Machine learning method for predicting mechanical properties of CuZr alloy

Through machine learning methods combined with molecular dynamics simulation, the mechanical properties of CuZr alloys are predicted, which solves the problem of time-consuming and low accuracy of predicting the mechanical properties of metal glass in the existing technology, and achieves fast and accurate prediction of mechanical properties, reducing experimental costs. , provides strong support for alloy design.

CN118983020APending Publication Date: 2024-11-19GUIZHOU UNIV

Patent Information

Application Number
CN202411070234.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-06
Publication Date
2024-11-19

AI Technical Summary

Technical Problem

Existing simulation-based and traditional experimental methods are difficult to quickly and effectively predict the mechanical properties of metal glass. Especially in the case of multi-component and multi-cold speeds, the experimental equipment is not very accurate and takes a long time.

Method used

Using machine learning method combined with molecular dynamics simulation, the mechanical properties prediction model of CuZr alloy is constructed. By performing molecular dynamics simulation in LAMMPS software, setting force field parameters and simulation parameters, quenching and stretching processes under different conditions are simulated, and the quenching and stretching processes are generated. Stress-strain curves are used to train models using algorithms such as linear regression and XGBoost, optimize parameters and use Shapley's additive interpretation technology to explain the importance of the model.

Benefits of technology

It realizes rapid and accurate prediction of the mechanical properties of CuZr alloys, reduces dependence on expensive experiments, provides strong guidance on alloy design, and improves material performance and application efficiency.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118983020A_ABST
    Figure CN118983020A_ABST
Patent Text Reader

Abstract

The invention provides a machine learning method for predicting the mechanical property of a CuZr alloy, which comprises the following steps: under different random conditions, constructing a plurality of groups of CuxZr100-x alloys with different proportions, performing force field parameter setting on a molecular dynamics simulation model in S1, performing programming, performing energy minimization and molecular dynamics balance on the model in S2 in LAMMPS software, and calculating the mechanical property of the CuZr alloy. The method comprises the following steps: establishing a model, enabling the established model to have a stable configuration, setting simulation parameters of the molecular dynamics simulation model, and obtaining different simulation conditions by setting the simulation parameters including the proportion of constituent elements, the stretching temperature, the cooling rate during quenching, the simulation dimension, the boundary condition, the simulation temperature, the time step length and the like. According to the machine learning method for predicting the mechanical property of the CuZr alloy, the data set is collected through molecular dynamics simulation, the mechanical property of the alloy is predicted through the machine learning method, and the method utilizes a large amount of molecular dynamics simulation data and changes the factors such as the composition, the preparation method and the temperature of the alloy.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention relates to the fields of computer molecular simulation and artificial intelligence technology, and in particular to a machine learning method for predicting the mechanical properties of a CuZr alloy. Background Art

[0002] In recent years, machine learning technology has developed rapidly. Because it can extract useful information from large amounts of complex and redundant data in an efficient, economical and time-saving manner, it has been widely used in the fields of medicine, environment, information communication, materials science, etc. At present, machine learning technology is rapidly developing in conjunction with molecular dynamics simulation, which can be used to distinguish different crystal phases, guide the development and design of materials, etc. The combination of the two has the advantages of fast computing speed and low computing cost.

[0003] Metallic glass has excellent properties, such as high strength, high elastic limit and excellent corrosion resistance, which gives it broad application potential. Especially in industry, the high strength and large elastic limit of bulk metallic glass is an extremely useful property. In recent years, scholars have carried out considerable research in this field.

[0004] The microstructure of a material determines its macroscopic properties. The microstructure of metallic glass is affected by its alloy composition, preparation method, cooling rate, etc., and these influences are very complex. A lot of research has been done on the microstructure and mechanical properties of metallic glass using molecular dynamics simulation. However, its scope of application is relatively single. It is difficult to quickly determine the mechanical properties of alloys in the case of multiple components and multiple cooling rates. Experiments have problems such as low equipment accuracy and long time consumption. Existing methods based on simulation and traditional experiments cannot provide efficient and rapid predictions of the mechanical properties of metallic glass.

[0005] Therefore, it is necessary to provide a machine learning method for predicting the mechanical properties of CuZr alloys to solve the above technical problems. Summary of the invention

[0006] The present invention provides a machine learning method for predicting the mechanical properties of CuZr alloy, which solves the problem that the existing methods based on simulation and traditional experiments cannot provide efficient and rapid prediction of the mechanical properties of metallic glass.

[0007] In order to solve the above technical problems, the present invention provides a machine learning method for predicting the mechanical properties of CuZr alloy, comprising the following steps:

[0008] S1. Under different random conditions, construct multiple groups of Cu with different proportions x Zr 100-x Alloys;

[0009] S2, setting the force field parameters for the molecular dynamics simulation model in S1;

[0010] S3, after programming, perform energy minimization and molecular dynamics equilibrium on the model in S2 in LAMMPS software to make the constructed model have a stable configuration;

[0011] S4, setting simulation parameters of the molecular dynamics simulation model, the simulation parameters including the ratio of constituent elements, stretching temperature, cooling rate during quenching, simulation dimension, boundary conditions, potential function, ensemble selection, system temperature control method, simulation temperature and time step, etc., to obtain different simulation conditions;

[0012] S5. Determine the ensemble for machining process simulation as NPT, and simulate quenching, relaxation and other processes;

[0013] S6. According to the above configuration, the model is stretched at a constant rate to obtain multiple sets of simulated stress-strain curves;

[0014] S7. Use machine learning methods to train the data set, select the optimal model and parameters, and explain them.

[0015] Preferably, the molecular dynamics simulation model of CuZr alloy stretching established by LAMMPS software in S1 has sizes of 17nm, 6nm, and 3nm, and Cu x Zr 100-x The x range is [36, 64].

[0016] Preferably, the force field used in setting the force field parameters of the molecular dynamics simulation model in S1 in S2 is the eam potential function, and the command is as follows:

[0017] pair_style eam / fs

[0018] pair_coeff**Cu-Zr_4.eam.fs Cu Zr.

[0019] Preferably, the model in S3 has multiple simulation conditions, including the ratio of constituent elements, stretching temperature, and cooling rate during quenching. The data are randomly generated using Python, with a temperature range of [10, 500] and a cooling rate range of [10 11 K / s-10 12 K / s].

[0020] Preferably, the model in S3 is simulated, comprising the following steps:

[0021] S31, set the initial temperature to 2500K, and relax for 10ns in the NVT ensemble;

[0022] S32, at a constant cooling rate R cAfter quenching to 50 K and relaxing at zero pressure, the temperature was increased from 50 K to 300 K.

[0023] Preferably, the model in S4 is subjected to a uniaxial tensile test, relevant mechanical performance parameters are calculated, and a data set is constructed, which specifically includes the following steps:

[0024] S41, with 10 8 The uniaxial tensile test was carried out at a constant strain rate of / s, with the tensile direction being the x direction, the y and z directions maintaining zero pressure, and the time step being 1.0fs;

[0025] S42. Based on the stress-strain curve obtained above, calculate the mechanical performance parameters such as Young's modulus, toughness, yield stress, flow stress, elasticity, drop stress, ultimate tensile stress, etc., and construct a data set.

[0026] Preferably, constructing a machine learning model using the data set in S6 comprises the following steps:

[0027] S61, randomly taking 70% of the data set as the training set and 30% as the test set, using nine different algorithms including linear regression, ridge regression, lasso, XGBoost, artificial neural network, Gaussian process regression, LGBoost, multilayer perceptron, and decision tree to predict the mechanical properties of CuZr;

[0028] S62. Use cross-validation, grid search and other techniques to optimize the machine learning model. After selecting the best model and parameters, use the Shapley additive interpretation (SHAP) technique to interpret the constructed model and analyze the importance of each feature to the model.

[0029] Preferably, in S6, mechanical property parameters of the material are calculated based on the stress-curve to construct a data set.

[0030] Compared with the related art, the machine learning method for predicting the mechanical properties of CuZr alloy provided by the present invention has the following beneficial effects:

[0031] The present invention provides a machine learning method for predicting the mechanical properties of CuZr alloy. A data set is collected through molecular dynamics simulation, and a machine learning method is used to predict the mechanical properties of the alloy. The method utilizes a large amount of molecular dynamics simulation data, and by changing factors such as the composition, preparation method, and temperature of the alloy, a data set and a feature parameter set required for machine learning are constructed, and multiple features such as Young's modulus and yield strength are calculated.

[0032] The nine most commonly used machine learning models, including Linear regression, XGBoost, DecisionTree, and Random Forest, were used to find models with better effects through feature screening, 10-fold cross validation, and grid search techniques. A machine learning method for predicting the mechanical properties of CuZr alloy was established. The proposed method provides a promising prospect for accelerating the design of alloys, reduces dependence on expensive experiments, and is conducive to further explaining and guiding the design of alloys. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 A structural schematic diagram of a preferred embodiment of a machine learning method for predicting mechanical properties of CuZr alloy provided by the present invention;

[0034] Figure 2 A molecular dynamics simulation model for specific simulation processing;

[0035] Figure 3 (a)-(f) are the correlation coefficients R when adjusting the characteristic parameters 2 changes;

[0036] Figure 4 (a)-(g) are the comparison results of the three best machine learning models predicting six properties (ultimate tensile stress, toughness, flow stress, drop stress, elasticity, and yield strength);

[0037] Figure 5 (a)-(f) are the performance of three machine learning models in predicting a certain performance on the validation set;

[0038] Figure 6 (a)-(f) are summary graphs of the Shapley Additive Explanations (SHAP) technique for the three features with the best prediction results in the machine learning model. DETAILED DESCRIPTION

[0039] The present invention will be further described below in conjunction with the accompanying drawings and implementation modes.

[0040] Please refer to Figure 1 , Figure 2 , Figure 3 , Figure 4 , Figure 5 and Figure 6 ,in, Figure 1 A structural schematic diagram of a preferred embodiment of a machine learning method for predicting mechanical properties of CuZr alloy provided by the present invention; Figure 2 A molecular dynamics simulation model for specific simulation processing; Figure 3 (a)-(f) are the correlation coefficients R when adjusting the characteristic parameters 2changes; Figure 4 (a)-(g) are the comparison results of the three best machine learning models predicting six properties (ultimate tensile stress, toughness, flow stress, drop stress, elasticity, and yield strength); Figure 5 (a)-(f) are the performance of three machine learning models in predicting a certain performance on the validation set; Figure 6 (a)-(f) are the Shapley Additive Explanation (SHAP) technique summary diagrams of the three features with the best prediction results in the machine learning model. A machine learning method for predicting the mechanical properties of CuZr alloys includes the following steps:

[0041] S1. Under different random conditions, construct multiple groups of Cu with different proportions x Zr 100-x Alloys;

[0042] S2, setting the force field parameters for the molecular dynamics simulation model in S1;

[0043] S3, after programming, perform energy minimization and molecular dynamics equilibrium on the model in S2 in LAMMPS software to make the constructed model have a stable configuration;

[0044] S4, setting simulation parameters of the molecular dynamics simulation model, the simulation parameters including the ratio of constituent elements, stretching temperature, cooling rate during quenching, simulation dimension, boundary conditions, potential function, ensemble selection, system temperature control method, simulation temperature and time step, etc., to obtain different simulation conditions;

[0045] S5. Determine the ensemble for machining process simulation as NPT, and simulate quenching, relaxation and other processes;

[0046] S6. According to the above configuration, the model is stretched at a constant rate to obtain multiple sets of simulated stress-strain curves;

[0047] S7. Use machine learning methods to train the data set, select the optimal model and parameters, and explain them.

[0048] Firstly, a molecular dynamics simulation model of CuZr alloy stretching was established, and then each model was simulated under different components, cooling rates, and stretching temperatures to obtain the stress-strain curves under different conditions, and then various mechanical properties (Young's modulus, toughness, yield stress, flow stress, drop stress, ultimate tensile stress, yield stress) were calculated. This provides a data set for the subsequent construction of a machine learning model. After the machine learning model is built, it can be used to quickly obtain the mechanical property parameters under different simulation conditions to improve material performance and application. The computing environment of the present invention is flexible, saving experimental time and experimental costs.

[0049] The molecular dynamics simulation model of CuZr alloy stretching established by LAMMPS software in S1 has the sizes of 17nm, 6nm, and 3nm. x Zr 100-x The x range is [36, 64].

[0050] The force field used in S2 to set the force field parameters of the molecular dynamics simulation model in S1 is the eam potential function, and the command is as follows:

[0051] pair_style eam / fs

[0052] pair_coeff**Cu-Zr_4.eam.fs Cu Zr.

[0053] The model in S3 has a variety of simulation conditions, including the ratio of constituent elements, stretching temperature, and cooling rate during quenching. The data are randomly generated using Python, with a temperature range of [10, 500] and a cooling rate range of [10 11 K / s-10 12 K / s].

[0054] The model in S3 is simulated, including the following steps:

[0055] S31, set the initial temperature to 2500K, and relax for 10ns in the NVT ensemble;

[0056] S32, at a constant cooling rate R c After quenching to 50 K and relaxing at zero pressure, the temperature was increased from 50 K to 300 K.

[0057] The model in S4 performs a uniaxial tensile test, calculates relevant mechanical performance parameters, and constructs a data set, specifically including the following steps:

[0058] S41, with 10 8 The uniaxial tensile test was carried out at a constant strain rate of / s, with the tensile direction being the x direction, the y and z directions maintaining zero pressure, and the time step being 1.0fs;

[0059] S42. Based on the stress-strain curve obtained above, calculate the mechanical performance parameters such as Young's modulus, toughness, yield stress, flow stress, elasticity, drop stress, ultimate tensile stress, etc., and construct a data set.

[0060] The data set in S6 is used to construct a machine learning model, including the following steps:

[0061] S61, randomly taking 70% of the data set as the training set and 30% as the test set, using nine different algorithms including linear regression, ridge regression, lasso, XGBoost, artificial neural network, Gaussian process regression, LGBoost, multilayer perceptron, and decision tree to predict the mechanical properties of CuZr;

[0062] Nine different algorithms, including linear regression, ridge regression, lasso, XGBoost, artificial neural network, Gaussian process regression, LGBoost, multilayer perceptron, and decision tree, were used to predict the mechanical properties of CuZr. 70% of the data set was randomly used as a training set and 30% as a test set to ensure that the training and test data sets have a reasonable distribution over the span of each variable. Several sets of validation data were prepared. The validation set does not participate in any training and testing, and is only used to evaluate the model performance after the training is completed.

[0063] S62. Use cross-validation, grid search and other techniques to optimize the machine learning model. After selecting the best model and parameters, use the Shapley additive interpretation (SHAP) technique to interpret the constructed model and analyze the importance of each feature to the model.

[0064] In the S6, the mechanical property parameters of the material are calculated according to the stress-strain curve, and a data set is constructed.

[0065] The data set was divided into 10 subsets, and one of them was selected as the validation set in turn without repetition. The 10-fold cross-validation method was used for validation to reduce overfitting and underfitting problems. The grid search method was used to adjust the parameters of each model until the optimal value was reached. During the evaluation process, the correlation coefficient R was used. 2 As an indicator to measure model performance, it ranges from 0 to 1. The closer its value is to 1, the stronger the ability of the variables in the equation to explain y is, and the better the model fits the data.

[0066] The prediction results on the test set and validation set were comprehensively compared to screen out the three most effective machine learning models for predicting each mechanical property. Finally, the Shapley additive interpretation (SHAP) technique was used to interpret the constructed models and analyze the importance of each feature to the model.

[0067] Please refer to Figure 3 (a)-(f) show the R corresponding to a parameter in the process of adjusting the characteristic parameters. 2 For the adjustment of two parameters, the z-axis represents the R after each parameter adjustment. 2 For the adjustment of the three parameters, the color change is used to represent the R 2 The value changes, there is a color map on the right side of each figure, from top to bottom represents R 2The change of the value from large to small, the darker the red, the better the prediction effect of the current parameter, the darker the blue, the worse the prediction effect of the current parameter. For each model, a unique red point is marked to indicate that it is the best parameter in the current model, and the values ​​of each parameter are marked in the lower left corner.

[0068] Figure 4 (a)-(g) show the results of using the three best performing models to predict six mechanical properties. After training nine models, the three best performing models were selected and their performance on 30 sets of test data was observed. Figure 4 (g) is the correlation coefficient R of each model's prediction performance 2 , Figure 5 (a)-(f) show the performance of each model on the validation set. The black dot in the figure represents the true value of the sample, and the red, green, and blue colors represent the prediction performance of different models. The closer to the black dot, the more accurate the model prediction. Figure 6 (a)-(f) are SHAP value analyses of flow stress, toughness, and ultimate tensile stress. The horizontal axis is the SHAP value of each property, each point represents a sample point, and the samples are spread out vertically. Different colors represent different characteristic values ​​(red corresponds to higher values, and blue corresponds to lower values).

[0069] LAMMPS was used to perform energy minimization and molecular dynamics equilibrium on the model in S2 to ensure that the constructed model had a stable configuration.

[0070] Use Python random numbers to set multiple groups of simulation parameters including the proportion of Cu, temperature during stretching, and cooling rate. Each group of simulation parameters corresponds to a model, and the remaining parameters such as boundary conditions, potential function, ensemble selection, simulation temperature, and time step remain consistent.

[0071] Based on the above conditions, the atomic composition, size, cooling rate and stretching temperature of each group of samples are different. The initial temperature is set to 2500K, and the relaxation is 10ns in the NVT ensemble. The cooling rate is constant. c (10 11 -10 12 K / s) to 50K, and after relaxation at zero pressure, the temperature was raised from 50 to 300K;

[0072] Take 10 8 / s constant strain rate for uniaxial tensile test, stretching in the x-direction, maintaining zero pressure in the y and z directions, the time step is 1.0fs, the simulation results include the three-dimensional coordinates, velocity and force of each atom in the simulation system. After the model is constructed, the simulated stress-strain curve is obtained. According to the stress-curve, the mechanical properties parameters such as Young's modulus, toughness, yield stress, flow stress, toughness, drop stress and ultimate tensile stress are calculated.

[0073] Nine different algorithms including linear regression, ridge regression, lasso, XGBoost, artificial neural network, Gaussian process regression, LGBoost, multilayer perceptron and decision tree were used to predict the mechanical properties of CuZr. 70% of the data set was randomly used as the training set and 30% as the test set to ensure that the training and test data sets have a reasonable distribution in the span of each variable. Several sets of validation data were prepared. The validation set does not participate in any training and testing, and is only used to evaluate the model performance after the training is completed.

[0074] Compared with the related art, the machine learning method for predicting the mechanical properties of CuZr alloy provided by the present invention has the following beneficial effects:

[0075] The present invention provides a machine learning method for predicting the mechanical properties of CuZr alloy. A data set is collected through molecular dynamics simulation, and a machine learning method is used to predict the mechanical properties of the alloy. The method utilizes a large amount of molecular dynamics simulation data, and by changing factors such as the composition, preparation method, and temperature of the alloy, a data set and a feature parameter set required for machine learning are constructed, and multiple features such as Young's modulus and yield strength are calculated.

[0076] The nine most commonly used machine learning models, including Linear regression, XGBoost, DecisionTree, and Random Forest, were used to find models with better effects through feature screening, 10-fold cross validation, and grid search techniques. A machine learning method for predicting the mechanical properties of CuZr alloy was established. The proposed method provides a promising prospect for accelerating the design of alloys, reduces dependence on expensive experiments, and is conducive to further explaining and guiding the design of alloys.

[0077] The above descriptions are merely embodiments of the present invention and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A machine learning method for predicting mechanical properties of CuZr alloys, characterized in that: The following steps are involved: S1. Under different random conditions, construct multiple groups of Cu with different proportions x Zr 100-x Alloys; S2, setting the force field parameters for the molecular dynamics simulation model in S1; S3, after programming, perform energy minimization and molecular dynamics equilibrium on the model in S2 in LAMMPS software to make the constructed model have a stable configuration; S4, setting simulation parameters of the molecular dynamics simulation model, the simulation parameters including the ratio of constituent elements, stretching temperature, cooling rate during quenching, simulation dimension, boundary conditions, potential function, ensemble selection, system temperature control method, simulation temperature and time step, etc., to obtain different simulation conditions; S5. Determine the ensemble for machining process simulation as NPT, and simulate quenching, relaxation and other processes; S6. According to the above configuration, the model is stretched at a constant rate to obtain multiple sets of simulated stress-strain curves; S7. Use machine learning methods to train the data set, select the optimal model and parameters, and explain them.

2. The machine learning method for predicting mechanical properties of CuZr alloy according to claim 1, characterized in that: The molecular dynamics simulation model of CuZr alloy stretching established by LAMMPS software in S1 has the sizes of 17nm, 6nm, and 3nm. x Zr 100-x The x range is [36, 64].

3. The machine learning method for predicting mechanical properties of CuZr alloy according to claim 2, characterized in that: The force field used in S2 to set the force field parameters of the molecular dynamics simulation model in S1 is the eam potential function, and the command is as follows: pair_style eam / fs pair_coeff**Cu-Zr_4.eam.fs Cu Zr.

4. The machine learning method for predicting mechanical properties of CuZr alloy according to claim 1, characterized in that: The model in S3 has a variety of simulation conditions, including the ratio of constituent elements, stretching temperature, and cooling rate during quenching. The data are randomly generated using Python, with a temperature range of [10, 500] and a cooling rate range of [10 11 K / s-10 12 K / s].

5. The machine learning method for predicting mechanical properties of CuZr alloy according to claim 1, characterized in that: The model in S3 is simulated, including the following steps: S31, set the initial temperature to 2500K, and relax for 10ns in the NVT ensemble; S32, at a constant cooling rate R c After quenching to 50 K and relaxing at zero pressure, the temperature was increased from 50 K to 300 K.

6. The machine learning method for predicting mechanical properties of CuZr alloy according to claim 1, characterized in that: The model in S4 performs a uniaxial tensile test, calculates relevant mechanical performance parameters, and constructs a data set, specifically including the following steps: S41, with 10 8 The uniaxial tensile test was carried out at a constant strain rate of / s, with the tensile direction being the x direction, the y and z directions maintaining zero pressure, and the time step being 1.0fs; S42. Based on the stress-strain curve obtained above, calculate the mechanical performance parameters such as Young's modulus, toughness, yield stress, flow stress, elasticity, drop stress, ultimate tensile stress, etc., and construct a data set.

7. The machine learning method for predicting mechanical properties of CuZr alloy according to claim 1, characterized in that: The data set in S6 is used to construct a machine learning model, including the following steps: S61. Randomly use 70% of the data set as the training set and 30% as the test set, and use nine different algorithms including linear regression, ridge regression, lasso, XGBoost, artificial neural network, Gaussian process regression, LGBoost, multilayer perceptron, and decision tree to predict the mechanical properties of CuZr alloy. S62. Use cross-validation, grid search and other techniques to optimize the machine learning model. After selecting the best model and parameters, use the Shapley additive interpretation (SHAP) technique to interpret the constructed model and analyze the importance of each feature to the model.

8. The machine learning method for predicting mechanical properties of CuZr alloy according to claim 1, characterized in that: In the S6, the mechanical property parameters of the material are calculated according to the stress-strain curve, and a data set is constructed.

Citation Information

Patent Citations

  • Simulation method for improving plasticity of amorphous alloy under compression condition

    CN116631541A

  • Molecular dynamics method for simulating microstructure interface formation of silica and silanes

    WO2022061485A1

Cited By

  • Aluminum alloy extrusion prediction optimization method based on machine learning, and related apparatus

    WO2026113047A1