Machine learning molecular dynamics simulation of vibration driven CO₂ hydrogenation to formate on Cu(111) surface

抄録

Methanol plays critical roles in chemical industries but the precise behavior of the synthesis under operating conditions remains unelucidated. The methanol synthesis is initiated by the CO₂ hydrogenation on Cu-based catalyst, a gas-surface reaction which proceeds by Eley-Rideal (E-R) type mechanism. Through molecular-beam experiment, it has been shown that the dissociative and associative gas-surface reactions can be precisely controlled by adjusting the translational, rotational, and vibrational energies of the molecular reactant. The result shows that the hydrogenation can be promoted by increasing the vibrational and translational energies of hot CO₂, in which the vibrational energy is thought to play more significant role [1]. However, the molecular beam experiment only allows the control of the average of CO₂ vibrational energy. Considering there are multiple vibrational modes of CO₂ molecule, the question is raised on which mode that actually contributes to the promotion of the reaction.

To elucidate the importance of vibrational mode in the CO₂ hydrogenation, a set of molecular dynamics (MD) simulations driven by machine-learning potential has been performed. Unlike the molecular beam experiment, MD allows precise control of initial vibrational mode of CO₂ (i.e., bending and stretching modes) combined with translational energy and incident angles. Ultimately, the dependence between reaction probability and specific mode of CO₂ can be obtained.

The active and on-the-fly learning scheme is implemented to generate the database efficiently. This scheme utilizes the quantification of uncertainty provided by the Gaussian Process framework that learn the energy and forces of various atomic environments in order to generate the accurate machine-learning potential. The results of MD shows that the successful hydrogenation of CO₂ (Fig. 1) as well as the desorption of CO₂ due to the Pauli repulsion (Fig. 2) can be simulated with satisfying accuracy under experimental surface temperature. Fig.3 and Fig.4 shows the high similarity of the kinetic energy profile obtained by machine-learning MD with the ab-initio MD. The details on the behavior of reaction as the result of varying the normal mode excitation will be presented in the conference presentation, especially the combination of initial states that might reduce the required translational energy for the hydrogenation.

[1] Quan, J. et al. Nat. Chem. 11, 722–729 (2019)