The unique electronic structure of metals may cause significant differences between classical adiabatic molecular dynamics (MD) and experimental results. Traditionally, molecular dynamics methods utilise the Born-Oppenheimer approximation, which assumes separation of electronic and nuclear degrees of freedom and enables that the nuclear dynamics can be described by a single potential energy surface (PES). However, on metallic surfaces often the energy exchange between adsorbate and surface atoms is significant and causes the breakdown of the Born-Oppenheimer approximation.

There are multiple methods to include such nonadiabatic effects, with one of the most efficient being molecular dynamics with electronic friction (MDEF) [1]. MDEF introduces the non-adiabatic effects by means of both electronic friction and random forces. However, a meaningful comparison between computational simulations and experiments demands the capability to run thousands of MDEF trajectories which are currently unfeasible through ab- initio MD. In this context, machine learning provides a lighter model to compute the electronic friction tensor and the PES allowing to overcome this computational limitation.

Currently, I work on developing machine-learning-based models that enable simulating nonadiabatic molecular dynamics of hydrogen molecules at different copper surfaces.

[1] Head-Gordon, M.; Tully, J. C. J. Chem. Phys. 103, 10137–10145 (1995)