(377c) Large-Scale Nonadiabatic Molecular Dynamics Enabled By Machine Learning | AIChE

(377c) Large-Scale Nonadiabatic Molecular Dynamics Enabled By Machine Learning


Wang, J. - Presenter, Virginia Polytechnic Institute and State University
Xin, H., Virginia Tech
Transferring energy from solid surfaces to chemical bonds of adsorbed species is a fundamental process in heterogeneous catalysis. To a great extent, the dynamics of molecule-surface interactions driven by thermal energy, i.e., heat, can be well described within the framework of the Born-Oppenheimer approximation. New reaction channels[1] often open up in response to electronic excitations via surface-mediated energy transfer. In this process, the energetic electrons or holes can scatter into surface species and heat up vibrational modes of adsorbates. This process involves not only the electronic ground state but also the excited state, and it is nonadiabatic with respect to the nuclear motions[2].

In the weak coupling limit, the nonadiabatic energy transfer on metal surfaces can be modeled using only the ground-state potential energy surfaces and the nonadiabacity is taken into account with electronic friction contributions in Langevin dynamics by solving Newtonian equations with a thermal fluctuation term[2], [3]. The electronic friction via partial populating and depopulating of the adsorbate ground state right above the Fermi level is governing the energy transfer. The crucial parameter here is the electronic friction of an excited charge with nuclear motions of surface adsorbates. However, it is extremely time consuming to compute electron-phonon coupling strength for large systems.

In this talk, we will discuss our most recent development of a machine-learning nonadiabatic molecular dynamics approach that uses predicted forces and electronic friction coefficients by ‘learning from data’. Machine learning algorithms, such as the artificial neural networks[4] can use past trajectories as training datasets for fast and accurate prediction of forces and electronic friction coefficients, thus allows us to perform statistical analysis of many trajectories. We benchmark the approach for activation of oxygen from hollow to bridge because of strong interests in understanding laser-induced chemistry of CO oxidation on metal surfaces. By comparing the Head-Gordon/Tully electronic friction with the Local Density Friction Approximation (LDFA), we found that the LDFA is mode unselective, thus giving identical coefficients along the lateral and vertical directions to surface and no oxygen activation is observed, while the Head-Gordon/Tully friction can differentiate the electron dragging force in vertical and lateral modes and capture the laser-induced oxygen activation on Ru [2].

[1] M. Bonn et al., “Phonon- Versus Electron-Mediated Desorption and Oxidation of CO on Ru(0001),” Science, vol. 285, no. 5430, pp. 1042–1045, Aug. 1999.

[2] M. Head‐Gordon and J. C. Tully, “Molecular dynamics with electronic frictions,” J. Chem. Phys., vol. 103, no. 23, pp. 10137–10145, Dec. 1995.

[3] M. Brandbyge, P. Hedegård, T. F. Heinz, J. A. Misewich, and D. M. Newns, “Electronically driven adsorbate excitation mechanism in femtosecond-pulse laser desorption,” Phys. Rev. B, vol. 52, no. 8, p. 6042, 1995.

[4] A. Khorshidi and A. A. Peterson, “Amp: A modular approach to machine learning in atomistic simulations,” Comput. Phys. Commun., vol. 207, pp. 310–324, Oct. 2016.