F. Bruckner:

"Multiscale simulation of magnetic nanostructures";

Supervisor, Reviewer: D. Süss, D. Praetorius; Institute for Solid State Physics, 2013; oral examination: 2013-05-27.

Due to the ongoing miniaturization of modern magnetic devices, ranging from GMR sensors,

magnetic write heads, to spintronic devices, micromagnetic simulations gain more and more

importance, since they are an essential tool to understand the behavior of magnetic materials

in the nanometer scale. The use of numerical simulations allows to optimize the

microstructure of such devices or to test new concepts prior to performing expensive

experimental tests.

The purpose of this work is to extend the micromagnetic model by additional macroscopic

parts which are described in an averaged sense. In contrast to the microscopic parts which are

described by Landau-Lifshitz-Gilbert (LLG) equations, these macroscopic parts are based on

classical magnetostatic Maxwell equations, which could be extended to a full Maxwell

description in a straight-forward way. The averaged description using Maxwell equations

allows to overcome the upper bound for the discrete element sizes, which is intrinsic

to the micromagnetic models, since the detailed domain structure of the ferromagnetic

material needs to be resolved. Combining microscopic and macroscopic models and solving

the corresponding equations simultaneously provides a multiscale method, which allows to

handle problems of a dimension, which would otherwise be far out of reach. A basic

prerequisite for the application of the method, is that microscopic and macroscopic parts can

be separated into disjoint regions. The performance of the implemented algorithm is

demonstrated by the simulation of the transfer curve of a magnetic recording read head, as it

is used in current hard drives.

Independent of the former approach the use of parallelized algorithms allows to handle larger

problems. This work especially deals with the shared memory parallelization of hierarchical

matrices, since these require a large amount of the storage consumption as well as of the

computation time for typical simulations. For the setup of the matrices a nearly perfect

parallelization could be reached, whereas for the matrix-vector-multiplication the

computation time stagnates at a few computation cores, due to the restricted main memory

bandwidth.

Created from the Publication Database of the Vienna University of Technology.