Reactivity at surfaces with high-dimensional neural network potentials

by Martin Liebetrau
Doctoral thesis
Date of Examination:2023-06-02
Date of issue:2023-07-24
Advisor:Prof. Dr. Jörg Behler
Referee:Prof. Dr. Jörg Behler
Referee:Prof. Dr. Alec Wodtke
Persistent Address: http://dx.doi.org/10.53846/goediss-9956

 

 

Files in this item

Name:Thesis_Final_ML_published.pdf
Size:40.8Mb
Format:PDF

The following license files are associated with this item:

Abstract

English

The surface of a material is critical for its properties, especially in heterogeneous catalysis. Surfaces properties can be probed, for example, by H atom scattering. However ab-inito theoretical investigations of H atom scattering are held back by the high cost for performing the large amount of necessary simulations. Machine learning potentials can be used to bridge the gap between the high accuray of ab-inito calculations and the low computational cost of force field calculations. In this thesis, a high-dimensional neural network potential (HDNNP) based on RPBE density funcitonal theory reference data has been constructed for the scattering of H atoms from the [0001] α-Al2O3 surface. This system is well accessible for theoretical studies and experimental benchmarks due to the simple nature of H atom scattering, lacking the involvement of steric or vibrational effects. The process of generating the reference data set and validating the constructed HDNNP, using system properties like the phonon band structure and the potential energy surface as a function of collective variables, is described in detail. Furthermore, the HDNNP is validated using experimental kinetic energy distributions and angular distributions from H atom scattering. Insights into the surface structure are gained by comparing the experimental and theoretical scattering distributions. Finally, the accuracy of the RPBE functional describing the interaction of the H atom with the surface is evaluated and discussed.
Keywords: Alumininium oxide; Theoretical Chemistry; Machine Learning; High-Dimensional Neural Network Potentials; Density Functional Theory; Atomistic Simulations; H Atom Scattering; Experimental Benchmark
 
Statistik