A method for determining constitutive parameters of metal materials based on machine learning

Abstract

The present invention discloses a method for determining the constitutive parameters of metal materials based on machine learning, and belongs to the technical field of material dynamic performance testing. The method performs static and dynamic tensile tests on metal materials through a test device, and performs real-time measurements based on an optical measurement approach; a cGAN conditional generative adversarial network is established based on a numerical analysis method and machine learning means, and the strain field cloud map that is difficult to obtain completely during the tensile damage stage of the material during the test is identified and the cloud map is completed based on a certain physical mechanism; a mapping relationship between the time-course data of the strain field cloud map of the material tensile process and the constitutive model parameters input into the numerical simulation platform is established through a CNN convolutional neural network, and the cloud map data obtained from the test is parameter inverted to obtain the constitutive model parameters that can be directly input into the numerical simulation platform. The present invention can quickly obtain the dynamic constitutive relationship of materials under complex stress states, and has the advantages of accuracy and convenience.

Description

Technical Field

[0001] The present invention belongs to the technical field of material dynamics performance testing, and specifically relates to a method for determining constitutive parameters of metal materials based on machine learning. Background Art

[0002] In order to characterize the mechanical behavior of metal materials under complex stress states and high strain rate loading, traditional methods are mainly based on static and dynamic tensile tests, obtained by directly fitting test data or inverting with the help of optimization algorithms. These methods have the following limitations: (1) Insufficient accuracy: Traditional methods require the design of a large number of specimens with different stress states, and by conducting a series of tests on each type of specimen, it is expected to obtain the material constitutive parameters of each corresponding stress state. However, existing test images and numerical analysis have clearly shown that the stress state of the specimen will continue to change during the loading stage from self-elasticity to damage, which will lead to the following two problems: a. Within the specimen, materials at different positions involve different deformation stages, and the stress state distribution is not uniform; b. The test data reflects the macroscopic behavior of the materials at various positions constituting the specimen under various stress states, but does not reflect the response characteristics under a specific stress state. It is obviously inappropriate to directly interpret such test data as corresponding stress states. The application of such erroneous data will inevitably lead to dangerous material design and structural response prediction. (2) Inefficiency: In order to invert material parameters, most existing studies have ignored the microscopic information of the material within the specimen, and based on the macroscopic measurement data of the material test (such as the relationship between force and displacement) as the benchmark, based on the optimization algorithm, and achieved through repeated iterations. However, based on the discussion in point (1), this approach cannot accurately obtain the material parameters under a given stress state. In fact, the responses at different positions of the specimen are not the same, which reflects the response characteristics of the material at the local position and represents the characteristics of the material under different states. However, these data resources have never been fully valued, resulting in the need to expand the data source by simply increasing the number of specimens. Summary of the invention

[0003] Based on the above technical problems, the present invention proposes a method for determining the constitutive parameters of metal materials based on machine learning.

[0004] The technical solution adopted by the present invention is:

[0005] A method for determining constitutive parameters of metal materials based on machine learning, comprising the following steps:

[0006] a. Process the test material into test specimens, conduct static and dynamic tests on the test specimens, record the deformation images of the test specimens during the entire loading process, analyze and assemble the strain field cloud map time history matrix ;

[0007] b. Establish the finite element model of the test specimen and obtain the strain field cloud diagram time history matrix under each parameter combination ; Based on the established neural network, the damage stage cloud map obtained in step a but with defects is completed to obtain the strain field cloud map time history matrix with complete information ;

[0008] c. Carry out elastic stage simulation controlled only by elastic parameters, and output the elastic stage strain field cloud map time history matrix obtained by simulation ; Based on CNN convolutional image recognition technology, the elastic parameters and the elastic stage strain field cloud map time history matrix obtained by the experiment are established Considering different combinations of plastic parameters, the time-history matrix of the strain field cloud diagram of the plastic stage obtained by simulation is output. , and then establish the plastic parameters and the plastic stage strain field cloud diagram time history matrix obtained by the experiment The mapping relationship;

[0009] d. Carry out numerical simulation controlled by damage parameters to the damage stage, and output the strain field cloud map time history matrix of the damage stage obtained by simulation ; Establish the time history matrix of damage parameters and strain field cloud diagram of damage stage obtained by test The mapping relationship is based on the strain field cloud map time history matrix with complete information in step b. Obtain the repaired and adjusted damage stage cloud map and invert to obtain the damage parameters;

[0010] e. Replace the test cloud diagrams of specimens in different stress states as input to obtain the dynamic constitutive parameters of the material under complex stress states.

[0011] The beneficial technical effects of the present invention are:

[0012] The present invention proposes a method for determining the constitutive parameters of metal materials based on machine learning. The method adopts a scheme combining material testing with deep learning, can obtain the damage parameters of materials under complex stress states, effectively characterize the mechanical behavior of metal materials under complex stress states and high strain rate loading, and has the advantages of high precision and high efficiency. The present invention can further establish a material data set that can be directly used for theoretical calculations and numerical simulations.

[0013] The present invention integrates the cutting-edge technologies of optical measurement technology, deep learning and numerical simulation, and proposes a new method for efficiently determining the constitutive parameters of metal materials. Through appropriate improvements, it can also be applied to the parameter determination of other types of materials (such as polymers, concrete materials, etc.). Taking into account the response differences at different positions and stages during the tensile process of the specimen, the full-field information of this whole process is captured by optical measurement and digital image recognition technology; based on deep learning and numerical simulation technology, by comparing the global field variable information at each stage, a large amount of data containing various stress states can be obtained from the test of a single specimen, thereby realizing the identification of material constitutive parameters under complex stress states and different strain rates. In addition, considering the actual situation that the images actually obtained during the test may have defects or data loss due to damage, the image repair and supplementation process is systematically considered in the present invention.

[0014] Specifically, the present invention has the following advantages:

[0015] (1) Efficient acquisition of full-field strain data: Using DIC technology and high-speed cameras, full-field strain data of materials under multiple strain rates and complex stress states can be obtained.

[0016] (2) Accurate inversion of constitutive parameters: Through deep learning technology, especially cGAN and CNN network models, the constitutive parameters of the material in the elastic, plastic and damage stages are accurately inverted.

[0017] (3) Data completion and analysis: Through deep learning models, missing data in the experiment can be completed to improve data integrity and accuracy.

[0018] The present invention not only improves the efficiency and accuracy of determining material constitutive parameters, but also provides strong support for material science research and engineering applications. The method of the present invention has important application prospects in improving the level of material design and analysis and optimizing the performance of engineering structures. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments:

[0020] Figure 1 A schematic diagram of a process for determining constitutive parameters of metal materials based on machine learning according to the present invention;

[0021] Figure 2 Schematic diagram of the shape of the specimen and the speckle area in different stress states involved in the embodiments of the present invention; Figure 2 (1) is the standard tensile specimen. Figure 2 (2) is a dog bone specimen with a notch radius of 50 mm. Figure 2 Middle (3) is a flat groove plate with a notch radius of 26 mm;

[0022] Figure 3It is a diagram of the test device corresponding to the static and dynamic tests in the embodiment of the present invention;

[0023] Figure 4 This is a schematic diagram of the process of generating cloud map data and assembling it into a five-dimensional matrix through DIC processing of the test photos in the embodiment of the present invention;

[0024] Figure 5 The corresponding relationship between the constitutive parameters inputted by the test piece in each stress state and the strain field cloud diagram time history data of the finite element analysis result in the embodiment of the present invention is shown;

[0025] Figure 6 The incomplete cloud graph process of the cGAN conditional generative adversarial network training completion experiment according to an embodiment of the present invention is shown;

[0026] Figure 7 This is a schematic diagram of the effect of using cGAN to complete the missing values ​​of the incomplete cloud map obtained by the experiment of the embodiment of the present invention;

[0027] Figure 8 The inversion workflow diagram of the constitutive parameters of a single stress state according to an embodiment of the present invention is as follows;

[0028] Figure 9 This is a schematic diagram of a numerical simulation constitutive curve of an embodiment of the present invention;

[0029] Figure 10 The macro load-displacement comparison diagram of the test and inversion of the embodiment of the present invention;

[0030] Figure 11 It is a comparison chart of the complex stress state inversion results of specimens with different stress states in the embodiments of the present invention. DETAILED DESCRIPTION

[0031] The present invention discloses a method for determining the constitutive parameters of metal materials based on machine learning. On the one hand, the method uses a universal testing machine and a drop hammer test device to perform static and dynamic tensile tests on metal materials, and performs real-time measurements based on optical measurement. On the other hand, a cGAN conditional generative adversarial network is established based on numerical analysis methods and machine learning means to identify the strain field cloud map that is difficult to fully obtain during the tensile damage stage of the material during the test and complete the cloud map based on a certain physical mechanism. Furthermore, a mapping relationship between the time-course data of the strain field cloud map of the material tensile process and the constitutive model parameters input into the numerical simulation platform is established through a CNN convolutional neural network, and the cloud map data obtained from the test is parameter inverted using this mapping relationship to obtain the constitutive model parameters that can be directly input into the numerical simulation platform. The present invention comprehensively applies experimental measurement, mechanical analysis, machine learning and inversion theory to quickly obtain the dynamic constitutive relationship of materials under complex stress states, and then completes the determination of the dynamic properties of such materials, with the advantages of accuracy and convenience.

[0032] like Figure 1 As shown, a method for determining constitutive parameters of metal materials based on machine learning specifically includes the following steps:

[0033] S1, such as Figure 2 As shown, the test material is finely machined to form test specimens with different geometric shapes and sizes. Each specimen corresponds to a different initial stress state, characterized by stress triaxial coefficient and Lode angle. In this embodiment, three stress states are provided for reference, including standard tensile specimens (stress triaxial coefficient: 1 / 3, Lode angle: 1), 50mm notch radius dog bone specimens (stress triaxial coefficient: 0.6, Lode angle: 1), 26mm notch radius flat groove plate (stress triaxial coefficient: 0.7, Lode angle: 0). The left side is a side view and the right side is a front view. Black paint spots of different sizes and random shapes are evenly sprayed in the gauge area of ​​each specimen to facilitate optical testing and data processing. 4 to 5 of the above-mentioned specimens are processed and treated in each group to carry out static and dynamic tests under different loading rates (characterized by strain rate).

[0034] S2, such as Figure 3 As shown, the universal testing machine and the drop hammer test device are used to carry out static and dynamic tensile tests on the specimens processed in step S1. Figure 3 The left picture in the middle is a static test device, and the right picture is a dynamic test device.

[0035] During the test, the static test uses displacement-controlled loading until the specimen breaks, and records the load-displacement time history data of the testing machine. In the dynamic test, the mass of the drop hammer in each test is controlled to be M, and the drop hammer release height is H. When conducting dynamic tests, the selection of the drop hammer mass M and the drop hammer release height H should ensure that the standard tensile specimen in S1 can be broken by impact. The strain rate in the dynamic test is calculated using formula (1-1). The strain rate at the moment of reaching the yield stress is taken as the calibration strain rate for this group of tests.

[0036] (1-1)

[0037] In the formula, are the velocity time histories of the upper and lower parts of the specimen, respectively, obtained from the analysis of step S4; is the dimension of the specimen in the tensile direction.

[0038] S3. In static and dynamic tests, industrial cameras and high-speed cameras are used to record the deformation images of the speckle area of ​​the specimen during the entire loading process.

[0039] S4. Apply image recognition and analysis technology to calculate the strain (displacement) field information of the speckle area and assemble it into a strain (displacement) field cloud map time-history matrix .

[0040] The test images recorded in step S3 are digitally analyzed by image analysis and computing software (such as DIC) to output the strain (displacement) field cloud map of the specimen from loading to fracture. N cloud maps are uniformly selected along the time dimension and assembled into a cloud map such as Figure 4 The five-dimensional matrix shown in Figure 1. The first and second dimensions of the matrix record the coordinate information of each data point in the cloud map; the third to fifth dimensions are the hue, response time, and strain rate information of the data point, respectively.

[0041] S5, such as Figure 5 As shown in the figure, based on the numerical simulation platform, a finite element model of the specimen designed in step S1 is established, corresponding to the loading system adopted in the test in step S2 (wherein the static simulation adopts the load time history data measured in the test), and the material constitutive parameters are used as control variables to carry out parametric numerical simulation to obtain the strain (displacement) field cloud diagram time history data under each parameter combination. .

[0042] S6, such as Figure 6 As shown in the figure, a U-NET neural network is established, and the deformation cloud map obtained in step S5 is used as a data set to train the machine to learn the evolution law of the cloud map from the elastic-plastic stage to the damage stage, and to master the ability to generate the corresponding damage stage cloud map from the given elastic-plastic stage cloud map.

[0043] S7, such as Figure 6 As shown, a PatchGAN neural network is established to compare the damage stage cloud maps obtained by numerical simulation and machine prediction in steps S5 and S6, respectively, and calculate the relative error δ between the two. If the error δ is less than the threshold (determined by the complexity of the experiment, 0.2% is recommended), the U-NET neural network is fed back to store the corresponding hyperparameters; otherwise, steps S6 and S7 are repeated until U-NET can predict and generate a damage stage cloud map that meets the accuracy requirements. Relative error The calculation is performed using the L2 loss function shown in formula (2-1).

[0044] (2-1)

[0045] in, , They are the hue values ​​(corresponding to the third dimension) at the cloud image pixel coordinates x and y (corresponding to dimensions 1 and 2) in the prediction and test sets, respectively.

[0046] S8. The damage stage cloud diagram (such as Figure 7 As shown in Figure 3), the machine trained in steps S6 and S7 is input to obtain a cloud matrix with complete information to prepare data for the subsequent inversion of constitutive parameters.

[0047] S9. Based on the finite element model generated in step S5, conduct elastic stage simulation controlled only by elastic parameters (elastic modulus E, Poisson's ratio ν), and output the strain (displacement) field cloud map time history matrix The constitutive control formula in the elastic stage is as follows:

[0048] e x = 1 E [ s x − n ( s y + s z ) ] (3-1)

[0049] e y = 1 E [ s y − n ( s x + s z ) ] (3-2)

[0050] e z = 1 E [ s z − n ( s y + s x ) ] (3-3)

[0051] Where E is the elastic modulus, is Poisson's ratio, , , are the strains in the x, y, and z directions, respectively. , , are the stresses in the x, y, and z directions respectively.

[0052] S10, such as Figure 8 As shown in the figure, a neural network CNN1 based on CNN convolutional image recognition technology is established, the simulation results of step S9 are input into CNN1 as a training set for learning, and the mapping relationship between the elastic parameters E, ν and the time-course matrix of the cloud diagram of step S9 is established. The mean absolute percentage error (MAPE) was used to evaluate the model prediction accuracy.

[0053] [ E , n ] = f C N N 1 ( x , y , e , t ) (4-1)

[0054] (4-2)

[0055] Where E is the elastic modulus, is Poisson's ratio, Represents a functional relationship, x and y are the directional coordinates of a point in the cloud map, is the strain value of a point in the cloud graph, t is the time when the cloud graph is in the stretching process; N is the number of machine output parameters, is the input parameter corresponding to the training set, is the prediction parameter corresponding to the training set. When MAPE is reduced to below 0.5%, the machine training is completed; otherwise, continue to increase the number of training sets and repeat steps S9 to S10 to achieve the desired effect.

[0056] S11. The elastic phase cloud diagram obtained in step S4 , input the trained machine CNN1, the machine identifies and inverts the material parameters (elastic modulus E and Poisson's ratio ν) under this working condition.

[0057] S12. Based on the model in step S9, the elastic parameters (E, ν) determined in step S11 are applied, based on the plastic stage constitutive model similar to that shown in formula (5-1) (a reasonable constitutive model can be selected according to the material type, and this application process is universal), considering different plastic parameters ( , , , , , , ) combination, carry out elastic-plastic simulation analysis, and output the time-history matrix of the strain (displacement) field cloud diagram in the plastic stage .

[0058] k = ( 1 − α ) [ k 0 + Q ( 1 − e − β e p ) ] + α [ A ( e p + e 0 ) n ] (5-1)

[0059] in is the flow stress, is the plastic strain, , is the yield stress and yield strain of the material; , , , , , , is the parameter to be predicted.

[0060] S13, similar to step S10, establish neural network CNN2, input the simulation results of S12 as training sets into CNN2 for learning, and establish plasticity parameters ( , , , , , , ) and the cloud map time matrix obtained in step S12 The mapping relationship is shown in formula (6-1). MAPE is also used to evaluate the prediction accuracy of the model. When MAPE is less than 0.5%, the machine training is completed.

[0061] [ α 、 A 、 Q 、 n 、 β 、 k 0 、 e 0 ] = f C N N 2 ( x , y , e , t ) (6-1)

[0062] S14, input the plastic phase cloud map obtained in the experiment in step S4 into the trained machine CNN2, and the machine identifies and inverts the plastic parameters of the material under this working condition , , , , , , .

[0063] S15, similar to S12, based on the calculation model of step S9, the elastic and plastic parameters determined in steps S11 and S14 are applied to carry out the calculation based on the parameters ( , , , , ) control is extended to the numerical simulation of the damage stage, and the strain (displacement) field cloud diagram time history matrix of this stage is output .

[0064] The behavior of the material in the damage stage depends on its stress state, which is reflected in the changes of material parameters with the stress state in experiments and numerical simulations. According to the research practice in this field, the stress state can be expressed by the stress triaxial coefficient as shown in the following formula: and Lord Point describe.

[0065] or = 1 3 ( s 1 + s 2 + s 3 ) 1 2 [ ( s 1 − s 2 ) 2 + ( s 2 − s 3 ) 2 + ( s 3 − s 1 ) 2 ] (7-1)

[0066] (7-2)

[0067] (7-3)

[0068] in, , , is the principal stress of the unit; is the Lord parameter, obtained by formula (7-2), is the normalized Lode angle; is the stress triaxial coefficient.

[0069] The damage stage constitutive structure starts from the damage (e.g. Figure 9 Point C, described by equations (7-4)-(7-7)) and damage evolution ( Figure 9 The CD segment is composed of two parts described by formula (7-8).

[0070] e ¯ f p r [ or , i ¯ ] = b 0 [ 1 + c ( 1 2 [ ( f 1 − f 2 ) a + ( f 2 − f 3 ) a + ( f 3 − f 1 ) a ] ) 1 a + c ( 2 or + f 1 + f 3 ) ] 1 m (7-4)

[0071] in , , is an intermediate variable, which can be obtained by the following formulas (7-5)-(7-7):

[0072] f 1 = 2 3 c o s [ π 6 ( 1 − i ¯ ) ] (7-5)

[0073] f 2 = 2 3 c o s [ π 6 ( 3 + i ¯ ) ] (7-6)

[0074] f 3 = − 2 3 c o s [ π 6 ( 1 + i ¯ ) ] (7-7)

[0075] (7-8)

[0076] in is the damage initiation strain, is the stress triaxial coefficient, is the normalized Lode angle, is the failure displacement under quasi-static conditions, is the characteristic size of the numerical simulation unit grid, is the effective plastic strain, is the relative displacement at failure, , , , , These are all parameters to be predicted.

[0077] S16, similar to step S10, establish neural network CNN3, input the simulation results of step S15 as training set into CNN3 for learning, and establish damage parameters ( , , , , ) and the cloud map time matrix obtained in step S15 The mapping relationship is shown in formula (8-1).

[0078] [ a 、 b 0 、 c 、 m 、 d 0 ] = f C N N 3 ( x , y , e , t ) (8-1)

[0079] S17, the damage stage cloud map obtained by the test in step S8 but after repair and adjustment Input the trained machine CNN3, the machine identifies and inverts the damage parameters of the material under this working condition , , , , .

[0080] S18, numerical simulation is carried out based on the constitutive parameters inverted in steps S11, S14, and S17, such as Figure 10 As shown, record the macroscopic response behavior of the specimen (such as the force and displacement time history curve) and compare it with the actual test data under the corresponding working conditions. As shown in formula (9-1), calculate the relative error between the two ,like If it is less than 0.5%, the prediction is completed; otherwise, increase the number of training times and repeat steps S9 to S17.

[0081] (9-1)

[0082] in, is the number of data points obtained in the experiment, is the i-th load value obtained in the test, The machine predicted load value corresponding to the displacement of the i-th data obtained in the test.

[0083] S19. Repeat the process of steps S9 to S18, replace the test cloud diagrams of specimens in different stress states as input, compare the results of the two, and verify the accuracy of the constitutive model under different stress states.

[0084] like Figure 11 As shown in the figure, the inversion results of different stress states are compared, and the MAPE between the constitutive parameters of different stress states is calculated. If the MAPE is less than 2%, the average value of the inversion results of different specimens is used as the constitutive parameter of the material under the complex stress state. If the MAPE is greater than 2%, S15 to S18 are repeated, the scale of the training set in step S15 is expanded, and more detailed parameter combinations are added for machine training.

[0085] ;

[0086] in, is the length of the strain dimension in the figure, is the length of the stress triaxial coefficient dimension; is the flow stress of the first specimen at the i-th strain value and the j-th stress triaxial coefficient state; is the flow stress of the second specimen at the i-th strain value and the j-th stress triaxial coefficient state.

[0087] S20. Repeat the process of steps S9 to S19, select the test cloud diagrams of different strain rates as input, and obtain the dynamic constitutive parameters of the material under complex stress state.

[0088] S21. Build a support vector machine model. According to the strain rate hardening model shown in formula (10-1), the damage initiation model shown in formula (10-2), and the damage evolution model shown in formula (10-4), generate training sets of unknown parameters and corresponding variables respectively, and establish mapping relationships as shown in formulas (10-5) to (10-7).

[0089] s c s = s 0 [ 1 + ( e ˙ D ) 1 q ] (10-1)

[0090] in, Flow stress at different strain rates calculated for the Cowper-Symonds model, is the flow stress under quasi-static conditions, is the strain rate, , To be determined parameter.

[0091] (10-2)

[0092] The calculation of parameter b is as follows:

[0093] (10-3)

[0094] in, is the quasi-static parameter predicted in step S16, is the strain rate under quasi-static conditions, is the equivalent average strain rate, c is the undetermined friction coefficient, are the parameters to be fitted. (10-4)

[0095] in, is the failure displacement at different strain rates, Predict the failure displacement under quasi-static conditions for step S16, To be determined parameter.

[0096] [ D , q ] = f 1 ( e ˙ , s c s ) (10-5)

[0097] [ c ] = f 2 ( e ¯ ˙ p , b ) (10-6)

[0098] [ l ] = f ( e ¯ ˙ p , d ) (10-7)

[0099] S22. Based on the parameters obtained by inversion in step S20, apply the mapping relationship obtained in step S21 to predict the material parameters corresponding to equations (10-1), (10-2), and (10-4).

[0100] S23. Encapsulate the material constitutive parameters obtained in steps S20 and S22 and assemble them into a material data set that can be directly used in theoretical calculations and numerical simulations.

[0101] The method of the present invention has broad application prospects in the fields of material science research, engineering structure analysis and optimization, new material design and development, etc. By providing an efficient and accurate means of determining material constitutive parameters, it can meet the needs of modern engineering and scientific research for efficient and accurate data, provide strong data support for material design and performance optimization, help improve the reliability and performance of engineering structures, and promote the progress of material science and engineering technology.

[0102] Parts not described in the above implementation manner can be implemented by adopting or drawing on existing technologies.

[0103] Of course, the above description is not a limitation of the present invention, and the present invention is not limited to the above implementation modes. Any changes, modifications, additions or substitutions made by those skilled in the art within the essential scope of the present invention should fall within the protection scope of the present invention.

Claims

1. A method for determining constitutive parameters of metal materials based on machine learning, characterized in that The following steps are involved: a. Process the test material into test specimens, conduct static and dynamic tests on the test specimens, record the deformation images of the test specimens during the entire loading process, analyze and assemble the strain field cloud map time history matrix ε exp (t); b. Establish the finite element model of the test specimen and obtain the strain field cloud diagram time history matrix ε under each parameter combination FEM (t); and based on the established neural network, the damage stage cloud map obtained in step a but with defects is completed to obtain the strain field cloud map time history matrix ε with complete information rep (t); c. Carry out elastic stage simulation controlled only by elastic parameters, and output the elastic stage strain field cloud map time history matrix obtained by simulation Based on CNN convolutional image recognition technology, the elastic parameters and the elastic stage strain field cloud map time history matrix obtained by the experiment are established. Considering different combinations of plastic parameters, the time-history matrix of the strain field cloud diagram of the plastic stage obtained by simulation is output. Then, the plastic parameters and the time history matrix of the strain field cloud diagram of the plastic stage obtained by the experiment are established. The mapping relationship; d. Carry out numerical simulation controlled by damage parameters to the damage stage, and output the strain field cloud map time history matrix of the damage stage obtained by simulation Establish the time history matrix of damage parameters and strain field cloud diagram of damage stage obtained by experiment The mapping relationship is based on the strain field cloud map time history matrix ε with complete information in step b. rep (t) Obtaining the repaired and adjusted damage stage cloud map and inverting to obtain damage parameters; e. Replace the test cloud diagrams of specimens in different stress states as input to obtain the dynamic constitutive parameters of the material under complex stress states; Step b includes the following steps: b1. Based on the numerical simulation platform, a finite element model of the test specimen is established. The constitutive parameters of the test material are used as control variables to carry out parametric numerical simulation and obtain the strain field cloud diagram time history matrix ε under each parameter combination. FEM (t); b2. Establish a U-NET neural network model, and use the strain field cloud map time history data obtained in step b1 as the data set to train the machine to learn the evolution law of the cloud map from the elastic-plastic stage to the damage stage, and master the ability to generate the corresponding damage stage cloud map from the given elastic-plastic stage cloud map; b3. Establish a PatchGAN neural network model, compare the damage stage cloud maps obtained by numerical simulation and machine prediction in step b1 and step b2 respectively, and calculate the relative error δ between the two; if the error δ is less than the threshold, then feed back the U-NET neural network model to store the corresponding hyperparameters; otherwise, repeat steps b1 and b2 until the PatchGAN neural network model can predict and generate a damage stage cloud map that meets the accuracy requirements; The relative error δ is calculated using the L2 loss function shown in formula (2-1): in, are the hue values ​​at the cloud image pixel coordinates x and y in the prediction and test sets, respectively; b4. Input the damage stage cloud map obtained by the test in step a but with defects into the neural network trained in steps b2 and b3 to obtain the cloud map matrix ε with complete information. rep (t); Step c includes the following steps: c1. Based on the finite element model established in step b1, carry out elastic stage simulation controlled only by elastic parameters, and output the elastic stage strain field cloud diagram time history matrix obtained by simulation The constitutive control formula in the elastic stage is as follows: Where E is the elastic modulus, ν is Poisson's ratio, ε x , ε y , ε z are strains in the x, y, and z directions, respectively, x , σ y , σ z are stresses in the x, y, and z directions respectively; c2. Establish a neural network model CNN1 based on CNN convolutional image recognition technology, input the elastic stage simulation results of step c1 as a training set into the neural network model CNN1 for learning, and establish the elastic parameters and the elastic stage strain field cloud map time history matrix obtained by the simulation output of step c1 The mapping relationship of c3. The strain field cloud map time history matrix ε obtained in step a4 is exp (t), input the trained neural network model CNN1, the machine identifies and inverts the material elastic parameters under this working condition; c4. Apply the elastic parameters determined in step c3, based on the plastic stage constitutive model shown in formula (5-1), consider different combinations of plastic parameters, carry out elastic-plastic simulation analysis, and output the plastic stage strain field cloud map time history matrix obtained by simulation Where k is the flow stress, ε p is the plastic strain; k0 and ε0 are the yield stress and yield strain of the material; A, α, Q, n, β, k0, and ε0 are the parameters to be predicted; c5. Establish a neural network model CNN2, input the simulation results of step c4 as a training set into the neural network model CNN2 for learning, and establish a time history matrix of the parameters to be predicted and the strain field cloud map of the plastic stage obtained in step c4 The mapping relationship; [α、A、Q、n、β、k0、ε0]=f CNN2 (x,y,e,t) (6-1) c6. The strain field cloud map time history matrix ε obtained in step a4 is converted into exp (t) Input the trained neural network model CNN2, and the machine identifies and inverts the plastic parameters of the material under this working condition; Step d includes the following steps: d1. Carry out numerical simulation extended to the damage stage controlled by damage parameters a, b0, c, m, d0, and output the strain field cloud map time history matrix of the damage stage obtained by simulation The behavior of a material in the damage stage depends on its stress state, which is reflected in the variation of material parameters with stress state in experiments and numerical simulations; the stress state is determined by the stress triaxial coefficient η and the normalized Lode angle as shown in the following formula: describe: Among them, σ1, σ2, σ3 are the principal stresses of the unit; ξ is the Lord parameter, which is obtained by formula (7-2); is the normalized Lode angle; The damage stage constitutive model consists of two parts: damage initiation and damage evolution: Among them, f1, f2, and f3 are intermediate variables, which are obtained by the following equations (7-5)-(7-7): in is the damage initiation strain, η is the stress triaxial coefficient, is the normalized Lode angle, d0 is the failure displacement under quasi-static conditions, L is the characteristic size of the numerical simulation unit grid, is the effective plastic strain, is the relative displacement at failure, a, b0, c, m, d0 are all parameters to be predicted; d2. Establish a neural network model CNN3, input the simulation results of step d1 as a training set into the neural network model CNN3 for learning, and establish the mapping relationship between the damage parameters a, b0, c, m, d0 and the strain field cloud map time history matrix obtained in step d1, as shown in formula (8-1): [a, b0, c, m, d0] = f CNN3 (x, y, ε, t) (8-1) where x and y are the directional coordinates of a point in the cloud map, ε is the strain value of a point in the cloud map, and t is the moment of the cloud map in the stretching process; d3, input the cloud map matrix obtained by the experiment in step b4 but after repair and adjustment into the trained neural network model CNN3, and the machine identifies and inverts the damage parameters a, b0, c, m, d0 of the material under this working condition; d4. Carry out numerical simulation based on the constitutive parameters inverted in step c and step d, record the macroscopic response behavior of the specimen, and compare it with the experimental measured data under the corresponding working conditions, as shown in formula (9-1), calculate the relative error Δ between the two, if Δ is less than 0.5%, the prediction is completed; otherwise, increase the number of training times and repeat step c and step d; Where N is the number of data points obtained in the experiment, is the i-th load value obtained in the test, is the machine predicted load value corresponding to the displacement of the i-th data obtained in the test; d5. Replace the test cloud diagrams of specimens in different stress states as input, compare the two results, and verify the accuracy of the constitutive model under different stress states; Compare the inversion results of different stress states, calculate the average absolute percentage error between the constitutive parameters of different stress states, and if the average absolute percentage error is less than 2%, use the average value of the inversion results of different specimens as the constitutive parameter of the material under the complex stress state; if the average absolute percentage error is greater than 2%, expand the scale of the training set in step d1 and perform machine training; Among them, I1 is the length of the strain dimension, and I2 is the length of the stress triaxial coefficient dimension; is the flow stress of the first specimen at the i-th strain value and the j-th stress triaxial coefficient state; is the flow stress of the second specimen at the i-th strain value and the j-th stress triaxial coefficient state.

2. The method for determining constitutive parameters of metal materials based on machine learning according to claim 1, characterized in that: Step a includes the following steps: a1. Process the test material into test specimens with different geometric shapes and sizes, each test specimen corresponds to a different initial stress state; spray speckles of different sizes and random shapes in the gauge area of ​​each test specimen; process multiple test specimens of each type to carry out static and dynamic tests at different loading rates; a2. Use a universal testing machine and a drop weight testing machine to perform static and dynamic tests on the test specimens processed in step a1 respectively; The static test adopts displacement-controlled loading mode, loading until the specimen breaks, and recording the load-displacement time history data of the testing machine; In the dynamic test, the mass of the drop weight in each test is controlled to be M, and the release height of the drop weight is controlled to be H. When conducting a dynamic test, the mass of the drop weight and the release height of the drop weight should be selected to ensure that the test specimen can be broken by the impact. The strain rate in the dynamic test is calculated using formula (1-1); the strain rate at the moment when the yield stress is reached is taken as the calibration strain rate for this group of tests: In the formula, v t (t),v b (t) are the speed history of the upper part and the lower part of the test specimen respectively; l spe is the dimension of the test specimen in the tensile direction; a3. In static tests and dynamic tests, industrial cameras and high-speed cameras are used to record the deformation images of the speckle area of ​​the test specimen during the whole loading process; a4. Apply image recognition and analysis technology to calculate the strain field information of the speckle area and assemble it into a strain field cloud map time history matrix ε exp (t).

3. The method for determining constitutive parameters of metal materials based on machine learning according to claim 2, characterized in that: Step a4 includes the following steps: digitally analyzing the deformation image of the entire loading process recorded in step a3 by image analysis and calculation software to output a strain field cloud map of the test specimen from loading to fracture; uniformly selecting N cloud maps along the time dimension to assemble into a five-dimensional matrix; the first and second dimensions of the matrix record the coordinate information of each data point in the cloud map, and the third to fifth dimensions are the hue, response time and strain rate information of the data point respectively.

4. The method for determining constitutive parameters of metal materials based on machine learning according to claim 1, characterized in that: In step c2, the mean absolute percentage error is used to evaluate the prediction accuracy of the model; [E,n]=f CNN1 (x,y,e,t) (4-1) Where E is the elastic modulus, ν is the Poisson's ratio, x and y are the directional coordinates of a point in the cloud map, ε is the strain value of a point in the cloud map, and t is the time when the cloud map is in the stretching process; N is the number of machine output parameters, is the input parameter corresponding to the training set, is the prediction parameter corresponding to the training set; when MAPE is reduced to below 0.5%, the machine training is completed; otherwise, continue to increase the number of training sets and repeat steps c1 and c2; In step c5, the mean absolute percentage error is also used to evaluate the prediction accuracy of the model. When the mean absolute percentage error is less than 0.5%, the machine training is completed.

5. The method for determining constitutive parameters of metal materials based on machine learning according to claim 1, characterized in that: The following steps are also included: g1. Build a support vector machine model. According to the strain rate hardening model shown in formula (10-1), the damage initiation model shown in formula (10-2), and the damage evolution model shown in formula (10-4), generate training sets of unknown parameters and corresponding variables respectively, and establish mapping relationships as shown in formulas (10-5) to (10-7): Among them, σ cs is the flow stress at different strain rates calculated by the Cowper-Symonds model, σ0 is the flow stress under quasi-static conditions, is the strain rate, D and q are unknown parameters; Among them, parameter b is calculated using formula (10-3): Where b0 is the quasi-static parameter predicted in step d2, is the strain rate under quasi-static conditions, is the equivalent average strain rate, c is the friction coefficient to be determined, γ is the parameter to be fitted, f1, f2, and f3 are determined by formulas (7-5) to (7-7), respectively; Where d is the failure displacement under different strain rates, d0 is the failure displacement under quasi-static conditions predicted in step d2, and λ is an undetermined parameter; g2. Based on the parameters obtained by inversion in step e, apply the mapping relationship obtained in step g1 to predict the material parameters corresponding to equations (10-1), (10-2), and (10-4); g3. Encapsulate the material parameters obtained in step e and step g2 and assemble them into a material data set that can be directly used in theoretical calculation and numerical simulation.

Images