High-flux alloy performance prediction method combining friction stir processing, molecular dynamics simulation and machine learning

Abstract

The invention belongs to the field of new material preparation and processing, and discloses a high-flux alloy performance prediction method combining friction stir processing, molecular dynamics simulation and machine learning, the method comprises the following steps: preparing an alloy with components in gradient distribution along a processing direction by using friction stir processing, establishing a relationship between the components and the performance by combining performance testing, and predicting the performance of the alloy. And constructing a test sample database. Then, the performance of different alloy components is obtained by changing the component proportion through molecular dynamics simulation, and a training sample database is constructed; and finally, the performance of alloys with different components is predicted through machine learning, the whole process of alloy rapid design and screening is achieved, and research, development and application of new alloy materials are promoted.

Description

Technical Field

[0001] The present invention belongs to the field of new material preparation and processing, and specifically relates to a high-throughput alloy performance prediction method combining stir friction processing, molecular dynamics simulation and machine learning. Background Art

[0002] The development of alloys has played a key role in the evolution of human society. Since ancient times, the emergence of alloys has greatly expanded the level of human production. However, with the advancement of science and technology and the increasingly stringent requirements of society on the comprehensive performance of alloys, traditional alloy design methods, such as trial and error methods that rely on experience, have problems such as long cycles and low design success rates in composition design and screening. Therefore, improving the efficiency of alloy composition design and screening has become the key to the research and development of new alloy materials.

[0003] Alloy design usually achieves expected performance by regulating alloy composition and organization. It is a comprehensive result based on the quantitative relationship between "composition-organization-process and performance". Such an alloy design process results in long preparation cycle and low screening success rate when preparing alloys. In recent years, machine learning has been proven to be applicable to the field of alloy design. Machine learning establishes a predictive model through the relationship between "alloy composition-preparation process-alloy performance", provides guidance for alloy design, and realizes rapid prediction and optimization of alloy material performance. However, sufficient data sets are required to train and verify machine learning models. The types of elements and alloy systems involved in alloy design are complex, and the data sample size is small, resulting in limited actual generalization effect of machine learning models. On the other hand, the prediction results of machine learning models need to be verified with experimental data. The existing alloy preparation technology has a long preparation process cycle, difficult to extract experimental data, and low verification efficiency. Summary of the invention

[0004] In view of the shortcomings of the above-mentioned prior art, the present invention provides a high-throughput alloy property prediction method combining stir friction processing, molecular dynamics simulation and machine learning. The present invention can solve the problem of insufficient model training data set in the prior art, and further solve the problem of long preparation cycle and low screening success rate in the preparation of alloys in the prior art.

[0005] To achieve the above object, the technical solution adopted by the present invention is as follows: A high-throughput alloy property prediction method combining friction stir processing, molecular dynamics simulation and machine learning includes the following processes: Prefabricating a gradient structure on a first substrate, pre-embedding a second metal in the gradient structure, and performing friction stir processing at the gradient structure to prepare a composition gradient alloy having a gradient distribution of components along a direction of the friction stir processing; Characterize the composition and properties of the composition gradient alloy, obtain yield strength and tensile strength data of different alloy compositions, establish the relationship between alloy composition and mechanical properties, and construct a test sample database based on the relationship between alloy composition and mechanical properties; Molecular dynamics simulation is used to obtain the yield strength and tensile strength data of different alloy components in composition gradient alloys, establish the relationship between alloy composition and mechanical properties, and construct a training sample database based on the relationship between alloy composition and mechanical properties; Dividing a training data set and a test data set from the training sample database and the test sample database respectively, performing variable screening on the training data set, and obtaining an optimal training data set; Using the optimal training data set to perform machine learning training on different machine learning models, selecting the optimal machine learning model through error index comparison and making predictions, using the hardness, tensile strength, and yield strength predicted by machine learning to verify with the test data set, performing repeated iterations, and training the optimal machine learning model again until it passes the verification; Use the validated optimal machine learning model to predict high-throughput alloy properties.

[0006] Preferably, the gradient structure adopts a groove having a gradient shape.

[0007] Preferably, the gradient structure adopts trapezoidal grooves or triangular grooves.

[0008] Preferably, when friction stir processing is performed at the gradient structure, the insertion depth of the stirring needle is the vertical distance from the bottom of the groove to the upper surface of the first substrate.

[0009] Preferably, the second metal is a metal plate having the same shape as the gradient structure or a second metal powder.

[0010] Preferably, when using molecular dynamics simulation to obtain yield strength and tensile strength data of different alloy components in a composition gradient alloy: Molecular dynamics simulation uses embedded atomic potential to describe the interaction between atoms, as shown in the following formula;

[0011]

[0012] in, For atoms i The total potential energy of is the two-body interaction potential between atoms; The local electron density generated for the atoms in the system; For atoms i Put the electron density into The energy required for the location; is the distance between two atoms; for j The electrons outside the nucleus of an atom i Contributing charge density at the atom.

[0013] Preferably, when using molecular dynamics simulation to obtain the yield strength and tensile strength data of different alloy components in the composition gradient alloy, the deformation test for obtaining the performance by molecular dynamics simulation is uniaxial tension.

[0014] Preferably, when performing variable screening on the training data set, Lasoo regression is used for variable screening, as shown in the following formula:

[0015] in, is a parameter variable; is the observation vector; is the design matrix; is the shrinkage parameter; It is the Lasso regression model.

[0016] Preferably, the different machine learning models include supporting random forest models, vector regression models and gradient boosting decision tree models.

[0017] Preferred: When selecting the optimal machine learning model through error index comparison, the mean square error MSE, mean absolute error MAE, root mean square error RMSE, mean absolute error MAPE and goodness of fit R are used. 2 To evaluate the accuracy of machine learning models; When the optimal machine learning model is trained until it passes the validation, the mean square error (MSE), mean absolute error (MAE), root mean square error (RMSE), mean absolute error (MAPE) and goodness of fit (R) are used. 2 To evaluate the accuracy of the optimal machine learning model.

[0018] Compared with the prior art, the present invention has the following beneficial effects: The present invention prefabricates a gradient structure on a first substrate, a second metal in the gradient structure, and stir friction processing at the gradient structure, so that a component gradient alloy with a gradient distribution of components along the stir friction processing direction can be prepared. Stir friction processing is a severe plastic deformation processing technology. The friction heat and plastic deformation heat generated by the rotation of the stirring head thermally plasticize the material around the stirring needle, causing the material to produce severe plastic rheology. Therefore, the present invention uses the high-temperature thermal coupling effect of stir friction processing to allow the added main element pure metal to alloy and violently plastically deform with the initial ingot, thereby achieving microstructure refinement, homogenization and densification. At the same time, the gradient structure of the pure main element metal ensures that the components along the stir friction processing direction have gradient characteristics, and high-throughput preparation can be achieved by coordinating the changes in processing parameters, providing experimental sample data for machine learning. In addition, molecular dynamics simulation can obtain the organizational structure of different alloys and their relationship with performance by changing the component ratio of the model, thereby obtaining a large number of data samples with different alloy components. Therefore, stir friction processing is used to provide experimental samples for machine learning, and molecular dynamics simulation is used to provide training samples for machine learning, thereby laying the foundation for realizing the optimal design of high entropy alloy components, improving the efficiency of machine learning verification, and achieving the purpose of guiding alloy design. It can be seen that the method of the present invention belongs to a short-process preparation technology, which can shorten the preparation cycle, reduce costs, and realize rapid screening and preparation of alloys. This method can realize the integrated regulation of alloy composition and microstructure through thermomechanical coupling, and prepare alloys with excellent comprehensive performance. This method prepares an alloy with a gradient distribution of components along the processing direction through stir friction processing technology, and combines molecular dynamics simulation and machine learning intelligent prediction to achieve high-throughput design, preparation and screening of alloys. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 It is a schematic diagram of the high-throughput alloy property prediction method combining stir friction processing, molecular dynamics simulation and machine learning of the present invention. DETAILED DESCRIPTION

[0020] The specific implementation of the embodiment of the present invention is described in detail below in conjunction with the accompanying drawings. It should be understood that the specific implementation described here is only used to illustrate and explain the present invention, and is not used to limit the present invention.

[0021] The present invention discloses a high-throughput alloy performance prediction method combining friction stir processing, molecular dynamics simulation and machine learning. The method uses friction stir processing to prepare alloys with gradient distribution of components along the processing direction, combines performance testing to establish the relationship between "components and performance", and constructs a test sample database. Subsequently, molecular dynamics simulation is used to obtain the performance of different alloy components by changing the component ratio, and a training sample database is constructed. Finally, machine learning is used to predict the performance of alloys with different components, realize the whole process of rapid alloy design and screening, and promote the research and development and application of new alloy materials.

[0022] One embodiment of the present invention discloses a high-throughput alloy property prediction method combining friction stir processing, molecular dynamics simulation and machine learning, which specifically includes the following steps: Step 1: Use sandpaper to polish and remove the oxide film on the surface of the alloy substrate (i.e., the first substrate), and use acetone to clean the surface of the alloy substrate, and then pre-set a gradient structure on the alloy substrate. The gradient structure uses a groove with a gradient shape, and specifically, a trapezoidal groove or a triangular groove can be used. For convenience, the above-mentioned trapezoid is preferably an isosceles trapezoid, and the triangle is preferably an isosceles triangle. Pre-fill the pure metal sheet or powder of the trapezoidal structure in the gradient structure on the alloy substrate, and stir friction processing is performed to prepare an alloy whose composition is gradiently distributed along the processing direction (i.e., a composition gradient alloy); when performing stir friction processing, insert the stirring needle into the alloy substrate, and stir friction gradient alloying treatment is performed on the area where the gradient structure (such as an isosceles triangle groove, the bottom side of the isosceles triangle groove and the vertex angle of the isosceles triangle groove are located in the same horizontal plane) to obtain an alloy plate whose alloy composition is gradiently distributed along the bottom side of the gradient structure (such as an isosceles triangle groove) to the vertex angle of the isosceles triangle groove, wherein the insertion depth of the stirring needle is the vertical distance from the bottom of the isosceles triangle groove to the upper surface of the base alloy plate; Specifically, the thickness of the alloy substrate is 4-10 mm. The top length of the alloy plate prefabricated groove (isosceles trapezoidal groove) is 0.5-1 mm, the bottom length is 2-8 mm, the height is 5-50 mm, and the depth is 1-2 mm. The purity of pure metal sheets / powders is ≥99.5%. The stirring head is made of W-Re alloy, the shoulder diameter is 10-16 mm, and the stirring needle is conical; the root diameter of the stirring needle is 4-8 mm, and the top diameter is 3-6 mm. The rotation speed of the stirring head is 200-1000 r / min, and the forward speed is 10-60 mm / min.

[0023] Step 2: Sampling is performed at preset intervals along the length direction of the surface of the alloy plate obtained in step 1 to obtain multiple alloy samples; the phase ratio, grain size, yield strength and tensile strength data of different alloy components are tested for each alloy sample obtained, and the relationship between "alloy composition and mechanical properties" is established, and a test sample database is constructed based on the relationship between "alloy composition and mechanical properties"; Step 3: Use molecular dynamics simulation to obtain the yield strength and tensile strength data of different alloy components, establish the relationship between "alloy composition and mechanical properties", and build a training sample database based on the relationship between "alloy composition and mechanical properties"; Specifically, the molecular dynamics simulation uses embedded atomic potential to describe the interaction between atoms, as shown in the following formula;

[0024]

[0025] in, For atoms i The total potential energy of is the two-body interaction potential between atoms; The local electron density generated for the atoms in the system; For atoms i Put the electron density into The energy required for the location; is the distance between two atoms; for j The electrons outside the nucleus of an atom i Contributing charge density at the atom.

[0026] Specifically, the deformation test for obtaining properties through molecular dynamics simulation is uniaxial tension.

[0027] In the above scheme, the molecular dynamics simulation path is: in file writing → running in Lammps system → output result file. The specific process is as follows:

[0028] The in file writing involves 6 parts. ① Initial simulation system settings: set the basic unit of the simulated alloy to metal, the boundary condition to periodic boundary condition P, and the atomic type to atomic; ② Constructing the geometric model: set the lattice constant of the alloy structure, FCC d FCC or BCC d BCC or HCP d HCP , where FCC is a face-centered cubic structure, BCC is a body-centered cubic structure, and HCP is a close-packed hexagonal structure. dis the lattice constant of the corresponding structure; set the basic size of the alloy, box 0 X 长 0Y 宽 0 Z; define the alloy as composed of several elements, use the substitution method to establish the atomic model, first fill the model with the main element atoms, and then use other atoms to replace them in proportion to obtain the final model. ③ Give the model physical properties: set the corresponding molar mass of the atoms and define the interaction relationship between atoms. ④ Minimize the model energy and initialize the temperature: set the energy convergence condition to 1×10 -10 ~1×10 -15 eV, the temperature is room temperature 293K, and the npt ensemble is selected for relaxation. npt is an isothermal and isobaric ensemble. ⑤ Define the basic parameters of deformation: including strain rate and stretching direction. ⑥ Result output: define the output parameters as strain-stress values ​​at every 10-100 steps, and finally obtain the tensile properties of alloys with different compositions.

[0029] Step 4: Divide the training data set and the test data set from the training sample database and the test sample database respectively, and perform variable screening on the collected training data set to reduce the influencing variables of the eigenvalues ​​and obtain the optimal training data set; Specifically, Lasoo regression is used for variable screening in the data preprocessing stage, as shown in the following formula:

[0030] in, is a parameter variable; is the observation vector; is the design matrix; is the shrinkage parameter; It is the Lasso regression model.

[0031] Specifically, the ratio of the number of samples in the model training set and the model test set is 5:1.

[0032] Step 5: Use random forest, vector regression model and gradient boosting decision tree model to perform machine learning training, input the optimal training data set into the model, select the optimal machine learning model through error index comparison and make predictions, use machine learning to predict the hardness, tensile strength and yield strength and test data set for verification, perform repeated iterations, and train the optimal machine learning model again until it passes the verification, so as to achieve the purpose of accurately predicting alloy composition and performance.

[0033] Specifically, the mean square error (MSE), mean absolute error (MAE), root mean square error (RMSE), mean absolute error (MAPE) and goodness of fit (R) are used. 2 To evaluate the accuracy of the model, when R 2 The iteration ends when >0.8.

[0034] In summary, the present invention prepares an alloy with a gradient distribution of components along the processing direction through stir friction processing technology, and combines molecular dynamics simulation and machine learning intelligent prediction to achieve high-throughput design, preparation and screening of alloy components.

[0035] The preferred embodiments of the present invention are described in detail above in conjunction with the accompanying drawings, but the present invention is not limited thereto. After considering the specification and practicing the invention disclosed herein, those skilled in the art will easily think of other embodiments of the present invention. Within the technical concept of the present invention, the technical solution of the present invention can be subjected to a variety of simple modifications, including the combination of various specific technical features in any suitable manner. In order to avoid unnecessary repetition, the present invention will not further describe various possible combinations. However, these simple modifications and combinations should also be regarded as the contents disclosed by the present invention, and all belong to the protection scope of the present invention.

Claims

1. A high-throughput alloy property prediction method combining friction stir processing, molecular dynamics simulation and machine learning, characterized in that: The process includes: Prefabricating a gradient structure on a first substrate, pre-embedding a second metal in the gradient structure, and performing friction stir processing at the gradient structure to prepare a composition gradient alloy having a gradient distribution of components along a direction of the friction stir processing; Characterize the composition and properties of the composition gradient alloy, obtain yield strength and tensile strength data of different alloy compositions, establish the relationship between alloy composition and mechanical properties, and construct a test sample database based on the relationship between alloy composition and mechanical properties; Molecular dynamics simulation is used to obtain the yield strength and tensile strength data of different alloy components in composition gradient alloys, establish the relationship between alloy composition and mechanical properties, and construct a training sample database based on the relationship between alloy composition and mechanical properties; Dividing a training data set and a test data set from the training sample database and the test sample database respectively, performing variable screening on the training data set, and obtaining an optimal training data set; Using the optimal training data set to perform machine learning training on different machine learning models, selecting the optimal machine learning model through error index comparison and making predictions, using the hardness, tensile strength, and yield strength predicted by machine learning to verify with the test data set, performing repeated iterations, and training the optimal machine learning model again until it passes the verification; Use the validated optimal machine learning model to predict high-throughput alloy properties.

2. A high-throughput alloy property prediction method combining friction stir processing, molecular dynamics simulation and machine learning according to claim 1, characterized in that: The gradient structure uses grooves having a gradient shape.

3. A high-throughput alloy property prediction method combining friction stir processing, molecular dynamics simulation and machine learning according to claim 2, characterized in that: The gradient structure adopts trapezoidal grooves or triangular grooves.

4. The high-throughput alloy property prediction method combining friction stir processing, molecular dynamics simulation and machine learning according to claim 2, characterized in that: When friction stir processing is performed at the gradient structure, the insertion depth of the stirring needle is the vertical distance from the bottom of the groove to the upper surface of the first substrate.

5. The high-throughput alloy property prediction method combining friction stir processing, molecular dynamics simulation and machine learning according to claim 1, characterized in that: The second metal is a metal plate having the same shape as the gradient structure or a second metal powder.

6. The high-throughput alloy property prediction method combining friction stir processing, molecular dynamics simulation and machine learning according to claim 1, characterized in that: When using molecular dynamics simulation to obtain the yield strength and tensile strength data of different alloy components in a composition gradient alloy: Molecular dynamics simulation uses embedded atomic potential to describe the interaction between atoms, as shown in the following formula; in, For atoms i The total potential energy of is the two-body interaction potential between atoms; The local electron density generated for the atoms in the system; For Atom i Put the electron density into The energy required for the location; is the distance between two atoms; for j The electrons outside the nucleus of an atom i Contributing charge density at the atom.

7. The high-throughput alloy property prediction method combining friction stir processing, molecular dynamics simulation and machine learning according to claim 1, characterized in that: When molecular dynamics simulation is used to obtain the yield strength and tensile strength data of different alloy components in a composition gradient alloy, the deformation test for obtaining the performance by molecular dynamics simulation is uniaxial tension.

8. The high-throughput alloy property prediction method combining friction stir processing, molecular dynamics simulation and machine learning according to claim 1, characterized in that: When performing variable screening on the training data set, Lasoo regression is used for variable screening, as shown in the following formula: in, is a parameter variable; is the observation vector; is the design matrix; is the shrinkage parameter; It is the Lasso regression model.

9. The high-throughput alloy property prediction method combining friction stir processing, molecular dynamics simulation and machine learning according to claim 1, characterized in that: The different machine learning models include supporting random forest models, vector regression models, and gradient boosting decision tree models.

10. The high-throughput alloy property prediction method combining friction stir processing, molecular dynamics simulation and machine learning according to claim 1, characterized in that: When selecting the optimal machine learning model by comparing error indicators, the mean square error MSE, mean absolute error MAE, root mean square error RMSE, mean absolute error MAPE and goodness of fit R are used. 2 To evaluate the accuracy of machine learning models; When the optimal machine learning model is trained until it passes the validation, the mean square error (MSE), mean absolute error (MAE), root mean square error (RMSE), mean absolute error (MAPE) and goodness of fit (R) are used. 2 To evaluate the accuracy of the optimal machine learning model.

Images