Data driven magnet design

Reducing greenhouse gas emissions has become a top priority on the global agenda. Electrification of transport and renewable energies rely on permanent magnets. Tailoring permanent magnets to the specific requirements of an application while reducing the content of critical elements is vital for green technologies. Micromagnetic simulations are a viable tool for finding new material compositions and structures at the microscopic length scale. Unfortunately, current simulations are hardly scalable to address design issues at the much larger length scale of applications. To achieve a breakthrough, machine learning will be used. We will improve the fundamental understanding of permanent magnet performance using a so-called inverse design approach. First, magnetic thin films with many different compositions of neodymium, dysprosium, lanthanum, and cerium with iron and boron will be fabricated. These films will be characterized by high-throughput analysis, and the data will be stored in a database that relates chemical composition, structure, and processing conditions to magnetic properties. Second, this database is then used to develop and train a machine learning algorithm that employs a graph network to predict the performance of the magnets. The machine learning model will be enhanced by combining measured data with results from micromagnetic simulations of the magnets. Finally, we will use the trained machine learning model inversely to search for promising magnet compositions and structures for the desired magnetic properties. The machine learning model will use graph neural networks. The granular structure of permanent magnets leads to a natural representation in graphs. Graph neural networks can make prediction for larger and more complex systems than used during training. The proposed technique allows prediction of magnetic properties for large systems including thousands of grains. The project will focus on tailoring the properties of (Nd,Dy,La,Ce)FeB magnets with greatly reduced Nd and Dy content, which are considered particularly critical in the coming years. This reduction will be achieved by exploring possible chemical compositions and microstructures.