Representing high-dimensional potential-energy surfaces for reactions at surfaces by neural networks

The determination of dissociative adsorption probabilities based on first-principles total-energy calculations requires a numerically efficient and accurate interpolation scheme in order to be able to run a sufficient number of trajectories. Here we present a neural network scheme for the construction of a continuous potential energy surface (PES). We illustrate the accuracy and efficiency of our method for H2 interacting with the (2 × 2) potassium covered Pd(1 0 0) surface. The sticking probability of H2/K(2 × 2)/Pd(1 0 0) is determined by molecular dynamics simulations on the neural network PES and compared to results using an independent analytical interpolation.