Leveraging available data for efficient exploration of materials space using Machine Learning: A case study for identifying rare earth-free permanent magnets

Abstract

It is shown that even the relatively small amount of available materials data can be innovatively utilized to explore the materials space in order to identify materials with desired target properties. As an example of this, data from the novomag and Novamag databases are used to train random forest and neural network models which can predict thermodynamic stability, and magnetic properties of materials. Performance of these models are tested thoroughly and are found satisfactory. These models are subsequently used to interpolate within the above databases, and to extrapolate to parts of the materials composition and structure space not covered in these databases, to identify stable, magnetic materials that have large saturation magnetization and large easy-axis anisotropy. Screening 686 materials via the trained models, and subsequently performing first principles calculations, 21 new candidate materials for rare earth free permanent magnet are identified. Some of these materials have anisotropy constants as large as 5 and 6 MJ m−3, larger than that of the most widely used permanent magnet Nd2Fe14B. This simple approach can be used to screen materials with other functionalities in future.