Presentation Information
[P2-4]Machine Learning links X-ray diffraction to Coercivity and Phase Analysis
*Qais Ali1, Alexander Kovacs2, Johann Fischbacher2, Harald Oezelt2, David Böhm2, Markus Gusenbauer2, Clemens Wager1, Leoni Breth2, Heisam Moustafa2, Masao Yano3, Noritsugu Sakuma3, Akihito Kinoshita3, Tetsuya Shoji3, Akira Kato3, Thomas Schrefl1,2 (1. Christian Doppler Laboratory for magnet design through physics informed machine learning, Department for Integrated Sensor Systems, University for Continuing Education Krems, Wiener Neustadt, Austria (Austria), 2. Department for Integrated Sensor Systems, University for Continuing Education Krems, Wiener Neustadt, Austria (Austria), 3. Advanced Materials Engineering Division, Toyota Motor Corporation, Susono, Japan (Japan))
Keywords:
Machine learning,Coercivity prediction,Accurate and interpretable,X-ray diffraction,Permanent magnets
High-performance permanent magnets face significant pressure for further development due to their reliance on rare-earth elements, which have high supply-chain risks [1]. Therefore, the main goal is to reduce the rare-earth content while maintaining high coercivity. Coercivity of a magnet is an extrinsic property. While coercivity varies with the operating temperature of the magnet, it can also be optimized by changes in microstructure and composition. However, during such attempts a magnet may also contain multiple phases. Phase analysis can be performed based on X-ray diffraction (XRD) patterns. Although some approaches for automated phase identification have been developed [2], direct prediction of magnetic properties, for instance coercivity using XRD patterns, has yet to be established. In our work, we introduce a solution to this missing link for grain boundary diffused Nd-reduced sintered magnets using machine learning. Infiltration is performed by NdCu solution of different weight percentage and atomic ratio into a base magnet with a low Nd concentration, whereas annealing condition is fixed for all samples. Assuming that grain boundary diffusion will generate core-shell geometry of grains, an optimization scheme for XRD patterns is developed. Systematic optimization of the composition and volume fraction of core, shell and potential secondary phase is then performed with composition optimizer model. The volume fraction of main and secondary phases is also automatically determined from scanning electron microscopy images using WAVEBASE, a cloud based materials development software developed by Toyota [3]. The XRD patterns are high dimensional dataset. By using XRD spectra, machine learning models can learn the relationships between the structure of material and its coercivity, leading to more accurate predictions. However, using the entire pattern for coercivity prediction may result into overfitting of a machine learning model. Therefore, feature importance is performed using predictable and interpretable machine learning models. Our model uses Genetic Algorithm based Partial Least Squares with only the First Component (GA-PLSFC) [4]. A PLS regressor with one component improves interpretability by treating regression coefficients as contributions to objective variables. Nevertheless, the model may lack predictive accuracy with many features. A genetic algorithm addresses this by selecting key features to enhance prediction. This coercivity predictor model also helps to identify the effect of secondary phases on coercivity based on missing peaks out of the main phase. Prediction of coercivity using our approach can help in fast performance screening of both heavy rare-earth lean and rare-earth free magnets. Furthermore, it can be an addition to non-destructive techniques for coercivity estimation. Our model can be trained based on small dataset; therefore, it cannot be applied to predict coercivity of all magnets. Figure 1 shows the graphical representation of this work.
Acknowledgments: The financial support by the Austrian Federal Ministry of Labour and Economy, the National Foundation for Research, Technology and Development and the Christian Doppler Research Association is gratefully acknowledged.
References:
[1] European Commission, Entrepreneurship Directorate-General for Internal Market, Industry, SMEs, M Grohol,
and C Veeh. Study on the critical raw materials for the EU 2023 – Final report. Publications Office of the
European Union, 2023.
[2] N. J. Szymanski, C. J. Bartel, Y. Zeng, Q. Tu, and G. Ceder. Probabilistic deep learning approach to automate
the interpretation of multi-phase diffraction spectra. Chemistry of Materials, 33(11):4204–4215, Jun 2021.
[3] Accessed on 31 Jan 2025. Available at: https://www.toyota.co.jp/wavebase/.
[4] H. Kaneko. Genetic algorithm-based partial least-squares with only the first component for model interpreta-
tion. 7(10):8968–8979, 2022.
Acknowledgments: The financial support by the Austrian Federal Ministry of Labour and Economy, the National Foundation for Research, Technology and Development and the Christian Doppler Research Association is gratefully acknowledged.
References:
[1] European Commission, Entrepreneurship Directorate-General for Internal Market, Industry, SMEs, M Grohol,
and C Veeh. Study on the critical raw materials for the EU 2023 – Final report. Publications Office of the
European Union, 2023.
[2] N. J. Szymanski, C. J. Bartel, Y. Zeng, Q. Tu, and G. Ceder. Probabilistic deep learning approach to automate
the interpretation of multi-phase diffraction spectra. Chemistry of Materials, 33(11):4204–4215, Jun 2021.
[3] Accessed on 31 Jan 2025. Available at: https://www.toyota.co.jp/wavebase/.
[4] H. Kaneko. Genetic algorithm-based partial least-squares with only the first component for model interpreta-
tion. 7(10):8968–8979, 2022.
