guest ::
| Home > Publications database > AngstromCube - Truncated Green’s functions meet real-space formalism |
| Home > Publications database > AngstromCube - Truncated Green’s functions meet real-space formalism |
| Talk (non-conference) (Other) | FZJ-2025-05494 |
;
2025
Abstract: Linear-scaling Density Functional Theory (DFT) calculations without a restriction to non-metals require a real-space representation of the Green's function.Range-truncated Green's functions have first been introduced in the context of the Korringa-Kohn-Rostoker multiple-scattering method (KKR).In contrast, wave function-based (standard) DFT has proven to achieve the same accuracies on uniform Cartesian real-space grids as with a plane-wave basis set.A grid-based representation of Green's function allows to perform range truncation while it comes with a high computational costs.As the memory footprint is orders of magnitude larger than that of the compact KKR basis, top-tier HPC installations are necessary to offer sufficient memory.Nevertheless, the method promises improved accuracy for the determination of forces. Proper scattering properties of the atomic nuclei on real-space grids are ensured by using pseudopotentials or, more general, the Projector Augmented Wave method (PAW).The localized projector functions require a higher computational effort in the grid regions around atomic positions, so that the workload distribution reflects the geometry of atoms and requires taylor-made load balancing solutions.We show how this issue is addressed in the GPU-accelerated linear-scaling Green's function DFT application AngstromCube.
|
The record appears in these collections: |
PDF