Presentation Information
[O15-4]Artificial Intelligence to support Permanent Magnet Research and Development – Intrinsic Magnetic Properties and Microstructure Analysis
*Gerhard Schneider1, Amit Kumar Choudhary1, Andreas Jansche1, Rajkumar Varaganti1, Dominic Hohs1, Dagmar Goll1 (1. Aalen University (Germany))
Keywords:
Machine Learning,Microstructure,Domain Pattern,Permanent Magnet,Intrinsic Properties
The properties of functional materials are determined by the intrinsic properties of the phases present and the microstructure.
In FeNdB permanent magnets, the main phase is the 14:2:1 phase. Other elements, which can substitute Fe, Nd or B in the ternary phase to some extent, modify the intrinsic magnetic properties. We have developed regression models to predict the density ρm, the Curie temperature TC as well as the saturation magnetization Ms and the anisotropy field Ha of 14:2:1 phases at room temperature from the chemical composition as input features. The dataset for the training and testing of the model was collected from literature. Ternary, quaternary, quinary and senary alloy systems have been taken into account. The models demonstrate a high accuracy in predicting ρm, TC, Ha and Ms. The ρm prediction model achieves an average prediction error of 0.5% on unseen phases, the average deviation for TC prediction is 3%, for Ms 3% and for Ha 10% when compared to literature values.
Artificial intelligence approaches are also helpful in characterizing the microstructure of sintered FeNdB magnets. Several different tools have been developed for digital image enhancement, phase segmentation, grain size analysis, analysis of crystal orientation of grains and defect detection. Based on a data-driven approach, a computer-assisted workflow for the quantitative analysis of optical Kerr microscopy images of sintered FeNdB-type permanent magnets was developed (Fig. 1). By analyzing the domain patterns visible in the Kerr image with data-driven approaches such as traditional machine learning and advanced deep learning, we can quantify grain orientation and size. The values from the trained machine learning models were compared to the measurements from EBSD and manual analysis for performance evaluation. The trained Deep Learning model for measuring grain orientation e.g. produces an error of about 2% when compared to the EBSD approach. Machine learning approaches are also developed to identify microstructural defects in sintered rare earth magnets as large pores, oxide agglomerates or abnormally grown large grains. The population of defects e.g. enables different manufacturers to be differentiated.
Reference
[1] A.K. Choudhary, T. Grubesa, A. Jansche, T. Bernthaler, D. Goll, G. Schneider, Deep learning and correlative microscopy for quantification of grain orientation in sintered FeNdB-type permanent magnets by domain pattern analysis, Acta Materialia 264 (2024) 119563, https://doi.org/10.1016/j.actamat.2023.119563
In FeNdB permanent magnets, the main phase is the 14:2:1 phase. Other elements, which can substitute Fe, Nd or B in the ternary phase to some extent, modify the intrinsic magnetic properties. We have developed regression models to predict the density ρm, the Curie temperature TC as well as the saturation magnetization Ms and the anisotropy field Ha of 14:2:1 phases at room temperature from the chemical composition as input features. The dataset for the training and testing of the model was collected from literature. Ternary, quaternary, quinary and senary alloy systems have been taken into account. The models demonstrate a high accuracy in predicting ρm, TC, Ha and Ms. The ρm prediction model achieves an average prediction error of 0.5% on unseen phases, the average deviation for TC prediction is 3%, for Ms 3% and for Ha 10% when compared to literature values.
Artificial intelligence approaches are also helpful in characterizing the microstructure of sintered FeNdB magnets. Several different tools have been developed for digital image enhancement, phase segmentation, grain size analysis, analysis of crystal orientation of grains and defect detection. Based on a data-driven approach, a computer-assisted workflow for the quantitative analysis of optical Kerr microscopy images of sintered FeNdB-type permanent magnets was developed (Fig. 1). By analyzing the domain patterns visible in the Kerr image with data-driven approaches such as traditional machine learning and advanced deep learning, we can quantify grain orientation and size. The values from the trained machine learning models were compared to the measurements from EBSD and manual analysis for performance evaluation. The trained Deep Learning model for measuring grain orientation e.g. produces an error of about 2% when compared to the EBSD approach. Machine learning approaches are also developed to identify microstructural defects in sintered rare earth magnets as large pores, oxide agglomerates or abnormally grown large grains. The population of defects e.g. enables different manufacturers to be differentiated.
Reference
[1] A.K. Choudhary, T. Grubesa, A. Jansche, T. Bernthaler, D. Goll, G. Schneider, Deep learning and correlative microscopy for quantification of grain orientation in sintered FeNdB-type permanent magnets by domain pattern analysis, Acta Materialia 264 (2024) 119563, https://doi.org/10.1016/j.actamat.2023.119563
