Design of Nanocomposite Magnets by Machine Learning

Permanent magnets are a key technology for sustainable technologies. Currently, high-performance magnets used for motors and generators depend on rare earth elements like Neodymium, Dysprosium or Terbium. To avoid rare earth shortages caused by the increasing demand for electrification of transport and power generation, alternative magnets with significantly lower rare earth content are needed. One solution is a two-phase magnet. It can withstand high external fields (hard magnetic regions) and shows high magnetisation (soft magnetic regions). These two properties are measured by the energy density product and is used as figure of merit for permanent magnets. Rare earth elements are needed for the hard magnetic regions. In this project we aim to find an optimal spatial distribution for magnetically hard and soft regions to reduce the rare earth content while maintaining a high magnetic performance. For this task we will combine fast, massively parallel micromagnetic simulations and artificial intelligence. A framework will be developed to automatically generate parametrisable finite element meshes and perform a large number of micromagnetic simulations. These results for various hard/soft magnetic distributions will be used as training data for a neural network. The network, called Predictor, learns the influence of material composition and geometrical properties on the overall energy density product. Learning happens by adjusting the parameters of the network, the weights, employing tailored mathematical methods. We will explore and adapt various methods for high dimensional optimization problems for this task. A copy of the trained Predictor network with fixed weights will be used inversely as a Designer network. The Designer will be used to optimise the material composition and geometrical properties for high energy density products. Newly found design parameters will then by validated by micromagnetic simulations and fed back to the Predictor as training data in a feedback loop. First, this active learning scheme will be developed and validated for simple, well-known magnetic structures. In a further step, we adapt this machine learning scheme to search for optimal material distribution with a resolution of a few hundred atoms. With this generative neural network for inverse design of high-performance, rare earth reduced permanent magnets we will push the boundaries of structural design strategies towards the theoretical limit. Our findings will provide new guidelines to produce competitive, eco-friendly permanent magnets for sustainable technologies.