2:17 type rare earth soft magnetic alloy multi-property prediction method
By constructing a multi-performance prediction model for rare-earth soft magnetic alloys through machine learning, the problem of time-consuming screening of rare-earth soft magnetic alloy components in traditional methods has been solved. This enables efficient screening of rare-earth soft magnetic alloy components with high saturation magnetization, suitable Curie temperature, and ideal magnetic anisotropy, thus shortening the research and development cycle.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- ZHEJIANG UNIV
- Filing Date
- 2026-04-07
- Publication Date
- 2026-06-12
Smart Images
Figure CN122201564A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of magnetic materials technology, specifically to a method for predicting the multiple properties of a 2:17 type rare earth soft magnetic alloy. Background Technology
[0002] As power electronic devices develop towards higher frequencies and higher power densities, the heat generation and power loss issues during operation become increasingly prominent. This places more stringent demands on the overall performance and thermal stability of high-frequency soft magnetic materials. 2:17 type rare earth-transition metal compounds (Re2TM) 17 These materials typically possess high saturation magnetization and Curie temperatures that meet the thermal stability requirements of power device operation. Through proper composition control, they can exhibit easily magnetized surface characteristics (planar or conical). This planar or conical magnetocrystalline anisotropy effectively suppresses domain wall resonance and natural resonance under high-frequency alternating magnetic fields, reducing high-frequency residual losses and enabling them to demonstrate soft magnetic application potential in the megahertz frequency band.
[0003] However, the macroscopic magnetic properties of 2:17 type rare-earth intermetallic compounds are closely related to their microscopic chemical composition. Basic binary 2:17 type alloys often struggle to simultaneously achieve high saturation magnetization, suitable Curie temperature, and ideal magnetocrystalline anisotropy. Therefore, multi-element doping with transition metals is typically required for modification. This multi-element co-doping significantly alters the electronic structure and interatomic interactions within the alloy, leading to complex changes in its various magnetic properties. To obtain excellent comprehensive magnetic properties, rational composition design and optimal proportioning are necessary within a vast and complex multi-element chemical composition space. Faced with this multi-element composition space, selecting superior compositions that combine high saturation magnetization, suitable Curie temperature, and ideal magnetocrystalline anisotropy is challenging. Traditional trial-and-error experimental synthesis methods are time-consuming and resource-intensive. First-principles calculation methods based on density functional theory also consume significant computational resources when dealing with complex multi-element doped systems, making it difficult to effectively meet the high-throughput screening requirements for massive amounts of unknown compositions. Data-driven methods such as machine learning can efficiently establish the relationship between the underlying properties of materials and their macroscopic performance, providing new ideas for the composition design of new materials.
[0004] Therefore, to solve the above problems, it is necessary to develop an efficient and accurate performance prediction and component screening method based on machine learning to achieve high-throughput rapid screening of multi-component combinations of 2:17 type rare earth soft magnetic alloys and accelerate the research and development process of new high-frequency soft magnetic materials. Summary of the Invention
[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide a multi-performance prediction method for 2:17 type rare earth soft magnetic alloys, so as to solve the problems of long development cycle and high cost of high frequency rare earth soft magnetic alloy materials, and realize the rapid prediction of saturation magnetization, Curie temperature and magnetocrystalline anisotropy of 2:17 type rare earth soft magnetic alloy system.
[0006] To achieve the above-mentioned technical objectives and effects, the present invention is implemented through the following technical solution: A method for predicting multiple properties of a 2:17 type rare earth soft magnetic alloy, comprising the following steps: S1: Collect Re2TM 17 The chemical composition data of rare earth soft magnetic alloys and the corresponding experimental values of target performance are used to form an initial dataset based on the composition characteristics of the alloys and the target performance data. Among them, the target performance includes saturation magnetization, Curie temperature and magnetocrystalline anisotropy. Re is selected from one or more combinations of La, Ce, Pr, Nd, Y, Lu, Ho, Er, Yb, Gd, Sm and Dy. TM is at least one of Fe and Co, or is composed of at least one of Fe and Co and one or more combinations of Cu, Ga, Ti, V, Cr, Mn, Mo and Si. S2: Extract the physicochemical properties of each element in the alloy's chemical formula. Based on the atomic percentage of each element in the chemical formula, calculate the maximum, minimum, weighted average, and average deviation of each physicochemical property at the overall alloy level to generate statistical characteristics. Calculate the thermodynamic characteristics of the alloy and merge the statistical and thermodynamic characteristics of the alloy into the initial dataset of S1 to form a feature dataset. S3: Based on the feature dataset constructed in S2, the feature data is standardized. S4: The Pearson correlation coefficient method is used to perform feature correlation analysis and dimensionality reduction on the statistical and thermodynamic features of the dataset after standardization in step S3, and feature subsets of saturation magnetization, Curie temperature and magnetocrystalline anisotropy are constructed respectively. S5: For the regression prediction tasks of saturation magnetization and Curie temperature, two independent artificial neural network regression models are constructed respectively; for the classification prediction task of magnetocrystalline anisotropy, a Vote classification model integrating three base classifiers, K-nearest neighbors, random forest and support vector machine, is constructed, and the hyperparameters of the Vote classification model are tuned using a grid search strategy; the three feature subsets constructed in S4 are used to train the model independently to obtain the optimal model; S6: Predict the Re2TM 17The compositional characteristics of the rare earth soft magnetic alloy, as well as the statistical and thermodynamic characteristics obtained through steps S2 and S3, are input into the optimal artificial neural network regression model and Vote classification model obtained in step S5, respectively, and the corresponding magnetic properties and anisotropy classification prediction results are output.
[0007] According to a preferred embodiment of the present invention, in S1, the chemical composition of all data in the initial dataset is a 2:17 type rare earth intermetallic compound, the sum of the Re element proportions is 2, and the sum of the TM element proportions is 17. The composition characteristics refer to the atomic percentage of each constituent element in the chemical formula.
[0008] According to a preferred embodiment of the present invention, in S2, the physicochemical properties include atomic number, Mendeleev number, melting point, covalent radius, number of valence electrons in p / d / f orbitals, number of unfilled electrons in p / d / f orbitals, total number of valence electrons, ground state magnetic moment, space group number, electronegativity, spin quantum number, orbital quantum number, and total angular momentum quantum number; The thermodynamic characteristics include enthalpy of mixing and entropy of mixing.
[0009] According to a preferred embodiment of the present invention, in S4, for all statistical and thermodynamic characteristics, the Pearson correlation coefficient method is used to calculate the correlation coefficient between characteristics and the correlation coefficient between characteristics and target performance; Among them, only one of two features whose correlation coefficient is greater than the first threshold is retained; if the correlation coefficient between a feature and the target performance is less than the second threshold, the feature is removed from the feature subset of the target performance.
[0010] Furthermore, in S4, the first threshold is set to 0.85, and the second threshold is set to 0.15.
[0011] According to a preferred embodiment of the present invention, in step S4, the compositional features do not participate in the feature correlation analysis and dimensionality reduction; the final feature subsets of saturation magnetization, Curie temperature and magnetocrystalline anisotropy include compositional features, dimensionality-reduced statistical features and thermodynamic features.
[0012] According to a preferred embodiment of the present invention, in step S5, both independent artificial neural network regression models have two hidden layers, and each hidden layer has 300 neurons.
[0013] According to a preferred embodiment of the present invention, in step S5, the Vote classification model comprehensively evaluates the results of the outputs of the three base classifiers, K-nearest neighbors, random forest and support vector machine, and uses a voting mechanism to output the final magnetocrystalline anisotropy category of the alloy under test.
[0014] According to a preferred embodiment of the present invention, the grid search strategy used in step S5 includes: S51: Configure the hyperparameter search range for the three base classifier models: K-nearest neighbors, random forest, and support vector machine. S52: Traverse all hyperparameter combinations within the range and evaluate the performance of each parameter combination using ten-fold cross-validation; S53: Train the model using the parameter combination with the highest cross-validation score to obtain the corresponding optimal model as the base classifier.
[0015] The prediction method proposed in this invention can be used to predict multiple properties of 2:17 type rare earth soft magnetic alloys and can be applied to the screening and optimization of high frequency soft magnetic materials.
[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention achieves multi-performance prediction of 2:17 type rare-earth soft magnetic alloys by constructing a high-precision machine learning model. The models used for regression prediction of saturation magnetization and Curie temperature, as well as magnetocrystalline anisotropy classification prediction, exhibit high accuracy and stability. From 4,488,640 compositions, 2:17 type rare-earth soft magnetic alloy components with high saturation magnetization, high Curie temperature, and planar anisotropy were rapidly screened, shortening the R&D cycle of rare-earth soft magnetic alloys and providing reliable candidate formulations for high-frequency soft magnetic materials. Attached Figure Description
[0017] To illustrate the invention more clearly, a brief description will be provided in conjunction with the accompanying drawings. It should be understood that the drawings, together with the embodiments of the invention, are used to explain the invention and do not constitute a limitation thereof.
[0018] Figure 1 The above is a flowchart of a method for predicting and screening multiple properties of a 2:17 type rare earth soft magnetic alloy, provided in an embodiment of the present invention.
[0019] Figure 2 This is a topology diagram of the artificial neural network model used for regression prediction in an embodiment of the present invention.
[0020] Figure 3 This is a structural diagram of the Vote model used for classification prediction in an embodiment of the present invention. Detailed Implementation
[0021] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only for explaining the present invention and are not intended to limit the scope of protection of the present invention.
[0022] Example 1 This embodiment describes a method for predicting the multiple properties of a 2:17 type rare earth soft magnetic alloy, as shown in the attached figure. Figure 1 As shown, it includes the following steps: (1) Collect Re2TM from published literature17 The chemical composition data of the alloy, as well as the experimental values of the alloy's saturation magnetization, Curie temperature, and magnetocrystalline anisotropy, are used to form an initial dataset based on the alloy's compositional characteristics (compositional characteristics refer to the atomic percentage of each constituent element in the chemical formula) and target performance data. Among them, the Re element is selected from one or more combinations of La, Ce, Pr, Nd, Y, Lu, Ho, Er, Yb, Gd, Sm, and Dy, and TM is at least one of Fe and Co, or is composed of at least one of Fe and Co and one or more combinations of Cu, Ga, Ti, V, Cr, Mn, Mo, and Si, and the stoichiometric ratio of the Re element to the TM element is strictly limited to 2:17.
[0023] In this embodiment, after data cleaning and aggregation, the initial dataset contains data on three target performance parameters: 538 experimental data points on saturation magnetization, 484 experimental data points on Curie temperature, and 227 experimental data points on magnetocrystalline anisotropy. Alloy composition distribution includes: 29 binary compounds, 389 ternary compounds, 115 quaternary compounds, and 5 pentagonal compounds.
[0024] (2) Extract the physicochemical properties of each element in the chemical formula of the alloy. Based on the atomic ratio of each element in the chemical formula, calculate the maximum, minimum, weighted average and average deviation of each physicochemical property at the overall level of the alloy to generate statistical features. Calculate the thermodynamic features of the alloy and merge the statistical features and thermodynamic features of the alloy into the initial dataset of S1 to form a feature dataset. Among them, the physicochemical properties include atomic number, Mendeleev number, melting point, covalent radius, number of valence electrons in p / d / f orbitals, number of unfilled electrons in p / d / f orbitals, total number of valence electrons, ground state magnetic moment, space group number, electronegativity, spin quantum number, orbital quantum number, and total angular momentum quantum number; The thermodynamic characteristics include enthalpy of mixing and entropy of mixing.
[0025] (3) Based on the feature dataset constructed by S2, the data is standardized using StandardScale.
[0026] (4) The Pearson correlation coefficient method is used to reduce the dimensionality of the standardized statistical and thermodynamic characteristics (excluding the component characteristics), as shown in the following formula:
[0027] in and It is the average of two features in the sample. and These represent the actual values of the two features, and These represent the standard deviations of the two features, n The number of data points. r The value ranges from -1 to 1. r A value of 1 indicates a perfect positive correlation (the two variables are linearly related in the same direction), -1 indicates a perfect negative correlation (linearly related in opposite directions), and 0 indicates no linear correlation. The correlation coefficient between each pair of features is calculated, with a threshold of 0.85. When the correlation coefficient between two features is greater than 0.85, a high degree of collinearity is determined, and one feature is randomly removed, with the rest removed. Next, for saturation magnetization, Curie temperature, and magnetocrystalline anisotropy, separate dedicated feature subsets are constructed. When constructing a specific feature subset, the correlation coefficient between statistical and thermodynamic features and the corresponding target performance is calculated, with a threshold of 0.15. Features with an absolute correlation coefficient less than 0.15 are directly removed from that feature subset. After the above dimensionality reduction processing, separate dedicated feature subsets are constructed for saturation magnetization, Curie temperature, and magnetocrystalline anisotropy, with the number of features retained (including component features and dimensionality-reduced statistical and thermodynamic features) being 71, 64, and 61, respectively.
[0028] Before training begins, each feature subset is randomly divided into 80% of the data to form the training set and the remaining 20% to form the test set. The training set is used for model training and parameter updates, while the validation set is used to evaluate model performance.
[0029] (5) For saturation magnetization and Curie temperature, construct independent artificial neural network regression prediction models (ANN models), such as Figure 2 As shown, the ANN model structure includes an input layer, two hidden layers, and an output layer. Each hidden layer has 300 neurons, and the number of iterations is set to 500. The performance of the regression prediction model is evaluated using the coefficient of determination (R²). 2 ) and root mean square error (RMSE).
[0030] The corresponding formula is as follows:
[0031]
[0032] in This represents the actual value on the test set. For predicted values, n This represents the number of samples.
[0033] The hyperparameters of the artificial neural network regression model are: activation='ReLU'; learning_rate=0.001; optimizer='Adam'.
[0034] For discretely distributed magnetocrystalline anisotropy (classification and prediction task), a Vote classification model integrating three heterogeneous base classifiers—K-Nearest Neighbors (KNN), Random Forest (RF), and Support Vector Machine (SVM)—was constructed, such as... Figure 3 As shown. To obtain the best generalization performance, a grid search strategy was used to tune the hyperparameters of the three heterogeneous base classifiers: first, hyperparameter search spaces were configured for KNN, RF, and SVM respectively; then, all hyperparameter combinations within the range were traversed, and the performance of each parameter combination was evaluated through ten-fold cross-validation; finally, the parameter combination with the highest cross-validation score was extracted to train the model, forming the optimal base classifier. The Vote ensemble model comprehensively evaluated the outputs of the three optimal base classifiers, and the final magnetocrystalline anisotropy category of the alloy under test was output using a majority voting mechanism. The classification prediction model used classification accuracy and F1 score as performance evaluation indicators.
[0035] The formula for accuracy is as follows:
[0036] The formula for the F1 score is as follows:
[0037] Where TP represents the number of true positive samples, TN represents the number of true negative samples, FP represents the number of false positive samples, and FN represents the number of false negative samples.
[0038] The scope of hyperparameter optimization for KNN classification models using grid search includes: n_neighbors is 5 to 15, with a step size of 1; weights='distance' or 'uniform'; p = 1 or 2.
[0039] The scope of hyperparameter optimization for RF classification models using grid search includes: n_estimators is between 100 and 300, with a step size of 5; max_depth is 5 to 16, and the step size is 1; min_samples_split is 1 to 5, with a step size of 1; class_weight='balanced'.
[0040] The scope of hyperparameter optimization for SVM classification models using grid search includes: kernel='rbf' or 'linear' or 'poly'; C ranges from 0.5 to 100, with a step size of 0.5; gamma='scale'; class_weight='balanced'.
[0041] The hyperparameters of the Vote classification model are: estimators=['rfc','knn','svc']; voting='hard'.
[0042] (6) In this embodiment, the general formula of the virtual alloy to be predicted is set to Re2(Fe,M). 17 Furthermore, each chemical formula is strictly limited to containing four elements. Specifically: Re is limited to one or two of La, Ce, Pr, Nd, and Y; M is limited to one or two of Al, Si, Ti, V, Mn, Co, Cu, and Ga. When generating virtual chemical formulas, the gradient evolution step size for each element is set to 0.1, and a hard constraint rule is applied: the sum of the proportions of Re elements in the same chemical formula must be 2, and the sum of the proportions of Fe and M elements must be 17.
[0043] (7) For the data in the virtual component space generated in step (6), perform the feature generation process in step (2) and the feature standardization process in step (3) in sequence; then, according to the feature retention index determined in step (4), directly extract the corresponding test feature vector; then input the feature vector into the artificial neural network regression model and the Vote classification model trained in step (5) respectively. Output the corresponding magnetic properties and anisotropy classification prediction results.
[0044] In practical applications, the virtual alloy components can be further screened based on the predicted results of the aforementioned magnetic properties and anisotropy classification. In one embodiment, the screening criteria for virtual components are set as follows: saturation magnetization greater than 150 emu / g, Curie temperature greater than 800 K, and magnetocrystalline anisotropy as planar anisotropy. In this embodiment, virtual alloy components whose predicted results meet the screening criteria are retained.
[0045] Following the screening process described above, five samples were randomly selected from the results for experimental verification. The results are as follows:
[0046] The experimental saturation magnetization, Curie temperature, and magnetocrystalline anisotropy type of the selected samples all meet the screening requirements. The root mean square error (RMSE) between the predicted and experimental values of saturation magnetization and Curie temperature is only 4.54 emu / g and 34.99 K, respectively, while the prediction accuracy of magnetocrystalline anisotropy is 100%.
[0047] As can be seen from the above examples, the method for constructing a multi-performance prediction model for 2:17 type rare earth soft magnetic alloys provided by this invention can achieve rapid and accurate screening of multiple properties of 2:17 type rare earth soft magnetic alloys, and can effectively shorten the research and development cycle of new rare earth soft magnetic alloys.
[0048] The above descriptions are merely specific embodiments of the present invention, intended only to illustrate the technical solutions and core concepts of the present invention. It should be noted that these embodiments do not exhaustively describe all details and are not intended to limit the present invention. Many modifications and variations can be made based on the content of this specification, and any modifications, additions, or similar substitutions made within the scope of the principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A method for predicting multiple properties of a 2:17 type rare earth soft magnetic alloy, characterized in that, The method includes the following steps: S1: Collect Re2TM 17 The chemical composition data of rare earth soft magnetic alloys and the corresponding experimental values of target performance are used to form an initial dataset based on the composition characteristics of the alloys and the target performance data. Among them, the target performance includes saturation magnetization, Curie temperature and magnetocrystalline anisotropy. Re is selected from one or more combinations of La, Ce, Pr, Nd, Y, Lu, Ho, Er, Yb, Gd, Sm and Dy. TM is at least one of Fe and Co, or is composed of at least one of Fe and Co and one or more combinations of Cu, Ga, Ti, V, Cr, Mn, Mo and Si. S2: Extract the physicochemical properties of each element in the alloy's chemical formula. Based on the atomic percentage of each element in the chemical formula, calculate the maximum, minimum, weighted average, and average deviation of each physicochemical property at the overall alloy level to generate statistical characteristics. Calculate the thermodynamic characteristics of the alloy and merge the statistical and thermodynamic characteristics of the alloy into the initial dataset of S1 to form a feature dataset. S3: Based on the feature dataset constructed from S2, the feature data is standardized. S4: The Pearson correlation coefficient method is used to perform feature correlation analysis and dimensionality reduction on the statistical and thermodynamic features of the dataset after standardization in step S3, and feature subsets of saturation magnetization, Curie temperature and magnetocrystalline anisotropy are constructed respectively. S5: For the regression prediction tasks of saturation magnetization and Curie temperature, two independent artificial neural network regression models are constructed respectively; for the classification prediction task of magnetocrystalline anisotropy, a Vote classification model integrating three base classifiers, K-nearest neighbors, random forest and support vector machine, is constructed, and the hyperparameters of the Vote classification model are tuned using a grid search strategy; the three feature subsets constructed in S4 are used to train the model independently to obtain the optimal model; S6: Predict the Re2TM 17 The compositional characteristics of the rare earth soft magnetic alloy, as well as the statistical and thermodynamic characteristics obtained through steps S2 and S3, are input into the optimal artificial neural network regression model and Vote classification model obtained in step S5, respectively, and the corresponding magnetic properties and anisotropy classification prediction results are output.
2. The method for predicting multiple properties of a 2:17 type rare earth soft magnetic alloy according to claim 1, characterized in that, In S1, the chemical composition of all data in the initial dataset is a 2:17 type rare earth intermetallic compound, the sum of the Re element proportions is 2, and the sum of the TM element proportions is 17. The composition characteristics refer to the atomic percentage of each constituent element in the chemical formula.
3. The method for predicting multiple properties of a 2:17 type rare earth soft magnetic alloy according to claim 1, characterized in that, In S2, the physicochemical properties include atomic number, Mendeleev number, melting point, covalent radius, number of valence electrons in p / d / f orbitals, number of unfilled electrons in p / d / f orbitals, total number of valence electrons, ground state magnetic moment, space group number, electronegativity, spin quantum number, orbital quantum number, and total angular momentum quantum number; the thermodynamic characteristics include enthalpy of mixing and entropy of mixing.
4. The method for predicting multiple properties of a 2:17 type rare earth soft magnetic alloy according to claim 1, characterized in that, In S4, for all statistical and thermodynamic features, the Pearson correlation coefficient method is used to calculate the correlation coefficient between features and the correlation coefficient between features and target performance. Among them, only one of two features whose correlation coefficient is greater than the first threshold is retained; if the correlation coefficient between a feature and the target performance is less than the second threshold, the feature is removed from the feature subset of the target performance.
5. The method for predicting multiple properties of a 2:17 type rare earth soft magnetic alloy according to claim 1, characterized in that, In step S4, the composition features are not involved in the feature correlation analysis and dimensionality reduction; the final feature subsets of saturation magnetization, Curie temperature and magnetocrystalline anisotropy include composition features, dimensionality-reduced statistical features and thermodynamic features.
6. The method for predicting multiple properties of a 2:17 type rare earth soft magnetic alloy according to claim 1, characterized in that, In step S5, both independent artificial neural network regression models have two hidden layers, with 300 neurons in each hidden layer.
7. The method for predicting multiple properties of a 2:17 type rare earth soft magnetic alloy according to claim 1, characterized in that, In step S5, the Vote classification model comprehensively evaluates the results of the three base classifiers, namely K-nearest neighbors, random forest, and support vector machine, and uses a voting mechanism to output the final magnetocrystalline anisotropy category of the alloy under test.
8. The method for predicting multiple properties of a 2:17 type rare earth soft magnetic alloy according to claim 1, characterized in that, The grid search strategy used in step S5 includes: S51: Configure the hyperparameter search range for the three base classifier models: K-nearest neighbors, random forest, and support vector machine. S52: Traverse all hyperparameter combinations within the range and evaluate the performance of each parameter combination using ten-fold cross-validation; S53: Train the model using the parameter combination with the highest cross-validation score to obtain the corresponding optimal model as the base classifier.