Femtosecond laser-induced desorption of hydrogen molecules from Ru(0001): A systematic study based on machine-learned potentials

Femtosecond laser-induced dynamics of molecules on metal surfaces can be seamlessly simulated with all nuclear degrees of freedom using ab-initio molecular dynamics with electronic friction (AIMDEF) and stochastic forces which are a function of a time-dependent electronic temperature. This has recently been demonstrated for hot-electron mediated desorption of hydrogen molecules from a Ru(0001) surface covered with H and D atoms [Juaristi et al., Phys. Rev. B 2017, 95, 125439]. Unfortunately, AIMDEF simulations come with a very large computational expense that severely limits statistics and propagation times. To keep ab-initio accuracy and allow for better statistical sampling, we have developed a neural network interatomic potential of hydrogen on the Ru(0001) surface based on data from ab-initio molecular dynamics simulations of recombinative desorption. Using this potential we simulated femtosecond laser-induced recombinative desorption using varying unit cells, coverages, laser fluences, and isotope ratios with reliable statistics. As a result, we can systematically study a wide range of these parameters and follow dynamics over longer times than hitherto possible, demonstrating that our methodology is a promising way to realistically simulate femtosecond laser-induced dynamics of molecules on metals. Moreover, we show that previously used cell sizes and propagation times were too small to obtain converged results.