Micromagnetic and Multiphysics Software
Principal Investigators
Dieter Suess, Florian Bruckner and Claas Abert
Predicting how magnetic materials behave at the nanoscale requires solving the equations of micromagnetism numerically. In real devices, however, magnetism rarely acts alone: magnetization dynamics interacts with elastic deformations, acoustic waves and electric currents.
Our group develops open-source simulation software that makes these complex computations fast, flexible and reproducible. Our latest codes, magnum.np and NeuralMag, build on machine-learning frameworks such as PyTorch and JAX. They run efficiently on modern computing hardware and are fully differentiable, enabling classical micromagnetic, multiphysics and inverse-design simulations.
Background and Motivation
Micromagnetics describes magnetization processes on length scales ranging from nanometres to micrometres. It provides an essential theoretical foundation for the development of magnetic sensors, data-storage technologies, spintronic devices and magnonic systems.
At the centre of micromagnetic modelling is the Landau–Lifshitz–Gilbert equation, which describes the dynamics of magnetization. For realistic geometries and material configurations, this equation must be solved numerically. The capabilities of the available simulation software therefore directly determine which physical systems and research questions can be investigated.
Modern magnetic devices are also rarely governed by magnetic interactions alone. Magnetization dynamics can couple to elastic deformations, surface acoustic waves and piezoelectric substrates. Accurately modelling these systems requires genuine multiphysics simulation methods that combine different physical equations and material properties.
Our group has a long tradition in developing such numerical tools, beginning with the finite-element software magnum.fe and complemented by extensive work on micromagnetic models and numerical methods.
Computational Methods
magnum.np
Our current developments build on machine-learning frameworks instead of hand-written low-level kernels.
magnum.np [Sci. Rep. 13, 12054 (2023)] is a finite-difference code written entirely in PyTorch: the device abstraction of the tensor library allows the same high-level Python code to run on CPUs and GPUs with performance competitive with established codes, while the compact code base makes it easy to extend with new interactions and solvers.
Magnetoelastic Solver
One such extension is a self-consistent magnetoelastic solver that integrates the Landau–Lifshitz–Gilbert equation, augmented by strain-induced effective fields, concurrently with the elastic equation of motion, including the correct treatment of stress and strain conditions at material interfaces [Phys. Rev. Applied 25, 034050 (2026)].
NeuralMag
NeuralMag [npj Comput. Mater. 11, 193 (2025)] introduces a novel nodal finite-difference discretization with improved accuracy over the traditional scheme at no additional cost and supports interchangeable PyTorch and JAX backends on CPUs, GPUs and TPUs.
Both codes are freely available under open-source licenses, installable via pip, and are developed openly on GitLab together with international partners.
Concept of a magnetically programmable surface acoustic wave filter: a SAW launched by interdigital transducers travels through an array of perpendicularly magnetized islets; switching the array between antiparallel (top) and parallel (bottom) states programs the transmission of the filter.
Key Results and Applications
Differentiable Micromagnetic Simulations
Because the machine-learning frameworks underlying magnum.np and NeuralMag provide automatic differentiation, entire micromagnetic simulations become differentiable with respect to any input — geometry, material parameters or external fields.
Gradients with respect to millions of design variables can thus be computed at the cost of roughly one additional simulation, which turns the codes into powerful engines for inverse problems such as the design of magnonic devices or the reconstruction of magnetization from measurements (see our research topic on inverse computational magnetism).
Predictive Device Modeling
The multiphysics extensions enable predictive device modeling: we recently proposed a surface acoustic wave filter whose transmission is programmed by its internal magnetic state rather than an external field — simulations predict transmission changes of 52 dB/mm at 3.8 GHz when arrays of exchange-decoupled magnetic islets are switched between parallel and antiparallel configurations [npj Spintronics 4, 13 (2026)].
Open-Source Impact
Beyond our own research, the codes serve a growing international user community in academia and industry as transparent, reproducible and extensible alternatives to closed simulation software.
Inverse design with NeuralMag: a topology-optimized permanent magnet (red) inside the design region (transparent blue) that maximizes the magnetic field at a target point above the structure.