Project Lead: Claas Abert
Lifetime: 2021-2025

Scientific Aims:

Computational micromagnetics has a long history as a valuable tool for the theoretical investigation of magnetic systems at the micron scale. While first micromagnetic computations in the 1960ties were performed on small two-dimensional systems with only a few degrees of freedom, the enormous increase of computing power and the development of improved numerical algorithms has led to a broad landscape of micromagnetic codes that are capable of handling millions of degrees of freedom in a reasonable time. Despite these advances, today’s micromagnetic simulations are still restricted to the micron sized systems since the micromagnetic model calls for a very fine spatial discretization in the single-digit nanometer regime.

Within this project, we aim to significantly push the boundaries of size restrictions and simulation speed by introducing algorithms that are particularly suited for distributed computations and by relying on software frameworks that a specialized on the handling of large amounts of data. Namely, we will implement the parallel-in-time integration scheme parareal which has already proved to lead to significant computational speedups in other scientific disciplines. In another subproject we will exploit the capabilities of modern finite-element libraries to implement a distributed higher-order micromagnetic code suited for high-performance computational clusters. In the third subproject, we will use the highly optimized tensor library TensorFlow to implement a novel finite-difference formulation to be solved on graphics processing units. This novel formulation will allow for the rigorous and accurate description of composite materials which will be particularly beneficial for the simulation of granular media as used in numerous magnetic applications. Combining the findings of the individual subprojects will lead to further performance enhancements.

For a long time, the evolution of computing power was driven by increasing processor clock speed. Due to physical limitations, this increase stopped in recent years leading to a stagnating performance of serial codes. Since then, exploiting parallel computing architectures has become a crucial task for the development of scientific software. Bringing established techniques for distributed computing from various scientific disciplines to micromagnetic simulation tools will significantly advance the field of largescale micromagnetic computations and will pave the way to realistic macroscopic simulations.