Computational Description of Chemical Bond Formation Using Neural-Network Potentials: H-Atom Scattering from Graphene
by Sebastian Wille
Date of Examination:2022-06-29
Date of issue:2022-08-01
Advisor:Prof. Dr. Alec M. Wodtke
Referee:Prof. Dr. Alec M. Wodtke
Referee:Prof. Dr. Jörg Behler
Referee:Prof. Dr. Ricardo A. Mata
Referee:Prof. Dr. Claus Ropers
Referee:Prof. Dr. Burkhard Geil
Referee:Prof. Dr. Martin A. Suhm
Files in this item
Name:PhD_thesis_SW.pdf
Size:44.0Mb
Format:PDF
Description:Dissertation_Sebastian_Wille_2022
Abstract
English
When a hydrogen or a deuterium atom is interacting with a graphene surface during scattering events, its energy loss determines if the projectile will be adsorbed or scattered. The energy loss highly depends on the initial conditions of the projectile and the surface. Unlike for metal and insulator surfaces, a barrier to C–H bond formation is involved, which is the movement of a single carbon atom out of the surface plane (puckering) and by that rehybridization of the C-atom from sp2 to sp3. In the experiments, a bimodal branching of the translational energy and scattering angle distribution is observed. One signal corresponds to a quasi-elastic channel, where the impinging atom loses only a minor fraction of its initial kinetic energy. The location of this channels’ maximum intensity can be reasonably well estimated with a simple binary collision model. The second signal corresponds to inelastic scattering events, where the projectile loses a large fraction of its initial kinetic energy. Previous studies using a reactive empirical bond order potential are in good agreement with incidence kinetic energies of 1 eV, but could not reproduce the distribution of scattering with 2 eV incidence kinetic energy at given incidence conditions. Machine learning potentials are best known to bridge the gap between accuracy and performance. To study the bond formation of H- and D-atoms with graphene with high dimensional neural network potentials, I implemented the existing description of neural network potentials in our developed molecular dynamics program and generated a high-dimensional neural network for H-atoms at a free-standing graphene surface. The performance of this new potential to accurately describe the experiment is much better than the previously reported reactive bond order potential. The neural network potential is able to capture the correct branching of the two channels observed in the experiment and can even reproduce subtle differences seen in energy loss distributions for different hydrogen isotopes. But still, systematic differences can be seen, especially for small incidence polar angles, where out-of-plane scattering is more likely. New experiments for a H-atom scattering along the surface normal indicate that nuclear quantum effects might play a more important role in the scattering dynamics than previously anticipated. Therefore, classical and path-integral based simulations are compared to fullquantum mechanical simulations based on wave-package propagation. Classical simulations are incapable to predict the correct energy loss and scattering distributions from experiment for higher incidence kinetic energies of the projectile. Path-integral based simulations improve the description, but only taking all quantum effects into account offers the best description of H-atom scattering from graphene with normal incidence direction.
Keywords: Atomic Scattering; Machine Learning; Molecular Dynamics