Method for accelerated design of iron-based bulk amorphous alloy with high magnetism

Abstract

The invention relates to a method for accelerated design of iron-based bulk amorphous alloy with high magnetism, which comprises the following steps: S1, acquiring components of n alloys and corresponding saturation magnetization intensity, and carrying out data preprocessing; s2, screening out features having important influence on prediction; s3, obtaining a final prediction model; s4, regressing the two basic models by utilizing a voting set learning method in combination with an extreme gradient boosting tree and a decision tree; s5, explaining the set model by using weighted average Sharpley addition explanation analysis; s6, analyzing important influence factors according to the model interpretation information; and S7, alloy components with specific saturation magnetization intensity are designed, and the block iron-based amorphous alloy is prepared. The method has an important acceleration effect on research and development of new materials, and is helpful for deeply understanding the structure-performance structure-activity relationship of the amorphous alloy.

Description

Technical Field

[0001] The present invention relates to a method for accelerating the design of an iron-based bulk amorphous alloy with high magnetism, and specifically to a method for accelerating the design of an iron-based bulk amorphous alloy with high magnetism based on machine learning. Background Art

[0002] Iron-based amorphous alloys are a class of materials with excellent soft magnetic properties. Their low core loss and high saturation magnetization make them a preferred material for high-efficiency, high-speed motors and energy-saving transformers. Their performance advantages and potential have been widely recognized in industries such as photovoltaics, power grids, and industrial motors, making them key materials for next-generation products. Research and development of iron-based amorphous alloys that combine high glass-forming ability with high saturation magnetization is of great value. The complexity of amorphous alloy composition makes composition design very difficult. Currently, composition design of iron-based amorphous alloys is still primarily based on trial-and-error experiments, resulting in long R&D cycles and high costs. Developing new composition design methods is crucial for the preparation of highly magnetic bulk iron-based amorphous alloys (with three-dimensional dimensions greater than 1.0 mm). Machine learning, an emerging big data processing technology, has been successfully applied to new material design. By building models of specific properties based on material databases, machine learning can rapidly predict material properties, thereby accelerating the design of new materials and shortening the development cycle.

[0003] Based on this, the present invention proposes a method for predicting the magnetic properties of bulk iron-based amorphous alloys based on machine learning, and analyzes the interpretable information of the machine learning model to explore the relationship between performance and structure, and accelerate the design of bulk amorphous alloy materials. Summary of the Invention

[0004] In response to the shortcomings of the existing technology, the present invention provides a method for accelerating the design of iron-based bulk amorphous alloys with high magnetism. The method of the present invention is based on machine learning, analyzes the interpretable information of the machine learning model, explores the relationship between performance and structure, accelerates the design of bulk amorphous alloy materials, greatly reduces the trial and error cost, and quickly realizes the prediction of material properties, thereby accelerating the design of new materials and shortening the material development cycle.

[0005] The present invention is achieved through the following technical solutions:

[0006] A method for accelerating the design of an iron-based bulk amorphous alloy with high magnetic properties, comprising the following steps:

[0007] S1. Collect experimental data of n types of bulk Fe-based amorphous alloys from existing literature, including chemical composition, saturation magnetization, and diameter. The composition and saturation magnetization of each alloy form a set of raw data. The n sets of raw data for n alloys constitute the original data set.

[0008] According to the chemical composition theory, the valence electron number (VEC), valence electron number (VEC1), average electron number (e), average atomic radius (r), electronegativity (X), theoretical molar volume (V), theoretical melting point (Tm), and theoretical density (ρ) are calculated as follows:

[0009]

[0010] Among them, c i ,VEC i , e i , r i , X i ,m i ,ρ i ,T mi and W i is the atomic percentage, number of valence electrons, number of electrons, atomic radius, electronegativity, mass fraction, density, melting point and atomic weight of the i-th element, and n is the number of chemical elements in the sample. After normalization, the results are combined with the magnetization intensity of the corresponding alloy to form a standard data set, which is divided into training set and test set in a ratio of 8:2;

[0011] S2, filter method combined with embedded feature selection to select features that have important influence on prediction;

[0012] S3. Establish an extreme gradient boosting tree model and a decision tree regression model based on the data set in step S1, perform hyperparameter tuning, and evaluate the model performance using evaluation indicators;

[0013] S4. Using the voting ensemble method, different weight combinations of the extreme gradient boosting tree model and the decision tree regression model are tested, the best weight combination is selected, and finally the optimal model for alloy magnetic property prediction is obtained;

[0014] S5. Perform weighted calculation on the SHAP absolute value, where the weight of the SHAP value corresponds to the weight of the basic model in the set;

[0015] S6. According to the weighted feature importance ranking, the important features affecting magnetic properties are obtained: VEC1 and iron content. The SHAP values ​​of the two features are calculated, and it is found that the alloy composition has a positive effect on magnetic properties within a certain range.

[0016] S7. Design an alloy system based on the range obtained in step S6, calculate the relevant characteristic values ​​in step S1 as input to the amorphous alloy magnetic property prediction model in step S3, and thus predict the saturation magnetization intensity of amorphous alloys with different element compositions under the alloy system, so as to achieve rapid design of amorphous alloys.

[0017] According to the present invention, preferably, in step S1 , the valence electron number (VEC1) does not contain Fe, Co, or Ni magnetic elements.

[0018] According to the present invention, preferably, in step S1, the composition includes the elements constituting the alloy and the content of the elements, and the original data set is used as effective data for predicting the magnetic properties of the alloy.

[0019] Preferably, according to the present invention, in step S2, feature selection technology: simplifies the model by reducing meaningless features, thereby improving the computing power of the model without affecting the accuracy, the filter method can help identify linear correlations between features, and embedded feature selection can reveal the nonlinear relationship between features and target variables.

[0020] Preferably, according to the present invention, in step S3, the hyperparameter tuning is specifically as follows: establishing an extreme gradient boosting tree model and a decision tree regression model according to the data set in step S1, setting the number of decision trees in the extreme gradient boosting tree model, the minimum sample weight sum of the leaf nodes of each tree, the minimum loss of node splitting, the learning rate, the L1 regularization term and the L2 regularization term range, as well as the maximum depth of the decision tree in the decision tree regression model, the minimum impurity reduction required for node splitting, the minimum number of samples for each node for splitting, the minimum number of samples for each leaf node and the maximum number of features used by each tree, and using the Bayesian optimization method to optimize the above hyperparameter set to ensure that the best parameter combination is obtained on the test set and the optimal basic model is obtained.

[0021] According to the preferred embodiment of the present invention, in step S3, the calculation formula of the evaluation index is:

[0022]

[0023] in is the predicted value, is the actual value y i The average value of .

[0024] Preferably, according to the present invention, in step S4, the voting integration method is: the core idea of ​​ensemble learning is to reduce the variance and bias of the model by training different basic models and combining their prediction results. By combining the opinions of multiple models, the robustness and prediction accuracy of the model can be improved. Voting integration integrates the prediction results of multiple independent regressors to make the final prediction, and can also be used in combination with other integration methods.

[0025] Preferably according to the present invention, in step S5, the weight is as follows: for voting integration, voting is divided into hard voting and soft voting. Hard voting simply takes the majority vote predicted by a single regressor as the prediction result, while soft voting uses the method of weighted average to obtain the final result. Among them, the weight determines the contribution degree of the prediction result of each basic model in the final integration result. The weight is very crucial in soft voting, and appropriate weight allocation can make the integrated model more robust and accurate.

[0026] Preferably according to the present invention, in step S6, it is obtained that within the range of VEC1 ≤ 0.955 and 71.6 < Fe < 78 at.%, the alloy composition has a positive effect on magnetic properties; and the generalization ability of the model is verified using literature data.

[0027] Preferably according to the present invention, an amorphous alloy with a diameter of 4.0 mm and a saturation magnetization intensity of 1.45 T is designed based on six elements of Fe, Co, B, P, C, and Cu according to the integrated learning model, and the atomic percentage of the designed alloy composition is Fe 76.1 Co 3.7 Cu 0.2 B 3.6 P 9.4 C7;

[0028] Preferably according to the present invention, raw materials are weighed according to the alloy molecular chemical formula. The purity of Fe, FeB, FeP, FeC, Co, and Cu is 99.99%; the raw materials are melted in a vacuum induction melting furnace under specific melting conditions to obtain a master alloy, and a bulk iron-based amorphous alloy is prepared from the master alloy; XRD testing and VSM testing are performed on the iron-based amorphous alloy ribbon.

[0029] The steps of preparing the bulk iron-based amorphous alloy from the master alloy include: putting the master alloy into a quartz tube, evacuating the spin coater and then filling it with argon as a protective gas, and heating the master alloy to a molten state through an induction coil under the protection of argon; a high-pressure argon gas is ejected from the stripping nozzle, and sampling is carried out by the spray casting method. The molten metal is quickly sprayed into a copper mold under the push of high-pressure inert gas to form a 4.0 mm rod-shaped amorphous sample.

[0030] Preferably according to the present invention, the specific melting conditions include a vacuum environment of 5×10 -3 Pa, and under the protective atmosphere of argon, the number of alloy melting times is 5 - 15 times.

[0031] The technical features and advantages of the present invention:

[0032] 1. The present method combines the extreme gradient boosting tree model with a voting ensemble learning method based on the decision tree regression model, making the model highly robust and accurate. Furthermore, this invention provides an efficient composition design method for highly magnetic iron-based bulk amorphous alloys, significantly accelerating the development of new iron-based soft magnetic amorphous alloys and furthering the understanding of the structure-property relationship of amorphous alloys.

[0033] 2. The present invention uses SHAP analysis to determine the importance of VEC1 and iron content in predicting Bs, and finds that when VEC1 < 0.955 and iron content > 71.6 at.%, they have a positive effect on magnetic properties. According to theoretical predictions, a 4.0 mm Fe 76.1 Co 3.7 Cu 0.2 B 3.6 P 9.4 C7 amorphous alloy rod, its B s It can be as high as 1.45T, which is better than other Fe-Co bulk amorphous alloys of the same size. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 This is a flowchart of the method for accelerating the design of iron-based bulk amorphous alloys with high magnetism based on machine learning according to the present invention.

[0035] Figure 2 This is the decision tree regression model result diagram of the present invention.

[0036] Figure 3 This is the result diagram of the extreme gradient boosting tree model of the present invention.

[0037] Figure 4 This is the result diagram of the integrated model of the present invention.

[0038] Figure 5 This is the SHAP weighted absolute value feature importance ranking diagram of the present invention.

[0039] Figure 6 This is a SHAP value distribution diagram of VEC1 and Fe content, which are important features of the present invention. DETAILED DESCRIPTION

[0040] To make the above-mentioned objects, features and advantages of the present invention more clearly understood, the technical solutions of the present invention are further described below with reference to the accompanying drawings and embodiments. However, the present invention is not limited to the embodiments listed, but also includes any other known modifications within the scope of the claimed invention.

[0041] First, the term "one embodiment" or "embodiment" herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in various places throughout this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive of other embodiments.

[0042] Example 1

[0043] A method based on machine learning to accelerate the design of high-magnetic Fe-based bulk amorphous alloys was proposed. A weighted average absolute SHAP analysis method was proposed to explain the ensemble model, and a "two-step feature selection" method was constructed based on feature sorting. Through SHAP analysis, the importance of VEC1 and iron content in predicting Bs was determined. It was found that when VEC1 < 0.955 and iron content > 71.6 at.%, it has a positive effect on magnetic properties. According to theoretical predictions, a 4.0 mm Fe 76.1 Co 3.7 Cu 0.2 B 3.6 P 9.4 C7 amorphous alloy rod, its B s It can be as high as 1.45T, which is better than other Fe-Co bulk amorphous alloys of the same size.

[0044] The specific steps include:

[0045] S1. Collect experimental data of n types of bulk Fe-based amorphous alloys from existing literature, including chemical composition, saturation magnetization, and diameter. The composition and saturation magnetization of each alloy form a set of raw data. The n sets of raw data for n alloys constitute the original data set.

[0046] According to the chemical composition theory, the valence electron number (VEC), valence electron number (VEC1), average electron number (e), average atomic radius (r), electronegativity (X), theoretical molar volume (V), theoretical melting point (Tm), and theoretical density (ρ) are calculated as follows:

[0047]

[0048] Among them, c i ,VEC i , e i , r i , X i , m i , ρ i , T mi and W iis the atomic percentage, number of valence electrons, number of electrons, atomic radius, electronegativity, mass fraction, density, melting point, and atomic weight of the i-th element, and n is the number of chemical elements in the sample. After normalization, the results are combined with the magnetization intensity of the corresponding alloy to form a standard data set, which is divided into training and test sets according to the ratio.

[0049] 303 sets of experimental data on bulk Fe-based amorphous alloys in published literature were collected, including the chemical composition of the materials (a total of 27 elements), saturation magnetization B s and its corresponding rod diameter. B s The data range is 0.35T~1.71T. The diameter of the bulk amorphous alloy can reflect the amorphous forming ability and can also be used to obtain the corresponding B of the specific size when predicting the new composition in the later stage. s The dataset is randomly divided into a training set and a test set in a ratio of 8:2. The model can be trained on the training set and the performance of the model can be evaluated on the test set. This allows a more objective assessment of the model's generalization ability to unseen data.

[0050] Step S2: A two-step feature selection method combining a filter method with embedded feature selection is used to select features with significant impact on prediction. Based on chemical composition theory, the valence electron number (VEC), the valence electron number excluding magnetic elements (Fe, Co, and Ni) (VEC1), the average electron number (e), the average atomic radius (r), the electronegativity (X), the theoretical molar volume (V), the theoretical melting point (Tm), and the theoretical density (ρ) are calculated, along with the top ten elements, to form the initial features. Feature selection techniques simplify the model by reducing insignificant features, improving computational power without compromising accuracy. Complex relationships exist between features, including both linear and complex nonlinear relationships. Therefore, a two-step feature selection method combining a filter method with embedded feature selection is considered. The PCC method is used for filtering. The PCC ranges from -1 to 1, with absolute values ​​closer to 1 indicating a stronger linear correlation between two features. If |PCC| of two features is ≥ 0.85, it indicates a strong correlation, and one of the features is removed. The five features B, C, P, volume, and electronegativity all showed strong correlations with each other. The |PCC| relationship between B and the remaining four features was greater than 0.85, suggesting that it may contain some information from the four features. B also ranked second among the elements in the dataset. According to theoretical calculation formulas, volume and electronegativity are significantly affected by B content, so C, P, volume, and electronegativity were removed. The SHAP weighted absolute value ranking of the features in the ensemble model showed that the importance of features "Co" and "Dy" in the model was almost zero. Therefore, we decided to remove these two features in the second step of feature selection, leaving the remaining 13 features as the final features.

[0051] Step S3: Use the gradient boosting tree model and decision tree regression model to train and test the data set to obtain the final prediction model.

[0052] The number of decision trees in the extreme gradient boosting tree model, the minimum sample weight sum of the leaf nodes of each tree, the minimum loss for node splitting, the learning rate, the range of L1 regularization terms and L2 regularization terms, as well as the maximum depth of the decision tree in the decision tree regression model, the minimum impurity reduction required for node splitting, the minimum number of samples per node for splitting, the minimum number of samples per leaf node and the maximum number of features used by each tree were set. The Bayesian optimization method was used to optimize the above hyperparameter sets to ensure that the best parameter combination was obtained on the test set and the optimal basic model was obtained.

[0053] Step S4: Using the voting ensemble learning method to combine the two basic models of extreme gradient boosting and decision tree regression;

[0054] Voting ensembles learn the predictions of multiple independent models to arrive at a final prediction. Voting methods can be categorized as hard voting or soft voting. Soft voting uses a weighted average to arrive at the final prediction, where the weights determine the contribution of each base model's prediction to the final ensemble learning result. In this work, soft voting is employed.

[0055] After defining the hyperparameters of the base model, R is calculated based on different weight combinations of the base model from 1 to 99%. 2 When the extreme gradient boosting tree model weight is 75% and the decision tree model weight is 25%, the R 2 The value is the highest, which is 97.36% and the RMSE is 0.0558T.

[0056] Step S5: using a weighted average Shapley additive interpretation (SHAP) analysis to interpret the ensemble model;

[0057] Because SHAP cannot explain ensemble models, this study proposes an interpretable method for ensemble learning. In this method, the absolute SHAP values ​​are weighted, with the weights corresponding to the weights of the underlying models in the ensemble. The ensemble model is interpreted by analyzing the contribution of each underlying model to the final prediction. Because different underlying models have different weights, the weights must be considered when using SHAP values ​​to analyze the impact of the same feature on the ensemble model. Using absolute SHAP values ​​can intuitively express the importance of a feature to the target variable and facilitate comparisons between different features. This method effectively measures the impact of each feature on the ensemble model output, taking both positive and negative effects into account.

[0058] Step S6: Analyze important influencing factors based on the model interpretation information;

[0059] According to the ranking of weighted feature importance, the important features affecting magnetic properties are obtained: VEC1 and iron content. Calculate the SHAP values of the two features, and it is obtained that the alloy composition has a positive effect on magnetic properties within the ranges of VEC1 ≤ 0.955 and 71.6 < Fe < 78 at.%.

[0060] Step S7: 14 groups of compositions published in the literature but not yet entered into the database were statistically predicted. The prediction results were compared with the experimental data, and the errors were all less than 10%. Among them, the errors of three groups of compositions were 0, indicating that the model has a high generalization ability.

[0061] Step S8: Design alloy compositions with specific saturation magnetization intensities, prepare bulk iron-based amorphous alloys, and perform XRD tests and VSM tests on them.

[0062] The above are only the preferred embodiments of this patent. It should be noted that for those of ordinary skill in the art, without departing from the technical principles of this patent, several improvements and substitutions can be made, and these improvements and substitutions should also be regarded as the protection scope of this patent.

Claims

1. A method for accelerating the design of an iron-based bulk amorphous alloy with high magnetic properties, characterized in that: The following steps are involved: S1. Collect experimental data of n types of bulk Fe-based amorphous alloys from existing literature, including chemical composition, saturation magnetization, and diameter. The composition and saturation magnetization of each alloy form a set of raw data. The n sets of raw data for n alloys constitute the original data set. According to the chemical composition theory, the valence electron number (VEC), valence electron number (VEC1), average electron number (e), average atomic radius (r), electronegativity (X), theoretical molar volume (V), theoretical melting point (Tm), and theoretical density (ρ) are calculated as follows: Among them, c i , VEC i , e i , r i , X i , m i ,ρ i , T mi and W i is the atomic percentage, number of valence electrons, number of electrons, atomic radius, electronegativity, mass fraction, density, melting point and atomic weight of the i-th element, and n is the number of chemical elements in the sample. After normalization, the results are combined with the magnetization intensity of the corresponding alloy to form a standard data set, which is divided into training set and test set in a ratio of 8:2; S2, filter method combined with embedded feature selection to select features that have important influence on prediction; S3. Establish an extreme gradient boosting tree model and a decision tree regression model based on the data set in step S1, perform hyperparameter tuning, and evaluate the model performance using evaluation indicators; S4. Using the voting ensemble method, different weight combinations of the extreme gradient boosting tree model and the decision tree regression model are tested, the best weight combination is selected, and finally the optimal model for alloy magnetic property prediction is obtained; S5. Perform weighted calculation on the SHAP absolute value, where the weight of the SHAP value corresponds to the weight of the basic model in the set; S6. According to the weighted feature importance ranking, the important features affecting magnetic properties are obtained: VEC1 and iron content. The SHAP values ​​of the two features are calculated, and it is found that the alloy composition has a positive effect on magnetic properties within a certain range. S7. Design an alloy system based on the range obtained in step S6, calculate the relevant characteristic values ​​in step S1 as input to the amorphous alloy magnetic property prediction model in step S3, and thus predict the saturation magnetization intensity of amorphous alloys with different element compositions under the alloy system, so as to achieve rapid design of amorphous alloys.

2. The method according to claim 1, characterized in that In step S1 , the valence electron count (VEC1) does not include magnetic elements such as Fe, Co, and Ni. The composition includes the elements constituting the alloy and the element contents. The original data set is used as effective data for predicting the magnetic properties of the alloy.

3. The method according to claim 1, characterized in that In step S2, feature selection techniques are used to simplify the model by reducing meaningless features, thereby improving the computational power of the model without affecting accuracy. The filter method can help identify linear correlations between features, while embedded feature selection can reveal nonlinear relationships between features and target variables.

4. The method according to claim 1, wherein In step S3, hyperparameter tuning is specifically as follows: based on the data set in step S1, an extreme gradient boosting tree model and a decision tree regression model are established, the number of decision trees in the extreme gradient boosting tree model, the minimum sample weight sum of the leaf nodes of each tree, the minimum loss of node splitting, the learning rate, the L1 regularization term and the L2 regularization term range are set, as well as the maximum depth of the decision tree in the decision tree regression model, the minimum impurity reduction required for node splitting, the minimum number of samples for each node for splitting, the minimum number of samples for each leaf node and the maximum number of features used by each tree, and the Bayesian optimization method is used to optimize the above hyperparameter set to ensure that the best parameter combination is obtained on the test set and the optimal basic model is obtained.

5. The method according to claim 1, wherein In step S3, the calculation formula of the evaluation index is: in is the predicted value, is the actual value y i The average value of .

6. The method according to claim 1, characterized in that In step S4, the method of voting integration is as follows: The core idea of ensemble learning is to train different basic models and combine their prediction results to reduce the variance and bias of the model. By combining the opinions of multiple models, the robustness and prediction accuracy of the model can be improved. Voting integration integrates the prediction results of multiple independent regressors for the final prediction and can also be used in combination with other integration methods.

7. The method according to claim 1, characterized in that In step S5, the weights are as follows: For voting integration, voting is divided into hard voting and soft voting. Hard voting simply takes the majority vote in the predictions of individual regressors as the prediction result, while soft voting uses the method of weighted average to obtain the final result. The weights determine the contribution degree of the prediction results of each basic model in the final integration result. Weights are very crucial in soft voting, and appropriate weight allocation can make the integrated model more robust and accurate.

8. The method according to claim 1, characterized in that In step S6, it is obtained that within the range of VEC1 ≤ 0.955 and 71.6 < Fe < 78 at.%, the alloy composition has a positive effect on magnetic properties; and the generalization ability of the model is verified using literature data.

9. The method according to claim 1, characterized in that According to the ensemble learning model, Fe, Co, B, P, C and Cu are selected to design an amorphous alloy with a diameter of 4.0 mm and a saturation magnetization of 1.45 T. The atomic percentage of the alloy composition is Fe. 76.1 Co 3.7 Cu 0.2 B 3.6 P 9.4 C7; weighing raw materials according to the alloy molecular chemical formula, wherein the purity of Fe, FeB, FeP, FeC, Co, and Cu is 99.99%; melting the raw materials in a vacuum induction melting furnace under specific melting conditions to obtain a master alloy, and preparing a bulk iron-based amorphous alloy based on the master alloy; performing XRD and VSM tests on the iron-based amorphous alloy ribbon; The steps of preparing the bulk iron-based amorphous alloy from the master alloy include: putting the master alloy into a quartz tube, evacuating the melt spinner and then filling it with argon as a protective gas, and heating the master alloy to a molten state through an induction coil under the protection of argon; a high-pressure argon gas is ejected from the stripping nozzle, and sampling is carried out by the spray casting method. The molten metal is rapidly sprayed into a copper mold under the push of high-pressure inert gas to form a 4.0 mm rod-shaped amorphous sample.

10. The method according to claim 1, characterized in that The specific melting conditions include 5×10 -3 Pa vacuum environment, and under the protective atmosphere of argon, the alloy is melted 5-15 times.

Images