Bridging the Gap between Direct Dynamics and Globally Accurate Reactive Potential Energy Surfaces Using Neural Networks.

Direct dynamics simulations become increasingly popular in studying reaction dynamics for complex systems where analytical potential energy surfaces (PESs) are unavailable. Yet, the number and/or the propagation time of trajectories are often limited by high computational costs, and numerous energies and forces generated on-the-fly become wasted after simulations. We demonstrate here an example of reusing only a very small portion of existing direct dynamics data to reconstruct a 90-dimensional globally accurate reactive PES describing the interaction of CO2 with a movable Ni(100) surface based on a machine learning approach. In addition to reproducing previous results with much better statistics, we predict scattering probabilities of CO2 at the state-to-state level, which is extremely demanding for direct dynamics. We propose this unified way to investigate gaseous and gas-surface reactions of medium size, initiating with hundreds of preliminary direct dynamics trajectories, followed by low-cost and high-quality simulations on full-dimensional analytical PESs.