Metal ordered phase property prediction model based on machine learning and modeling method and application thereof

Through the metal ordered phase properties prediction model based on machine learning, the problem of time-consuming and low accuracy in predicting the metal ordered phase properties in traditional methods is solved, and fast and accurate orderly phase properties prediction and multivariate system properties prediction are achieved. Improve material design efficiency.

CN118942592APending Publication Date: 2024-11-12CENT SOUTH UNIV

Patent Information

Application Number
CN202411039812.7
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-31
Publication Date
2024-11-12

AI Technical Summary

Technical Problem

Traditional methods have problems of time-consuming, high cost and low accuracy when predicting the properties of metal ordered phases, especially in the design of multiple ordered phases, which leads to the extension of new material development cycle and the increase in material design risks.

Method used

Using a metal ordered phase property prediction model based on machine learning, the metal ordered phase is constructed by collecting and processing original data, machine learning feature engineering and model training, and combining quantum chemistry calculations to predict the thermal and mechanical properties of the ordered phases. database.

Benefits of technology

It achieves rapid and accurate prediction of ordered phase properties, avoids blind experiments, saves resources and time, improves material design efficiency, and overcomes the dimensional disaster problem in large-scale quantum chemocomputing.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118942592A_ABST
    Figure CN118942592A_ABST
Patent Text Reader

Abstract

The invention discloses a metal ordered phase property prediction model based on machine learning and a modeling method and application thereof. According to the model, firstly, metal ordered phase original data are collected, different ordered phase crystal structures are obtained in batches through Pymatgen, first-principle high-throughput calculation is carried out, formation energy and mechanical property data of the different ordered phase crystal structures are obtained, then machine learning feature engineering processing and data preprocessing are carried out according to the obtained data, a data set is obtained, and the data set is used for data processing. Training, testing, optimizing and evaluating a plurality of machine learning models, selecting the model with the optimal comprehensive evaluation as a prediction model, and finally comparing the prediction result of the model with the calculation result of the first high-throughput principle to verify the accuracy of the model. The ordered phase database constructed by using the model provided by the invention has the advantages of small calculation amount, quick response and the like on the premise of quantum chemistry accuracy, and provides theoretical guidance for metal material design.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to a metal ordered phase property prediction model, and in particular to a metal ordered phase property prediction model based on machine learning, a modeling method and an application thereof, and belongs to the field of metal ordered phase design. Background Art

[0002] Metal ordered phase refers to a structural variant with a specific atomic arrangement order. This ordered arrangement can give the material unique properties and characteristics, and usually exhibits excellent mechanical and thermal properties. Metal ordered phase can enhance the mechanical properties of the material and improve its strength, hardness and toughness. For example, in tungsten solid solution, different ordered phase structures and components have different effects on its mechanical properties. Some ordered phases can soften the tungsten solid solution, while some ordered phases will harden the tungsten solid solution, limiting the wide application of tungsten alloys. Research on the stability and mechanical properties of the ordered phase can provide theoretical guidance for material design to improve the toughness of tungsten alloys. Therefore, the study and control of metal ordered phases is of great significance for the development of new materials and the improvement of material properties, and has a far-reaching impact on the development of materials engineering and advanced manufacturing technology.

[0003] However, traditional experimental and theoretical calculation methods have limitations in predicting the properties of metal ordered phases, and are time-consuming, costly, and inaccurate, especially in the design of multi-ordered phases. These problems extend the development cycle of new materials and increase risks in the material design process. Summary of the invention

[0004] In view of the problems existing in the prior art, the first purpose of the present invention is to provide a prediction model for the properties of metal ordered phases based on machine learning. The model is based on a machine learning prediction model to assist in predicting the heat of formation and mechanical properties of ordered phases of any composition, which can search more quickly, avoid blind experiments, save resources and time, and provide theoretical guidance for material design and preparation, thereby improving efficiency.

[0005] The second object of the present invention is to provide a modeling method based on machine learning metal ordered phase property prediction model. This method can perform quantum chemical calculations with a small number of samples, and then combine the machine learning training results to help predict the properties and performance of more atomic numbers and multivariate conditions. This method overcomes the problem of difficulty in realizing high atomic number calculations in large-scale quantum chemical calculations due to the dimensionality disaster.

[0006] The third object of the present invention is to provide an application of a metal ordered phase property prediction model based on machine learning to batch predict data on the formation energy and mechanical properties of different ordered phases and construct a metal ordered phase database.

[0007] In order to achieve the above technical objectives, the present invention provides a modeling method for predicting the properties of metal ordered phases based on machine learning, which is characterized by comprising:

[0008] Step S1: Collect the original data of metal ordered phases, obtain the crystal structures of different ordered phases in batches through Pymatgen, and perform first-principles high-throughput calculations to obtain their formation energy and mechanical properties data;

[0009] Step S2: Based on the formation energy and mechanical property data obtained in S1, machine learning feature engineering processing and data preprocessing are performed to obtain a data set;

[0010] Step S3: Use the data set obtained in S2 to train, test, optimize and evaluate multiple machine learning models, and select the model with the best comprehensive evaluation as the prediction model;

[0011] Step S4: Compare the prediction results of the model obtained in S3 with the first-principle high-throughput principle calculation results to verify the accuracy of the model.

[0012] As a preferred solution, the process of collecting the original data is as follows:

[0013] Step S1-1 analyzes the crystal structure information of the metal ordered phase including space group, lattice constant and Wyckoff position, and constructs a POSCAR file based on the ordered phase space group, lattice constant and Wyckoff position by Pymatgen;

[0014] Step S1-2 uses VASP calculation software based on plane wave method density functional theory to perform structural optimization on POSCAR and calculate the energy of its structure, and then performs first-principles high-throughput calculation to obtain its formation energy and mechanical properties data.

[0015] As a preferred solution, the construction process of the Wyckoff position POSCAR file is: different Wyckoff positions are used as different atomic position inputs, and the atomic types include 24 transition elements, wherein the element types are any combination of the 24 elements.

[0016] As a preferred solution, the formation energy calculation process is: the energy calculator of all structures obtained by VASP calculation software generates heat ΔH f , the calculation process is:

[0017] Formula 1:

[0018] The mechanical properties include bulk modulus B, shear modulus G, Young's modulus E and G / B value, and the calculation process is: using ULICS strain matrix to perform VASP calculation to obtain its strain-strain relationship, and then fitting and calculating the single crystal elastic constant matrix Cij through Hooke's law; the elastic constants of the polycrystalline can be obtained by the elastic constants of the single crystal through Voigt and Reuss formula to obtain BR, GR and Bv, Gv, and then as the upper and lower bounds of the modulus, they are substituted into the Hill formula to obtain the average values ​​BH and GH. The specific calculation formula is as follows:

[0019] Formula 2:

[0020] Formula 3: 1 / R =(S 11 +S 22 +S 33 )+2(S 12 +S 13 +S 23 );

[0021] Formula 4:

[0022] Formula 5:

[0023] Formula 6: 15 / R =4(S 11 +S 22 +S 33 )-4(S 12 +S 13 +S 23 )+3(S 44 +S 55 +S 66 );

[0024] Formula 7:

[0025] Formula 8: E = 9BG / (3B + G);

[0026] According to the above calculation results, the hardness Hv value can also be calculated by the empirical formula, specifically:

[0027] Formula 9:

[0028] In formulas 1 to 9: E total (A x B y C z ) is the total free energy of the intermetallic compound, E bulk (A), E bulk (B) and E bulk(C) are the free energies of A, B and C respectively, and x, y and z are the element numbers of A, B and C respectively.

[0029] As a preferred solution, the process of the machine learning feature engineering process is:

[0030] Step S2-1: selecting at least 23 features including atomic number, radius, valence electron, energy, electronegativity, weight, density, space family, family, period, lattice constant, melting point, boiling point, boiling point-melting point, thermal conductivity, thermal expansion coefficient, hardness, bulk modulus, shear modulus, Young's modulus, Poisson's ratio, entropy and heat capacity from n single elements, and constructing n×23 sets of feature vectors;

[0031] Step S2-2: Based on the obtained characteristic vector, the ordered phase characteristic vector including the ordered phase atomic radius difference Δr, electron difference Δχ, valence electron concentration VEC, intensity matching η and Peierls-Nabarro factor F is calculated. The calculation process is as follows:

[0032] Formula 10:

[0033] Formula 11:

[0034] Formula 12:

[0035] Formula 13:

[0036] Formula 14:

[0037] In formulas 10 to 14: C i , r i , χ i , VEC i , G i They represent the molar ratio, atomic radius ratio, electronegativity ratio, valence electron concentration, and shear modulus ratio of each component respectively; r, χ, G, and ν represent the average values ​​of atomic radius, electronegativity, shear modulus, Young's modulus, and Poisson's ratio of all elements respectively.

[0038] As a preferred solution, the data preprocessing process is: the data obtained through feature engineering is processed with missing values, outliers and standardization in turn, and then the Pearson correlation coefficient is used to eliminate the feature data with high correlation, wherein the calculation process of the Pearson correlation coefficient PCC is:

[0039] Formula 15:

[0040] When |PCC|>0.95, it means the correlation is very high, and the group of features is deleted;

[0041] In formula 15: F ki , F li Represents two different features in the n×23 feature vectors of the dataset, F k , F l Indicates its average value.

[0042] As a preferred solution, the machine learning model includes: KNN regression model, linear regression model, ridge regression model, support vector machine regression model, decision tree regression model, and three ensemble learning algorithms of decision trees: random forest regression model, Adaboost and gradient ascent regression model.

[0043] As a preferred solution, the optimization process of the machine learning model is as follows: the optimization of the regression model includes sample set splitting and hyperparameter adjustment, the sample set is split into training set / test set of 7 to 8:2 to 3, the hyperparameter adjustment includes selecting the number of random seeds of 80 to 120, adjusting the k value of the KNN regression model, selecting the Gaussian radial basis kernel function to adjust the γ value of the support vector machine regression model, and adjusting the max_depth and n_estimators values ​​of the decision tree regression model and the three integrated learning regression models.

[0044] As a preferred solution, the model evaluation is performed by the mean absolute error (MAE) and R 2 Score execution, the smaller the MAE value and the R 2 The closer the value is to 1, the better the comprehensive evaluation of the model is. The calculation process is:

[0045] Formula 16:

[0046] Formula 17:

[0047] In formulas 16 and 17: pre ,y i , They represent the machine learning predicted value, the true value calculated by the first principles, and the average of all the true values ​​calculated by the first principles.

[0048] The present invention also provides a machine learning-based prediction model for metal ordered phase properties, which is obtained by any of the modeling methods described above.

[0049] The present invention also provides an application of a metal ordered phase property prediction model based on machine learning, which is characterized by: batch predicting data on the formation energy and mechanical properties of different ordered phases to build a metal ordered phase database.

[0050] In the technical solution provided by the present invention, machine learning can discover the patterns and laws hidden behind the data by analyzing a large amount of material structure and performance data based on existing data sets and models; by building and training prediction models, machine learning can effectively predict the heat of formation and mechanical properties of ordered phases of metals, avoiding the blindness and waste of resources of traditional trial and error methods. In addition, machine learning can also use the results of quantum chemical calculations as training data to predict the properties of more complex multivariate systems under conditions of few samples, thereby overcoming the challenges of the dimensionality curse in traditional quantum chemical calculations. Therefore, predicting the properties of metal ordered phases and multivariate ordered phases, including heat of formation and mechanical properties, through machine learning methods can provide more reliable and economical theoretical guidance for the design, optimization and application of metal materials.

[0051] Compared with the prior art, the beneficial technical effects of the technical solution of the present invention are:

[0052] 1) The method of the present invention is based on a machine learning prediction model to assist in predicting the heat of formation and mechanical properties of ordered phases of arbitrary components, which can search more quickly, avoid blind experiments, save resources and time, and provide theoretical guidance for material design and preparation, thereby improving efficiency;

[0053] 2) The method of the present invention can perform quantum chemical calculations using a small number of samples, and combined with the results of machine learning training, it can help predict the properties and performance in cases with a larger number of atoms and multiple elements. This method overcomes the problem of difficulty in achieving high-atom-number calculations in large-scale quantum chemical calculations due to the dimensionality curse. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 The crystal structure information of the D03 ordered phase in Example 1 of the present invention;

[0055] Figure 2 is the Pearson correlation coefficient of the structural characteristic vectors of the (a) binary, (b) ternary, and (c) quaternary D03 ordered phase in Example 1 of the present invention;

[0056] Figure 3 Parameter selection diagram for the KNN regression model and the support vector machine regression model in Example 1 of the present invention;

[0057] Figure 4 A comparison chart of the calorific value and the first principles calculated value generated by machine learning prediction of the decision tree regression model, the random forest regression model, the Adaboost and the gradient ascent regression model in Example 1 of the present invention;

[0058] Figure 5Comparison diagram of the machine learning predicted values ​​and first principles calculated values ​​of the mechanical properties (a) bulk modulus B, (b) shear modulus G, (c) Young's modulus E, (d) G / B value, and (e) hardness Hv value in Example 1 of the present invention;

[0059] Figure 6 The crystal structure information of the C11b ordered phase in Example 2 of the present invention;

[0060] Figure 7 A comparison chart of the machine learning predicted values ​​and the first principles calculated values ​​of the mechanical properties (a) bulk modulus B, (b) shear modulus G, (c) Young's modulus E, (d) G / B value, and (e) hardness Hv value in Example 2 of the present invention;

[0061] Figure 8 The invention provides an application of machine learning to predict and generate thermal models in the implementation of the present invention. DETAILED DESCRIPTION

[0062] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0063] Example 1

[0064] This embodiment provides a method for predicting the properties of metal ordered phases based on machine learning. Taking the BCC structure D03 ordered phase as an example, the prediction method of the present invention is described in detail, including the following steps:

[0065] Step S1: Collect the original data set, use Pymatgen to batch obtain POSCAR files of different ordered phase crystal structures, and perform first-principles high-throughput calculations of their formation energy and mechanical properties data;

[0066] Step S2: Perform machine learning feature engineering processing, select features related to the material, and perform data preprocessing, such as missing value processing, outlier processing, data standardization, similarity measurement, and data cleaning, to obtain a machine learning data set;

[0067] Step S3: construct a machine learning algorithm model, optimize multiple learning models, train, test and evaluate them, and select the machine learning model with the best comprehensive training and testing effect as the machine learning prediction model;

[0068] Step S4: Verify the accuracy of the machine learning prediction model, and compare and verify the results of the ordered phase finally predicted by machine learning with the results of first-principles calculations;

[0069] Step S5: Apply the machine learning prediction model to batch predict the formation energy and mechanical properties of different ordered phases to form a database.

[0070] Reference Figure 1In step S1, the main steps of collecting the original data set include:

[0071] Step 1: Figure 1 , the space group of the ordered phase D03 crystal structure is The Wyckoff positions are 4a, 4b, 8c1, and 8c2. These four different Wyckoff positions are used as the positions of atoms A, B, C, and D. Therefore, the D03 ordered phase can form six groups of crystal structures, including binary (AB3), ternary (AABC, ABBC, ABCC, ABCA), and quaternary (ABCD). The lattice constant is The 24 transition elements were randomly arranged and combined with A, B, C, and D atoms, and the Pymatgen library in Python software was used to construct POSCAR files. Several groups of POSCAR files were obtained for each group of ordered phase restrictions for 6 different crystal structures.

[0072] Step 2: Use first-principles calculations to optimize the structure of POSCAR and calculate the heat of formation ΔH of all the ordered phase structures. f , the calculation formula is:

[0073]

[0074] Where E total (A x B y C z D l ) is the total free energy of the intermetallic compound, E bulk (A), E bulk (B), E bulk (C) and E bulk (D) are the free energies of A, B, C and D respectively, and x, y, z and l are the number of elements in A, B, C and D respectively;

[0075] Step 3: Further calculate the mechanical properties of the optimized POSCAR file, including bulk modulus B, shear modulus G, Young's modulus E, G / B value, and hardness Hv value;

[0076] The D03 ordered phase formation energy and a series of mechanical properties obtained through first-principles high-throughput calculations, and the four groups of data of ternary AABC, ABBC, ABCC, and ABCA are summed into one group of data, obtaining the original data set of machine learning consisting of three groups of binary, ternary, and quaternary structures, which is also the target quantity of machine learning.

[0077] In step S2, the material-related features include 23 characteristic atomic numbers n, radius r, valence electron ve, energy e, electronegativity pe, weight w, density d, space family sp, family g, period p, lattice constant a, melting point Tm, boiling point Tb, boiling point-melting point T, thermal conductivity c, thermal expansion coefficient α, hardness Hv, bulk modulus B, shear modulus G, Young's modulus E, Poisson's ratio v, entropy S, and heat capacity Cp of each element, and the D03 ordered phase has 4 Wyckoffs in total, thus forming 4×23=92 sets of characteristic vectors; The calculated ordered phase features include ordered phase atomic radius difference Δr, electron difference Δχ, valence electron concentration VEC, intensity matching η, and Peierls-Nabarro factor F, a total of 5 sets of eigenvectors, so there are a total of 97 sets of eigenvectors at this time; missing value processing, outlier processing, and standardization processing are performed on all the above data, and finally 275, 132, and 88 sets of binary, ternary, and quaternary data are obtained respectively; the Pearson correlation coefficient (PCC) is used to measure the data similarity. The PCC coefficients of binary, ternary, and quaternary data features are as follows Figure 2 As shown in (a)-(c), after deleting the highly correlated features with |PCC|>0.95, namely electronegativity pe, weight w, group g, lattice constant a, melting point Tm, and shear modulus G, there are 97-4×6=73 sets of eigenvectors left.

[0078] In step S3, the machine learning algorithm model mainly selects 8 regression models, including KNN regression model, linear regression model (Linear), ridge regression model (Ridge), support vector machine regression model (SVR), decision tree regression model (DT), random forest regression model (RF), Adaboost and gradient ascent regression model (GB); the sample set is split into training set / test set = 8 / 2, and the number of random seeds is 100; hyperparameters of each model are adjusted. For the prediction model of generating heat, the test results of the parameters k and γ values ​​of the KNN regression model and the SVR model are as follows: Figure 3 (ab) and (cd) show that the mean absolute error (MAE) is smaller and R 2 Parameters with values ​​closer to 1 are Figure 3 Select appropriate parameter values. The results are shown in Table 1. When the number of feature vectors is 97, k = 4 and γ = 0.26826958; when the number of feature vectors is 92, k = 8 and γ = 0.754312006; when the number of feature vectors is 73, k = 9 and γ = 0.754312006. The decision tree model and the three ensemble learning models are adjusted to obtain max_depth = 15 and n_estimators = 100. After selecting the optimal hyperparameters for different models, the data are trained, tested and evaluated. The model evaluation is mainly based on MAE value and R 2The test results are shown in Table 2 and Table 3. The analysis shows that the decision tree + Adaboost ensemble learning regression model with 97 sets of feature vectors performs best on the training and test sets, with the lowest MAE value and the R closest to 1. 2 Value, so it is selected as the model for subsequent prediction of heat generation. The machine learning prediction model of mechanical properties bulk modulus B, shear modulus G, Young's modulus E, G / B value, and hardness Hv value continued to select the decision tree + Adaboost ensemble learning regression model of 97 sets of feature vectors based on the above analysis, with parameters of max_depth=15 and n_estimators=100.

[0079] In step S4, Figure 4 As shown in Table 2 and Table 3, the machine learning predicted heat generation values ​​of the decision tree regression model, random forest regression model, Adaboost and gradient ascent regression model, which have better model performance, are compared with the first principle calculated values. The machine learning predicted values ​​and the first principle calculated values ​​are in good agreement; Figure 5 As shown, the machine learning predicted values ​​and the first-principles calculated values ​​of the mechanical properties bulk modulus B, shear modulus G, Young's modulus E, Poisson's ratio V, G / B value, and hardness Hv value are also in good agreement. This model can be used as a model for predicting mechanical properties.

[0080] Table 1D03 ordered phase machine learning prediction model parameters

[0081]

[0082] Table 2D03 MAE values ​​of the training and test sets of the ordered phase machine learning prediction model

[0083]

[0084]

[0085] Table 3D03 R of the training and test sets of the ordered phase machine learning prediction model 2 Score

[0086]

[0087]

[0088] Example 2

[0089] This embodiment provides a method for predicting the properties of metal ordered phases based on machine learning, using the BCC structure C11 b Taking the ordered phase as an example, the prediction method of the present invention is described in detail, comprising the following steps:

[0090] Step S1: Collect the original data set, use Pymatgen to batch obtain POSCAR files of different ordered phase crystal structures, and perform first-principles high-throughput calculations of their formation energy and mechanical properties data;

[0091] Step S2: Perform machine learning feature engineering processing, select features related to the material, and perform data preprocessing, such as missing value processing, outlier processing, data standardization, similarity measurement, and data cleaning, to obtain a machine learning data set;

[0092] Step S3: construct a machine learning algorithm model, optimize multiple learning models, train, test and evaluate them, and select the machine learning model with the best comprehensive training and testing effect as the machine learning prediction model;

[0093] Step S4: Verify the accuracy of the machine learning prediction model, and compare and verify the results of the ordered phase finally predicted by machine learning with the results of first-principles calculations;

[0094] Step S5: Apply the machine learning prediction model to batch predict the formation energy and mechanical properties of different ordered phases to form a database.

[0095] Reference Figure 6 In step S1, the main steps of collecting the original data set include:

[0096] Step 1: Figure 1 , ordered phase C11 b The space group of the crystal structure is 139 4 / , and the Wyckoff positions are 2a, 4e1, and 4e2. These three different Wyckoff positions are used as the positions of atoms A, B, and C. Therefore, the D03 ordered phase can form two groups of crystal structures, binary (AB2) and ternary (ABC), with lattice constants of The 24 transition elements were randomly arranged and combined with A, B, and C atoms, and the Pymatgen library in Python software was used to construct POSCAR files. Several groups of POSCAR files were obtained with two different crystal structures and each group of ordered phase restrictions.

[0097] Step 2: Use first-principles calculations to optimize the structure of POSCAR and calculate the heat of formation ΔH of all the ordered phase structures. f , the calculation formula is:

[0098]

[0099] Where E total (A x B y C z ) is the total free energy of the intermetallic compound, Ebulk (A), E bulk (B), E bulk (C) and are the free energies of A, B, and C respectively, and x, y, and z are the number of elements of A, B, and C respectively; Step 3: Further calculate the mechanical properties of the optimized POSCAR file, which include bulk modulus B, shear modulus G, Young's modulus E, G / B value, and hardness Hv value;

[0100] C11 obtained by first-principles high-throughput calculations b The ordered phase formation energy and a series of mechanical properties are used to obtain the original data set for machine learning, which is two groups of binary and ternary structures and is also the target quantity for machine learning.

[0101] In step S2, the material-related features are selected, including 23 features of each element, C11 b There are 3 Wyckoffs in the ordered phase, thus constituting 3×23=69 groups of eigenvectors. The calculated characteristics of the ordered phase include the ordered phase atomic radius difference Δr, electron difference Δχ, valence electron concentration VEC, intensity matching η, and Peierls-Nabarro factor F, for a total of 5 groups of eigenvectors. Therefore, there are 69+5=74 groups of eigenvectors at this time. Missing value processing, outlier processing, and standardization processing were performed on all the above data, and finally 354 and 330 groups of binary and ternary data were obtained, respectively.

[0102] In step S3, the machine learning algorithm model mainly selected 8 regression models, including KNN regression model, linear regression model (Linear), ridge regression model (Ridge), support vector machine regression model (SVR), decision tree regression model (DT), random forest regression model (RF), Adaboost and gradient ascent regression model (GB); the sample set was split into training set / test set = 8 / 2, and the number of random seeds was 100; for the prediction model of generating heat, the hyperparameters of each model were adjusted, among which the parameters k and γ values ​​of the KNN regression model and the SVR model were tested to be the best when γ = 2 and γ = 0.0954095476, and the decision tree model and the three integrated learning models were adjusted to obtain max_depth = 15 and n_estimators = 100; after selecting the optimal hyperparameters for different models, the data were trained, tested and evaluated, and the model evaluation was mainly conducted through MAE value and R 2 The test results are shown in Table 4. After analysis, it is found that the decision tree + Adaboost ensemble learning regression model performs best on the training and test sets, with the lowest MAE value and the R closest to 1. 2The machine learning prediction model of mechanical properties bulk modulus B, shear modulus G, Young's modulus E, G / B value, and hardness Hv value continued to select the decision tree + Adaboost ensemble learning regression model based on the above analysis, with parameters of max_depth = 14 and n_estimators = 200.

[0103] In step S4, Figure 7 As shown in the figure, the machine learning predicted values ​​of the mechanical properties bulk modulus B, shear modulus G, Young's modulus E, G / B value, and hardness Hv value are in good agreement with the first-principles calculated values. This model can be used as a model for predicting mechanical properties.

[0104] Table 4C11 b MAE values ​​and R of the training and test sets of the ordered phase machine learning prediction model 2 Score

[0105]

[0106] Reference Figure 8 In step S5, the machine learning prediction model is simply applied to predict the ordered phases of different components of eight systems of W, Ti, Fe, Co, Ni, Cu, Ta, and Re, such as D03 AB3, A3B, C11b AB3, A2B, and B2 / B32 AB, and their stability is compared to select the most stable ordered phase.

[0107] The above describes the method embodiments of the present invention in conjunction with the accompanying drawings, but the present invention is not limited to the above embodiments, and various changes can be made according to the purpose of the invention of the present invention. Any parameter changes or calculation simplifications made according to the principle of the technical solution of the present invention, as long as they meet the purpose of the invention of the present invention and do not deviate from the principle and concept of the method for predicting the properties of metal ordered phases based on machine learning of the present invention, belong to the protection scope of the present invention.

Claims

1. A modeling method for predicting the properties of ordered phases of metals based on machine learning, characterized in that: include: Step S1: Collect the original data of metal ordered phases, obtain the crystal structures of different ordered phases in batches through Pymatgen, and perform first-principles high-throughput calculations to obtain their formation energy and mechanical properties data; Step S2: Based on the formation energy and mechanical property data obtained in S1, machine learning feature engineering processing and data preprocessing are performed to obtain a data set; Step S3: Use the data set obtained in S2 to train, test, optimize and evaluate multiple machine learning models, and select the model with the best comprehensive evaluation as the prediction model; Step S4: Compare the prediction results of the model obtained in S3 with the first-principle high-throughput principle calculation results to verify the accuracy of the model.

2. A modeling method for predicting the properties of ordered metal phases based on machine learning according to claim 1, characterized in that: The original data collection process is as follows: Step S1-1 analyzes the crystal structure information of the metal ordered phase including space group, lattice constant and Wyckoff position, and constructs a POSCAR file based on the ordered phase space group, lattice constant and Wyckoff position by Pymatgen; Step S1-2 uses VASP calculation software based on plane wave method density functional theory to perform structural optimization on POSCAR and calculate the energy of its structure, and then performs first-principles high-throughput calculation to obtain its formation energy and mechanical properties data.

3. A modeling method for predicting the properties of ordered metal phases based on machine learning according to claim 1, characterized in that: The construction process of the Wyckoff position POSCAR file is as follows: different Wyckoff positions are used as different atomic position inputs, and the atomic types include 24 transition elements, wherein the element types are any combination of the 24 elements.

4. A modeling method for predicting the properties of ordered metal phases based on machine learning according to claim 3, characterized in that: The formation energy calculation process is as follows: the energy calculator of all structures obtained by VASP calculation software generates heat ΔH f , the calculation process is: Formula 1: The mechanical properties include bulk modulus B, shear modulus G, Young's modulus E and G / B value, and the calculation process is: using ULICS strain matrix to perform VASP calculation to obtain its strain-strain relationship, and then fitting and calculating the single crystal elastic constant matrix Cij through Hooke's law; the elastic constants of the polycrystalline can be obtained by the elastic constants of the single crystal through Voigt and Reuss formula to obtain BR, GR and Bv, Gv, and then as the upper and lower bounds of the modulus, they are substituted into the Hill formula to obtain the average values ​​BH and GH. The specific calculation formula is as follows: Formula 2: Formula 3: 1 / B R =(S 11 +S 22 +S 33 )+2(S 12 +S 13 +S 23 ); Formula 4: Formula 5: Formula 6: 15 / G R =4(S 11 +S 22 +S 33 )-4(S 12 +S 13 +S 23 )+3(S 44 +S 55 +S 66 ); Formula 7: Formula 8: E = 9BG / (3B + G); According to the above calculation results, the hardness Hv value can also be calculated by the empirical formula, specifically: Formula 9: In formulas 1 to 9: E total (A x B y C z ) is the total free energy of the intermetallic compound, E bulk (A), E bulk (B) and E bulk (C) are the free energies of A, B and C respectively, and x, y and z are the element numbers of A, B and C respectively.

5. The modeling method for predicting the properties of ordered metal phases based on machine learning according to claim 1, characterized in that: The process of machine learning feature engineering is as follows: Step S2-1: selecting at least 23 features including atomic number, radius, valence electron, energy, electronegativity, weight, density, space family, family, period, lattice constant, melting point, boiling point, boiling point-melting point, thermal conductivity, thermal expansion coefficient, hardness, bulk modulus, shear modulus, Young's modulus, Poisson's ratio, entropy and heat capacity from n single elements, and constructing n×23 sets of feature vectors; Step S2-2: Based on the obtained characteristic vector, the ordered phase characteristic vector including the ordered phase atomic radius difference Δr, electron difference ΔX, valence electron concentration VEC, intensity matching η and Peierls-Nabarro factor F is calculated. The calculation process is as follows: Formula 10: Formula 11: Formula 12: Formula 13: Formula 14: In formulas 10 to 14: C i , r i , χ i , VEC i , G i They represent the molar ratio, atomic radius ratio, electronegativity ratio, valence electron concentration, and shear modulus ratio of each component respectively; r, X, G, and v represent the average values ​​of atomic radius, electronegativity, shear modulus, Young's modulus, and Poisson's ratio of all elements respectively.

6. A modeling method for predicting the properties of metal ordered phases based on machine learning according to claim 1 or 5, characterized in that: The process of data preprocessing is: the data obtained through feature engineering are processed with missing values, outliers and standardization in turn, and then the Pearson correlation coefficient is used to eliminate the feature data with high correlation, wherein the calculation process of the Pearson correlation coefficient PCC is: Formula 15: When |PCC|>0.95, it means the correlation is very high, and the group of features is deleted; In formula 15: F ki , F li Represents two different features in the n×23 feature vectors of the dataset, F k , F l Indicates its average value.

7. The modeling method for predicting the properties of ordered metal phases based on machine learning according to claim 1, characterized in that: The machine learning models include: KNN regression model, linear regression model, ridge regression model, support vector machine regression model, decision tree regression model, and three ensemble learning algorithms of decision trees: random forest regression model, Adaboost and gradient ascent regression model; The optimization process of the machine learning model is as follows: the optimization of the regression model includes sample set splitting and hyperparameter adjustment. The sample set is split into a training set / test set of 7 to 8:2 to 3. The hyperparameter adjustment includes selecting a random seed number of 80 to 120, adjusting the k value of the KNN regression model, adjusting the γ value by selecting the Gaussian radial basis kernel function for the support vector machine regression model, and adjusting the max_depth and n_estimators values ​​of the decision tree regression model and the three integrated learning regression models.

8. A modeling method for predicting the properties of metal ordered phases based on machine learning according to claim 1 or 7, characterized in that: The model is evaluated by mean absolute error (MAE) and R 2 Score execution, the smaller the MAE value and the R 2 The closer the value is to 1, the better the comprehensive evaluation of the model is. The calculation process is: Formula 16: Formula 17: In formulas 16 and 17: pre ,y i , They represent the machine learning predicted value, the true value calculated by the first principles, and the average of all the true values ​​calculated by the first principles.

9. A prediction model for metal ordered phase properties based on machine learning, characterized in that: Obtained by the modeling method according to any one of claims 1 to 8.

10. The application of the machine learning-based metal ordered phase property prediction model according to claim 9, characterized in that: Batch predict the formation energy and mechanical properties of different ordered phases and build a metal ordered phase database.

Citation Information

Patent Citations

  • Titanium alloy mechanical property prediction method and device based on machine learning

    CN116230136A

  • Full-size alloy phase prediction method based on machine learning

    CN117672422A

  • High-entropy alloy hardness prediction method based on machine learning and interpretive analysis

    CN117766069A

Cited By

  • Microalgae floating bead flotation harvesting efficiency prediction and optimization method based on machine learning

    CN119397212A

  • Component design method of metastable beta titanium alloy with low Young modulus

    CN119479851A

  • Selection and optimization method and device for machine learning force field model

    CN119670551A