the required number of NN training energies for obtaining a high-quality PES that provides a reliable description of the dissociation and adsorption dynamics is orders of magnitude smaller than the number of total-energy calculations needed in traditional ab initio on the fly molecular dynamics. Our analysis also demonstrates the importance of a reliable, high-dimensional PES to describe reaction rates for dissociative adsorption of molecules at surfaces.