CD Laboratory for Magnet design through physics informed machine learning

In order to prevent supply shortages of rare earth elements, this CD Laboratory is developing new strategies for the material design of magnets. Simulation models on supercomputers enable optimisation beyond human understanding.

High-performance magnets play a key role in green technologies such as sustainable energy generation and environmentally friendly transport. In the latter, the ambitious climate policy target of a maximum average temperature increase of two degrees could be achieved through rapid and comprehensive electrification of the powertrain. This will have a significant impact on the demand for critical materials, with an eight-fold increase in the average demand for neodymium for energy technologies, cars and appliances forecast between 2015 and 2050. Neodymium and heavy rare earth elements such as terbium and dysprosium are essential components for high-temperature permanent magnets. In order to prevent supply shortages of rare earth elements, the CD Laboratory is developing new material design strategies for magnets without terbium and dysprosium and with a reduced neodymium content. This is achieved using machine learning methods that support computer-aided magnet design by integrating physical models across all relevant length scales, with material optimisation also taking raw material costs into account.

This CD Laboratory is based on a software suite developed in-house, i.e. a combination of several computer programs. This is now used worldwide for the development of magnetic data storage devices, magnetic sensors and permanent magnetic materials. This software suite, based on the computing power of many computers working in parallel, i.e. massively parallel hardware, allows the study of magnetisation processes on the basis of the smallest magnetic units, the core-shell grains - and can extrapolate their properties to the level of the entire magnet using machine learning.

Such micromagnetic simulations can be used to calculate the magnetic field strength required to demagnetise a permanent magnet (the so-called coercive field). The aim is to train a regression model that links the coercive field of a grain with its geometry and rare earth content. The single grain model forms the basis for calculating the remagnetisation of many interacting grains of the magnetic material. On a macroscopic scale, a neural network is developed for fast magnetic field estimation.

The proposed methodology has a twofold optimisation potential: on a microscopic level, the chemical composition and the geometry of the core-shell grain are optimised. On the other hand, at the device level, the multi-material optimisation assigns highly coercive (and expensive) materials only to regions that are exposed to strong demagnetisation fields. The project results can dramatically improve magnetic properties beyond human understanding and support research and development for a rapid market introduction of permanent magnets for environmental preservation.