A machine learning-based method for predicting iron-based nanocrystalline alloys

Abstract

The application discloses a prediction method of an iron-based nanocrystalline alloy based on machine learning, an initial data set is constructed based on soft magnetic performance data of the iron-based nanocrystalline alloy, the initial data set is processed through a machine learning method, an XGBoost model is trained based on the processed data to obtain a magnetic induction intensity prediction model and a coercive force prediction model, the magnetic induction intensity prediction model and the coercive force prediction model are taken as objective functions, and alloy composition data with the highest magnetic induction intensity and the lowest coercive force are obtained through iteration of a genetic algorithm NSGA-II, so that a component with excellent performance can be quickly screened out.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of iron-based nanocrystalline soft magnetic alloys, and particularly relates to a prediction method for iron-based nanocrystalline alloys based on machine learning. BACKGROUND

[0002] Iron-based nanocrystalline soft magnetic alloys refer to a new type of soft magnetic material having excellent magnetic properties, which is obtained by precipitating nanoscale α-Fe grains in the amorphous matrix through appropriate heat treatment based on amorphous precursors.

[0003] Iron-based nanocrystalline alloys have high saturation magnetic induction (Bs), low coercivity (Hc), and low loss due to the coupling effect of amorphous and nanocrystalline phases. However, the currently industrialized nanocrystalline soft magnetic alloy is still the FINEMET system alloy developed in the 1980s, but the saturation magnetic induction of this alloy system is relatively low, only 1.24T. The high saturation magnetic induction of nanocrystalline alloys is difficult to be compatible with good amorphous forming ability, and the high saturation magnetic induction is difficult to be compatible with low loss / low coercivity, which is a difficult problem that needs to be solved in the industry.

[0004] Currently, the design and research of iron-based nanocrystalline soft magnetic alloys still mainly rely on traditional "trial and error type", which is to design alloy composition based on experience, and then to prepare samples and characterize / evaluate performance for each alloy composition by experimental means. The experimental period is long, the efficiency is low, the cost is high, and there is a certain randomness. Therefore, developing a high-precision and high-efficiency alloy composition design method has become the current research focus.

[0005] In recent years, machine learning has become a frontier and hot field in the field of materials. Machine learning can explore complex implicit relationships between various parameters from a large amount of experimental data, establish accurate prediction models, and greatly shorten the development cycle of materials.

[0006] Chinese patent CN115527625A discloses a hardness prediction method and system for high-entropy alloys. The method includes obtaining candidate features corresponding to the hardness data of AlCoCrCuFeNi system high-entropy alloys, and establishing a data set; training a Stacking integrated model using the data set; screening the candidate features using Pearson correlation coefficient, XGBoost evaluation model, random forest, genetic algorithm, XGBoost-based recursive feature elimination method, and exhaustive method to determine the screened features; establishing a classifier using principal component analysis and logistic regression based on the screened features; constructing an alloy composition search space based on the trained Stacking integrated model and the classifier; and predicting the hardness based on the alloy composition search space.

[0007] However, most of the machine learning researches are now focused on the prediction performance rather than promoting alloy design. There is a wide search space for various component proportions, and it is difficult to quickly screen out components with excellent performance. SUMMARY

[0008] The application provides a prediction method for iron-based nanocrystalline alloys based on machine learning, which can quickly screen out components with excellent performance.

[0009] A prediction method for iron-based nanocrystalline alloys based on machine learning, comprising:

[0010] An initial data set and a label are obtained, an initial data set is constructed based on soft magnetic performance data of iron-based nanocrystalline alloys, and a final data set is obtained by preprocessing the initial data set, the final data set including a training sample set and a test sample set, and the label including a magnetic induction intensity label and a coercive force label;

[0011] The importance of the magnetic induction intensity and the coercive force of the iron-based nanocrystalline alloy features in the data set is sorted and screened by the Pearson correlation coefficient, the random forest and the forward selection method in sequence to obtain a first feature set and a second feature set, the first feature set being a key feature set corresponding to the magnetic induction intensity, and the second feature set being a key feature set corresponding to the coercive force;

[0012] An XGBoost model is trained based on the training sample set corresponding to the first feature set and the training sample set corresponding to the second feature set, and an optimal hyperparameter is adjusted by a grid search method to obtain a magnetic induction intensity prediction model and a coercive force prediction model;

[0013] The magnetic induction intensity prediction model and the coercive force prediction model are used as objective functions, and at least one iteration of a genetic algorithm NSGA-II is used to screen alloy component data meeting the highest magnetic induction intensity and the lowest coercive force from an alloy component data set of the iron-based nanocrystalline, and an iron-based nanocrystalline alloy component is obtained based on the screened alloy component data, the alloy component data set of the iron-based nanocrystalline being an alloy component data set of the iron-based nanocrystalline in the first feature set and the second feature set.

[0014] Further, a key feature with the highest importance of the magnetic induction intensity in the first feature set is used as a first key feature, the first key feature being a non-ferromagnetic electron concentration, and a linear regression method is used to obtain a negative linear relationship between the non-ferromagnetic electron concentration and the magnetic induction intensity.

[0015] Further, the negative linear equation between the non-ferromagnetic electron concentration and the magnetic induction intensity is:

[0016] B s = 2.46 - 1.004 x VEC1

[0017] wherein, Bs is the magnetic induction intensity, VEC1 is the non-ferromagnetic electron concentration.

[0018] Further, the key feature with the highest importance ranking of the magnetic induction intensity obtained from the first feature set is taken as the first key feature, which is the non-ferromagnetic electron concentration;

[0019] The key feature with the highest importance ranking of the coercivity obtained from the second feature set is taken as the second key feature, which is the electron concentration;

[0020] Different non-ferromagnetic electron concentrations are input into the magnetic induction intensity prediction model to obtain the corresponding predicted magnetic induction intensity, and through SHAP analysis method, it is obtained that when the non-ferromagnetic electron concentration VEC1≤0.78, the SHAP Value is greater than 0, and the predicted magnetic induction intensity tends to increase;

[0021] Different electron concentrations are input into the coercivity prediction model to obtain the corresponding coercivity, and through SHAP analysis method, it is obtained that when the electron concentration VEC1≤7.12, the SHAP Value is less than 0, and the predicted coercivity tends to decrease.

[0022] Further, the features of the iron-based nanocrystalline alloy in the initial data set include alloy elements, component proportions of the alloy elements, process parameters for preparing the iron-based nanocrystalline alloy, and characteristic parameters of the iron-based nanocrystalline alloy.

[0023] Further, the preprocessing of the initial data set to obtain the data set comprises:

[0024] (1) screening the data set corresponding to the amorphous state at the quenched state in the initial data set, so as to match the amorphous precursor strip formed by the single-roller rapid quenching with the amorphous state;

[0025] (2) deleting the data with missing annealing temperature and annealing time in the data set obtained in step (1);

[0026] (3) deleting the data set with a heating rate of 1.6 K / s or more in the data set obtained in step (2) to obtain the final data set.

[0027] Further, the magnetic induction intensity label and the coercivity label are obtained by the following method, comprising: logarithmically transforming the real coercivity to obtain the coercivity label, and taking the real magnetic induction intensity as the magnetic induction intensity label.

[0028] Further, the importance rankings and screening of the features of the iron-based nanocrystalline alloy in the data set for the magnetic induction intensity and the coercivity are sequentially performed by the Pearson correlation coefficient, the random forest, and the forward selection method, comprising:

[0029] A feature with co-linearity in the iron-based nanocrystalline alloy characteristics is obtained through a Pearson correlation coefficient, and one of the features with co-linearity reserved in the iron-based nanocrystalline alloy characteristics is obtained as a key feature set;

[0030] A key feature sequence is obtained by performing feature importance sorting on the key feature set through a random forest;

[0031] According to the feature importance from high to low, the key features in the key feature sequence are gradually input into the random forest by using a forward selection method, and the trained random forest outputs a first predicted magnetic induction intensity and a first predicted coercive force;

[0032] A first magnetic induction intensity RMSE index is constructed based on the first predicted magnetic induction intensity and the magnetic induction intensity label, and a first coercive force RMSE index is constructed based on the first predicted coercive force and the coercive force label, when the first magnetic induction intensity RMSE index reaches a magnetic induction intensity threshold, the key feature set corresponding to the first magnetic induction intensity RMSE index is taken as a first feature set, and when the first coercive force RMSE index reaches a coercive force threshold, the key feature set corresponding to the first coercive force RMSE index is taken as a second feature set.

[0033] Compared with the prior art, the beneficial effects of the present application are:

[0034] The present application constructs an initial data set based on the soft magnetic performance data of the iron-based nanocrystalline alloy, processes the initial data set through a machine learning method, trains an XGBoost model based on the processed data to obtain a magnetic induction intensity prediction model and a coercive force prediction model, takes the magnetic induction intensity prediction model and the coercive force prediction model as objective functions, and obtains alloy composition data with the highest magnetic induction intensity and the lowest coercive force through iteration of a genetic algorithm NSGA-II, so that excellent components can be quickly screened. BRIEF DESCRIPTION OF DRAWINGS

[0035] Figure 1 A flowchart of the prediction method of the iron-based nanocrystalline alloy based on machine learning provided for Example 1;

[0036] Figure 2 XRD diffraction patterns of the iron-based nanocrystalline alloys predicted in Examples 1 and 2;

[0037] Figure 3 The magnetic induction intensity B predicted by the magnetic induction intensity prediction model provided for Example 1 s The fitting graph between the real magnetic induction intensity B s

[0038] Figure 4 The coercive force H predicted by the coercive force prediction model provided for Example 1 c The fitting graph between the real coercive force H​c Fitting plot between the real magnetic induction B

[0039] Figure 5 Predicted magnetic induction B s Linear relationship plot between the non-ferromagnetic electron concentration VEC1 and the real magnetic induction B

[0040] Figure 6 Coercivity H s SHAP relationship plot between the non-ferromagnetic electron concentration VEC1 and the real magnetic induction B

[0041] Figure 7 Coercivity H c SHAP relationship plot between the electron concentration VEC and the real magnetic induction B. DETAILED DESCRIPTION

[0042] To make the technical problems, technical solutions and advantages to be solved by the present application clearer, specific embodiments will be described in detail below with reference to the drawings.

[0043] The present application aims to provide a method for efficiently designing iron-based nanocrystalline soft magnetic alloys with high saturation magnetic induction (B s ) and low coercivity (H c ) based on a machine learning model. The machine learning model is used to learn the screened experimental data set, and the key features are selected through feature engineering. Linear regression and explainability SHAP are combined to quantitatively and qualitatively explain the machine learning model, and a feasible composition design principle based on iron-based nanocrystalline soft magnetic alloys is proposed.

[0044] Embodiment 1

[0045] The present application provides a prediction method for iron-based nanocrystalline alloys based on machine learning, as shown in Figure 1 , which comprises:

[0046] (1) Data collection and cleaning:

[0047] (1.1) Establishing an initial data set: Based on the obtained iron-based nanocrystalline alloy soft magnetic performance data from existing experiments, an initial data set is established based on the relevant experimental data. The characteristics of the iron-based nanocrystalline alloy in the initial data set include alloy elements, composition ratio of alloy elements, process parameters for preparing the iron-based nanocrystalline alloy, and characteristic parameters of the iron-based nanocrystalline alloy. The process parameters for preparing the iron-based nanocrystalline alloy are annealing temperature (T a ), annealing time (T t), copper roller rotation speed (v), the characteristic parameters of the iron-based nanocrystalline alloy are valence electron concentration (VEC), non-ferromagnetic valence electron concentration (VEC1), mixing entropy (ΔS), electronegativity (χ), atomic radius difference (δ), a total of 20 characteristics, the coercive force label is obtained by logarithmic transformation of the real coercive force, and the real magnetic induction intensity is taken as the magnetic induction intensity label. A total of 348 data, and the alloy element screening is Fe, Co, Ni, Si, P, B, C, Cu, Nb, Al, Cr and Mn.

[0048] (1.2) Establishing the final data set: preprocessing the initial data set to obtain the final data set, the final data set including the training sample set and the test sample set, and the specific steps of constructing the final data set are:

[0049] Step 1, screening the data set corresponding to the non-crystalline state in the initial data set in the quenched state, so as to match the non-crystalline precursor strip formed by single-roller rapid quenching with the non-crystalline state;

[0050] Step 2, deleting the data with missing annealing temperature and annealing time in the data set obtained in step 1;

[0051] Step 3, deleting the data set corresponding to the heating rate reaching 1.6K / s or more in the data set obtained in step 2 to obtain the final data set.

[0052] (2) Screening key features: sequentially sorting and screening the important degrees of the iron-based nanocrystalline alloy features in the data set for magnetic induction intensity and coercivity by Pearson correlation coefficient, random forest and forward selection method to obtain the first feature set and the second feature set, and the specific steps are:

[0053] Through the Pearson correlation coefficient (PCC), the collinear features in the iron-based nanocrystalline alloy features are obtained, i.e. Si and VCE, ΔS and Fe are collinear, and Si and ΔS are removed.

[0054] Through the random forest algorithm (RF), the feature importance of the remaining 18 features is sorted, and the feature importance sequence is obtained, i.e. the feature importance sequence of the B s of the nanocrystalline soft magnetic alloy is VEC1>Fe>χ>VEC>Nb>Cu>T a >C>P>B>δ>T t >v>Co>Cr>Mn>Al>Ni, and the feature importance sequence of the H c of the nanocrystalline soft magnetic alloy is VEC>T a >χ>VEC1>Fe>B>δ>Cu>P>T t>C > V > Nb > Co > Al > Cr > Mn > Ni. The specific steps are: taking the Gini Index (GI) as the evaluation index, calculating the average value of the contribution degree of each feature on the decision tree in the random forest, then comparing the contribution degree between different features, and finally sorting the importance of each feature.

[0055] According to the importance of the features from high to low, the key features in the key feature sequence are input into the random forest by using the forward selection method, and the trained random forest outputs the first predicted magnetic induction intensity and the first predicted coercivity. The first magnetic induction intensity RMSE index is constructed based on the first predicted magnetic induction intensity and the magnetic induction intensity label, and the first coercivity RMSE index is constructed based on the first predicted coercivity and the coercivity label. When the first magnetic induction intensity RMSE index reaches the magnetic induction intensity threshold, the key feature set corresponding to the first magnetic induction intensity RMSE index is taken as the first feature set, and the features included in the first feature set are VEC1>Fe>χ>VEC>Nb>Cu. When the first coercivity RMSE index reaches the coercivity threshold, the key feature set corresponding to the first coercivity RMSE index is taken as the second feature set, and the features included in the second feature set are VEC>T a >χ>VEC1>Fe>B>δ>Cu>P>T t . The feature that most affects the saturation magnetic induction intensity (B s ) is VEC1, and the feature that most affects the coercivity (H c ) is VEC.

[0056] (3) An XGBoost model is established, the XGBoost model is trained based on the training sample set corresponding to the first feature set and the training sample set corresponding to the second feature set, the optimal hyperparameters are adjusted by the grid search method to obtain a magnetic induction intensity prediction model and a coercivity prediction model, and test samples are input into the magnetic induction intensity prediction model and the coercivity prediction model to obtain predicted magnetic induction intensity and predicted coercivity. The model evaluation indexes of mean square error (RMSE) and determination coefficient (R 2 ) are used to evaluate the predicted magnetic induction intensity and the predicted coercivity obtained by the magnetic induction intensity prediction model and the coercivity prediction model.

[0057] R 2 is a measure of the ability of the model to explain the variance of the dependent variable, and the value of R 2 is between negative infinity and 1; RMSE is a common index for measuring prediction error. The closer the value of R 2 is to 1 or the smaller the value of RMSE is, the better the fitting effect of the model is.

[0058] The results are shown in Figure 3 , Figure 4 , the predicted saturation magnetic induction intensity (Bs RMSE and R of the model 2 The values ​​were 0.039T and 0.948, respectively, predicting coercivity (H). c RMSE and R of the model 2 The values ​​are 0.4 A / m and 0.823, respectively.

[0059] (4) NSGA-II iterative optimization: Using the magnetic flux density prediction model and the coercivity prediction model as objective functions, the NSGA-II genetic algorithm iteratively filters out alloy composition data that meets the criteria of highest magnetic flux density and lowest coercivity from the iron-based nanocrystalline alloy composition dataset. Based on the filtered alloy composition data, the iron-based nanocrystalline alloy composition, i.e., Fe, is obtained. 81.4 Si 6.8 B 8.6 Cu 0.7 P 1.9 Nb 0.6 The alloy composition dataset of iron-based nanocrystals is the alloy composition dataset of iron-based nanocrystals in the first feature set and the second feature set.

[0060] (5) Model Interpretation: The key features VEC1 and B, which have the highest feature importance, are ranked using linear regression and SHAP analysis. s ,VEC and H c The relationships between them were analyzed. During the feature selection process, factors influencing B were identified. s The most important feature is VEC1, which affects H c The most important feature is VEC.

[0061] To observe B more clearly s The relationship with VEC1 will bring B s Project VEC1 into a two-dimensional space. Figure 5 The saturation magnetic induction intensity B s The relationship between VEC1 and nonferromagnetic valence electron concentration. (B) s It shows a clear negative linear relationship with VEC1, which is consistent with reality.

[0062] The negative linear equation for the nonferromagnetic electron concentration and magnetic induction intensity is as follows:

[0063] B s =2.46 - 1.004 × VEC1

[0064] Among them, B s is the magnetic flux density, and VEC1 is the nonferromagnetic electron concentration.

[0065] The SHAP analysis method is used to perform explainable analysis on the magnetic induction intensity prediction model and the coercivity prediction model. The SHAP method is used to calculate feature importance and evaluate the influence of the feature on the model prediction. The positive and negative values of SHAP represent that the feature improves or weakens the performance. For B s , the smaller the VEC1 value is, the larger the SHAP Value is, and the larger the B s value is.

[0066] When VEC1≤0.78, the SHAP Value is greater than 0, and the B s value tends to increase. For H c , the larger the VEC value is, the larger the SHAP Value is, and the larger the H c value is. It is found that when VEC≤7.12, the SHAP Value is less than 0, and the H c value tends to decrease, as shown in Figure 6 . Figure 7 .

[0067] wherein VEC1 and B s present obvious negative correlation, VEC and H c have obvious trend relationship at a certain critical point. Comprehensive consideration obtains the following prediction criteria:

[0068] When VEC1≤0.78, B s tends to increase; when VEC≤7.12, H c tends to decrease.

[0069] (6) Experimental verification:

[0070] a. After raw materials such as Fe, Si, B, Nb, Cu and P with a purity greater than 99% are weighed according to the alloy composition and the atomic percentage of each component, the alloy ingot with uniform composition is obtained by placing the raw materials into a ceramic crucible of an induction melting furnace and melting under a high vacuum argon atmosphere through induction melting;

[0071] b. The alloy ingot obtained in step a is crushed and loaded into a quartz tube, and an amorphous precursor tape with a width of about 1 mm and a thickness of about 22 μm is prepared by using a single-roll rapid quenching tape casting process at a speed of 40 m / s in an argon atmosphere;

[0072] c. The precursor tape prepared in step b is loaded into a vacuum tube-type annealing furnace at 520℃, and annealed for 5 min, and then quenched to room temperature to obtain an iron-based nanocrystalline soft magnetic alloy. In order to detect the composition and structure of the alloy, XRD test is performed. As shown in Figure 2 , there are two sharp α-Fe crystallization peaks at 2θ = 45° and 65°, respectively, indicating that the precipitated phase is α-Fe phase;

[0073] d. The saturation magnetic induction intensity (B) of the alloy was measured using a vibrating sample magnetometer (VSM) and a DC BH meter, respectively. s ) and coercivity (H c Upon testing, B in this embodiment... s and H c The experimental values ​​were 1.72T and 3.8A / m, respectively.

[0074] The VEC1 obtained in this embodiment 1 experimental verification is 0.73, and the predicted B s =2.46 - 1.004 × 0.73 = 1.72T, which is equal to the experimental value of 1.72T. VEC1 = 0.73 ≤ 0.78. Compared with the value obtained from Comparative Example 1, its B s The results are largely consistent with the predictions, demonstrating the feasibility of the component design principles proposed based on model interpretation.

[0075] Example 2

[0076] The machine learning model design steps (1)-(3) and (5)-(6) in this embodiment are the same as those in embodiment 1, except for step (4). Step (4) in this embodiment is: based on the genetic algorithm NSGA-II, the iron-based nanocrystalline soft magnetic alloy composition with target performance is screened from the search space through two rounds of iteration; the alloy composition is Fe 81.6 Si 7.3 B 8.5 Cu 0.7 P 1.8 Nb 0.1 .

[0077] The saturation magnetic flux density (B) of this embodiment was measured using a VSM and a DC BH meter. s ) and coercivity (H c The experimental values ​​were 1.71T and 3.3A / m, respectively.

[0078] The VEC1 obtained in this embodiment 2 experimental verification is 0.72, and the predicted B s =2.46 - 1.004 × 0.73 = 1.73T, which is close to the experimental value of 1.71T, and VEC1 = 0.72 ≤ 0.78. This is basically consistent with the predicted results, proving the feasibility of the component design principle proposed based on model interpretation.

[0079] The preferred embodiments of the present invention have been described above. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the present invention. Any modifications and improvements made within the scope of the principles of the present invention should be included within the protection scope of the present invention.

[0080] Most of the current machine learning algorithms have the problem of "black box", the relationship between the performance of Fe-based nanocrystalline soft magnetic alloy and the parameters such as composition and process is not clear. In the embodiment of the present application, a method for designing Fe-based nanocrystalline soft magnetic alloy with high saturation magnetic induction (B s ) and low coercivity (H c ) based on XGBoost machine learning model optimized by interpretable analysis genetic algorithm is developed, not only the soft magnetic alloy with excellent performance is quickly developed, but also the black box model inherent in machine learning can be explained, the high-accuracy composition design principle is established, which plays an important guiding significance for developing soft magnetic nanocrystalline alloy.

[0081] Comparative example 1

[0082] The alloy composition of comparative example 1 is Fe 81 Si4B 10 Cu1P2Nb2, the VEC is 7.25, VEC1 is 0.77, and the experimental values of the saturation magnetic induction (B s ) and coercivity (H c ) of the present comparative example are 1.64T and 7.3A / m respectively, measured by VSM and DCB-H instrument.

Claims

1. A method of predicting an iron-based nanocrystalline alloy based on machine learning, characterized by, The method comprises the following steps: obtaining an initial data set and a label, constructing an initial data set based on the soft magnetic performance data of the iron-based nanocrystalline alloy, and preprocessing the initial data set to obtain a final data set, the final data set comprising a training sample set and a test sample set, and the label comprising a magnetic induction intensity label and a coercive force label; sequentially performing magnetic induction intensity and coercive force importance sorting and screening on the iron-based nanocrystalline alloy features in the data set by using Pearson correlation coefficient, random forest and forward selection method to obtain a first feature set and a second feature set, the first feature set being a key feature set corresponding to the magnetic induction intensity, and the second feature set being a key feature set corresponding to the coercive force; training an XGBoost model based on the training sample set corresponding to the first feature set and the training sample set corresponding to the second feature set, and adjusting the optimal hyperparameters by using a grid search method to obtain a magnetic induction intensity prediction model and a coercive force prediction model; using the magnetic induction intensity prediction model and the coercive force prediction model as objective functions, and screening out alloy composition data meeting the highest magnetic induction intensity and the lowest coercive force from the alloy composition data set of the iron-based nanocrystalline alloy through at least one iteration of genetic algorithm NSGA-II, and obtaining the alloy composition of the iron-based nanocrystalline alloy based on the screened alloy composition data; the alloy composition data set of the iron-based nanocrystalline alloy is the alloy composition data set of the iron-based nanocrystalline alloy in the first feature set and the second feature set; using the key feature with the highest magnetic induction intensity importance degree in the first feature set as a first key feature, the first key feature being non-ferromagnetic electron concentration, and obtaining a negative linear relationship between the non-ferromagnetic electron concentration and the magnetic induction intensity by using a linear regression method; the negative linear equation between the non-ferromagnetic electron concentration and the magnetic induction intensity is: B s = 2.46 - 1.004 x VEC1 wherein, B s B is the magnetic induction, and VEC1 is the non-ferromagnetic electron concentration.

2. The machine learning-based prediction method of ferrous nanocrystalline alloy according to claim 1, wherein, using the key feature with the highest magnetic induction intensity importance degree in the first feature set as a first key feature, the first key feature being non-ferromagnetic electron concentration; using the key feature with the highest coercive force importance degree in the second feature set as a second key feature, the second key feature being electron concentration; inputting different non-ferromagnetic electron concentrations into the magnetic induction intensity prediction model to obtain corresponding predicted magnetic induction intensities, and obtaining, by using SHAP analysis method, that when the non-ferromagnetic electron concentration VEC1 is less than or equal to 0.78, the SHAP Value is greater than 0, and the predicted magnetic induction intensity tends to increase; inputting different electron concentrations into the coercive force prediction model to obtain corresponding coercive forces, and obtaining, by using SHAP analysis method, that when the electron concentration VEC1 is less than or equal to 7.12, the SHAP Value is less than 0, and the predicted coercive force tends to decrease. 3.The method of claim 1, wherein, The iron-based nanocrystalline alloy features in the initial data set include alloy elements, component proportions of the alloy elements, process parameters for preparing the iron-based nanocrystalline alloy, and characteristic parameters of the iron-based nanocrystalline alloy. 4.The method of claim 1, wherein, The preprocessing of the initial data set to obtain the data set comprises the following steps: (1) screening the data set corresponding to the non-crystalline state in the quenched state in the initial data set, so as to match the non-crystalline precursor strip formed by single-roller rapid cooling and winding; (2) deleting the data with missing annealing temperature and annealing time in the data set obtained in step (1). (3) deleting the data set corresponding to the temperature rising rate of 1.6 K / s or more from the data set obtained in step (2) to obtain a final data set. 5.The method of claim 1, wherein, The magnetic induction intensity label and the coercive force label are obtained by: logarithmically transforming the real coercive force to obtain the coercive force label, and taking the real magnetic induction intensity as the magnetic induction intensity label. 6.The method of claim 1, wherein, The important degree of the iron-based nanocrystalline alloy features in the data set is sorted and screened by the Pearson correlation coefficient, the random forest and the forward selection method in sequence for the magnetic induction intensity and the coercive force, including: collinear features in the iron-based nanocrystalline alloy features are obtained by the Pearson correlation coefficient, and one feature in the collinear features is retained to obtain a key feature set; a key feature sequence is obtained by sorting the feature importance of the key feature set by the random forest; according to the feature importance from high to low, the key features in the key feature sequence are input into the random forest step by step by the forward selection method, and the random forest after training outputs a first predicted magnetic induction intensity and a first predicted coercive force; a first magnetic induction intensity RMSE index is constructed based on the first predicted magnetic induction intensity and the magnetic induction intensity label, a first coercive force RMSE index is constructed based on the first predicted coercive force and the coercive force label, when the first magnetic induction intensity RMSE index reaches a magnetic induction intensity threshold, the key feature set corresponding to the first magnetic induction intensity RMSE index is taken as a first feature set, and when the first coercive force RMSE index reaches a coercive force threshold, the key feature set corresponding to the first coercive force RMSE index is taken as a second feature set.

Images