Development of a Generally Applicable Machine Learning Potential with Accurate Long-Range Electrostatic Interactions
by Tsz Wai Ko
Date of Examination:2022-06-23
Date of issue:2022-08-16
Advisor:Prof. Dr. Jörg Behler
Referee:Prof. Dr. Jörg Behler
Referee:Prof. Dr. Ricardo Mata
Referee:Prof. Dr. Dominik Marx
Persistent Address:
http://dx.doi.org/10.53846/goediss-9342
Files in this item
Name:PhD_thesis_Tsz_Wai_Ko.pdf
Size:11.2Mb
Format:PDF
Description:This thesis summerises the development of a generally applicable machine learning potential, which describes accurate long-range electrostatic interactions
The following license files are associated with this item:
Abstract
English
Machine learning potentials (MLPs) have become an indispensable tool for large-scale atomistic simulations, due to their accuracy comparable with ab-initio methods at considerably reduced computational cost. The development of MLPs has attracted increasing attention and numerous relevant applications in materials science, physics and chemistry have been reported. Most MLPs up to date are based on the approximation of locality, meaning that only short-range atomic interactions are considered. The total energy of the system can be decomposed into a sum of environment-dependent atomic energies. This approximation works well for the majority of systems and allows the MLPs to describe systems containing thousands of atoms with very high accuracy by just training on configurations of small systems. Moreover, they can incorporate long-range electrostatic interactions by employing fixed charges or more flexible environment-dependent charges. Despite countless encouraging developments of MLPs, they are unable to describe non-local effects arising from long-range charge transfer and multiple charge states. This shortcoming prevents the study of many interesting phenomena such as chemical interactions involving protonation/deprotonation and biological processes. A new generation of MLPs such as charge equilibration via neural network technique (CENT) and Becke population neural network (BpopNN) is now beginning to emerge in an effort to address these long standing challenges. In this thesis, the limitations of conventional MLPs are overcome by introducing a fourth-generation high-dimensional neural network potential (4G-HDNNP), which combines accurate atomic energies with a charge equilibration scheme relying on environment dependent atomic electronegativities. 4G-HDNNP describes the correct global charge distribution of the system, resulting in a markedly improved potential energy surface. The capabilities of the method have been demonstrated for a set of benchmark systems that involves non-local charge transfer, where existing methods fail even at the qualitative level. Finally, an extension of the 4G-HDNNP, namely the electrostatically embedded 4G-HDNNP (ee4G-HDNNP), is proposed to further enhance the description of non-local effects, and the general transferability to different configurations that are not covered in the reference data set. The promising improvements of ee4G-HDNNP compared to the 4G-HDNNP have been shown on a large data set of both neutral and charged sodium chloride clusters with large structural diversity. This novel method is anticipated to become a reliable tool for the study of many complex biological and electrochemical problems, while existing ab-initio methods combined with modern computer technology are still computationally demanding for large-scale atomistic simulations.
Keywords: Atomistic Simulations; Density Functional Theory; Electrostatic Interactions; Machine Learning Potentials
