Efficient Dynamic Computational Strategy for Heterogeneous Catalysis Based on Neural Network Potential Energy Surface: A Case Study of Temperature-Dependent Thermodynamics and Kinetics for the Chemisorbed on-surface CO

As a favorable alternative and complement of experimental techniques, computational tools on top of ab initio calculations have played an indispensable role in revealing the molecular details, thermodynamics and kinetics in catalytic reactions. The static computational strategy, which recovers the reaction thermodynamics and kinetics based on the calculations of a few stationary geometries at zero temperature and some ideal statistic mechanics models, is the most popular approach in theoretical catalysis due to its simplicity. In comparison, the ab initio molecular dynamics (AIMD) is a well-tested approach to provide more precise descriptions of catalytic processes, however, experiencing a significantly expensive computational cost in the direct ab initio calculation of potential energy and gradients. Here we proposed a highly efficient dynamic computational strategy for the calculation of thermodynamic and kinetic properties in heterogeneous catalysis on the basis of neural network potential energy surface (NN PES) and MD simulations. Taking CO adsorbate on Ru(0001) surface as the illustrative model catalytic system, we demonstrated that our NN-PES-based MD simulations can efficiently generate the reliable smooth two-dimensional potential-of-mean-force (2-D PMF) surfaces in a wide range of temperatures (from 300 to 900 K), and thus temperature-dependent thermodynamic properties can be obtained in a comprehensive investigation on the whole PMF surface rather than a rough estimation using ideal models based on a few optimized geometries. Moreover, MD simulations offer an effective way to describe the surface kinetics such as the CO adsorbate on-surface movement, which goes beyond the most popular static estimation based on calculated free energy barrier and transition state theory (TST). By comparing the results obtained in the dynamic and static approaches, we further revealed that the dynamic strategy significantly improves the predictions of both thermodynamic and kinetic properties as compared to the popular ideal statistic mechanics approaches such as harmonic analysis and TST. It is expected that this accurate yet efficient dynamic strategy can be a powerful tool in understanding reaction mechanisms and reactivity of a catalytic surface system, and further guides the rational design of heterogeneous catalysts.