High-performance computing approach to hybrid functionals in the all-electron DFT Code FLEUR
| Autor: | Redies, Matthias |
|---|---|
| Přispěvatelé: | Blügel, Stefan, Müller, Matthias S. |
| Jazyk: | angličtina |
| Rok vydání: | 2022 |
| Předmět: | |
| Zdroj: | Aachen : RWTH Aachen University 1 Online-Ressource : Illustrationen (2022). doi:10.18154/RWTH-2022-07156 = Dissertation, RWTH Aachen University, 2022 Jülich : Forschungszentrum Jülich GmbH Zentralbibliothek, Verlag, Schriften des Forschungszentrums Jülich Reihe Schlüsseltechnologien / Key Technologies 257, xi, 109 (2022). = Dissertation, RWTH Aachen University, 2022 |
| DOI: | 10.18154/RWTH-2022-07156 |
| Popis: | Dissertation, RWTH Aachen University, 2022; Aachen : RWTH Aachen University 1 Online-Ressource : Illustrationen (2022). = Dissertation, RWTH Aachen University, 2022 Virtual materials design attempts to use computational methods to discover new materials with superior properties within the vast space of all conceivable materials. Density-functional theory (DFT) is central to this field, enabling scientists to predict material properties from first principles, i.e. without relying on external parameters or experimental values. While standard DFT is capable of predicting many materials with satisfying accuracy, it struggles with some properties such as details of the electronic structure or certain material classes, e.g. materials exhibiting strongly correlated electrons. This has created a need for methods with greater predictive power. One such class of methods are hybrid exchange-correlation functionals which combine the exact Hartree-Fock exchange with local exchange-correlation functionals, resulting in highly accurate predictions for many insulating or semiconductor materials. However, the computational cost of hybrid functionals increases rapidly with system size and limits their application to small systems. This thesis aims to solve the computational challenge posed by hybrid functionals in large systems by utilizing the massive computational power of today's supercomputers. This thesis presents the improved implementation of hybrid exchange-correlation functionals in FLEUR, an all-electron full-potential linearized augmented planewave code. The improved CPU and a new GPU implementations allow users to make efficient use of modern compute nodes and a highly-scalable MPI implementation distributes calculations with a single k-point to 3000 cores or 64 GPUs and far beyond that for calculations with multiple k-points. This work promotes hybrid functionals to systems with hundreds of atoms, opening up their application to many new material classes and properties. We demonstrate the power of this algorithm by applying it to garnets, a class of complex magnetic materials with large unit-cells, which have promising applications in fields such as spintronics or quantum computing. Garnets are rare-earth oxides that exhibit strongly correlated electrons in localized 3d- or 4f-states, making the combination of hybrid functionals and FLAPW ideally suited to investigate these materials. After benchmarking our method against other highly predictive methods and experimental results using yttrium iron garnet as a reference system, we shift our focus the rare-earth-iron garnets gadolinium iron garnet and thulium iron garnet. For these materials we perform the first-ever hybrid exchange-correlation functional calculations of their electronic structure and magnetic moments, establishing the predictive power of the hybrid functionals in FLEUR for large complex magnets. Published by RWTH Aachen University, Aachen |
| Databáze: | OpenAIRE |
| Externí odkaz: |
|