Low-fault-energy high-entropy alloy-oriented component design method based on machine learning

Abstract

The invention provides a low-fault-energy high-entropy alloy-oriented component design method based on machine learning, and the method comprises the steps: building a high-entropy alloy component-fault energy data set through literature investigation, and building physical and thermodynamic feature descriptors; carrying out Pearson correlation analysis and feature importance analysis on the feature descriptors, and screening out key features which have the front-rank influence on alloy stacking fault energy; then, an optimal feature set of the model is screened out through a forward selection strategy, first three machine learning models with optimal performance are screened out, and then a machine learning model used for constructing an integrated algorithm framework is determined; and the component space of the candidate alloy is determined according to the component range of the alloy, the candidate alloy components used for prediction of the integrated algorithm framework are finally screened out, the candidate alloy components are predicted through the integrated algorithm framework, and the high-entropy alloy components with the low stacking fault energy are screened out. According to the method, the design accuracy and efficiency of the low-fault-energy high-entropy alloy are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of alloy composition design, and in particular to a composition design method for low stacking fault energy and high entropy alloys based on machine learning. Background Art

[0002] With the rapid development of productivity, modern industries such as aerospace, energy, military defense, transportation and other fields have put forward higher requirements for metal materials, especially for structural materials in extreme low temperature environments. Brittle fracture often occurs when they are used at low temperatures, resulting in casualties and waste of resources. However, traditional alloying strategies limit the number of possible alloy element combinations, and the improvement of alloy performance is increasingly approaching a bottleneck. Therefore, it is imperative to find new alloy design concepts and develop high-performance low-temperature structural materials. As a pioneer in the study of high-entropy alloys, face-centered cubic (FCC) structured CoCrFeNiMn-based high-entropy alloys have received widespread attention. This type of alloy has a wide adjustable range of stacking fault energy (SFE), and at low temperatures, the stacking fault energy of the alloy will be further reduced, and the low stacking fault energy can introduce a rich deformation mechanism into the alloy, thereby improving the low-temperature toughness of the alloy.

[0003] However, this vast composition space also brings complexity and uncertainty to the design of alloy composition. At present, the composition design of low stacking fault energy CoCrFeNiMn-based high entropy alloys mainly relies on the method of fine-tuning the content of alloying elements and empirical formulas. There is blindness in the estimation of stacking fault energy and a large amount of experimental data is required, which not only increases the cost of alloy design, but also makes it difficult to quickly and accurately explore the composition of low stacking fault energy alloys.

[0004] In recent years, with the development of computer science and artificial intelligence, new ideas have been provided for the composition design of alloys. As a data-driven science, machine learning has been widely used in composition design, process optimization and other aspects in the field of materials. Its active learning strategy can push the alloy composition from low dimension to high dimension in a short time, greatly reducing the cost of alloy development. However, due to the very limited experimental data on the stacking fault energy of CoCrFeNiMn-based high entropy alloys, there is very little research on the prediction of stacking fault energy of high entropy alloys by machine learning. Summary of the invention

[0005] In view of the problems existing in the above-mentioned background technology, the present invention provides a composition design method for low stacking fault energy high entropy alloys based on machine learning. Based on the composition-stacking fault energy simulation data in the literature, the physical and thermodynamic parameters obtained by screening are used to improve the accuracy of the machine learning model, and the phases are initially screened according to the VEC criterion. Finally, a machine learning integrated algorithm framework is used to screen out high entropy alloy components with low stacking fault energy, thereby improving the accuracy and efficiency of the composition design of CoCrFeNiMn-based low stacking fault energy high entropy alloys.

[0006] Specifically, the present invention provides a composition design method for low stacking fault energy high entropy alloys based on machine learning, comprising the following steps: Step 1: Collect stacking fault energy simulation data, establish a simulation data set of alloy composition-stacking fault energy, calculate corresponding physical and thermodynamic characteristic parameters according to the alloy composition, and analyze the alloy composition distribution in the simulation data set; Step 2: Analyze the correlation between the characteristic parameters pairwise by the Pearson correlation coefficient, remove the characteristic parameters with large correlation coefficient, and then use the random forest feature importance analysis method to screen out the key characteristic parameters affecting the stacking fault energy; Step 3: using the alloy composition and the key characteristic parameters as inputs of the machine learning model, using the stacking fault energy as output of the machine learning model, adopting a forward selection strategy for the key characteristic parameters, screening out the best feature sets of different machine learning models, thereby obtaining a variety of high-precision machine learning models; Step 4: Filter out the models with the highest performance among the various high-precision machine learning models obtained, adopt the idea of ​​parallel prediction, and build an integrated algorithm prediction framework for subsequent predictions; Step 5: According to the distribution of alloy components obtained in step 1, determine the preliminary composition space of the candidate alloys to form a candidate alloy component data set, and then use the integrated algorithm prediction framework to predict the stacking fault energy of the candidate alloy component data set, rank the alloy components according to their scores, and finally screen out candidate alloy components with low stacking fault energy.

[0007] As a further illustration of the present invention, the simulation data set of alloy composition-stacking fault energy includes alloy elements Co, Cr, Fe, Ni, Mn, V, Al and stacking fault energy values.

[0008] As a further illustration of the present invention, the key characteristic parameters include: specific heat, Pauling electronegativity variance, lattice stability calculated by first principles, elastic constants, variance of lattice stability calculated by phase diagram, lattice stability calculated by phase diagram, elastic constants, ionization energy, atomic weight, Allen electronegativity variance, and Pauling electronegativity.

[0009] As a further illustration of the present invention, the machine learning models include: multilayer perceptron regression, support vector machine regression model, extreme gradient boosting regression model, gradient boosting regression model and random forest regression model.

[0010] As a further illustration of the present invention, the optimal feature set includes: 6 features of the multilayer perceptron regression model, 8 features of the support vector machine regression model, 9 features of the extreme gradient boosting regression model, 11 features of the gradient boosting regression model, and 4 features of the random forest regression model.

[0011] As a further illustration of the present invention, the models with the highest performance rankings selected in step 4 are specifically: a multilayer perceptron regression model, a support vector machine regression model, and an extreme gradient boosting regression model.

[0012] As a further illustration of the present invention, the prediction accuracy of the models screened in step 4 is greater than 90%, and the selected performance evaluation indicators include goodness of fit and root mean square error. When evaluating the performance of the machine learning model, the validation set adopts a 10-fold cross-validation strategy, and the test set directly evaluates its prediction accuracy.

[0013] As a further illustration of the present invention, in step 5, the composition space is restricted using the VEC criterion to obtain an alloy composition that does not contain a BCC harmful phase.

[0014] As a further illustration of the present invention, in step 5, the integrated algorithm framework calculates the weighted average and standard deviation of the stacking fault energy prediction values ​​of the top performance ranking models screened out in step 4, and then calculates the scores of the candidate alloy components using a strategy that balances model stability and uncertainty, and ultimately screens out high-entropy alloy components with low stacking fault energy and excellent strength and plasticity.

[0015] As a further illustration of the present invention, when the score of each candidate alloy component is obtained through the integrated algorithm framework, the specific formula of the score function is: ; Among them, Mean is the weighted mean and Std is the standard deviation.

[0016] Compared with the prior art, the present invention has the following beneficial technical effects: The composition design method provided by the present invention solves the problem of the current lack of experimental data on stacking fault energy based on the collected simulation data of alloy stacking fault energy; additional physical and thermodynamic descriptors are established according to the chemical composition of the alloy, and key features affecting the stacking fault energy of the alloy are screened out through Pearson correlation analysis and random forest feature importance analysis; the best feature set of each machine learning model is obtained through the forward selection strategy of features, and the complex nonlinear relationship between alloy composition, characteristic parameters and alloy stacking fault energy is successfully established, and this strategy improves the prediction accuracy of the machine learning model; the VEC criterion is used to limit the phase composition of the alloy in the original composition space, which narrows the candidate alloy composition space and avoids the model from falling into the local optimal trap; an integrated algorithm framework is successfully created using multiple high-precision machine learning models. Thanks to its unique scoring strategy, the instability of single model prediction is avoided, and it is expected to obtain a high-entropy alloy composition with low stacking fault energy and excellent strength-ductility matching, thereby improving the development efficiency of new high-entropy alloys for low-temperature applications.

[0017] Other features and advantages of the technical solution will be described in the subsequent description, and partly become apparent from the description, or understood by implementing the technical solution. The purpose and other advantages of the technical solution can be achieved and obtained through the structures specifically pointed out in the written description and the drawings.

[0018] The technical solution of the present technical solution is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] The accompanying drawings are used to provide a further understanding of the technical solution and constitute a part of the specification. Together with the embodiments of the technical solution, they are used to explain the technical solution and do not constitute a limitation of the technical solution. In the accompanying drawings: Figure 1 This is a flow chart of the composition design of the low stacking fault energy single-phase FCC high entropy alloy in Example 1 of the present invention.

[0020] Figure 2 This is a distribution diagram of alloy components in the original stacking fault energy data set of the present invention.

[0021] Figure 3 It is the Pearson correlation analysis diagram of all physical and thermodynamic parameters in the present invention.

[0022] Figure 4 This is a feature importance ranking diagram obtained by using the random forest algorithm in the present invention.

[0023] Figure 5 Performance evaluation chart of various machine learning models selected in the present invention.

[0024] Figure 6These are scatter plots of the various machine learning models used in the present invention, wherein Figure (a) to Figure (e) are scatter plots of five different machine learning models, respectively.

[0025] Figure 7 is the XRD pattern of the candidate alloy in the present invention.

[0026] Figure 8 1 and 2 are engineering stress-strain curves of candidate alloys of the present invention, wherein Figure (a) to Figure (e) are engineering stress-strain curves of alloy 1 to alloy 5, respectively.

[0027] Fig. 9 1 is a work hardening rate curve diagram of the candidate alloys in the present invention, wherein Figure (a) to Figure (e) are work hardening rate curve diagrams from alloy 1 to alloy 5, respectively.

[0028] Fig.10 TEM analysis diagrams of the candidate alloys of the present invention after being pulled apart, and Figure (a) to Figure (e) are TEM analysis diagrams of alloy 1 to alloy 5, respectively. DETAILED DESCRIPTION

[0029] The preferred embodiments of the present technical solution are described below in conjunction with the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present technical solution, and are not used to limit the present technical solution. Example 1

[0030] like Figure 1 As shown, this embodiment provides a composition design method for low stacking fault energy high entropy alloys based on machine learning, comprising the following steps: Step 1: Collect stacking fault energy simulation data, establish a simulation data set of alloy composition-stacking fault energy, calculate the corresponding physical and thermodynamic characteristic parameters according to the alloy composition, and analyze the alloy composition distribution in the simulation data set. The results are as follows: Figure 2 shown.

[0031] Due to the lack of experimental data on stacking fault energy of high entropy alloys, the stacking fault energy data contained in the created data set are all from the results of stacking fault energy simulations such as first principles and molecular dynamics. Specifically, the stacking fault energy simulation data are obtained by collecting stacking fault energy simulation data from the literature. The alloy composition-stacking fault energy simulation data set includes Co, Cr, Fe, Ni, Mn, V, Al alloy elements and stacking fault energy values. The data set contains data on the stacking fault energy of CoCrFeNiMn-based high entropy alloys, and a total of 498 sets of data were collected.

[0032] Specifically, there are 34 corresponding physical and thermodynamic characteristic parameters calculated according to the alloy composition, including VASP-PAW-GGA LSE, SGTE LSE, Allen EN, Pauling EN, Density, Atomic Weight, Valencee, IE, Melting Point, Specific Heat, Metallic Radius, Shear Modulus, Total Electrons, Atomic Planar Density, C_11, C_12, C' and their respective variances.

[0033] Step 2: Analyze the correlation between the two characteristic parameters through the Pearson correlation coefficient, eliminate the characteristic parameters with large correlation coefficients, and then use the random forest feature importance analysis method to screen out the key characteristic parameters that affect the stacking fault energy.

[0034] Specific results such as Figure 3 and Figure 4 As shown, the key characteristic parameters selected (sorted by feature importance) include: specific heat (Specific Heat), Pauling electronegativity variance (Pauling EN_var), lattice stability calculated by first principles (VASP-PAW-GGA LSE), elastic constant C_12, variance of lattice stability calculated by phase diagram (SGTE LSE_var), lattice stability calculated by phase diagram (SGTE LSE), elastic constant C_11, ionization energy (IE), atomic weight (AtomicWeight), Allen electronegativity variance (Allen EN_var), and Pauling electronegativity (Pauling EN).

[0035] Specifically, when removing features through Pearson correlation analysis, if a feature is highly correlated with multiple features (> 0.9), it will be deleted to avoid multicollinearity problems; when analyzing feature importance, the features are summed up according to their importance ranking, and all features whose sum of feature importance is greater than 0.98 are retained.

[0036] Step 3: Use the alloy composition and key characteristic parameters as the input of the machine learning model, and the stacking fault energy as the output of the machine learning model. Use a forward selection strategy for the key characteristic parameters to screen out the best feature sets of different machine learning models, thereby obtaining a variety of high-precision machine learning models.

[0037] Specifically, the machine learning models used include: Multilayer Perceptron Regression (MLP Regression, MLPR), Support Vector Regression Model (Support Vector Regression, SVR), Extreme Gradient Boosting Regression Model (XGBoostRegression, XGBR), Gradient Boosting Regression Model (Gradient Boosting Regression, GBR), Random Forest Regression Model (Random Forest Regression, RFR).

[0038] Specifically, the forward selection strategy is a method that reduces the number of features one by one according to the order of feature importance. For example, a feature set with 5 features deletes the feature with the sixth highest feature importance compared to a feature set with 6 features. Using this strategy, we get the feature set with the best model performance. We use 5 machine learning models to fit the nonlinear relationship between input and output. The performance of the models on the validation set (10-CV) and the test set is shown in the figure below. Figure 5 and Figure 6 The specific performance indicators are shown in Table 1: Table 1 Performance indicators of five machine learning models on the validation set (10-CV) and test set

[0039] In summary, the optimal number of features (sorted by feature importance) is: MLPR (6), SVR (8), XGBR (9), GBR (11), and RFR (4).

[0040] Step 4: For the various high-precision machine learning models obtained, select the models with the highest performance rankings, adopt the idea of ​​parallel prediction, and build an integrated algorithm prediction framework for subsequent predictions.

[0041] Preferably, the top three models with excellent performance are screened out, specifically: MLPR model, SVR model and XGBR model.

[0042] The prediction accuracy of the models selected in step 4 is greater than 90%. The performance evaluation indicators selected include goodness of fit and root mean square error. When evaluating the performance of machine learning models, the validation set adopts a 10-fold cross-validation strategy, and the test set directly evaluates its prediction accuracy.

[0043] The parallel prediction method adopted by the present invention avoids the uncertainty of single model prediction and improves the stability of model prediction, thereby facilitating the screening of candidate alloys that meet the performance requirements.

[0044] Step 5: Based on the distribution of alloy components obtained in step 1, determine the preliminary composition space of the candidate alloys to form a candidate alloy composition data set. Then, use the integrated algorithm prediction framework to predict the stacking fault energy of the candidate alloy composition data set, rank the alloy components according to their scores, and finally screen out candidate alloy components with low stacking fault energy.

[0045] Furthermore, in order to obtain an alloy composition without the harmful BCC phase, the VEC criterion is used to limit the composition space. Specifically, when the VEC of the alloy is greater than 8, the brittle phase BCC phase will not appear in the alloy. By limiting the VEC criterion, the alloy composition space is reduced from 500,000 groups to 295,363 groups.

[0046] The stacking fault energy of these 295363 groups of candidate alloys was predicted using the integrated algorithm framework established above. According to the framework's unique scoring strategy (calculation of weighted average and standard deviation), the score of each candidate alloy was obtained, and the alloys were sorted and screened. Finally, five candidate alloy compositions with low stacking fault energy and excellent strength and plasticity were obtained. The specific alloy compositions, weighted average values ​​of stacking fault energy and scores are shown in Table 2.

[0047] Table 2 Nominal compositions (at.%) of five low stacking fault energy candidate alloys

[0048] When calculating the score for each candidate alloy composition, the specific formula of the scoring function is: ; Among them, Mean is the weighted mean and Std is the standard deviation. Example 2

[0049] The deformation mechanism of the candidate alloy components obtained by the design method provided in Example 1 was verified: Step 6: According to the candidate alloy composition in step 5, a candidate alloy ingot is prepared by vacuum arc melting method.

[0050] Specifically, pure metals of Fe, Mn, Ni, Co, Cr, V, and Al with a purity of more than 99.99% wt.% were selected for the preparation of ingots. All alloys were prepared by arc melting under Ti aspiration and high-purity argon atmosphere. According to the alloy composition recommended by the model, 50g ingots were produced according to their atomic ratios. In order to ensure the uniformity of the composition of the casting alloy, the raw materials must be repeatedly turned over 5 times during melting.

[0051] Step 7: Conduct phase identification and mechanical property testing on the candidate high entropy alloy ingot prepared in step 6.

[0052] Specifically, a Bruker D8 X-ray diffractometer was used for phase analysis, using Cu-Kα radiation, a scanning speed of 4° / min, and a scanning range of 20°-100°. Figure 7 As shown, the brittle phase BCC does not exist in any of the five candidate alloys. A small amount of HCP phase appears in alloy 1. The other four candidate alloys are all single-phase FCC. The phase composition of the candidate alloys meets the design expectations. The samples used for the tensile test were cut from the ingots and processed into rectangular dog-bone specimens with a geometric shape of (26.0 mm × 6 mm × 1.5 mm). A universal testing machine (Shimadzu AG-X Plus) was used to perform room temperature tensile tests at least 3 times, and the strain rate was selected as 0.001 / s. The experimental results are shown in Figure 8 and Fig. 9 As shown in the figure, the yield strength of the cast alloy is between 200-400MPa, the tensile strength is between 500-650MPa, and the alloy has excellent plasticity. Alloy 1 has premature stress concentration at the phase interface due to the presence of martensite, resulting in premature fracture. From the work hardening rate curve of the five candidate alloys, it can be seen that the work hardening ability of the five candidate alloys is excellent, and the single-phase FCC alloy has a secondary increase in the work hardening rate, indicating that other deformation mechanisms occur during the tensile process.

[0053] Step 8: Perform TEM analysis on the five candidate alloy tensile parts after tensile fracture to determine the deformation mechanism during the tensile process.

[0054] Specifically, a Talos F200X transmission electron microscope (TEM) was used for microstructure analysis. The TEM sample preparation process was as follows: the sample was polished to a thickness of less than 100 μm, and then a Struers Tenupol-5 electrolytic double-spray instrument was used for double-spray thinning (the double-spray liquid was a perchloric acid + ethanol solution with a volume ratio of 1:9, the operating voltage was 30 V, and the operating temperature was -30 °C). The experimental results are shown in Figure 2. Fig.10 As shown in the figure, low stacking fault energy alloys will have deformation twins, HCP phases and stacking faults during tensile deformation. From the TEM results, it can be seen that alloy 1 has HCP phase during tensile deformation, that is, its deformation mechanism is transition induced plasticity (TRIP). For the other four single-phase FCC candidate alloys, with the increase of predicted stacking fault energy, the deformation mechanism in the alloy changes from crossed deformation twins (alloy 3) to deformation twins and stacking faults (alloys 2 and alloy 4) and then to high-density stacking faults (alloy 5). It can be seen that the deformation mechanism of the alloy corresponds to the predicted value of the stacking fault energy of the model, indicating that the constructed machine learning integrated algorithm framework has excellent generalization and stability.

[0055] Obviously, those skilled in the art can make various changes and modifications to the technical solution without departing from the spirit and scope of the technical solution. Thus, if these modifications and variations of the technical solution fall within the scope of the claims of the technical solution and their equivalents, the technical solution is also intended to include these modifications and variations.

Claims

1. A composition design method for low stacking fault energy high entropy alloys based on machine learning, characterized in that: The steps include: Step 1: Collect stacking fault energy simulation data, establish a simulation data set of alloy composition-stacking fault energy, calculate corresponding physical and thermodynamic characteristic parameters according to the alloy composition, and analyze the alloy composition distribution in the simulation data set; Step 2: Analyze the correlation between the characteristic parameters pairwise by the Pearson correlation coefficient, remove the characteristic parameters with large correlation coefficient, and then use the random forest feature importance analysis method to screen out the key characteristic parameters affecting the stacking fault energy; Step 3: using the alloy composition and the key characteristic parameters as inputs of the machine learning model, using the stacking fault energy as output of the machine learning model, adopting a forward selection strategy for the key characteristic parameters, screening out the best feature sets of different machine learning models, thereby obtaining a variety of high-precision machine learning models; Step 4: Filter out the models with the highest performance among the various high-precision machine learning models obtained, adopt the idea of ​​parallel prediction, and build an integrated algorithm prediction framework for subsequent predictions; Step 5: According to the distribution of alloy components obtained in step 1, determine the preliminary composition space of the candidate alloys to form a candidate alloy component data set, and then use the integrated algorithm prediction framework to predict the stacking fault energy of the candidate alloy component data set, rank the alloy components according to their scores, and finally screen out candidate alloy components with low stacking fault energy.

2. The composition design method for low stacking fault energy high entropy alloys based on machine learning as claimed in claim 1, characterized in that: The alloy composition-stacking fault energy simulation data set includes Co, Cr, Fe, Ni, Mn, V, Al alloy elements and stacking fault energy values.

3. The composition design method for low stacking fault energy high entropy alloy based on machine learning as claimed in claim 1, characterized in that: The key characteristic parameters include: specific heat, Pauling electronegativity variance, lattice stability calculated by first principles, elastic constants, variance of lattice stability calculated by phase diagram, lattice stability calculated by phase diagram, elastic constants, ionization energy, atomic weight, Allan electronegativity variance, and Pauling electronegativity.

4. The composition design method for low stacking fault energy high entropy alloys based on machine learning as claimed in claim 1, characterized in that: The machine learning models include: multi-layer perceptron regression, support vector machine regression model, extreme gradient boosting regression model, gradient boosting regression model and random forest regression model.

5. The composition design method for low stacking fault energy high entropy alloy based on machine learning as claimed in claim 1, characterized in that: The optimal feature set includes: features of the multilayer perceptron regression model, 8 features of the support vector machine regression model, 9 features of the extreme gradient boosting regression model, 11 features of the gradient boosting regression model, and 4 features of the random forest regression model.

6. The composition design method for low stacking fault energy high entropy alloy based on machine learning as claimed in claim 1, characterized in that: The models with the highest performance rankings selected in step 4 are specifically: multilayer perceptron regression model, support vector machine regression model, and extreme gradient boosting regression model.

7. The composition design method for low stacking fault energy high entropy alloy based on machine learning as claimed in claim 1, characterized in that: The prediction accuracy of the models selected in step 4 is greater than 90%. The performance evaluation indicators selected include goodness of fit and root mean square error. When evaluating the performance of machine learning models, the validation set adopts a 10-fold cross-validation strategy, and the test set directly evaluates its prediction accuracy.

8. The composition design method for low stacking fault energy high entropy alloy based on machine learning as claimed in claim 1, characterized in that: In step 5, the VEC criterion is used to restrict the composition space to obtain an alloy composition that does not contain BCC harmful phases.

9. The composition design method for low stacking fault energy high entropy alloy based on machine learning as claimed in claim 1, characterized in that: In step 5, the integrated algorithm framework calculates the weighted average and standard deviation of the stacking fault energy prediction values ​​of the top performance models screened out in step 4, and then calculates the scores of the candidate alloy components using a strategy that balances model stability and uncertainty, and finally screens out high-entropy alloy components with low stacking fault energy and excellent strength and plasticity.

10. The composition design method for low stacking fault energy high entropy alloy based on machine learning as claimed in claim 9, characterized in that: When the score of each candidate alloy composition is obtained through the integrated algorithm framework, the specific formula of the score function is: ； Among them, Mean is the weighted mean and Std is the standard deviation.

Images