Presentation Information
[P2-3]Multi-objective optimization of magnet compositions by machine learning
*Hyuga Hosoi1, Yamano Hayate1, Kinoshita Akihito1, Sakuma Noritsugu1, Thomas Schrefl2, Umetani Yusuke1, Shoji Tetsuya1 (1. Toyota Motor Corporation (Japan), 2. Donau University Krems (Austria))
Keywords:
NdFeB,Machine Learning,Genetic Algorithm utilize magnets that maintain equivalent performance to,Multi–objective optimization,WAVEBASE
With the rapid expansion of vehicle electrification, the demand for electric motors is increasing. In the future, there is a concern about the supply risk of magnetic materials used in motors. For example, not only heavy rare earth elements such as Dy and Tb added to high–performance Nd–Fe–B magnet but also rare earth elements such as Nd and Pr could face imbalances in supply and demand. To mitigate this risk, it is necessary to develop and utilize magnets that maintain equivalent performance to conventional magnets while reducing the amount of expensive rare earth elements. Since the required magnetic properties vary depending on the motor development changes and products, it is essential to optimize the magnet composition accordingly. In this research, one of the machine learning techniques was utilized to quickly propose the optimal magnet composition. Over 170 kinds of (Nd,Ce,La,Pr,Dy,Tb)13.55–(Fe,Co,Ni)80.54–(B,C)5.91(at%) alloys were prepared by arc melting. These alloys were annealed at 1373 K for 24 h in Ar atmosphere. Annealed alloys were pulverized and sorted into particles with diameters of under 20 um in an inert atmosphere to make magnetically anisotropic powder. The physical properties of the produced powders, namely saturation magnetization (Ms) and anisotropic magnetic field (Ha), were evaluated at temperatures ranging from 300 to 453 K. Regression models were created with the magnet composition and evaluation temperatures as explanatory variables, and Ms and Ha as objective variables. As a result, good enough regression models were created, achieving an R2 score (a value for evaluating prediction accuracy) greater than 0.9. Additionally, it became possible to predict the values of saturation magnetization and anisotropic magnetic field for any of these magnetic compositions at any temperature between 300 K and 453 K. Next, the magnet compositions that maximizes Ms and Ha while minimizing magnet costs were examined. The created regression models were used with a genetic algorithm to efficiently search for the optimal magnet compositions(1). Alternative magnet compositions, which are more cost-effective yet comparable in magnetic properties to the high–performance magnets containing a small amount of expensive Tb, were found. Experiments were conducted to create these magnetic powders and evaluate them, confirming that the 2–14–1 crystal structure was formed as expected and that the predicted Ms and Ha were achieved. By utilizing machine learning, it is possible to quickly suggest a variety of magnet compositions suited to different applications. In the presentation, a case study of data analysis using Toyota's material development cloud platform, 'WAVEBASE,' will also be introduced(2). By using WAVEBASE, it is easy to carry out analyses ranging from regressions to multi-objective optimization and so on. (1) H. Yamano, et al: Efficient optimization approach for designing power device structure using machine learning, Japanese Journal of Applied Physics, 62, SC1050 (2023).(2) M. Yano, et al: Material data analysis cloud service “WAVEBASE”, TOYOTA Technical Review, Vol.69. (2023), 48-61.
