An accurate quantitative method for NdFeB magnetism based on laser-induced breakdown spectroscopy

Abstract

NdFeB magnetic materials are widely used in daily life, such as in permanent magnet motors, loudspeakers and computer disks. The NdFeB magnetic material has excellent magnetic properties, and its magnetic properties are also the key to judge the production quality of NdFeB. Therefore, the precise quantification of the magnetic properties of NdFeB magnetic materials is crucial. Laser induced breakdown spectroscopy (LIBS) is a technique to obtain the spectrum of chemical elements by excitation of plasma on the surface of a sample with a high energy laser. In this paper, a precise classification and magnetic quantification method for NdFeB magnetic materials based on laser-induced breakdown spectroscopy is designed, which is different from the traditional direct magnetic property detection method and uses element detection to quantitatively analyze the magnetic properties indirectly. A laser-induced breakdown spectroscopy system was used to collect the characteristic spectrum of NdFeB magnetic materials, and the sliding window minimum removal base method was independently designed to further optimize the detection accuracy. A classification model and quantitative analysis method model were further established and optimized. The random forest method was used to preliminarily classify NdFeB magnetic materials, and the GA-ELM method was used to conduct quantitative analysis of magnetic properties. Quantitative magnetic properties include Br, Hcj, Hcb and (BH)max. The error analysis of the final quantitative analysis is as follows: RMSE of Br reaches 0.0001526, RMSE of Hcj reaches 0.0001937, RMSE of Hcb reaches 0.00197, and RMSE of (BH)max reaches 0.00785. It is verified that the magnetic quantification method for NdFeB magnetic materials based on laser-induced breakdown spectroscopy can effectively conduct accurate quantitative analysis of the magnetic properties of NdFeB magnetic materials and provide a fast, convenient, accurate and economical detection method for the quality control of magnetic materials workshops.

---

1 Introduction

As the most representative magnetic material, NdFeB magnets are the most commonly used rare earth magnets. The magnetic properties of NdFeB magnets are the basis for judging their quality. The production process of NdFeB magnets involves batching, smelting, milling, profiling, sintering and tempering, magnetic measurement and processing of finished products, in which ingredients are the basis and sintering and tempering is the key. After sintering and tempering, a rough embryo of the NdFeB magnet will be generated, and once the rough embryo is formed, its magnetic properties will not change, and the rough embryo will only be further morphologically processed according to the product requirements. Therefore, the quality control of the rough embryo is particularly important for the evaluation of the magnetic properties of the NdFeB magnet.^1–7

According to the quality monitoring requirements of coarse embryos, we urgently need a method that can identify the quality of the coarse embryo quickly and accurately. Traditional crude embryo detection methods, such as using a vibrating sample magnetometer or superconducting quantum magnetometer, use the detection of its magnetic properties to detect its production quality. This detection requires the NdFeB crude embryo in a cylindrical sample with a fixed thickness and a fixed diameter, if the shape varies, additional topography processing is needed. In addition, the demagnetization measurement process of magnetic materials will make the material brittle, easily damaging the material. Generally, by using the same batch of samples to test on behalf of this batch of overall evaluation, it is impossible to detect the magnetic properties of each magnetic material.^8–11

In response to this issue, this article uses laser-induced breakdown spectroscopy (LIBS) technology, which has advantages such as no need for complex sample pretreatment, simultaneous analysis of multiple elements, low sample loss, and fast detection speed. It can be applied to the quality monitoring and detection requirements of neodymium iron boron rough embryos. LIBS technology uses a high-energy pulsed laser to focus on the surface of a sample, excite it, and generate a high brightness and high heat plasma. The elemental content of the sample is inferred from the spectral information of the plasma. We found that the magnetic properties of neodymium iron boron magnets are highly correlated with their elemental ratios,^12–14 while laser-induced breakdown spectra are highly correlated with the elemental ratios of the sample. We utilized the powerful predictive ability of artificial intelligence machine learning methods to establish a correlation between the magnetic properties of neodymium iron boron magnets and laser-induced breakdown spectra. Quality monitoring of neodymium iron boron rough blanks is performed to achieve precise quantification of magnetic properties for all magnetic materials.^15,16

In this paper, according to the uneven surface of the NdFeB coarse embryo material, a long focal optical path is designed to ensure the stability of the laser-induced breakdown spectrum when the defocusing range is ±5 mm. A self-designed minimum sliding window preprocessing method was used to remove background noise. An appropriate stoichiometric method was selected for the preliminary classification of magnetic materials, and the method was optimized to improve the classification accuracy. Finally, according to the types of magnetic materials, an appropriate stoichiometric method was selected to accurately analyze the four magnetic parameters.^17,18

2 Experimental

2.1. Sample preparation

Since there are many types of NdFeB magnetic materials, such as N35, N38, N40 and so on, each sample has different magnetic parameters, and the magnetic parameters of the same kind of sample fluctuate in a small range. In order to achieve the accurate quantitative analysis of the magnetic properties of NdFeB samples, it is necessary to first classify the samples. Then the magnetic properties of each type of sample were analyzed accurately and quantitatively. Therefore, we prepared two batches of NdFeB magnetic materials, one batch for sample classification and the other batch for precise quantification of magnetic properties.

2.1.1. Preparation of classified samples. In order to accurately quantify the magnetic properties of magnetic materials, the premise is that the sample types can be accurately separated. We selected five samples, namely N35, N38, N40, N35H and N35M. According to the standard sample information provided by the manufacturer, their magnetic properties are different as shown in Table 1.

Table 1 Magnetic properties of samples of different grades

Brand	Br	Hcj	Hcb	(BH)max
N35	11.7–12.2	12	10.9	33–36
N38	12.2–12.5	12	11.3	36–39
N40	12.5–12.8	12	11.6	38–41
N35M	11.7–12.2	14	10.9	33–36
N35H	11.7–12.2	17	10.9	33–36

---

Four magnetic properties are listed in the table, namely remanence (Br), intrinsic coercivity (Hcj), magnetic inductance coercivity (Hcb) and maximum magnetic energy product ((BH)max). It can be seen that their 4 parameters are different, and the 4 magnetic parameters of the sample of the same grade are also different, and it is necessary to conduct accurate quantitative analysis of magnetic parameters after sample grade classification.

2.1.2. Sample preparation for magnetic quantitative analysis. In order to conduct accurate quantitative analysis of magnetic properties after grade classification, for N40 grade samples, 20 samples with magnetic property gradients were selected for production testing and marked with numbers 1–20. Specific magnetic property parameters of the samples are shown in Table 2.

Table 2 Magnetic properties of samples of the same grade

Label	Br	Hcj	Hcb	(BH)max
1	12.803	10.882	11.359	38.047
2	12.841	10.842	11.314	38.384
3	12.815	10.799	11.245	38.246
4	12.822	10.535	11.386	38.263
5	12.878	10.761	11.317	38.424
6	12.813	10.965	11.436	38.171
7	12.784	10.528	11.121	38.878
8	12.621	10.757	11.307	37.135
9	12.615	10.711	11.254	37.251
10	12.862	10.492	11.204	37.605
11	12.842	10.926	11.401	38.233
12	12.72	10.706	11.196	37.51
13	12.807	10.762	11.296	38.053
14	12.804	10.62	11.026	38.171
15	13.052	10.641	11.128	39.531
16	12.831	10.993	11.444	38.151
17	12.802	10.618	11.324	37.997
18	12.936	10.904	11.345	38.318
19	12.813	10.908	11.335	38.058
20	12.73	10.781	11.386	37.216

---

For accurate quantitative analysis of magnetic properties, we divided the 20 samples into two parts: a training set and test set. In order to ensure the quality of the analysis model, samples with the largest 2 values, the smallest 2 values and the middle 2 values of the 4 kinds of magnetic property parameters were selected as the training set data, as shown in Table 3.

Table 3 Sample quantitative data grouping

Br		Hcj		Hcb		(BH)max
12.615	9	10.492	10	11.026	14	37.135	8
12.621	8	10.528	7	11.121	7	37.216	20
12.72	12	10.535	4	11.128	15	37.251	9
12.73	20	10.618	17	11.196	12	37.51	12
12.784	7	10.641	14	11.204	10	37.605	10
12.802	17	10.62	15	11.245	3	37.878	7
12.803	1	10.706	12	11.254	9	37.997	17
12.804	14	10.711	9	11.296	13	38.047	1
12.807	13	10.757	8	11.307	8	38.053	13
12.813	6	10.761	5	11.314	2	38.058	19
12.813	19	10.762	13	11.317	5	38.151	16
12.815	3	10.781	20	11.324	17	38.171	6
12.822	4	10.799	3	11.335	19	38.171	14
12.831	16	10.842	2	11.345	18	38.233	11
12.836	18	10.882	1	11.359	1	38.246	3
12.841	2	10.904	18	11.386	4	38.263	4
12.842	11	10.908	19	11.386	20	38.318	18
12.862	10	10.926	11	11.410	11	38.384	2
12.878	5	10.965	6	11.436	6	38.424	5
13.052	15	10.993	16	11.444	16	39.531	15

---

Finally, samples 2, 5, 6, 7, 8, 9, 10, 13, 14, 15, 16, 19 and 20 were selected as training set data and samples 1, 3, 4, 11, 12, 17 and 18 were selected as test set data.

2.2. Laser induced breakdown path

2.2.1. Design and verification of the long focal coaxial receiving optical path. Due to the uneven appearance of the NdFeB crude embryo, the focusing error is about ±3 mm, and it is easy to defocus the optical path detected by conventional LIBS, resulting in distortion of spectral information acquisition. If the automatic focusing method is applied, the height of the sample is adjusted by means of distance measurement, image sharpness measurement, etc., and the purpose of focusing is achieved using the translation table. Because quality monitoring has certain efficiency requirements, automatic focusing mode will greatly reduce the detection efficiency.^19,20

In this paper, a long focal coaxial optical path is designed, and the acquisition depth of field is increased by using a long focal lens to compensate for the spectral information acquisition distortion of samples in a small range of defocus. Since the optical path is a long focal coaxial receiving optical path, the overall optical path is the same as the camera imaging system, and the defocus degree is the same as the camera depth of field. We use the camera depth of field calculation method to calculate the overall optical path defocusing acceptance, and the camera optical path is shown in Fig. 1.

Fig. 1 Camera optical path.

The depth of field is calculated as follows:

where ΔL is the depth of field, f is the lens focal length, F is the lens aperture value, L is the shooting distance, and δ is the diameter of the confusion circle. According to the depth of field calculation formula, the hardware parameters of our optical path are respectively brought in. The lens focal length selected is 50 mm; the lens aperture value is calculated by dividing the lens focal length by the lens aperture, which is 3.9; the shooting distance is 150 mm; the diameter of the confusion circle indicates that the optical fiber numerical aperture is 0.22 mm. The system depth of field ΔL = 15.66 mm is calculated by bringing the parameters into the calculation, which is far enough to satisfy the sample focusing error of ±3 mm.

Next, the function of the telephoto coaxial receiving optical path is verified by means of spectrum acquisition. First, a telephoto coaxial optical path system is built.

The telephoto coaxial receiving optical path is constructed with a high-precision cage system, which can simultaneously ensure the stability of the experiment and the convenience of the optical path adjustment. A laser focusing mirror with a focal length of 100 mm, a fiber focusing mirror with a focal length of 50 mm and a short-wave dichroic mirror with a cut-off wavelength of 805 nm were selected to complete the functions of laser focusing, excitation sample and plasma spectrum coaxial acquisition.

In order to reduce the error caused by the uneven surface of the sample itself, we selected an aluminum alloy sample with a smooth and dense surface for laser-induced breakdown spectrum acquisition. 10 points were selected for each defocusing distance, and each point was collected 3 times for an average to represent the spectrum of this point, and the spectral intensity fluctuation of the characteristic wavelength of the sample was verified within the range of the focal length of the system ±3 mm (Fig. 2).

Fig. 2 Characteristic peak intensity in the defocusing experiment.

From the figure, we can see that within the defocusing range of ±3 mm, the fluctuation of characteristic peak intensity at different defocusing distances is small, even smaller than the spectral intensity fluctuation at different points of the same defocusing distance. It can be verified that the long focal coaxial receiving optical path designed by us can eliminate the spectral intensity fluctuation caused by the defocusing caused by the uneven surface of the sample.

2.2.2. Measurement and control system design and experimental platform construction. Next, the design of the measurement and control system was carried out, in which the laser was produced by Changchun New Industry Optoelectronic Technology Co., Ltd, with parameters of 1064 nm center wavelength, 6–8 ns laser pulse width, laser energy continuously adjustable from 0–150 mJ, a 1–20 Hz repetition rate, and a frequency accuracy of 0.01 Hz. After experimental testing, this laser parameter can effectively excite rare earth magnet samples, meeting the detection requirements. Referring to the NIST database, the characteristic spectral line wavelengths of rare earth magnet theme elements were searched to determine the coverage wavelength range of the spectrometer. Finally, the spectrometer selected was a three channel spectrometer from the AVANTES manufacturer. In the experiment, the three channels covered 197–952 nm, with a spectral resolution of 0.05 nm and an integration time of 2 ms. The 3D translation stage control system consists of a high-precision 3D translation stage and an electrically controlled 3D translation stage control box, mainly used to control the coordinate transformation of the laser focusing position on the sample surface, control the excitation position of the sample, and assist in focusing and spectral needle scanning. The high-precision 3D translation stage adopts precision grade ball screws, matched with linear slider guide rails, high-quality couplings and other accessories to ensure suitability for high-precision control, with a stroke range of 100 mm and a resolution of 20 μm. The control box of the electronic control three-dimensional translation stage can finely control the three-dimensional motion direction, motion speed, and motion acceleration of the three-dimensional translation stage. The delay pulse generator independently developed by the laboratory is used for overall timing control of various functional components. It can accurately control the timing of each functional component by receiving and sending precise timing pulse signals. The delay accuracy reaches 10 ns. Hardware support must be provided for the optimization and analysis of spectral acquisition delay parameters in the future. An experimental platform was built using these functional components, as shown in Fig. 3.

Fig. 3 Experimental platform for quantitative analysis of magnetic properties of laser-induced breakdown spectroscopy.

3 Results and discussion

3.1. Spectral analysis

We conducted laser-induced breakdown spectroscopy measurements based on rare earth magnet LIBS measuring instruments and neodymium iron boron samples. Due to the poor reproducibility of LIBS technology, multiple data points need to be averaged to eliminate the impact of the technology itself. Due to the small area detection of LIBS technology, in order to prevent uneven distribution of sample elements, multiple points need to be taken for detection. Our measurement collection method uses an average of 3 excitations per excitation point, with 10 excitations per sample, for a total of 20 sets of samples. Ultimately, 10 spectra can be obtained for each sample, for a total of 200 spectra.

In order to avoid the interference of bremsstrahlung radiation generated in the early stage of plasma generation during data collection, due to the fact that the attenuation rate of bremsstrahlung radiation is much higher than that of spectral signals, we adopt a delay measurement method. We delay until the bremsstrahlung radiation is completely attenuated before collecting the signal, which can obtain a signal with a high signal-to-noise ratio.

The final collected spectrum is shown in Fig. 4. From the spectrum, it can be seen that the elements contained in the neodymium iron boron sample have been excited, and corresponding characteristic curves can be found in the spectrum, such as Nd: 383.89 nm, Fe: 373.48 nm, and B: 345.13 nm.

Fig. 4 Characteristic spectral lines in NdFeB spectra.

From the spectrogram, it can be seen that there are still some issues such as baseline drift and background noise in the graph. It is necessary to select an appropriate analysis method and perform spectral preprocessing to further improve the final analysis accuracy.

3.2. Classification method selection

Before conducting chemometric analysis, it is necessary to choose a suitable chemometric method for neodymium iron boron samples. We have selected three machine learning methods for comparative analysis, namely partial least squares (PLS), random forest (RF), and extreme learning machine (ELM).^21–26 Using these three algorithms to classify the five neodymium iron boron samples, the results are shown in Fig. 5 represented by the confusion matrix.^21,22 The classification results using three algorithms that have not been further optimized are as follows: PLS is 87.5%, RF is 91.0%, and ELM is 89.5%.

Fig. 5 Classification results of PLS, RF, and ELM.

From the comparison of the classification results, it can be seen that RF classification accuracy is higher than those of the other three algorithms and the classification error data are more concentrated, making it easier to improve classification accuracy through algorithm optimization. Therefore, we choose the RF algorithm for the next optimization analysis.

3.3. Spectral pre-treatment research

We independently designed a Siding Window Minimum (SWM) to base method. First, we selected an appropriate sliding window size, took the minimum value of the data within the window as the window value, slid the window to obtain the window values of all wavelengths, and subtracted the corresponding window value from the data within the window to obtain denoised spectral data. The comparison of spectral data before and after processing is shown in Fig. 6. We compared the denoising effect by adjusting the window size to 5–50 data points. When the data length is small, although the denoising effect is good, the spectral intensity of the characteristic spectral lines is severely attenuated. When the data length is long, although the attenuation of feature Puxian intensity is small, the denoising effect is average. The preprocessed spectrum has removed a small amount of baseline drift and background noise issues.

Fig. 6 Pre-processing optimization results for sliding window lengths of 5, 10, 20 and 50 data points.

As shown in Table 4, we validated the classification results using the RF algorithm. When the sliding window lengths are 5 data points, 10 data points, 20 data points, and 50 data points, the classification accuracies are 0.920, 0.955, 0.970, and 0.945, respectively. Compared to the classification accuracy of 0.910 without preprocessing, there is a certain improvement in both. Finally, we chose a sliding window length of 20 data points for the next step of chemometric analysis.

Table 4 Correlation between sliding window size and classification accuracy

Sliding window size	Classification accuracy
5 data	0.920
10 data	0.955
20 data	0.970
50 data	0.945

---

Next, the SWM method was compared with common baseline correction methods, BEADS method and AIRPLS method. The classification accuracy pairs are shown in Table 5. As can be seen from the data in the table, the classification accuracy of BEADS is the highest at 0.975 and the classification accuracy of SWM and AIRPLS is 0.97. There is little difference between the three classifications.

Table 5 Comparison of classification accuracy of SWM, BEADS and AIRPLS

Pretreatment method	Classification accuracy
SWM	0.970
BEADS	0.975
AIRPLS	0.970

---

The running time of the three methods is compared as shown in Table 6. SWM run time is 4.487 s, BEADS run time is 57.534 s, and AIRPLS run time is 12.053 s.

Table 6 Comparison of classification accuracy of SWM, BEADS and AIRPLS

Pretreatment method	Running time (s)
SWM	4.487
BEADS	57.534
AIRPLS	12.053

---

Finally, the SWM method is selected after comprehensive consideration.

3.4. Optimization of classification methods

Next, we will optimize the main parameters of the algorithm. In the process of random forest regression modeling, two important parameters are involved: the number of randomly selected attributes (m_try) and the number of decision trees (n_tree). Optimizing the ntree value can improve the stability of the RF calibration model, and optimizing the mtry value can effectively improve the prediction accuracy of the calibration model. At the same time, these two parameters are closely related to the efficiency of the model. It is necessary to adjust the m_try while minimizing the size of the n_tree, as the computer's runtime increases geometrically with the increase of the n_tree.

As shown in Fig. 7, M is the number of data samples, and the final optimal parameters selected are n_tree = 50 and m_try = M/8, with a classification accuracy of 1. The classification results are shown in Fig. 8. Next, we will optimize the main parameters of the algorithm.

Fig. 7 Random forest parameter optimization.

Fig. 8 Error variation of Br, Hcb, Hcj, and (BH)max with genetic algebra.

In the process of random forest regression modeling, two important parameters are involved: the number of randomly selected attributes (m_try) and the number of decision trees (n_tree). Optimizing the ntree value can improve the stability of the RF calibration model, and optimizing the mtry value can effectively improve the prediction accuracy of the calibration model. At the same time, these two parameters are closely related to the efficiency of the model. It is necessary to adjust the m_try while minimizing the size of the n_tree, as the computer's runtime increases geometrically with the increase of the n_tree. As shown in Fig. 7, M is the number of data samples, and the final optimal parameters selected are n_tree = 50 and m_try = M/8, with a classification accuracy of 1.

3.5. Quantitative method selection

Next, we will conduct further quantitative analysis of the magnetic properties of the N40 sample, which includes four parameters: Br, Hcj, Hcb, and (BH)max.

First, it is necessary to select a chemometric method suitable for neodymium iron boron samples. We still chose partial least squares (PLS), random forest (RF), and extreme learning machine (ELM).^23–28 Using these three algorithms for quantitative analysis of the five neodymium iron boron samples, the RMSE results were used as the basis for evaluating the quality of the quantitative results. The RMSE results obtained using the three algorithms that have not been further optimized are shown in Table 5.^29,30

From the results, it can be seen that PLS has much lower quantitative analysis ability than RF and ELM. Overall, ELM has better quantitative analysis ability than RF. Therefore, we choose the ELM algorithm for further optimization analysis (Table 7).

Table 7 Quantitative analysis of RMSE errors for four magnetic parameters, Br, Hcj, Hcb, and (BH)max, using PLS, RF, and ELM

Magnetic parameter	PLS_RMSE	RF_RMSE	ELM_RMSE
Br	0.9693	0.1846	0.1098
Hcj	0.8094	0.1834	0.1180
Hcb	0.7667	0.1725	0.1846
(BH)max	2.7756	0.5389	0.5076

---

3.6. Optimization of quantitative methods

In order to further optimize the accuracy of quantitative analysis, the ELM algorithm establishes a model by randomly setting weights and biases. When determining the optimal model solely by adjusting the number of hidden layer neurons, it may lead to randomness in the results. Therefore, parameter optimization algorithms can be introduced to obtain the optimal model and further improve prediction accuracy. We introduce genetic algorithms (GAs) to determine the weights and biases of the ELM model. The RMSE of GAs-ELM (Genetic Algorithms-Extreme Learning Machine) changes with the number of iterations as shown in Fig. 8 and Table 8.^31–33

Table 8 Quantitative analysis of RMSE errors for four magnetic parameters, Br, Hcj, Hcb, and (BH)max, using ELM and GA-ELM

Magnetic parameter	ELM_RMSE	GA-ELM_RMSE
Br	0.1098	0.0001526
Hcj	0.1180	0.0001937
Hcb	0.1846	0.00197
(BH)max	0.5076	0.00785

---

It can be seen that the quantitative analysis errors of the four magnetic property parameters decrease with the increase of genetic generations and eventually tend to remain unchanged. Br reaches its optimal value at genetic generation 39, Hcb reaches its optimal value at genetic generation 63, Hcj reaches its optimal value at genetic generation 70, and (BH)max reaches its optimal value at genetic generation 48. Finally, we chose genetic generation 70. The full hyperparameters of GA are a population size of 20, a crossover probability of 0.7, a mutation probability of 0.01, an ‘elite selection’ selection strategy, and genetic algebra 70.

The final quantitative analysis results are shown in Fig. 9 and Table 6. The RMSE of Br reached 0.0001526, the RMSE of Hcj reached 0.0001937, the RMSE of Hcb reached 0.00197, and the RMSE of (BH)max reached 0.00785. The magnetic quantification method of neodymium iron boron magnetic materials based on laser-induced breakdown spectroscopy has been validated to effectively perform accurate magnetic property quantification analysis of neodymium iron boron magnetic materials, providing a fast, convenient, accurate, and economical detection method for quality monitoring in magnetic material workshops.

Fig. 9 Error variation of Br, Hcb, Hcj, and (BH)max with genetic algebra (A): comparison of the Br quantitative results, (B): comparison of the Hcj quantitative results, (C): comparison of the Hcb quantitative results, (D): comparison of the (BH)max quantitative results.

4 Conclusions

We designed a precise classification and magnetic quantification method for neodymium iron boron magnetic materials based on laser-induced breakdown spectroscopy. We collected laser-induced breakdown spectra of different types of neodymium iron boron magnetic materials with magnetic property gradients and classified and quantitatively analyzed them separately. After selecting the classification method and optimizing the independently designed spectrum and processing methods and classification methods, the final classification accuracy reaches 100%. Further selection of N40 samples with a magnetic gradient was conducted for quantitative analysis of magnetic properties.

Through quantitative method selection and optimization of ELM algorithm weights and biases using the genetic algorithm, the final quantitative accuracy of the RMSE of Br reached 0.0001526, the RMSE of Hcj reached 0.0001937, the RMSE of Hcb reached 0.00197, and the RMSE of (BH)max reached 0.00785. The magnetic quantification method of neodymium iron boron magnetic materials based on laser-induced breakdown spectroscopy has been validated to effectively perform accurate magnetic property quantification analysis of neodymium iron boron magnetic materials, providing a fast, convenient, accurate, and economical detection method for quality monitoring in magnetic material workshops.

Data availability

Data underlying the results presented in this paper are not publicly available at this time but maybe obtained from the authors upon reasonable request.

Author contributions

Guanyu Chen: methodology, data curation, writing – original draft. Jing Chen: writing – original draft, writing – review & editing. Dongming Qu: resources, formal analysis. Yuting Fu: validation, investigation. Guang Yang: conceptualization, funding acquisition, project administration. Huihui Sun: data curation, validation.

Conflicts of interest

The authors declare no competing financial interest.

Acknowledgements

This research was gratefully supported by the National Nature Science Foundation of China (Grant No. 62275099) and the National Key Research and Development Program of China (Grant No. 2023YFF0714103). They also appreciate the support from the Beijing Triumph Technology Co., Ltd for providing all the devices and materials in this work.

Notes and references

1. S. Y. Zhou, L. Zhou, S. K. Chen and J. J. Luo, Rare Met. Mater. Eng., 2006, 35, 1006–1008.
2. H. J. Zheng, J. G. Huang and C. A. Ma, 14^th Congress of International Federation for Heat Treatment and Surface Engineering, 2004, vol. 1, pp. 69–72.
3. Y. Liu and M. J. Tu, J. Rare Earths, 1999, 17, 119–121.
4. K. S. Yu, C. W. Cheng, A. C. Lee, W. Y. J. Jian, W. C. Chang, T. W. Chang and M. C. Tsai, Opt. Laser Technol., 2021, 146, 1–8.
5. J. Chen, H. Yang, G. Xu, P. Zhang, J. Lv, W. Sun, B. Li, J. Huang, D. Wang, X. Shu and Y. Wu, J. Rare Earths, 2022, 40, 302–308.
6. H. Xie, J. Zhao and Y. Yu, J. Iron Steel Res. Int., 2006, 13, 324–330.
7. J. Tao, B. Liu, P. J. Zhang, G. Q. Xu, J. Lv, J. Huang, J. Yan, W. Sun, B. S. Li, D. M. Wang and Y. C. Wu, J. Rare Earths, 2023, 41, 1203–1210.
8. D. Koelle, J. Electroceram., 1999, 3, 195–212.
9. B. Buford, P. Dhagat and A. Jander, J. Appl. Phys., 2014, 115, 1–3.
10. A. L. Ribeiro, 8^th IEEE Africon Conference, 2007, 1–3, 665–668.
11. A. Vettoliere and C. Granata, Rev. Sci. Instrum., 2014, 85, 1–6.
12. X. Y. Liu, Y. P. Li and L. X. Hu, Rare Met. Mater. Eng., 2012, 41, 1875–1880.
13. H. Chen, X. Yang, L. Sun, P. Yu, X. Zhang and L. Luo, J. Magn. Magn. Mater., 2019, 485, 49–53.
14. Y. Zhang and X. J. Xu, J. Supercond. Novel Magn., 2021, 34, 2711–2715.
15. Y. Rao, W. X. Ren, W. H. Kong, L. W. Zeng, M. F. Wu, X. Wang, J. Wang, Q. W. Fan, Y. Pan, J. B. Yang and Y. X. Duan, J. Anal. At. Spectrom., 2024, 39, 925–934.
16. V. N. Lednev, P. A. Sdvizhenskii, R. D. Asyutin and A. V. Perestoronin, J. Anal. At. Spectrom., 2021, 36, 540–547.
17. M. Ghezelbash, S. M. R. Darbani, A. E. Majd and A. Ghasemi, J. Supercond. Nov. Magn., 2017, 30, 1893–1898.
18. M. Ghezelbash, A. E. Majd, S. M. R. Darbani and A. Ghasemi, Ceram. Int., 2017, 43, 8356–8363.
19. R. Nishi, Y. Moriyama, K. Yoshida, S. Isakozawa, H. Mogako, M. Ozawa and S. Isakozawa, Microscopy, 2013, 62, 515–519.
20. M. Rudnaya, R. Mattheij, J. Maubach and H. Termorsche, World Congr. Eng., 2011, 1, 301–306.
21. N. M. Froelich, S. M. Azcarate, H. C. Goicoechea and A. D. Campiglia, Appl. Spectrosc., 2024, 1–12.
22. K. P. Chi, J. R. Lin, M. Chen, J. J. Chen, Y. M. Chen and T. Pan, Spectrochim. Acta– A: Mol. Biomol. Spectrosc., 2023, 308, 1–10.
23. Y. Y. Chen, J. J. Xia, Y. Wei, W. K. Xu, X. R. Mao, Y. F. Chen, S. G. Min and Y. M. Xiong, Chin. J. Anal. Chem., 2023, 51, 454–462.
24. T. T. Chen, T. L. Zhang, H. S. Tang, X. M. Cheng and H. Li, Anal. Chem., 2023, 95, 4819–4827.
25. C. W. Zhang, W. R. Song, Z. Y. Hou and Z. Wang, J. Anal. At. Spectrom., 2023, 38, 2554–2561.
26. G. R. Fu, W. Hu, W. P. Xie, X. Y. Yao, J. Xu, P. Yang and M. Y. Yao, Anal. Methods, 2023, 15, 5867–5874.
27. G. P. Zeng, Commun. Stat.– Theory Methods, 2020, 49, 2080–2093.
28. D. Krstinic, L. Seric and L. Slapnicar, IEEE Access, 2023, 11, 40692–40697.
29. S. Y. Zhao, Y. C. Zhao, Z. Y. Hou and Z. Wang, Spectrochim. Acta B, 2023, 203, 1–7.
30. B. H. Xu, Z. J. Lin, Z. Y. Xu, X. T. Yan, X. Y. Peng, C. K. Li, Y. Z. Qin, J. M. Li, N. Zhao and Q. M. Zhang, J. Anal. At. Spectrom., 2023, 38, 1032–1042.
31. L. Huang, Y. F. Chen, J. B. Wang, Z. D. Cheng, L. Tao, H. M. Zhou, J. Xu, M. Y. Yao, M. H. Liu and T. B. Chen, Anal. Methods, 2023, 15, 738–745.
32. T. T. Chen, T. L. Zhang, H. S. Tang, X. M. Cheng and H. Li, Anal. Chem., 2023, 95, 4819–4827.
33. H. Y. Jin, X. J. Hao and B. M. Mo, J. Anal. At. Spectrom., 2023, 38, 2280–2290.

---