Hyperelastic constitutive model method based on deep regression and physical information network

Through deep regression and physical information network methods, the constitutive model of hyperelastic materials is automatically constructed, which solves the problems of complexity and insufficient prediction of constructing and selecting constitutive models in the prior art, and achieves efficient performance reflection of hyperelastic materials and complex behavior prediction.

CN118428211BActive Publication Date: 2024-11-12NINGXIA UNIVERSITY
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410492047.8
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-04-23
Publication Date
2024-11-12
Estimated Expiration
2044-04-23

AI Technical Summary

Technical Problem

The prior art requires a lot of experimental testing and experience when constructing and selecting constitutive models of hyperelastic materials, which is difficult to automatically reflect the superelastic properties of the material, and the existing models are insufficient in predicting complex hyperelastic behaviors.

Method used

Using a method based on deep regression and physical information network, deep regression network and physical information network are trained through experimental data sets, constitutive models of hyperelastic materials are automatically derived, explicit expressions of strain energy functions are generated, and model parameters are determined.

Benefits of technology

The automated modeling of the hyperelastic material constitutive model is implemented, which simplifies the testing process, can predict complex hyperelastic behaviors that existing models cannot be competent, and reduces the problem of model inapplicability due to insufficient experience.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118428211B_ABST
    Figure CN118428211B_ABST
Patent Text Reader

Abstract

The present invention discloses a hyperelastic constitutive model method based on deep regression and physical information network, comprising the following steps: step one: experimental data set acquisition, used to generate input data of the deep regression network of step two and step three; step two: physical information network construction, used to predict the mapping relationship between invariants and strain energy functions, so as to perform the optimization training of the deep regression network in step three; step three: deep regression network construction, used to generate a hyperelastic constitutive model that can reflect the nonlinear behavior of a hyperelastic material plate; step four: neural network training based on the experimental data set of the hyperelastic material obtained in step one, and joint training of the physical information network in step two and the deep regression network in step three to find the optimal network parameters. The present invention adopts a machine learning method to automatically derive the constitutive model of the hyperelastic material from the experimental data set to reflect the hyperelastic performance of the material.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention belongs to the technical field of constitutive modeling of hyperelastic materials in solid mechanics, and in particular relates to a hyperelastic constitutive model method based on deep regression and physical information network. Background Art

[0002] Hyperelastic materials such as rubber are widely used in medical equipment, industry, construction, and national defense for seismic isolation and insulation. Hyperelastic materials are typical nonlinear elastic materials. Considering the differences in the performance of hyperelastic materials under different production processes and even different batches, hyperelastic constitutive modeling is carried out for a given material to reflect the hyperelastic properties of the material, which becomes the key to quantitative analysis of material deformation under different loads. The Neo-Hookean model was proposed by mechanic Ronald Rivlin in 1948. Over the past seventy years, mechanics have proposed many constitutive models of hyperelastic materials, such as the famous Ogden, Hill, Druker-Prager and other models.

[0003] When conducting hyperelastic constitutive modeling for a given material, a constitutive model of a certain expression must be selected first, and then relevant mechanical tests must be carried out, such as uniaxial tension and compression tests, equibiaxial tension and compression tests, etc., and based on the obtained experimental data set, combined with the optimization algorithm, the model parameters of the pre-selected constitutive model are determined. It is worth noting that engineers and technicians must have rich experience in hyperelastic constitutive and experimental testing to select a suitable constitutive model. In more cases, engineers and technicians often need to calibrate several or even dozens of constitutive models to find a hyperelastic constitutive model that can reflect the nonlinear behavior of the tested material as much as possible. Of course, each calibration also needs to rely on the constitutive test experience of hyperelastic materials and select an appropriate parameter calibration algorithm. For situations where the hyperelastic behavior of materials is very complex, the existing constitutive models cannot well predict the experimental data laws obtained from mechanical tests. In summary, how to construct and select a reasonable constitutive model to accurately describe the mechanical response of hyperelastic materials is still one of the difficult issues in the study of hyperelastic materials. Summary of the invention

[0004] In order to overcome the deficiencies of the above-mentioned prior art, the purpose of the present invention is to provide a hyperelastic constitutive model method based on deep regression and physical information network, which adopts a machine learning method to automatically derive the constitutive model of hyperelastic materials from experimental data sets to reflect the hyperelastic properties of the material.

[0005] In order to achieve the above object, the technical solution adopted by the present invention is:

[0006] A hyperelastic constitutive model method based on deep regression and physical information network includes the following steps;

[0007] Step 1: Acquire the experimental data set to generate input data for the deep regression network in steps 2 and 3;

[0008] Step 2: Building a physical information network; used to predict the mapping relationship between the invariants I1, I2 and the strain energy function W(I1, I2), and used for the optimization training of the deep regression network in step 3;

[0009] Step 3: Building a deep regression network to generate a hyperelastic constitutive model that can reflect the nonlinear behavior of the hyperelastic material plate;

[0010] Step 4: Perform neural network training based on the experimental data set of hyperelastic materials obtained in step 1, and jointly train the physical information network in step 2 and the deep regression network in step 3 to find the optimal network parameters.

[0011] The step 1 is specifically as follows:

[0012] Firstly, a mechanical test was carried out on an incompressible hyperelastic rectangular plate, and uniaxial tension and compression tests and equibiaxial tension and compression tests were performed to obtain the experimental data set. and nominal stress Experimental datasets of constructed hyperelastic materials;

[0013] Based on hyperelastic theory, calculation of different stretching ratios The corresponding deformation invariant under the combined parameters

[0014]

[0015] Among them, I1 and I2 are two invariants of the Cauchy Green deformation tensor.

[0016] The step 2 is specifically as follows:

[0017] Build a physical information network Wnet to predict the strain energy function W(I1,I2);

[0018] W pred (I1,I2)=W net (I1,I2;w,b) (2)

[0019] Among them, W pred is the predicted strain energy function based on the two invariants I1 and I2, W net represents a fully connected neural network, w represents the weight of the fully connected neural network, and b represents the bias of the fully connected neural network;

[0020] Combining the automatic differentiation computing capability of neural networks, the following parameters of the constitutive model are calculated:

[0021]

[0022] Further, the constitutive model parameters are combined to calculate the different stretching ratios Nominal stress prediction value under combination

[0023]

[0024] The loss function of the physical information network is defined as the measurement point stretching ratio The true value of the nominal stress at and the predicted value P pred Mean square error Loss

[0025]

[0026] in is the true value of the nominal stress at the measuring point, P pred is the predicted value of the nominal stress at the measuring point, N P is the nominal stress data point The number of sampling points.

[0027] The step three is specifically as follows:

[0028] Construct a deep regression network to analyze the invariants I1, I2 and the strain energy function W * (I1, I2), for deep regression networks, symbolic expressions are treated as binary trees, where each node represents a symbol of an expression in the available symbol library, i.e., the input variables are the deformation invariants (I1, I2) of the hyperelastic plate, constants (e.g., C 10 ,C 20 ,C 30 ) or operation (e.g., +, -, ×, / , sin, log, exp…), the final output of the deep regression network is the predicted strain energy function W * Explicit expression of (I1,I2);

[0029] For each candidate W * (I1, I2), the fitting quality reward R = 1 / (1 + NRMSE) is calculated in the deep regression network, where NRMSE is calculated by the target σ W Bias-normalized root mean square error:

[0030]

[0031] where σ W is the variance of the strain energy data set calculated by the target strain energy function, and W(I1,I2) is the physical information network W net The predicted strain energy function, W *(I1,I2) is the strain energy function predicted by the deep regression network.

[0032] The step 4 is specifically as follows:

[0033] When the loss function Loss of the physical information network and the reward R of the fitting quality of the deep regression network meet the conditions at the same time, stop the network training;

[0034] Different thresholds are set for the physical information network and the deep regression network. That is, when the loss function Loss of the physical information network is less than a certain threshold (such as 10 -4 ), and the fitting quality reward R exceeds a certain threshold (such as 99%) at the same time, the network training is stopped. Taking into account the differences in training efficiency and convergence effect between deep regression and physical information networks, an iterative alternating training scheme is adopted. First, the loss function Loss of the physical information network is reduced, and then the fitting quality reward R of the deep regression network is trained. The network joint training is realized through an iterative algorithm.

[0035] Through the above steps, the final output of the deep regression network is the expression W of the strain energy function of the hyperelastic constitutive model obtained by this constitutive modeling method * (I1, I2), that is, the hyperelastic constitutive model that reflects the nonlinear behavior of the hyperelastic material plate, and the strain energy function W of the constitutive model * The model parameters in (I1, I2) are determined simultaneously by a deep regression network. The obtained hyperelastic constitutive model can be verified by comparing with the experimental data set.

[0036] Beneficial effects of the present invention:

[0037] In summary, the method proposed in the present invention realizes the experimental data set obtained through mechanical testing experiments, combines deep regression with the training of physical information networks, directly obtains the explicit expression of the strain energy function in the hyperelastic constitutive model, and simultaneously determines the model parameters in the constitutive model.

[0038] The present invention realizes a neural network automated modeling scheme for a hyperelastic constitutive model based on an experimental data set. The constitutive model is completely obtained by neural network training, and the strain energy function expression of the model result may be different from all traditional hyperelastic constitutive models. The more possibilities of the constitutive model enable it to automatically discover new hyperelastic constitutive models and have the ability to predict complex hyperelastic behaviors that existing constitutive models are unable to cope with.

[0039] Engineering technicians often need to calibrate several or even dozens of constitutive models to find the hyperelastic constitutive model that reflects the nonlinear behavior of the material as much as possible. And each calibration needs to rely on relevant test experience and appropriate parameter calibration optimization algorithm. The hyperelastic constitutive model method based on deep regression and physical information network provided by the present invention does not require engineering technicians to perform repeated tests, greatly simplifies the test process of the hyperelastic constitutive model, and greatly reduces the problem of model inapplicability caused by lack of experience in the test of the hyperelastic constitutive model. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 Schematic diagram of the process flow of the hyperelastic constitutive model scheme based on deep regression and physical information network.

[0041] Figure 2 Schematic diagram of the hyperelastic constitutive and deformation analysis of a material based on a hyperelastic test case of this method. DETAILED DESCRIPTION

[0042] The present invention is further described in detail below in conjunction with embodiments.

[0043] The present invention realizes an artificial intelligence neural network automated modeling scheme of a hyperelastic constitutive model based on a deep regression and physical information network.

[0044] The following is a detailed description of the present invention in steps. Figure 1 As shown:

[0045] Step 1: Acquisition of experimental data set, which is used to generate input data for steps 2 and 3; First, relevant mechanical test experiments are carried out on an incompressible hyperelastic rectangular plate, and uniaxial tension and compression tests and equibiaxial tension and compression tests are carried out to obtain experimental data sets, and the tensile ratio is obtained. and nominal stress The experimental data set of hyperelastic materials composed of biaxial tensile test was used as an example to prepare the flat specimens for the test and carry out biaxial tensile test to obtain the tensile ratio of and nominal stress The experimental data set of hyperelastic materials composed of Figure 1 Specifically, a relevant tensile test is carried out on a rectangular superelastic plate specimen with a length of x (m) and a width of y (m), and the tensile strengths of different tensile ratios are obtained. Nominal stress under combined parameters The experimental data of the stretching ratio and nominal stress Experimental data set of hyperelastic materials composed of etc.

[0046] Based on hyperelastic theory, calculation of different stretching ratios The corresponding deformation invariant under the combined parameters. Taking the biaxial tensile test as an example, different stretching ratios The calculation formula of the corresponding deformation invariant under the combined parameters is as follows:

[0047]

[0048] Among them, I1 and I2 are two invariants of the Cauchy Green deformation tensor;

[0049] Step 2: Build a physical information network; used to predict the mapping relationship between I1, I2 and strain energy function W(I1, I2); build a physical information network Wnet to predict the strain energy function W(I1, I2).

[0050] First, a fully connected neural network Wnet(w,b) is established to predict the strain energy function W(I1,I2), where the neural network input is the deformation invariants I1,I2 of the hyperelastic material, and the output is the strain energy function W pred (I1,I2), as attached Figure 1 shown.

[0051] W pred (I1,I2)=W net (I1,I2;w,b) (2)

[0052] Among them, W pred is the predicted strain energy function based on the two invariants I1 and I2, W net represents a fully connected neural network, w represents the weight of the fully connected neural network, and b represents the bias of the fully connected neural network;

[0053] Combining the automatic differentiation computing capability of neural networks, the following parameters of the constitutive model are calculated:

[0054]

[0055] Further, the constitutive model parameters are combined to calculate the different stretching ratios Nominal stress prediction value under combination

[0056]

[0057] The loss function of the physical information network is defined as the measurement point stretching ratio The true value of the nominal stress at and the predicted value P pred Mean square error Loss

[0058]

[0059] in is the true value of the nominal stress at the measuring point, P pred is the predicted value of the nominal stress at the measuring point, N P is the nominal stress data point The number of sampling points under

[0060] Step 3: Building a deep regression network to generate a hyperelastic constitutive model that can reflect the nonlinear behavior of the hyperelastic material plate;

[0061] Construct a deep regression network to parse the strain energy function expression W represented by the invariants I1 and I2 * (I1,I2), as attached Figure 1 For the deep regression network, the symbolic expression is viewed as a binary tree, where each node represents a symbol of the expression in the available symbol library, i.e., the input variables are the deformation invariants (I1, I2) of the hyperelastic plate, constants (e.g., C 10 ,C 20 ,C 30 ) or operations (e.g., +, -, ×, / , sin, log, exp, ...), in this representation, input variables and constants are called terminal nodes or symbols (without children), operations that accept a single argument (e.g., sin, log, ...) are unary symbols (with one child node), and operations that accept two arguments (e.g., +, -, ×, =, ...) are binary symbols. By first considering each node in depth and then from left to right, a one-dimensional list can be computed. In this representation, operators are placed in front of the corresponding operands in the expression, reducing the need for parentheses. Using prefix notation and treating symbols (called tokens) as categories allows any expression to be treated as a pure sequence of category vectors. The token sequence here is generated by using an RNN, and in order to greatly reduce the search space, the probability of certain tokens can be zeroed based on the context encoded in the generated expression tree. In addition, we adopt a priori that forces expressions to be less than 30 tokens long, contain no more than 2 levels of nested trigonometric operations (e.g., disallow cos(f.t+sin(x / x0+tan(□))), but still allow cos(f.t+sin(□))), and do not contain self-nesting of exponential and logarithmic operators (e.g., disallow e e□ ) and disallow useless unary inverse operations (e.g., disallow e log(□) ). Finally, the deep regression network outputs the predicted strain energy function W * An explicit expression for (I1,I2).

[0062] For each candidate W * (I1, I2), the fitting quality reward R = 1 / (1 + NRMSE) is calculated in the deep regression network, where NRMSE is calculated by the target σ WBias-normalized root mean square error:

[0063]

[0064] where σ W is the variance of the strain energy data set calculated by the target strain energy function, and W(I1,I2) is the physical information network W net The predicted strain energy function, W * (I1,I2) is the strain energy function predicted by the deep regression network.

[0065] Step 4: Neural network training to find the optimal network parameters;

[0066] Perform network training based on the experimental data set of hyperelastic materials obtained in step 1, and jointly train the physical information network in step 2 and the deep regression network in step 3;

[0067] When the loss function Loss of the physical information network and the reward R of the fitting quality of the deep regression network meet the conditions at the same time, stop the network training;

[0068] Different thresholds are set for the physical information network and the deep regression network. That is, when the loss function Loss of the physical information network is less than a certain threshold (such as 10 -4 ), and the fitting quality reward R exceeds a certain threshold (such as 99%) at the same time, stop network training. Considering the differences in training efficiency and convergence effect between deep regression and physical information networks, an iterative alternating training scheme can be adopted to first reduce the loss function Loss of the physical information network, and then train the fitting quality reward R of the deep regression network, and realize network joint training through an iterative algorithm.

[0069] Through the above steps, the final output of the deep regression network is the strain energy function W of the hyperelastic constitutive model obtained by this constitutive modeling method. * The expression of (I1,I2) generates a hyperelastic constitutive model that can reflect the nonlinear behavior of the hyperelastic material plate. And the strain energy function W of the constitutive model is * The model parameters in (I1, I2) are determined simultaneously by a deep regression network. The obtained hyperelastic constitutive model can be verified by comparing with the experimental data set.

[0070] In summary, the method proposed in the present invention realizes the experimental data set obtained through mechanical testing experiments, combines deep regression with the training of physical information networks, directly obtains the explicit expression of the strain energy function in the hyperelastic constitutive model, and simultaneously determines the model parameters in the constitutive model.

[0071] Step 1 is used to generate input data for the physical information network in step 2 and the deep regression network in step 3; the physical information network in step 2 obtains the mapping relationship between I1, I2 and the strain energy function W(I1, I2), which is used for the optimization training of the deep regression network in step 3; the deep regression network in step 3 finally generates a hyperelastic constitutive model that can reflect the nonlinear behavior of the hyperelastic material plate;

[0072] Example:

[0073] The hyperelastic constitutive model scheme of the present invention is applied to the hyperelastic constitutive and deformation analysis of a certain material.

[0074] A compression test was conducted on a hyperelastic material of a human brain tissue to obtain the tensile strength of the and nominal stress The experimental data set of hyperelastic materials composed of Figure 2 As shown in the point (a).

[0075] Through the hyperelastic constitutive model scheme of the present invention, combined with deep regression and physical information network training, the explicit expression of the strain energy function in the hyperelastic constitutive model is directly obtained as follows:

[0076]

[0077] At the same time, the model parameters in the constitutive model are determined as a = 0.3815; b = 1.31; c = 0.2464;

[0078] Based on the training of the network model, the hyperelastic constitutive model can be further calculated to obtain the nominal stress. The results are shown in the attached figure. Figure 2 (a) is shown. Compared with the experimental data set, as shown in the attached Figure 2 As shown, it can be seen that the constitutive model predicts the experimental data obtained from mechanical tests very well, which verifies the obtained hyperelastic constitutive model.

[0079] The hyperelastic constitutive model obtained based on network model training can be further combined with finite element analysis to calculate the stress and strain distribution of the material under different loads, as shown in the attached figure. Figure 2 As shown in (b), quantitative analysis of the hyperelastic deformation mechanics of the material under different loads is carried out.

[0080] The method proposed in the present invention directly obtains the explicit expression of the strain energy function in the hyperelastic constitutive model through the experimental data set obtained from the mechanical test, combined with deep regression and physical information network training, and simultaneously determines the model parameters in the constitutive model. In particular, the constitutive model is completely obtained through neural network training, and the constitutive model expression has more possibilities, making it capable of predicting some complex hyperelastic behaviors that existing constitutive models are unable to cope with.

Claims

1. A hyperelastic constitutive model method based on deep regression and physical information network, characterized by: The steps include: Step 1: Acquire the experimental data set to generate input data for the deep regression network in steps 2 and 3; Step 2: Building a physical information network; used to predict the mapping relationship between the invariants I1, I2 and the strain energy function W(I1, I2), and used for the optimization training of the deep regression network in step 3; Step 3: Building a deep regression network to generate a hyperelastic constitutive model that can reflect the nonlinear behavior of the hyperelastic material plate; Step 4: Perform neural network training based on the experimental data set of hyperelastic materials obtained in step 1, and jointly train the physical information network in step 2 and the deep regression network in step 3 to find the optimal network parameters; The step 1 is specifically as follows: Firstly, a mechanical test was carried out on an incompressible hyperelastic rectangular plate, and uniaxial tension and compression tests and equibiaxial tension and compression tests were performed to obtain the experimental data set. and nominal stress Experimental datasets of constructed hyperelastic materials; Based on hyperelastic theory, calculation of different stretching ratios The corresponding deformation invariant under the combined parameters Among them, I1 and I2 are two invariants of the Cauchy Green deformation tensor; The step 2 is specifically as follows: Building a physical information network W net , to predict the strain energy function W(I1,I2); W pred (I1,I2)=W net (I1,I2;w,b) (2) Among them, W pred is the predicted strain energy function based on the two invariants I1 and I2, W net represents a fully connected neural network, w represents the weight of the fully connected neural network, and b represents the bias of the fully connected neural network; Combining the automatic differentiation computing capability of neural networks, the following parameters of the constitutive model are calculated: Further, the constitutive model parameters are combined to calculate the different stretching ratios Nominal stress prediction value under combination The loss function of the physical information network is defined as the measurement point stretching ratio The true value of the nominal stress at and the predicted value P pred Mean square error Loss in is the true value of the nominal stress at the measuring point, P pred is the predicted value of the nominal stress at the measuring point, N P is the nominal stress data point The number of sampling points under ; The step three is specifically as follows: Construct a deep regression network to analyze the invariants I1, I2 and the strain energy function W * (I1, I2). For the deep regression network, the symbolic expression is regarded as a binary tree, where each node represents a symbol of the expression in the available symbol library, that is, the input variables are the deformation invariants (I1, I2) of the hyperelastic plate, constants or operations, and the final output of the deep regression network is the predicted strain energy function W * Explicit expression of (I1,I2); For each candidate W * (I1, I2), the fitting quality reward R = 1 / (1 + NRMSE) is calculated in the deep regression network, where NRMSE is calculated by the target σ W Bias-normalized root mean square error: where σ W The variance of the target strain energy function, W(I1,I2), is given by the physical information network W net The predicted strain energy function W(I1,I2), W * (I1, I2) is the strain energy function predicted by the deep regression network; The step 4 is specifically as follows: When the loss function Loss of the physical information network and the reward R of the fitting quality of the deep regression network meet the conditions at the same time, stop the network training; Different thresholds are set for the physical information network and the deep regression network. That is, when the loss function Loss of the physical information network is less than a certain threshold and the fitting quality reward R exceeds a certain threshold at the same time, the network training is stopped. Considering the differences in training efficiency and convergence effect between deep regression and physical information networks, an iterative alternating training scheme is adopted to first reduce the loss function Loss of the physical information network, and then train the fitting quality reward R of the deep regression network. The network joint training is achieved through an iterative algorithm.

2. The hyperelastic constitutive model method based on deep regression and physical information network according to claim 1 is characterized in that: The final output of the deep regression network is the strain energy function W of the hyperelastic constitutive model obtained by this constitutive modeling method. * (I1, I2), that is, a hyperelastic constitutive model that can reflect the nonlinear behavior of the hyperelastic material plate is generated, and the strain energy function W of the constitutive model is * The model parameters in (I1, I2) are determined simultaneously by a deep regression network and compared with the experimental data set to verify the obtained hyperelastic constitutive model.