Machine learning for the prediction of intrinsic magnetic properties and microstructure characterization of sintered FeNdB type permanent magnets

The properties of material, which are the measurable responses of a material under various conditions, are determined by the properties of its phases and microstructure. Microstructure is the arrangement of phases and defects within a material characterized by various attributes, including type, size, orientation, shape, and amount. While intrinsic properties are studied through experiments and simulations, microstructure characterization typically relies on the microscopy image analysis. These factors collectively influence the overall properties and behavior. The extrinsic properties of a material arise from the intrinsic properties of its individual phases and the microstructure of the material. However, traditional methods for determining these properties can be tedious, expensive, subjective, and error prone due to manual efforts. Therefore, to address these challenges, a data-driven approach has been developed to obtain a better trade-off between accuracy and time-efficiency.

In this thesis, the generalization capabilities of machine learning approaches for intrinsic property prediction tasks and quantitative microstructure characterization using microscopy image data are demonstrated. ... mehrMore specifically, the use of Newtonian space of chemical composition of the 14:2:1 hard magnetic phase (TM14RE2B; TM = transition metal, RE = rare earth, B = boron) to predict the intrinsic magnetic properties at room temperature, while Kerr microscopy (KM) is used for microstructure characterization of the sintered FeNdB type permanent magnet. The 14:2:1 phase enables permanent magnets with excellent magnetic properties, and from an application perspective, Curie temperature (Tc), saturation magnetization (Ms), and anisotropy constant (K1) are essential parameters. The regression models to predict the mass density in [g/cm3], Tc in [K], and Ms in [µB/f.u.] of the 14:2:1 phase at room temperature are developed using a supervised learning approach and a dataset that includes the experimentally reported values across various literature sources. These models show strong predictive performance and are generalized to be valid across ternary to senary alloys, with an average deviation of 1.5% for mass density, a mean absolute error of less than 16 K which accounts for an average deviation of less than 3% for Tc, and an average deviation of less than 3% for Ms when compared against reference values.

For microstructure characterization, KM holds vital information for analyzing the distribution of grains, crystal orientation, and magnetic domain structures. It is less expensive and requires less time for sample preparation, image acquisition, and material analysis compared to the electron backscatter diffraction (EBSD) approach, especially for a large cross-section of samples. Within this thesis, deep learning models were developed using supervised learning approaches to analyze grain size and crystallographic grain orientation in FeNdB magnets from KM images. These models achieved a better trade-off between accuracy and time-efficiency, being 10x faster with a prediction error of less than 4% for grain size and less than 5º deviation for orientation analysis when compared to the EBSD approach, with the comparison including image acquisition, processing, and material analysis.

Finally, the developed data-driven approaches could serve as the basis for developing novel magnetic 14:2:1 phases from chemical composition and emerge as a cost effective tool for rapidly analyzing large cross-sectional areas of the samples to quantify heterogeneities in grain size and crystallographic grain orientation.