Constructive model parameter calibration method based on machine learning and numerical simulation

Abstract

The invention discloses a constitutive model parameter calibration method based on machine learning and numerical simulation, and the method comprises the following steps: carrying out an indoor triaxial test, and recording and observing a stress-strain test result obtained through the test; integrating the selected constitutive model to obtain a unit stress-strain curve, and fitting the unit stress-strain curve with a test result to obtain an initial constitutive model parameter; carrying out whole-process numerical simulation on the indoor triaxial test by adopting a numerical simulation means; comparing a whole-process simulation result with a test result; and the constructed machine learning agent model adjusts the initially determined constitutive model parameters and the initially determined non-local parameters to obtain real constitutive model parameters and non-local parameters. According to the constitutive model parameter calibration method based on machine learning and numerical simulation, the influence of strain localization in a triaxial test is eliminated through proofreading of a numerical simulation result and a real triaxial test result based on a machine learning synergy technology, and the efficiency and precision of geotechnical engineering analysis are improved.

Description

Technical Field

[0001] The present invention relates to the field of geotechnical engineering technology, and in particular to a constitutive model parameter calibration method based on machine learning and numerical simulation. Background Art

[0002] The triaxial test is a widely used test for studying the strength and deformation characteristics of soil. The test is simple to operate and can reflect the mechanical properties of soil under complex stress conditions. Ideally, it is a unit test, and theoretically, the stress and strain conditions at all locations inside the specimen should be the same. However, the soil with softening characteristics will have shear bands during the test, resulting in uneven stress and strain distribution, which is the strain localization of the triaxial specimen. Due to the existence of strain localization, the test results obtained from conventional triaxial tests cannot reflect the true strength and deformation characteristics of a point in the soil.

[0003] Scholars at home and abroad have tried to study the problem of strain localization in triaxial tests in different ways to provide a basis for calibrating the parameters of the constitutive model. Existing studies on the influence of strain localization in triaxial tests mostly focus on measuring the deformation characteristics of the specimen, and interpreting and correcting the test data based on the measurement situation. Local deformation measurement based on local displacement sensors in triaxial tests is an earlier method for measuring strain localization. This method arranges micro-displacement sensors on the surface of the specimen to record the displacement changes of the specimen in different directions. LVDT (Linear Variable Differential Displacement Transducer) is a commonly used high-precision displacement sensor that can be used to accurately collect displacement data of local points of triaxial specimens. However, the technology based on local displacement sensors can only obtain the deformation characteristics of a certain part of the specimen, and it is difficult to fully describe the non-uniform deformation of all areas on the triaxial specimen. In addition, the operation requirements of this equipment are high, which is easy to affect the triaxial test process.

[0004] With the continuous advancement of computer graphics and digital image processing technology, image recognition technology has been gradually introduced into the measurement method of triaxial test. Using image recognition technology, the radial and axial displacements of each point on the surface of the triaxial specimen can be measured, thereby obtaining the strain distribution and local strain field of the entire specimen. This method can comprehensively collect information on the specimen surface, performs well in observation accuracy and range, and can effectively measure the strain localization of triaxial specimens. However, this technology has high requirements for equipment, has limitations when measuring larger specimens, and has strict requirements on specimen size.

[0005] In summary, the existing research on strain localization in triaxial tests is mostly focused on measuring the local strain size and correcting and interpreting the test data based on the measurement situation. However, the method of eliminating the influence of strain localization based on the local deformation characteristics of the specimen is complicated, and the quality of data correction is difficult to guarantee, which cannot effectively solve the problem of strain localization in triaxial tests. When the stress-strain law of triaxial tests that does not eliminate the influence of strain localization is used to calibrate the constitutive model parameters, a stress result that is lower than the actual situation will be obtained, which cannot reasonably reflect the true mechanical properties of the soil. In addition, previous studies have not proposed an effective solution to the problem of how to calibrate the constitutive model parameters using the corrected triaxial test data. Summary of the invention

[0006] The purpose of the present invention is to provide a constitutive model parameter calibration method based on machine learning and numerical simulation, which eliminates the influence of strain localization in triaxial tests and improves the efficiency and accuracy of geotechnical engineering analysis by calibrating numerical simulation results with actual triaxial test results based on machine learning synergy enhancement technology.

[0007] To achieve the above object, the present invention provides a constitutive model parameter calibration method based on machine learning and numerical simulation, comprising the following steps: Step S1, performing an indoor triaxial test, recording the stress-strain test results obtained from the test, and observing the test stress-strain curve; Step S2, according to the initial stress state, void ratio and loading path of the actual triaxial test, the selected constitutive model is subjected to constitutive integral calculation to obtain the unit stress-strain curve; and the constitutive model parameters are continuously adjusted to fit the test stress-strain curve to obtain the initial constitutive model parameters; Step S3, using a numerical simulation method based on a non-local regularized constitutive model and a fluid-solid coupling numerical analysis method as the core, to perform a full-process numerical simulation of the indoor triaxial test; Step S4, comparing the stress-strain results obtained by simulating the whole process of step S3 with the experimental stress-strain curve in step S1; if the comparison results are similar, the initial constitutive model parameters and the initial non-local parameters are the final calibration parameters; if the comparison results are different, proceed to step S5; Step S5, make the stress-strain curve calculated based on numerical simulation consistent with the test data, construct a machine learning proxy model, adjust the initial constitutive model parameters and the initial non-local parameters, so that the stress-strain curve obtained based on the machine learning proxy model is consistent with the test stress-strain curve, and finally obtain the real constitutive model parameters and non-local parameters.

[0008] Preferably, in step S1, the test stress-strain curve is observed. If strain hardening occurs, the constitutive model parameters that truly reflect the mechanical properties of the soil can be obtained based on the test stress-strain curve without recalibration; if strain softening occurs, proceed to step S2.

[0009] Preferably, in step S3, the non-local regularized constitutive model introduces two non-local parameters in addition to its own constitutive model parameters: and In the numerical simulation, the constitutive model parameters adopt the initial constitutive model parameters of step S2, and the two non-local parameters are initially determined as follows: =twice the cell size; =1.

[0010] Preferably, the whole process of indoor triaxial test is numerically simulated, and the specific steps are as follows: Step S321, according to the actual size of the indoor triaxial test specimen, a triaxial specimen model considering fluid-solid coupling characteristics is established in the numerical analysis software; Step S322, applying a stress boundary condition that increases linearly from the top to the bottom of the sample according to the gravity gradient to the side of the simulated sample to simulate water injection; Step S323, using the back pressure saturation method to simulate the soil saturation process; Step S324, applying different total stress increments to the sample in the horizontal and vertical directions to simulate the K0 stress state of the soil in the original stratum; Step S325, simulating the shearing process of the triaxial test; Step S326, post-processing the results of the numerical simulation to obtain the stress-strain curve of the sample under the boundary conditions, constraint conditions and parameter value settings specified in step S1.

[0011] Preferably, in step S321, first, adopt The weight function performs a weighted average of the strains in the spatial neighborhood of the Gaussian point, and uses the obtained weighted strain to replace the original strain of the point; then, according to the in-situ conditions of the soil sample used in the test, the initial stress distribution in the sample is set, including vertical effective stress, horizontal effective stress and pore water pressure; finally, by adjusting the non-local parameters and To simulate the strain softening phenomenon of triaxial test.

[0012] Preferably, in step S323, the back pressure saturation method is used to simulate the soil saturation process. In order to change the pore water pressure in the soil from the suction in the unsaturated state to the pressure in the saturated state, while ensuring that the effective stress of the soil remains unchanged during the process, the same total stress and pore water pressure increment are applied to the top and bottom of the sample.

[0013] Preferably, in step S325, without changing the radial confining pressure of the sample, for the undrained shear process, the upper and lower surfaces and the side edges of the sample are set as impermeable interfaces; for the drained shear process, the upper and lower surfaces and the side edges of the sample are set as permeable interfaces; At the same time, in order to consider the effect of permeable stone in the experiment, the pore water pressure degrees of freedom on the upper and lower surfaces of the specimen were constrained; vertical displacement was applied to all nodes on the upper surface of the specimen, and the time required for each loading step was set according to the loading rate of the actual test.

[0014] Preferably, in step S326, the method of obtaining the shear stress and axial strain by post-processing is as follows: The stress of the specimen in the vertical direction is obtained by dividing the sum of the vertical reaction forces at all nodes on the top surface by the cross-sectional area of ​​the specimen. In order to be consistent with the small strain assumption used in the numerical simulation, the cross-sectional area of ​​the specimen is taken as the cross-sectional area of ​​the initial specimen. If the large strain assumption is adopted, the specimen cross-sectional area correction is required. The shear stress can be obtained by subtracting the vertical stress from the confining pressure. The axial strain is determined by dividing the total displacement applied to the top of the specimen by the initial height of the specimen.

[0015] Preferably, in step S5, a machine learning agent model is established, and the specific steps are as follows: Step S51, dimensionality reduction process: using principal component analysis method to extract data features of the original high-dimensional output space , based on a predefined threshold , choose the best reduction number , the original output space Dimensionality reduction to a simplified output space middle; Step S52, constructing a proxy model: according to the simplified output space obtained from the dimensionality reduction process, the input parameters and the output results are fitted using a polynomial chaos expansion method, thereby constructing a machine learning proxy model; Step S53: construct a training data set for the machine learning agent model using a numerical simulation software analysis method, and train the machine learning agent model in step S52; Step S54, continuously adjust the size of the constitutive model input parameters, and use the trained machine learning agent model to output the simulation results, compare the simulation results with the experimental stress-strain curve, so as to obtain the constitutive model input parameters that make the two highly fit.

[0016] Therefore, the present invention adopts the above-mentioned constitutive model parameter calibration method based on machine learning and numerical simulation, and the beneficial effects are as follows: (1) This invention proposes for the first time a method for calibrating parameters of a soil constitutive model under the influence of strain localization in triaxial tests. It uses a method for interactively correcting triaxial test results of soil with softening characteristics and numerical simulation results to solve the problem of incorrect calibration of constitutive model parameters due to soil softening in triaxial tests. (2) The present invention adopts a non-local regularization method when performing numerical simulation, which avoids the dependence of stress-strain results on the grid during simulation and can accurately and efficiently calibrate the constitutive model parameters; (3) The present invention uses the principal component analysis method and the polynomial chaos expansion method to construct a machine learning proxy model, so that the stress-strain output results corresponding to the input parameters of the constitutive model can be obtained through training with a small number of samples. The introduction of the machine learning proxy model greatly reduces the computing time and computing resource consumption required for fitting the test results; (4) The present invention utilizes a machine learning proxy model to obtain output results with different input parameters. The entire method has high repeatability, low operational difficulty, and does not require any additional experimental equipment, thereby greatly reducing costs.

[0017] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 It is a technical flow chart of a constitutive model parameter calibration method based on machine learning and numerical simulation of the present invention; Figure 2 It is a numerical simulation process diagram of the present invention; Figure 3 It is the technical roadmap of the machine learning agent model of the present invention. DETAILED DESCRIPTION

[0019] The technical solution of the present invention is further described below through the accompanying drawings and embodiments.

[0020] like Figure 1 As shown, a constitutive model parameter calibration method based on machine learning and numerical simulation includes the following steps: Step S1, perform an indoor triaxial test and record the stress-strain test results obtained in the test; wherein the stress-strain test results include: ① stress-strain curve and ② deformation of the sample.

[0021] In the present invention, the test stress-strain curve, i.e., test result ①, is observed. If strain hardening occurs, the constitutive model parameters that truly reflect the mechanical properties of the soil can be obtained based on the test result ① without recalibration. If strain softening occurs, the process proceeds to step S2.

[0022] Step S2: According to the initial stress state, void ratio and loading path of the actual triaxial test, the constitutive integral calculation is performed on the selected constitutive model to obtain the unit stress-strain curve. The constitutive model parameters are continuously adjusted, and the test results are fitted to obtain the initial constitutive model parameters.

[0023] Step S3: using a numerical simulation method based on a non-local regularized constitutive model and a fluid-solid coupling numerical analysis method as the core, a full-process numerical simulation of the indoor triaxial test is performed.

[0024] Step S4, obtain a stress-strain curve based on numerical simulation calculation by simulating the whole process of step S3. Compare the result with the test result ① in step S1; if the comparison results are similar, the initial constitutive model parameters and the initial non-local parameters are the final calibration parameters; if the comparison results are different, proceed to step S5.

[0025] Step S5: In order to make the stress-strain curve calculated based on the numerical simulation consistent with the test data, the initial constitutive model parameters and the initial non-local parameters are adjusted.

[0026] The peak and residual phases of the stress-strain curve are calibrated by adjusting the constitutive model parameters, and the nonlocal parameters The softening rate calibration is completed by adjustment, and the constructed machine learning proxy model is used to output the stress-strain simulation results corresponding to different input parameters, until the stress-strain curve obtained based on the machine learning proxy model is consistent with the experimental results①, and finally the real constitutive model parameters and non-local parameters are obtained.

[0027] Example Step S1, conduct an indoor triaxial test and record the stress-strain test results obtained in the test; wherein the stress-strain test results include: ① stress-strain curve and ② deformation of the sample.

[0028] In the present invention, the test stress-strain curve, i.e., test result ①, is observed. If strain hardening occurs, the constitutive model parameters that truly reflect the mechanical properties of the soil can be obtained based on the test result ① without recalibration. If strain softening occurs, the process proceeds to step S2.

[0029] Step S2: According to the initial stress state, void ratio and loading path of the actual triaxial test, the constitutive integral calculation is performed on the selected constitutive model to obtain the unit stress-strain curve. The constitutive model parameters are continuously adjusted, and the test results are fitted to obtain the initial constitutive model parameters.

[0030] Step S3: using a numerical simulation method based on a non-local regularized constitutive model and a fluid-solid coupling numerical analysis method as the core, a full-process numerical simulation of the indoor triaxial test is performed.

[0031] Step S31: The constitutive model of non-local regularization introduces two non-local parameters in addition to its own constitutive model parameters: and In the numerical simulation, the constitutive model parameters adopt the initial constitutive model parameters of step S2, and two non-local parameters are initially determined as follows: =twice the cell size; =1; Step S32: Based on the Geotechnical Test Method Standard GB / T 50123-2019, a full-process numerical simulation of the indoor triaxial test is performed. Figure 2 As shown, the specific steps are as follows: Step S321: According to the actual size of the indoor triaxial test specimen, a triaxial specimen model considering fluid-solid coupling characteristics is established in the numerical analysis software.

[0032] In order to solve the grid dependence and possible convergence problems of the finite element method in numerical simulation analysis, the present invention adopts the method of The weight function performs weighted average of the strains in the spatial neighborhood of a Gaussian point and uses the obtained weighted strain to replace the original strain of the point.

[0033] According to the in-situ conditions of the soil sample used in the test, the initial stress distribution in the sample is set, including vertical effective stress, horizontal effective stress and pore water pressure. The distribution of each stress should take into account the influence of gravity gradient. Then, by adjusting the non-local parameters and To accurately simulate the strain softening phenomenon of triaxial test.

[0034] Step S322: Apply a stress boundary condition that increases linearly from the top to the bottom of the sample according to the gravity gradient to the side of the simulated sample to simulate water injection.

[0035] Step S323: adopt the back pressure saturation method to simulate the soil saturation process. The specific process is as follows: In order to change the pore water pressure in the soil from suction in an unsaturated state to pressure in a saturated state, while ensuring that the effective stress of the soil remains unchanged during the process, the same total stress and pore water pressure increment are applied to the top and bottom of the specimen.

[0036] Step S324, simulating the K0 stress state of the soil in the original stratum, the specific process is as follows: Different total stress increments are applied to the specimens in the horizontal and vertical directions (a sufficiently long time step is set in the process to allow the soil to be fully drained) to simulate the K0 stress state of the soil in the original stratum.

[0037] Step S325, simulating the shearing process of the triaxial test, the specific process is as follows: Without changing the radial confining pressure of the specimen, for the undrained shear process, the upper and lower surfaces and sides of the specimen are set as impermeable interfaces; for the drained shear process, the upper and lower surfaces and sides of the specimen are set as permeable interfaces.

[0038] At the same time, in order to consider the effect of permeable stone in the experiment, the pore water pressure freedom on the upper and lower surfaces of the specimen was constrained.

[0039] Apply vertical displacement to all nodes on the upper surface of the specimen, and set the time required for each loading step according to the loading rate of the actual test.

[0040] Step S326, post-processing the results of the numerical simulation to obtain the stress-strain curve of the sample under the boundary conditions, constraint conditions and parameter value settings specified in step S1.

[0041] The post-processing method to obtain shear stress and axial strain is: The stress of the specimen in the vertical direction is obtained by dividing the sum of the vertical reaction forces at all nodes on the top surface by the cross-sectional area of ​​the specimen. In order to be consistent with the small strain assumption used in the numerical simulation, the cross-sectional area of ​​the specimen is taken as the cross-sectional area of ​​the initial specimen. If the large strain assumption is adopted, the specimen cross-sectional area correction is required. The shear stress can be obtained by subtracting the vertical stress from the confining pressure.

[0042] The axial strain is determined by dividing the total displacement applied to the top of the specimen by the initial height of the specimen.

[0043] Step S4: obtain a stress-strain curve based on numerical simulation calculation by simulating the whole process of step S3, and compare the result with the test result ① in step S1.

[0044] If the comparison results are similar, the initial constitutive model parameters and the initial non-local parameters are the final calibration parameters. If the comparison results are different, the process proceeds to step S5.

[0045] It should be noted that due to the influence of stress history (overconsolidation, etc.) and boundary conditions, the unit stress-strain curve will be different from the stress-strain curve based on numerical simulation, especially in the softening stage and residual stage where strain localization occurs in the specimen, the two results will be very different.

[0046] Step S5: In order to make the stress-strain curve calculated based on the numerical simulation consistent with the test data, the initial constitutive model parameters and the initial non-local parameters are adjusted.

[0047] The peak and residual phases of the stress-strain curve are calibrated by adjusting the constitutive model parameters, and the nonlocal parameters The softening rate calibration is completed by adjustment, and the constructed machine learning proxy model is used to output the stress-strain simulation results corresponding to different input parameters, until the stress-strain curve obtained based on the machine learning proxy model is consistent with the experimental results①, and finally the real constitutive model parameters and non-local parameters are obtained.

[0048] Among them, in order to increase the computing efficiency, a machine learning agent model is established such as Figure 3 As shown, the specific steps are as follows: Step S51, dimensionality reduction process: using principal component analysis (PCA) to extract the data features (such as stress-strain curves) of the original high-dimensional output space , based on a predefined threshold , choose the best reduction number , the original output space Dimensionality reduction to a simplified output space middle.

[0049] Step S52, constructing a proxy model: according to the simplified output space obtained from the dimensionality reduction process, the input parameters and output results are fitted using the Polynomial chaos expansions (PCE) method to construct a machine learning proxy model.

[0050] Step S53: construct a training data set for the machine learning agent model using a numerical simulation software analysis method, and train the machine learning agent model in step S52.

[0051] Step S54, continuously adjust the size of the constitutive model input parameters, and use the trained machine learning agent model to output the simulation results, compare the simulation results with the test results①, so as to obtain the constitutive model input parameters that make the two highly fit.

[0052] Therefore, the present invention adopts the above-mentioned constitutive model parameter calibration method based on machine learning and numerical simulation, based on the idea of ​​"digital twin", adopts the numerical simulation means with non-local regularized constitutive model and fluid-solid coupling numerical analysis method as the core, and carries out refined numerical simulation inversion for the whole process of indoor soil test, and adopts the construction technology of machine learning proxy model combining PCA with PCE method, which greatly increases the calculation efficiency and reduces the calculation cost. Through the interactive correction of the output results of the machine learning proxy model and the actual test results, the influence mechanism of the non-uniform deformation of the specimen when softening in the triaxial test on the calibration of the constitutive model parameters is analyzed, and finally a model parameter calibration method suitable for different test conditions and deformation characteristics of different types of soil is established, so as to realize the correct description of the true stress-strain relationship of the soil by the constitutive model.

[0053] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solution of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solution to deviate from the spirit and scope of the technical solution of the present invention.

Claims

1. A constitutive model parameter calibration method based on machine learning and numerical simulation, characterized in that: The following steps are involved: Step S1, conducting an indoor triaxial test, recording the stress-strain test results obtained from the test, and observing the test stress-strain curve; Step S2, according to the initial stress state, void ratio and loading path of the actual triaxial test, the selected constitutive model is subjected to constitutive integral calculation to obtain the unit stress-strain curve; and the constitutive model parameters are continuously adjusted to fit the test stress-strain curve to obtain the initial constitutive model parameters; Step S3, using a numerical simulation method based on a non-local regularized constitutive model and a fluid-solid coupling numerical analysis method as the core, to perform a full-process numerical simulation of the indoor triaxial test; Step S4, comparing the stress-strain results obtained by simulating the whole process of step S3 with the experimental stress-strain curve in step S1; if the comparison results are similar, the initial constitutive model parameters and the initial non-local parameters are the final calibration parameters; if the comparison results are different, proceed to step S5; Step S5, make the stress-strain curve calculated based on numerical simulation consistent with the test data, construct a machine learning proxy model, adjust the initial constitutive model parameters and the initial non-local parameters, so that the stress-strain curve obtained based on the machine learning proxy model is consistent with the test stress-strain curve, and finally obtain the real constitutive model parameters and non-local parameters.

2. The constitutive model parameter calibration method based on machine learning and numerical simulation according to claim 1, characterized in that: In step S1, the test stress-strain curve is observed. If strain hardening occurs, the constitutive model parameters that truly reflect the mechanical properties of the soil can be obtained based on the test stress-strain curve without recalibration; if strain softening occurs, the process proceeds to step S2.

3. The constitutive model parameter calibration method based on machine learning and numerical simulation according to claim 1, characterized in that: In step S3, the nonlocal regularized constitutive model introduces two nonlocal parameters in addition to its own constitutive model parameters: and In the numerical simulation, the constitutive model parameters adopt the initial constitutive model parameters of step S2, and the two non-local parameters are initially determined as follows: =twice the cell size; =1.

4. The constitutive model parameter calibration method based on machine learning and numerical simulation according to claim 3, characterized in that: The whole process of indoor triaxial test is numerically simulated. The specific steps are as follows: Step S321, according to the actual size of the indoor triaxial test specimen, a triaxial specimen model considering fluid-solid coupling characteristics is established in the numerical analysis software; Step S322, applying a stress boundary condition that increases linearly from the top to the bottom of the sample according to the gravity gradient to the side of the simulated sample to simulate water injection; Step S323, using the back pressure saturation method to simulate the soil saturation process; Step S324, applying different total stress increments to the sample in the horizontal and vertical directions to simulate the K0 stress state of the soil in the original stratum; Step S325, simulating the shearing process of the triaxial test; Step S326, post-processing the results of the numerical simulation to obtain the stress-strain curve of the sample under the boundary conditions, constraint conditions and parameter value settings specified in step S1.

5. The constitutive model parameter calibration method based on machine learning and numerical simulation according to claim 4, characterized in that: In step S321, first, adopt The weight function performs a weighted average of the strains in the spatial neighborhood of the Gaussian point, and uses the obtained weighted strain to replace the original strain of the point; then, according to the in-situ conditions of the soil sample used in the test, the initial stress distribution in the sample is set, including vertical effective stress, horizontal effective stress and pore water pressure; finally, by adjusting the non-local parameters and To simulate the strain softening phenomenon of triaxial test.

6. The constitutive model parameter calibration method based on machine learning and numerical simulation according to claim 4, characterized in that: In step S323, the back pressure saturation method is used to simulate the soil saturation process. In order to change the pore water pressure in the soil from the suction in the unsaturated state to the pressure in the saturated state, while ensuring that the effective stress of the soil remains unchanged during the process, the same total stress and pore water pressure increment are applied to the top and bottom of the sample.

7. The constitutive model parameter calibration method based on machine learning and numerical simulation according to claim 4, characterized in that: In step S325, without changing the radial confining pressure of the sample, for the undrained shear process, the upper and lower surfaces and the side edges of the sample are set as impermeable interfaces; for the drained shear process, the upper and lower surfaces and the side edges of the sample are set as permeable interfaces; At the same time, in order to consider the effect of permeable stone in the experiment, the pore water pressure degrees of freedom on the upper and lower surfaces of the specimen were constrained; vertical displacement was applied to all nodes on the upper surface of the specimen, and the time required for each loading step was set according to the loading rate of the actual test.

8. The constitutive model parameter calibration method based on machine learning and numerical simulation according to claim 4, characterized in that: In step S326, the method of obtaining the shear stress and axial strain by post-processing is as follows: The stress of the specimen in the vertical direction is obtained by dividing the sum of the vertical reaction forces at all nodes on the top surface by the cross-sectional area of ​​the specimen. In order to be consistent with the small strain assumption used in the numerical simulation, the cross-sectional area of ​​the specimen is taken as the cross-sectional area of ​​the initial specimen. If the large strain assumption is adopted, the specimen cross-sectional area correction is required. The shear stress can be obtained by subtracting the vertical stress from the confining pressure. The axial strain is determined by dividing the total displacement applied to the top of the specimen by the initial height of the specimen.

9. The constitutive model parameter calibration method based on machine learning and numerical simulation according to claim 1, characterized in that: In step S5, a machine learning agent model is established. The specific steps are as follows: Step S51, dimensionality reduction process: using principal component analysis method to extract data features of the original high-dimensional output space , based on a predefined threshold , choose the best reduction number , the original output space Dimensionality reduction to a simplified output space middle; Step S52, constructing a proxy model: according to the simplified output space obtained from the dimensionality reduction process, the input parameters and the output results are fitted using a polynomial chaos expansion method, thereby constructing a machine learning proxy model; Step S53: construct a training data set for the machine learning agent model using a numerical simulation software analysis method, and train the machine learning agent model in step S52; Step S54, continuously adjust the size of the constitutive model input parameters, and use the trained machine learning agent model to output the simulation results, compare the simulation results with the experimental stress-strain curve, so as to obtain the constitutive model input parameters that make the two highly fit.

Images