Development of a surrogate machine learning model for the acceleration of density functional calculations with the Korringa-Kohn-Rostoker method
Development of a surrogate machine learning model for the acceleration of density functional calculations with the Korringa-Kohn-Rostoker method
Density functional theory (DFT) has become an indispensable tool in materials science. Specialized DFT methods like the Korringa-Kohan Rostoker Green Function (KKR) method are predestined to investigate the technologically relevant effects of crystallographic defects on the electronic and magnetic s...
Saved in:
Personal Name(s): | Wasmer, Johannes (Corresponding author) |
---|---|
Blügel, Stefan (Thesis advisor) / Rüssmann, Philipp (Thesis advisor) / Berkels, Benjamin (Reviewer) | |
Contributing Institute: |
Quanten-Theorie der Materialien; PGI-1 Quanten-Theorie der Materialien; IAS-1 |
Imprint: |
2021
|
Physical Description: |
113 |
Dissertation Note: |
Masterarbeit, RWTH Aachen University, 2022 |
DOI: |
10.34734/FZJ-2023-05854 |
Document Type: |
Master Thesis |
Research Program: |
Joint Virtual Laboratort for AI, Data Analytics and Scalable Simulation EXC 2004: Matter and Light for Quantum Computing (ML4Q) EXC 2004: Materie und Licht für Quanteninformation (ML4Q) Topological Matter |
Link: |
Get full text Restricted OpenAccess |
Publikationsportal JuSER |
Density functional theory (DFT) has become an indispensable tool in materials science. Specialized DFT methods like the Korringa-Kohan Rostoker Green Function (KKR) method are predestined to investigate the technologically relevant effects of crystallographic defects on the electronic and magnetic structure of host materials. This thesis lays the groundwork for answering the question of whether surrogate machine learning (ML) models have the potential to accelerate such DFT calculations since their computational complexity severely limits them to systems sizes of about a thousand atoms in practice. To that end, a versatile suite of software tools that facilitates the generation and analysis of high-throughput computing DFT datasets with the JuKKR DFT codes and the AiiDA workflow engine is presented. We demonstrate its use by generating a database of 8,760 converged KKR DFT calculations of single impurity embeddings into elemental crystals with 60 different chemical elements and varying lattice constants and that preserves the full data provenance of each calculation. Finally, we use the single-impurity database to compare the Coulomb Matrix and the Smooth Overlap of Atomic Positions (SOAP) as structural descriptors of the local atomic environment for materials defects. Their potential use in surrogate ML models is showcased in a simple example of host crystal structure prediction that achieves 93 percent accuracy. |