Photoinduced desorption dynamics of CO from Pd(111): a neural network approach

A novel approach based on a neural network (NN)-generated potential energy surface (PES) is developed to describe the dynamics of the femtosecond laser-induced desorption of CO from Pd(111). Using trajectories computed with (T_e,T_l) ab-initio molecular dynamics with electronic friction (AIMDEF)¹ as input data, the NN-PES is trained within the embedded atom neural network framework using the atomic configurational energies and forces². The NN-PES robustness is checked by studying the errors in energies and forces, and also by testing its performance in complex molecular dynamics simulations. The (T_e,T_l)-AIMDEF results¹ are reproduced with a remarkable level of accuracy. This shows the outstanding performance of the obtained NN-PES that can cover an extensive range of surface temperatures (90-1000 K) and a large amount of degrees of freedom -those corresponding to multiple adsorbates and surface atoms. Application of this NN-PES for future computational tests of the system dynamics under different initial conditions should be straightforward, as well as the utilization of this methodological framework for development of accurate NN-PESs for other complex gas-solid interfaces.


¹M. Alducin et al. Phys. Rev. Lett. 123, 246802(2019)
²Y. Zhang et al., J. Phys. Chem. Lett. 10, 4962(2019)