Presentation Information
[P2-7]Low-Cobalt/Gallium High-Performance Nd-Fe-B Permanent Magnets Discovered by Machine-Learning Based Modeling
*Zhaozhe Zhong1, Jian Liu1, Huamin Tang2, Shiwei Kang2, Haopeng Jiang3, Pengfei Wang3, Cong Wang3 (1. Toyota Motor Technical Research and Service (Shanghai) Co., Ltd. (China), 2. Intelligent ElectroMobility R&D Center by TOYOTA (China) Co., Ltd. (China), 3. Yantai Zhenghai Magnetic Material Co., Ltd. (China))
Keywords:
Machine learning,Sintered Nd-Fe-B permanent magnet,Remanence,Coercivity,Cobalt,Gallium
Adjusting the compositions of Nd-Fe-B permanent magnets (PMs) is a primary strategy for developing new materials that enhance performance and/or reduce costs. The commonly used trial-and-error method, which is time-consuming and resource-intense, can only explore a very limited territory in the vastly possible compositional space. The challenge is further compounded by the intricate interactions of multiple elements, which significantly hinder the discovery of new materials with desired properties. Recently we applied Bayesian optimization based active learning loop to guide the search for new PMs and successfully identified a new composition with competitive performance but substantially lower material cost [1]. However, since the study was based on data from small-scale lab samples, the feasibility of applying these findings to mass production is uncertain, given the differences in fabrication equipment and processes. Herein we tackle this hurdle by leveraging a unique, industrially sourced dataset comprising 79 sintered Nd-Fe-B PM samples, consisting of 10 elements with nominal compositions of RE (29~34wt%)-Fe (balance)-M(0.5~4wt%)-B(0.88~1wt%), where RE represents neodymium and praseodymium, and M for cobalt (Co), copper, gallium (Ga), aluminum, zirconium, and titanium. All the samples were prepared following standardized manufacturing protocols with batch sizes larger than 600 kg.
Several machine learning algorithms, including Random Forest, XGBoost, Support Vector Regression, Ridge Regression, and Gaussian Process Regression (GPR), were evaluated to correlate elemental compositions with magnetic properties, specifically remanence (Br) and coercivity (Hcj). After systematic optimization of hyperparameters, GPR emerged as the most accurate model in the testing dataset with R2 = 0.92 for Br and 0.84 for Hcj, as shown in Fig. (a) and (b), respectively.
The GPR model was then utilized in an extensive grid search across a much larger compositional space than the initial dataset. By varying the elemental ratios at a step size of 0.05 wt% for RE and 0.01 wt% for other elements , the magnetic properties of roughly 2.34x108 compositions were screened, of which two novel compositions with substantially reduced Co content (0.22 & 0.30 wt%) and one with ultralow Ga content (0.10 wt%) were identified as the candidates of interest, because 1) the predicted comprehensive performance of these samples were among the highest of all training and calculated datasets; 2) the predicted Co and Ga contents were significantly smaller than the lowest values of the training dataset; and 3) low Co and Ga compositions are highly preferred due to their relatively high cost and potential supply chain vulnerabilities. Furthermore, to interrogate the model’s predication accuracy, we made up another composition, by taking a sample from the training dataset (RE31.83Fe64.9B0.92Co0.47Ga0.78Al0.04Cu0.78Zr0.28, which has the highest Hcj value in the 79 training samples) and purposely lowering its Ga content from 0.78 down to 0.1 wt%. The above four compositions were subsequently used for the fabrication of new PMs via the same industrial-scale manufacturing protocols. The experimentally measured Br and Hcj values of the resulting samples, to our surprise, were in remarkably good agreement with the model predictions as highlighted in Fig. (c), with errors only 5% or less.
This high level of prediction accuracy demonstrates that machine learning models, when trained on high-quality datasets, can serve as a powerful tool for predicting magnetic properties from compositional data, thereby accelerating the discovery of novel magnetic materials. The low Co and Ga contents of the recommended compositions outside the element ratio ranges of the training dataset also showcases their capability for extrapolation, which is of great value for guiding researcher into unexplored compositional spaces.
1. Chen, J., et al., Accelerated discovery of cost-effective Nd–Fe–B magnets through adaptive learning. Journal of Materials Chemistry A, 2023. 11(16): p. 8988-9001.
Several machine learning algorithms, including Random Forest, XGBoost, Support Vector Regression, Ridge Regression, and Gaussian Process Regression (GPR), were evaluated to correlate elemental compositions with magnetic properties, specifically remanence (Br) and coercivity (Hcj). After systematic optimization of hyperparameters, GPR emerged as the most accurate model in the testing dataset with R2 = 0.92 for Br and 0.84 for Hcj, as shown in Fig. (a) and (b), respectively.
The GPR model was then utilized in an extensive grid search across a much larger compositional space than the initial dataset. By varying the elemental ratios at a step size of 0.05 wt% for RE and 0.01 wt% for other elements , the magnetic properties of roughly 2.34x108 compositions were screened, of which two novel compositions with substantially reduced Co content (0.22 & 0.30 wt%) and one with ultralow Ga content (0.10 wt%) were identified as the candidates of interest, because 1) the predicted comprehensive performance of these samples were among the highest of all training and calculated datasets; 2) the predicted Co and Ga contents were significantly smaller than the lowest values of the training dataset; and 3) low Co and Ga compositions are highly preferred due to their relatively high cost and potential supply chain vulnerabilities. Furthermore, to interrogate the model’s predication accuracy, we made up another composition, by taking a sample from the training dataset (RE31.83Fe64.9B0.92Co0.47Ga0.78Al0.04Cu0.78Zr0.28, which has the highest Hcj value in the 79 training samples) and purposely lowering its Ga content from 0.78 down to 0.1 wt%. The above four compositions were subsequently used for the fabrication of new PMs via the same industrial-scale manufacturing protocols. The experimentally measured Br and Hcj values of the resulting samples, to our surprise, were in remarkably good agreement with the model predictions as highlighted in Fig. (c), with errors only 5% or less.
This high level of prediction accuracy demonstrates that machine learning models, when trained on high-quality datasets, can serve as a powerful tool for predicting magnetic properties from compositional data, thereby accelerating the discovery of novel magnetic materials. The low Co and Ga contents of the recommended compositions outside the element ratio ranges of the training dataset also showcases their capability for extrapolation, which is of great value for guiding researcher into unexplored compositional spaces.
1. Chen, J., et al., Accelerated discovery of cost-effective Nd–Fe–B magnets through adaptive learning. Journal of Materials Chemistry A, 2023. 11(16): p. 8988-9001.
