Presentation Information
[O15-1]Artificial Intelligence Assisted Optimization of Permanent Magnets
*Thomas Schrefl1, Johann Fischbacher2, Harald Oezelt2, Alexander Kovacs2, Qais Ali1, Clemens Wager1, Felix Lasthofer1, Hyuga Hosoi3, Yusuke Umetani3, Akihito Kinoshita3, Masao Yano3, Noritsugu Sakuma3, Akira Kato3, Tetsuya Shoji3 (1. Christian Doppler Laboratory for Magnet design through physics informed machine learning (Austria), 2. Department for Integrated Sensor Systems, University for Continuing Education Krems (Austria), 3. Advanced Materials Engineering Division, Toyota Motor Corporation (Japan))
Keywords:
micromagnetics,machine learning,graph neural network,optimization
To mitigate the supply risk of rare-earth elements, we are developing magnets that eliminate the use of terbium and dysprosium and reduce the neodymium content. Achieving these goals requires new strategies for material design, including tuning materials composition and controlling microstructure. We have developed machine learning methods that assist in materials design by integrating physical models to bridge the gap between length scales. Magnet design guidelines have been computed, considering raw material costs.
Starting from scanning electron microscopy (SEM) images, we reconstructed three-dimensional synthetic microstructures that replicate the statistical grain distributions observed in two-dimensional images. A Bayesian optimizer was used to minimize the difference between the power spectra of SEM images—computed using the materials platform WAVEBASE [1]—and those from slices through synthetically generated grain structures.To efficiently generate training data for machine learning models, we developed a reduced-order model for magnetization reversal, enabling simulations of large-grained magnets beyond the computational limits of traditional micromagnetic simulations [2]. The synthetically generated microstructures serve as input for this model, allowing for realistic simulations of magnetization processes. With this approach, we analyzed domain evolution during magnetization reversal in large-grained magnets.
Both synthetic microstructure generation and the reduced-order model were used to build a fast model for estimating demagnetization curves. We represented synthetic granular structures as graphs and trained a graph neural network [3] with demagnetization curves computed with the reduced-order model. The trained network accurately predicted demagnetization curves, matching results from full micromagnetic simulations and the reduced-order model (see Figure 1). The graph neural networks offer a way to represent and analyze complex, three-dimensional polycrystalline structures.
Finally, microstructural parameters derived from the hysteresis models discussed above were used in a genetic optimization algorithm that optimizes the chemical composition for minimum price and maximum coercivity. The financial support by the Austrian Federal Ministry for Labour and Economy and the National Foundation for Research, Technology and Development and the Christian Doppler Research Association is gratefully acknowledged.
[1] Accessed on 31 Jan 2025. Available at: https://www.toyota.co.jp/wavebase/
[2] Kovacs, A.; Fischbacher, J.; Oezelt, H.; Kornell, A.; Ali, Q.; Gusenbauer, M.; Yano, M.; Sakuma, N.; Kinoshita, A.; Shoji, T.; Kato, A.; Hong, Y.; Grenier, S.; Devillers, T.; Dempsey, N. M.; Fukushima, T.; Akai, H.; Kawashima, N.; Miyake, T.; Schrefl, T. Physics-Informed Machine Learning Combining Experiment and Simulation for the Design of Neodymium-Iron-Boron Permanent Magnets with Reduced Critical-Elements Content. Front. Mater., 9, 1094055 (2023)
[3] Battaglia, P. W.; Hamrick, J. B.; Bapst, V.; Sanchez-Gonzalez, A.; Zambaldi, V.; Malinowski, M.; Tacchetti, A.; Raposo, D.; Santoro, A.; Faulkner, R.; Gulcehre, C.; Song, F.; Ballard, A.; Gilmer, J.; Dahl, G.; Vaswani, A.; Allen, K.; Nash, C.; Langston, V.; Dyer, C.; Heess, N.; Wierstra, D.; Kohli, P.; Botvinick, M.; Vinyals, O.; Li, Y.; Pascanu, R. Relational Inductive Biases, Deep Learning, and Graph Networks. arXiv October 17, 2018.
Starting from scanning electron microscopy (SEM) images, we reconstructed three-dimensional synthetic microstructures that replicate the statistical grain distributions observed in two-dimensional images. A Bayesian optimizer was used to minimize the difference between the power spectra of SEM images—computed using the materials platform WAVEBASE [1]—and those from slices through synthetically generated grain structures.To efficiently generate training data for machine learning models, we developed a reduced-order model for magnetization reversal, enabling simulations of large-grained magnets beyond the computational limits of traditional micromagnetic simulations [2]. The synthetically generated microstructures serve as input for this model, allowing for realistic simulations of magnetization processes. With this approach, we analyzed domain evolution during magnetization reversal in large-grained magnets.
Both synthetic microstructure generation and the reduced-order model were used to build a fast model for estimating demagnetization curves. We represented synthetic granular structures as graphs and trained a graph neural network [3] with demagnetization curves computed with the reduced-order model. The trained network accurately predicted demagnetization curves, matching results from full micromagnetic simulations and the reduced-order model (see Figure 1). The graph neural networks offer a way to represent and analyze complex, three-dimensional polycrystalline structures.
Finally, microstructural parameters derived from the hysteresis models discussed above were used in a genetic optimization algorithm that optimizes the chemical composition for minimum price and maximum coercivity. The financial support by the Austrian Federal Ministry for Labour and Economy and the National Foundation for Research, Technology and Development and the Christian Doppler Research Association is gratefully acknowledged.
[1] Accessed on 31 Jan 2025. Available at: https://www.toyota.co.jp/wavebase/
[2] Kovacs, A.; Fischbacher, J.; Oezelt, H.; Kornell, A.; Ali, Q.; Gusenbauer, M.; Yano, M.; Sakuma, N.; Kinoshita, A.; Shoji, T.; Kato, A.; Hong, Y.; Grenier, S.; Devillers, T.; Dempsey, N. M.; Fukushima, T.; Akai, H.; Kawashima, N.; Miyake, T.; Schrefl, T. Physics-Informed Machine Learning Combining Experiment and Simulation for the Design of Neodymium-Iron-Boron Permanent Magnets with Reduced Critical-Elements Content. Front. Mater., 9, 1094055 (2023)
[3] Battaglia, P. W.; Hamrick, J. B.; Bapst, V.; Sanchez-Gonzalez, A.; Zambaldi, V.; Malinowski, M.; Tacchetti, A.; Raposo, D.; Santoro, A.; Faulkner, R.; Gulcehre, C.; Song, F.; Ballard, A.; Gilmer, J.; Dahl, G.; Vaswani, A.; Allen, K.; Nash, C.; Langston, V.; Dyer, C.; Heess, N.; Wierstra, D.; Kohli, P.; Botvinick, M.; Vinyals, O.; Li, Y.; Pascanu, R. Relational Inductive Biases, Deep Learning, and Graph Networks. arXiv October 17, 2018.
