Three-dimensional woven composite reverse design method based on generative adversarial neural network

Abstract

The application discloses a kind of three-dimensional woven composite reverse design methods based on generative adversarial neural network, comprising: selecting three-dimensional woven composite sample of different process structure by hyper-latin sampling;Python script simultaneously TexGen textile modeling software generates three-dimensional woven composite cell model and imports Abaqus to calculate material mechanics performance;The process structure of three-dimensional woven composite is normalized coding combined with material mechanics performance as data set, training generative adversarial neural network and passing through bayesian optimization hyperparameter.The application can improve the design and development efficiency of three-dimensional woven composite by more than 50% by mapping between three-dimensional woven composite process parameters and mechanics performance of the generative adversarial neural network model, improve the structure design accuracy and performance prediction accuracy of three-dimensional woven composite.

Description

Technical Field

[0001] This invention relates to the field of three-dimensional woven composite material design and manufacturing, and in particular to a reverse design method for three-dimensional woven composite materials based on generative adversarial neural networks. Background Technology

[0002] Textile structural composites possess advantages such as light weight, high strength, and corrosion resistance, and are widely used in aerospace, transportation, wind power generation, and military industries. Among them, three-dimensional woven composites exhibit better impact resistance and interlaminar properties compared to traditional two-dimensional woven composites. Furthermore, three-dimensional woven preforms offer greater design flexibility and near-net-shape capability, and have already been practically applied in key aerospace components such as aero-engine blades, engine throat liners, and space radomes.

[0003] While the high designability of 3D woven composites enhances their application potential, it also presents a key challenge: how to design the optimal structure based on the requirements of the target application. Traditional trial-and-error methods follow a structural design-material preparation-performance evaluation route to meet the target application requirements. This involves iterative improvements based on identified problems until a satisfactory composite material is designed. However, this method requires multiple experimental iterations to approximate the target performance, resulting in high cost and low efficiency. To address this issue, numerical simulation-based design and machine learning-based design methods have been applied to the design of 3D woven composites. Numerical simulation-based design involves establishing geometric and mechanical models of the 3D woven composite material and using numerical calculation methods such as the finite element method to simulate the material's mechanical response under different stress conditions, thereby predicting the material's performance and behavior. While numerical simulation can predict the composite material's performance before actual manufacturing, providing guidance for material design and optimization and reducing the number of experiments and R&D costs to some extent, it also requires simplifications and assumptions when building numerical models, leading to biases in the simulation results. Furthermore, it is computationally expensive and highly dependent on material parameters. Machine learning design methods establish a mapping relationship between input and output by learning and training on a large number of data samples, thereby enabling the prediction and classification of unknown data. This method can improve design efficiency and can also build predictive models based on existing data to predict the performance of new materials. However, the performance of machine learning is heavily dependent on the quality and quantity of training data, the interpretability of the model is poor, and there is also the problem of overfitting.

[0004] Although design methods based on numerical simulation and machine learning can reduce costs and improve design efficiency to some extent, they are essentially still based on repeated experiments to continuously approach the design conditions. Therefore, it is still impossible to avoid iterative design for different processes to select the optimal structure.

[0005] Currently, there is no design method for three-dimensional woven composite materials that can reduce design costs and improve design efficiency. Summary of the Invention

[0006] The purpose of this invention is to address the shortcomings of traditional forward design of three-dimensional woven composite materials, which relies on the designer's experience and cannot achieve the target performance design in one go, resulting in high cost and low efficiency. This invention uses a generative adversarial neural network to establish a mapping relationship between the three-dimensional woven process structure and mechanical properties, and generates the optimal process structure according to the target performance to achieve reverse customization of material properties and structure, thereby improving the design and development efficiency of three-dimensional woven composite materials.

[0007] To achieve the above objectives, this invention provides a reverse design method for three-dimensional woven composite materials based on generative adversarial neural networks, specifically including the following steps:

[0008] Step 1: Using the yarn type, yarn fineness, fabric warp and weft density, and yarn path of the three-dimensional woven structure as the design space, select three-dimensional woven composite materials with different process structures as samples through super Latin sampling;

[0009] Step 2: Using Python scripts in conjunction with TexGen textile modeling software, automatically generate three-dimensional woven composite material unit cell models with different structures based on the samples obtained in Step 1.

[0010] Step 3: Using Python scripts, import the unit cell models generated by TexGen into Abaqus to perform mechanical calculations on the three-dimensional woven composite material unit cell models with different structures, and obtain the material elastic modulus, ultimate strength and density corresponding to different structures;

[0011] Step 4: Normalize and encode the process structure of different yarn types, yarn fineness, fabric warp and weft density and yarn path. Use the material elastic modulus, ultimate strength and density corresponding to different structures as labels. Organize the data obtained after normalization and encoding into a three-dimensional structure-performance dataset of woven composite materials and divide it into training set and validation set, with the ratio of training set to validation set being 7:3.

[0012] Step 5: Construct a generative adversarial neural network model that includes a generator and a recognizer. The generator uses a fully connected neural network to generate a high-dimensional pseudo-3D woven structure from low-dimensional noisy data. The recognizer uses a convolutional neural network to identify whether the generated 3D woven structure meets the weaving conditions and target performance accuracy.

[0013] Step 6: Train the generative adversarial neural network model from Step 5 using the training set from Step 4;

[0014] Step 7: Validate the generative adversarial neural network model trained in Step 6 using the validation set from Step 4. If the generated yarn path has interlacing warp and weft yarns and the loss function of the model optimization is less than 0.1, then the generative adversarial neural network is used for the design of three-dimensional woven composite materials. If the requirements are not met, repeat Step 6 and retrain the generative adversarial neural network by combining Bayesian optimization of hyperparameters until the requirements are met.

[0015] Step 8: Input the target mechanical performance requirements into the trained generative adversarial neural network model to generate the corresponding three-dimensional woven composite material process structure parameters; perform numerical simulation on the three-dimensional woven composite material with these process structure parameters to ensure that the mechanical performance meets the target requirements;

[0016] Step 9: Design a three-dimensional woven composite material that meets the target requirements based on the process structure parameters obtained in Step 8.

[0017] Furthermore, the yarn types in step one include high-performance fibers such as carbon fiber, glass fiber, aramid fiber, and ultra-high molecular weight polyethylene fiber. The yarn fineness, fabric warp and weft density, and upper and lower limits of the warp path are determined based on actual conditions and different process combinations. Within the upper and lower limits of the process parameters, 50-400 uniformly distributed process structure samples are selected using ultra-Latin sampling.

[0018] Furthermore, the yarn fineness, fabric warp and weft density, and yarn path in step two are expressed as follows:

[0019]

[0020] f end (x)=(-o2) n ×(xl×d x ) 2

[0021] f inside (x)=(-o 33 ) n ×(xl×d x ) 3 -o 31 ×(xl×d x )

[0022] Where T represents yarn fineness, ρ is fiber bulk density, S is yarn cross-sectional area, a and b are the radii of the major and minor axes of the yarn cross-section, respectively, P represents the warp and weft density of the fabric, and L is the spacing between yarns; f end The Y-axis coordinate represents the interlacing area of ​​the warp yarns outside the weft yarns, x represents the X-axis coordinate of the warp yarns, o2 represents the coefficient of the quadratic term of the quadratic function, n is the number of weft yarn rows, and l×d xThis represents the offset between two warp yarns in the X direction, where l represents the structural variable in the X direction of the three-dimensional woven structure, and d... x f represents the offset of two adjacent weft yarns in the X direction; inside The Y-axis coordinate of the area where the warp yarns interweave within the weft yarns, o 33 The coefficient of the cubic term in a cubic function is denoted by o. 31 The coefficients of a cubic function can be calculated using the formula above, and the corresponding yarn geometry can be generated using TexGen to create the corresponding unit cell model.

[0023] Furthermore, in step three, the material's elastic modulus includes the tensile modulus in the X, Y, and Z directions and the shear modulus in the XY, XZ, and YZ directions, and the ultimate strength includes the tensile strength in the X, Y, and Z directions and the shear strength in the XY, XZ, and YZ directions.

[0024] Furthermore, in step three, the Abaqus mechanical calculation process includes assigning material properties, mesh generation, setting element properties, adding periodic boundaries, setting displacement / load boundaries, and using the three-dimensional Hashin failure criterion as the failure criterion for the fiber bundles, while the Von-Mises failure criterion is used for the matrix. Stiffness reduction is used as the degradation scheme for damaged elements. The expressions for the failure criteria of the fiber bundles and the matrix are as follows:

[0025] Fiber tensile failure (σ1>0):

[0026]

[0027] Fiber compression (σ1<0):

[0028]

[0029] Transverse tensile failure of the matrix within the fiber bundle (σ2>0):

[0030]

[0031] Transverse compressive failure of the matrix within the fiber bundle (σ2<0):

[0032]

[0033] Fiber bundle matrix normal tensile failure (σ3>0):

[0034]

[0035] Normal compressive failure of the matrix within the fiber bundle (σ3<0):

[0036]

[0037] Fiber matrix shear failure (σ 12 / 13<0):

[0038]

[0039] Matrix failure:

[0040]

[0041] Where, σ i τ is the stress in the principal directions (X, Y, Z directions) of the material. ij The shear stress of the material is given by S, where X, Y, and Z are the principal directions (X, Y, Z directions) of the fiber bundle, respectively. The subscripts T and C represent tension and compression, respectively. ij G represents the shear strength of the fiber bundle. ij Let α be the initial shear modulus of the fiber bundle, i,j∈[1,3]. When only subscript i is used, 1, 2, and 3 represent the X, Y, and Z directions of the material, respectively. When the subscript ij is used (i≠j), 12, 13, and 23 represent the XY, XZ, and YZ plane directions of the material, respectively. α is the longitudinal shear nonlinearity coefficient of the fiber bundle, and σ is the initial shear modulus of the fiber bundle. m It represents the ultimate strength of the matrix.

[0042] Furthermore, in step four, the normalization coding of yarn type, yarn fineness, fabric warp density and fabric weft density is achieved by normalizing the selected design range at equal intervals. That is, when the total design amount of any process structure parameter is u, the vth process parameter is normalized to v / u.

[0043] The normalized encoding of yarn paths is based on the cross-section along the warp direction within a representative volume unit of the three-dimensional woven structure. It forms x [m×n] matrices based on the number of warp yarns along different paths in the Y and Z directions within the representative volume unit, and the number of weft yarns along the X and Z directions within the representative volume unit. Here, x represents the total number of warp yarns along different paths within the representative volume unit, m represents the number of weft yarns in the X direction within the representative volume unit, and n represents the number of weft yarns in the Z direction within the representative volume unit. The interlacing points of the warp and weft yarns are used as elements within the matrices; elements where the warp yarn is above the weft yarn at an interlacing point are defined as 0, and elements where the warp yarn is below the weft yarn at an interlacing point are defined as 1.

[0044] Furthermore, the structure-performance dataset not only includes data obtained through the numerical simulation methods described in steps two to three, but can also be obtained through experimental preparation and testing to obtain the corresponding structural mechanical properties.

[0045] Furthermore, in step five, the generative adversarial neural network model includes a generator and a recognizer. The generator uses a fully connected neural network to generate a high-dimensional pseudo-3D woven structure from low-dimensional noisy data, and the recognizer uses a convolutional neural network to identify whether the generated 3D woven structure meets the weaving conditions and target performance accuracy.

[0046] The fully connected neural network includes an input layer, a ReLU activation function layer, a Sigmoid activation function layer, a hidden layer, a fully connected layer, and an output layer.

[0047] A convolutional neural network consists of an input layer, a convolutional layer, a pooling layer, a ReLU activation function layer, a fully connected layer, and an output layer.

[0048] Furthermore, the generative adversarial neural network trained in step six realizes the structure-property mapping expression for the three-dimensional woven composite material as follows:

[0049] (σ ij E ij ,Φ)=f(T,P,Len)

[0050] Where, σ ij E represents the strength of composite materials. ij The modulus of the composite material is represented by i,j∈[1,3]. When i=j, 11, 22, and 33 represent the tensile properties of the material in the X, Y, and Z directions, respectively. When i≠j, 12, 13, and 23 represent the shear properties of the material in the XY, XZ, and YZ plane directions, respectively. Φ represents the bulk density of the composite material, and Len is the yarn path.

[0051] The loss function V(G,D) for model optimization is:

[0052]

[0053] Where G represents the generator, D represents the recognizer, F represents the expected value of the distribution function, u represents the real sample, v represents the noise sample, and p data (u) represents the distribution of the real sample, p noise G(v) is a noise distribution defined in a low dimension; G(v) represents the output sample generated by the generator G after receiving a random noise sample v; D(u) represents the probability that the recognizer identifies u as a real sample; D(G(v)) represents the probability that the recognizer identifies the generated sample G(v) as a real sample; and 1-D(G(v)) represents the probability that the recognizer identifies the generated sample G(v) as a fake sample.

[0054] Furthermore, the hyperparameters in step seven include learning rate, minimum batch size, number of training epochs, number of neural network units, number of neural network layers, and optimizer parameters.

[0055] The beneficial effects of this invention are as follows:

[0056] This invention provides a reverse design method for three-dimensional woven composite materials based on generative adversarial neural networks (GANs). By establishing a GAN model, the method establishes the relationship between process parameters such as yarn type, yarn fineness, fabric warp and weft density, and yarn path, and the mechanical properties of the three-dimensional woven composite material. The GAN model directly outputs the optimal three-dimensional woven structure that meets the target performance requirements, improving the design and development efficiency of three-dimensional woven composite materials by more than 50%. Furthermore, it overcomes the problem of insufficient generalization of traditional mathematical-physical models, achieving performance-structure customization of three-dimensional woven structures and improving the design accuracy of structural performance. It also provides more design space for the development of new three-dimensional woven structure products. Attached Figure Description

[0057] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0058] Figure 1 This is a flowchart of the reverse design of three-dimensional woven composite materials in the reverse design method of three-dimensional woven composite materials based on generative adversarial neural networks provided by the present invention;

[0059] Figure 2 This is a schematic diagram of the coding method for yarn type, yarn fineness, warp density and weft density in the reverse design method for three-dimensional woven composite materials based on generative adversarial neural networks provided by the present invention.

[0060] Figure 3 This is a schematic diagram of the encoding method of the yarn path in the three-dimensional shallow cross-bend structure in the reverse design method of three-dimensional woven composite materials based on generative adversarial neural network provided by the present invention;

[0061] Figure 4 This is a schematic diagram of the encoding method of the yarn path of the three-dimensional orthogonal structure in the reverse design method of three-dimensional woven composite materials based on generative adversarial neural network provided by the present invention;

[0062] Figure 5 This is a flowchart illustrating the training and application of the generative adversarial neural network model in the reverse design method for three-dimensional woven composite materials based on generative adversarial neural networks provided by this invention.

[0063] Figure 6 This is a schematic diagram of a three-dimensional woven structure with different yarn paths in the reverse design method for three-dimensional woven composite materials based on generative adversarial neural networks provided by the present invention. Detailed Implementation

[0064] To more clearly illustrate the objectives, technical solutions, and advantages of this invention, a detailed description will be provided below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the invention. Furthermore, the technical features in the following embodiments can be combined with each other as long as they do not conflict with each other.

[0065] Example 1

[0066] This embodiment provides a reverse design method for three-dimensional woven composite materials based on generative adversarial neural networks. The specific process of this method is as follows: Figure 1 As shown, firstly, samples of different process structure parameters of three-dimensional woven composite materials are selected based on super Latin sampling. Secondly, unit cell models of the corresponding process structure parameters are generated by combining Python with TexGen. Thirdly, the mechanical properties of three-dimensional woven composite materials with different process structure parameters are calculated based on Abaqus. Then, the yarn type, yarn fineness, fabric warp and weft density and warp path are normalized and encoded to train the generative adversarial neural network. Finally, the desired target mechanical properties are input into the trained adversarial neural network to generate the optimal three-dimensional woven process structure parameters that meet the target mechanical properties.

[0067] The method includes:

[0068] Step 1: Using the yarn type, yarn fineness, fabric warp and weft density, and yarn path of the three-dimensional woven structure as the design space, select three-dimensional woven composite materials with different process structures as samples through super Latin sampling;

[0069] Step 2: Using Python scripts in conjunction with TexGen textile modeling software, automatically generate three-dimensional woven composite material unit cell models with different structures based on the samples obtained in Step 1.

[0070] Step 3: Using Python scripts, import the unit cell models generated by TexGen into Abaqus to perform mechanical calculations on the three-dimensional woven composite material unit cell models with different structures, and obtain the material elastic modulus, ultimate strength and density corresponding to different structures;

[0071] Step 4: Normalize and encode the process structure of different yarn types, yarn fineness, fabric warp and weft density and yarn path. Use the material elastic modulus, ultimate strength and density corresponding to different structures as labels to organize into a three-dimensional structure-performance dataset of woven composite materials and divide it into training set and validation set, with the ratio of training set to validation set being 7:3.

[0072] Step 5: Construct a generative adversarial neural network model that includes a generator and a recognizer. The generator uses a fully connected neural network to generate a high-dimensional pseudo-3D woven structure from low-dimensional noisy data. The recognizer uses a convolutional neural network to identify whether the generated 3D woven structure meets the weaving conditions and target performance accuracy.

[0073] Step 6: Train the generative adversarial neural network model from Step 5 using the training set from Step 4;

[0074] Step 7: Validate the generative adversarial neural network model trained in Step 6 using the validation set from Step 4. If the generated yarn path has interlacing warp and weft yarns and the loss function of the model optimization is less than 0.1, then the generative adversarial neural network is used for the design of three-dimensional woven composite materials. If the requirements are not met, repeat Step 6 and retrain the generative adversarial neural network by combining Bayesian optimization of hyperparameters until the requirements are met.

[0075] Step 8: Input the target mechanical performance requirements into the trained generative adversarial neural network model to generate the corresponding three-dimensional woven composite material process structure parameters; perform numerical simulation on the three-dimensional woven composite material with these process structure parameters to ensure that the mechanical performance meets the target requirements;

[0076] Step 9: Design a three-dimensional woven composite material that meets the target requirements based on the process structure parameters obtained in Step 8.

[0077] The yarn types used in Step 1 include high-performance fibers such as carbon fiber, glass fiber, aramid fiber, and ultra-high molecular weight polyethylene fiber. The yarn fineness, fabric warp and weft density, and upper and lower limits of the warp path are determined based on actual conditions and different process combinations. Within the upper and lower limits of the process parameters, 50-400 uniformly distributed process structure samples are selected using ultra-Latin sampling.

[0078] The yarn fineness, fabric warp and weft density, and yarn path in step two are expressed as follows:

[0079]

[0080] f end (x)=(-o2) n ×(xl×d x ) 2

[0081] f inside (x)=(-o 33 ) n ×(xl×d x ) 3 -o 31 ×(xl×d x )

[0082] Where T represents yarn fineness, ρ is fiber bulk density, S is yarn cross-sectional area, a and b are the radii of the major and minor axes of the yarn cross-section, respectively, P represents the warp and weft density of the fabric, and L is the spacing between yarns; f end The Y-axis coordinate represents the interlacing area of ​​the warp yarns outside the weft yarns, x represents the X-axis coordinate of the warp yarns, o2 represents the coefficient of the quadratic term of the quadratic function, n is the number of weft yarn rows, and l×d x This represents the offset between two warp yarns in the X direction, where l represents the structural variable in the X direction of the three-dimensional woven structure, and d... x f represents the offset of two adjacent weft yarns in the X direction; inside The Y-axis coordinate of the area where the warp yarns interweave within the weft yarns, o 33 The coefficient of the cubic term in a cubic function is denoted by o. 31 The coefficients of a cubic function can be calculated using the formula above, and the corresponding yarn geometry can be generated using TexGen to create the corresponding unit cell model.

[0083] In step three, the elastic modulus of the material includes the tensile modulus in the X, Y, and Z directions and the shear modulus in the XY, XZ, and YZ directions, and the ultimate strength includes the tensile strength in the X, Y, and Z directions and the shear strength in the XY, XZ, and YZ directions.

[0084] Step three of the mechanical calculation process in Abaqus includes: importing the unit cell model generated by TexGen into Abaqus using a Python script, applying corresponding material properties to the carbon fibers and carbon matrix, meshing with C3D8 element properties, and then applying periodic boundaries using EasyPBC. Load boundaries are applied along the X, Y, Z, XY, XZ, and YZ directions of the material before submitting the calculation to obtain the tensile modulus and tensile strength in the X, Y, and Z directions, and the shear modulus and shear strength in the XY, XZ, and YZ directions of the three-dimensional woven composite materials with different structures. The three-dimensional Hashin failure criterion is used as the failure criterion for the fiber bundles, the Von-Mises failure criterion for the matrix, and stiffness reduction as the degradation scheme for damaged elements. The expressions for the fiber bundle and matrix failure criteria are:

[0085] Fiber tensile failure (σ1>0):

[0086]

[0087] Fiber compression (σ1<0):

[0088]

[0089] Transverse tensile failure of the matrix within the fiber bundle (σ2>0):

[0090]

[0091] Transverse compressive failure of the matrix within the fiber bundle (σ2<0):

[0092]

[0093] Fiber bundle matrix normal tensile failure (σ3>0):

[0094]

[0095] Normal compressive failure of the matrix within the fiber bundle (σ3<0):

[0096]

[0097] Fiber matrix shear failure (σ 12 / 13 <0):

[0098]

[0099] Matrix failure:

[0100]

[0101] Where, σ i τ is the stress in the principal directions (X, Y, Z directions) of the material. ij The shear stress of the material is given by S, where X, Y, and Z are the principal directions (X, Y, Z directions) of the fiber bundle, respectively. The subscripts T and C represent tension and compression, respectively. ij G represents the shear strength of the fiber bundle. ij Let α be the initial shear modulus of the fiber bundle, i,j∈[1,3]. When only subscript i is used, 1, 2, and 3 represent the X, Y, and Z directions of the material, respectively. When the subscript ij is used (i≠j), 12, 13, and 23 represent the XY, XZ, and YZ plane directions of the material, respectively. α is the longitudinal shear nonlinearity coefficient of the fiber bundle, and σ is the initial shear modulus of the fiber bundle. m It represents the ultimate strength of the matrix.

[0102] The normalized codes for yarn type, yarn fineness, warp density, and weft density in step four are as follows: Figure 2 As shown, different process structures in the three-dimensional weaving structure are set as columns of a matrix, and different parameters in the process structure are set as rows of the matrix; different columns of the matrix correspond to four process structures, namely yarn fineness, fabric warp density, fabric weft density, and yarn type, and different rows of the matrix correspond to three different parameters in the process structure; the yarn type is directly normalized to 0, 0.5, and 1 at equal intervals.

[0103] The normalized codes for the yarn paths in a three-dimensional shallow bend structure are as follows: Figure 3As shown, the representative volume unit contains two cross-sections with two different warp paths, totaling eight warp yarns. Each warp path includes four warp yarns. Therefore, the two warp cross-sections are represented by eight matrices: {a1,b1,c1,d1; a2,b2,c2,d2}. The numbers a, b, c, and d represent warp yarns distributed differently in the Z direction on the same cross-section, while {a1,b1,c1,d1} and {a2,b2,c2,d2} represent the distribution of different warp yarns in the Y direction cross-section, respectively. The representative volume unit contains four weft yarns distributed along the X direction and five weft yarns distributed along the Z direction, thus forming a total of eight [4×5] matrices to represent the yarn paths of the three-dimensional shallow cross-bending structure.

[0104] The normalized encoding of the yarn path in the three-dimensional orthogonal structure is as follows: Figure 4 As shown, the representative volume unit has two different warp paths in its cross section. There is one warp yarn along the Z direction, so the two warp cross sections are represented by two matrices {a1; a2}. There are 4 weft yarns distributed along the X direction and 5 weft yarns distributed along the Z direction in the representative volume unit, so a total of 2 [4×5] matrices are formed to represent the yarn paths of the three-dimensional orthogonal structure.

[0105] The normalized coding of yarn type, yarn fineness, fabric warp density and fabric weft density is achieved by normalizing the selected design range at equal intervals. That is, when the total design amount of any process structure parameter is u, the vth process parameter is normalized to v / u.

[0106] The normalized encoding of yarn paths is based on the cross-section along the warp direction within a representative volume unit of the three-dimensional woven structure. It forms x [m×n] matrices based on the number of warp yarns along different paths in the Z direction within the representative volume unit, and the number of weft yarns along the X and Z directions within the representative volume unit. Here, x is the total number of warp yarns along different paths within the representative volume unit, m is the number of weft yarns in the X direction within the representative volume unit, and n is the number of weft yarns in the Z direction within the representative volume unit. The interlacing points of the warp and weft yarns are used as elements in the matrices; elements where the warp yarn is above the weft yarn at an interlacing point are defined as 0, and elements where the warp yarn is below the weft yarn at an interlacing point are defined as 1.

[0107] The structure-performance dataset not only includes data obtained through numerical simulation methods described in steps two to three, but also data obtained through experimental preparation and testing of the corresponding structural mechanical properties.

[0108] In step five, the generative adversarial neural network model includes a generator and a recognizer. The generator uses a fully connected neural network to generate a high-dimensional pseudo-3D woven structure from low-dimensional noisy data. The recognizer uses a convolutional neural network to identify whether the generated 3D woven structure meets the weaving conditions and target performance accuracy.

[0109] The fully connected neural network includes an input layer, a ReLU activation function layer, a Sigmoid activation function layer, a hidden layer, a fully connected layer, and an output layer.

[0110] A convolutional neural network consists of an input layer, a convolutional layer, a pooling layer, a ReLU activation function layer, a fully connected layer, and an output layer.

[0111] The training and application process of a generative adversarial neural network model is as follows: Figure 5 As shown, firstly, a "structure-performance" database of three-dimensional woven composite materials was collected through experiments and simulations. The dataset used the aforementioned yarn fineness, fabric warp and weft density, yarn type, and yarn path encoding, with the corresponding structural mechanical properties serving as the real data labels. Secondly, the data was divided into a training set and a validation set in a 7:3 ratio. The training set was used to train an adversarial generative neural network, while the validation set was used to verify the model's reliability. The model was considered successfully trained when warp and weft yarns intertwined in the generated yarn paths and the model's optimization loss function was less than 0.1. Then, the target mechanical performance requirements were input into the trained reverse engineering model, which output the optimal process structure parameters that met these requirements. Finally, a corresponding unit cell model was established based on the process structure parameters output by the reverse engineering model for numerical simulation verification, further validating the mechanical properties of the generated three-dimensional woven structure and thus achieving reverse customization of three-dimensional woven composite materials with the target mechanical properties.

[0112] The generative adversarial neural network model trained in step six realizes the structure-property mapping expression for the three-dimensional woven composite material as follows:

[0113] (σ ij E ij ,Φ)=f(T,P,Len)

[0114] Where, σ ij E represents the strength of composite materials. ij The modulus of the composite material is represented by i,j∈[1,3]. When i=j, 11, 22, and 33 represent the tensile properties of the material in the X, Y, and Z directions, respectively. When i≠j, 12, 13, and 23 represent the shear properties of the material in the XY, XZ, and YZ plane directions, respectively. Φ represents the bulk density of the composite material, and Len is the yarn path.

[0115] The loss function V(G,D) for model optimization is:

[0116]

[0117] Where G represents the generator, D represents the recognizer, F represents the expected value of the distribution function, u represents the real sample, v represents the noise sample, and p data(u) represents the distribution of the real sample, p noise G(v) is a noise distribution defined in a low dimension; G(v) represents the output sample generated by the generator G after receiving a random noise sample v; D(u) represents the probability that the recognizer identifies u as a real sample; D(G(v)) represents the probability that the recognizer identifies the generated sample G(v) as a real sample; and 1-D(G(v)) represents the probability that the recognizer identifies the generated sample G(v) as a fake sample.

[0118] The hyperparameters in step seven include learning rate, minimum batch size, number of training epochs, number of neural network units, number of neural network layers, and optimizer parameters.

[0119] Example 2

[0120] This embodiment provides a reverse design method for three-dimensional woven composite materials based on generative adversarial neural networks. Taking the design of a three-dimensional shallow cross-linked woven composite material without weft insertion as an example, the method includes:

[0121] Step 1: Select carbon fiber as the yarn type and carbon as the matrix. The carbon fiber types are 3K, 6K, 12K, and 24K, corresponding to yarn fineness of 200Tex, 400Tex, 800Tex, and 1600Tex respectively. The fabric warp and weft density is 2-5 threads / cm. This serves as the design space for training a generative adversarial neural network model. Specifically, only carbon fiber is set to 1 among the yarn types; yarn fineness is represented by [0, 0.33, 0.67, 1] for 200Tex, 600Tex, 1200Tex, and 2400Tex respectively; and fabric warp and weft density is represented by [0, 0.33, 0.67, 1] for 2 threads / cm, 3 threads / cm, 4 threads / cm, and 5 threads / cm respectively.

[0122] Step 2: Determine the target mechanical performance requirements based on application needs, and input the target mechanical performance into the generative adversarial neural network model trained in Step 1. Specifically, the tensile modulus and strength in the X, Y, and Z directions are 60 GPa, 75 GPa, 20 GPa and 80 MPa, 110 MPa, 60 MPa, respectively; the shear modulus and strength in the XY, XZ, and YZ directions are 12 GPa, 5 GPa, 8 GPa and 60 MPa, 20 MPa, 30 MPa, respectively. The target performance is input into the trained generative adversarial neural network model using a matrix representation. The first column and different rows represent the tensile modulus in the X, Y, and Z directions, the second column and different rows represent the tensile strength in the X, Y, and Z directions, the third column and different rows represent the shear modulus in the XY, XZ, and YZ directions, respectively, and the fourth column and different rows represent the shear strength in the XY, XZ, and YZ directions, respectively. The expression is as follows:

[0123]

[0124] Step 3: The trained generative adversarial neural network outputs the encoded process structure parameters that satisfy the target mechanical properties, as follows:

[0125]

[0126]

[0127] In the above matrices, the matrix with subscript 's' represents the generated yarn fineness, warp density, weft density, and yarn type encoding. The matrix with subscript 'a1' represents the path encoding of the first warp yarn distributed along the Z-direction in the first warp section, and the matrix with subscript 'a2' represents the path encoding of the first warp yarn distributed along the Z-direction in the second warp section. Similarly, matrices b1, c1, d1, and e1 represent the path encodings of the second, third, fourth, and fifth warp yarns distributed along the Z-direction in the first warp section, respectively, and matrices b2, c2, d2, and e2 represent the path encodings of the second, third, fourth, and fifth warp yarns distributed along the Z-direction in the second warp section, respectively. The output 3D woven structure is as follows: Figure 6 As shown in (a), the corresponding yarn fineness is 1200Tex, the warp density of the fabric is 4 threads / cm, the weft density of the fabric is 3 threads / cm, and the yarn type is carbon fiber.

[0128] Step 4: Based on the output three-dimensional woven structure parameters, prepare three-dimensional woven composite materials with corresponding structural parameters, and conduct corresponding tensile tests in the X and Y directions and shear tests in the XY, XZ, and YZ directions. The calculated results are shown in Table 1.

[0129] Table 1 Mechanical property parameters of three-dimensional woven composite materials with carbon matrix

[0130] X 58.0 77.9 Y 76.0 120.0 XY 13.0 61.0 XZ 4.7 21.3 YZ 8.6 27.5

[0131] The experimentally tested mechanical properties showed an error of less than 10% between the mechanical properties and the target mechanical properties, demonstrating the accuracy and efficiency of the reverse design method based on generative adversarial neural networks.

[0132] Example 3

[0133] This embodiment provides a reverse design method for three-dimensional woven composite materials based on generative adversarial neural networks. Taking the design of a three-dimensional shallow cross-linked woven composite material with weft insertion as an example, the method includes:

[0134] Step 1: Select aramid fiber, glass fiber, and ultra-high molecular weight polyethylene fiber as the yarn types, epoxy resin as the matrix, yarn fineness of 200-1200 Tex, and fabric warp and weft density of 2-12 threads / cm as the design space for training a generative adversarial neural network model. Here, yarn types are represented by [0,0,5,1] to represent aramid fiber, glass fiber, and ultra-high molecular weight polyethylene fiber, respectively; yarn fineness is represented by [0,0.1,0.2,0.3,0.4,0.5,0.6,0.7,0.8,0.9,1] to represent 200 Tex-1200 Tex, with each 0.1 increment representing 100 Tex; fabric warp and weft density is represented by [0,0.1,0.2,0.3,0.4,0.5,0.6,0.7,0.8,0.9,1] to represent 2 threads / cm-12 threads / cm, respectively.

[0135] Step 2: Determine the target mechanical performance requirements based on application needs, and input the target mechanical performance into the generative adversarial neural network model trained in Step 1. Specifically, the tensile modulus and strength in the X, Y, and Z directions are 54 GPa, 60 GPa, 25 GPa and 620 MPa, 860 MPa, 240 MPa, respectively; the shear modulus and strength in the XY, XZ, and YZ directions are 15 GPa, 8 GPa, 10 GPa and 120 MPa, 60 MPa, 80 MPa, respectively. The target performance is input into the trained generative adversarial neural network model using a matrix representation. The first column and different rows represent the tensile modulus in the X, Y, and Z directions, the second column and different rows represent the tensile strength in the X, Y, and Z directions, the third column and different rows represent the shear modulus in the XY, XZ, and YZ directions, and the fourth column and different rows represent the shear strength in the XY, XZ, and YZ directions, respectively. The specific representation is as follows:

[0136]

[0137] Step 3: The trained generative adversarial neural network outputs the encoded process structure parameters that satisfy the target mechanical properties, as follows:

[0138]

[0139]

[0140] In the above matrices, the matrix with subscript 's' represents the generated yarn fineness, warp density, weft density, and yarn type codes. The matrices with subscripts a1, b1, c1, and d1 represent the path codes for the first, second, third, and fourth warp yarns distributed along the Z-direction in the first warp section, respectively. The matrices with subscripts a2, b2, c2, d2, and e2 represent the path codes for the first, second, third, and fourth warp yarns distributed along the Z-direction in the second warp section, respectively. The output 3D woven structure is as follows: Figure 6As shown in (b), the corresponding yarn fineness is 800Tex, the fabric warp density is 8 threads / cm, the fabric weft density is 9 threads / cm, and the yarn type is glass fiber.

[0141] Step 4: Based on the output three-dimensional woven structure parameters, prepare three-dimensional woven composite materials with corresponding structural parameters, and conduct corresponding tensile tests in the X and Y directions and shear tests in the XY, XZ, and YZ directions. The calculated results are shown in Table 2.

[0142] Table 2 Mechanical property parameters of three-dimensional woven composite materials with epoxy resin as the matrix.

[0143] X 52.8 608.7 Y 60.7 877.5 XY 15.2 126.3 XZ 7.5 57.6 YZ 10.3 74.5

[0144] The experimentally tested mechanical properties showed an error of less than 10% between the mechanical properties and the target mechanical properties, demonstrating the accuracy and efficiency of the reverse design method based on generative adversarial neural networks.

[0145] Some steps in the embodiments of the present invention can be implemented using software, and the corresponding software program can be stored in a readable storage medium, such as an optical disc or a hard disk.

[0146] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A three-dimensional woven composite material reverse design method based on a generative adversarial neural network, characterized in that, The method includes: Step 1: Using the yarn type, yarn fineness, fabric warp and weft density, and yarn path of the three-dimensional woven structure as the design space, select three-dimensional woven composite materials with different process structures as samples through super Latin sampling; Step 2: Using Python scripts in conjunction with TexGen textile modeling software, automatically generate three-dimensional woven composite material unit cell models with different structures based on the samples obtained in Step 1; Step 3: Using a Python script, import the unit cell model generated by TexGen into Abaqus to perform mechanical calculations on the three-dimensional woven composite material unit cell models with different structures, and obtain the mechanical performance parameters corresponding to different structures; Step 4: Normalize and encode the process structure of different yarn types, yarn fineness, fabric warp and weft density and yarn path. Use the mechanical performance parameters corresponding to different structures as labels. Organize the data obtained after normalization and encoding into a three-dimensional woven composite material structure-performance dataset and divide it into training set and validation set. Step 5: Construct a generative adversarial neural network model that includes a generator and a recognizer; Step Six: Train the generative adversarial neural network model from Step Five using the training set from Step Four; Step 7: Validate the trained generative adversarial neural network model using the validation set from Step 4 to obtain the optimal generative adversarial neural network model. Step 8: Input the target mechanical performance requirements into the generative adversarial neural network model obtained in Step 7 to generate the corresponding three-dimensional woven composite material process structure parameters; perform numerical simulation on the three-dimensional woven composite material with these process structure parameters to ensure that the mechanical performance meets the target requirements; Step 9: Design a three-dimensional woven composite material that meets the target requirements based on the process structure parameters obtained in Step 8.

2. The method of claim 1, wherein, In step four, the normalization encoding of the yarn path is characterized by the cross-section in the warp direction within the representative volume unit of the three-dimensional woven structure. Based on the number of warp yarns along different paths in the Y and Z directions within the representative volume unit, and the number of weft yarns along the X and Z directions within the representative volume unit, x [m×n] matrices are formed. Where x is the total number of warp yarns in different paths within the representative volume unit, m is the number of weft yarns in the X direction within the representative volume unit, and n is the number of weft yarns in the Z direction within the representative volume unit; the interlacing point characteristics of warp and weft yarns are used as elements in the matrix, and elements where the warp yarn is above the weft yarn in the interlacing point are defined as 0, and elements where the warp yarn is below the weft yarn in the interlacing point are defined as 1. In step four, the normalization coding of yarn type, yarn fineness, and fabric warp and weft density is achieved by normalizing the selected design range at equal intervals. That is, when the total design amount of any process structure parameter is u, the vth process parameter is normalized to v / u.

3. The method of claim 2, wherein, The expressions for yarn fineness, fabric warp and weft density, and yarn path in step one are as follows: f end (x) = (-o2) n ×(x-l×d x ) 2 f inside (x) = (-o 33 ) n ×(x-l×d x ) 3 -o 31 ×(x-l×d x ) Where T represents yarn fineness, ρ is fiber bulk density, S is yarn cross-sectional area, a and b are the radii of the major and minor axes of the yarn cross-section, respectively, P represents the warp and weft density of the fabric, and L is the spacing between yarns; f end The Y-axis coordinate represents the interlacing area of ​​the warp yarns outside the weft yarns, x represents the X-axis coordinate of the warp yarns, o2 represents the coefficient of the quadratic term of the quadratic function, n is the number of weft yarn rows, and l×d x This represents the offset between two warp yarns in the X direction, where l represents the structural variable in the X direction of the three-dimensional woven structure, and d... x f represents the offset of two adjacent weft yarns in the X direction; inside The Y-axis coordinate of the area where the warp yarns interweave within the weft yarns, o 33 The coefficient of the cubic term in a cubic function is denoted by o. 31 This represents the coefficient of a cubic function.

4. The method of claim 3, wherein, The mechanical calculation process in Abaqus in step three includes assigning material properties, meshing, setting element properties, adding periodic boundaries, setting displacement / load boundaries, and using the three-dimensional Hashin failure criterion as the failure criterion for fiber bundles, Von-Mises failure criterion for the matrix, and stiffness reduction as the degradation scheme for damaged elements. The mechanical performance parameters include the material's elastic modulus, ultimate strength, and density.

5. The method of claim 4, wherein, In step five, the generator uses a fully connected neural network to generate a high-dimensional pseudo-three-dimensional woven structure from low-dimensional noisy data, and the recognizer uses a convolutional neural network to identify whether the generated three-dimensional woven structure meets the weaving conditions and target performance accuracy. The fully connected neural network includes an input layer, a ReLU activation function layer, a Sigmoid activation function layer, a hidden layer, a fully connected layer, and an output layer. The convolutional neural network includes an input layer, a convolutional layer, a pooling layer, a ReLU activation function layer, a fully connected layer, and an output layer.

6. The method of claim 5, wherein, The output expression of the generative adversarial neural network model obtained in step seven is: (σ ij ,E ij ,Φ)=f(T,P,Len) Where, σ ij E represents the strength of composite materials. ij The modulus of the composite material is represented by i,j∈[1,3]. When i=j, 11, 22, and 33 represent the tensile properties of the material in the X, Y, and Z directions, respectively. When i≠j, 12, 13, and 23 represent the shear properties of the material in the XY, XZ, and YZ plane directions, respectively. Φ represents the bulk density of the composite material, and Len is the yarn path. The loss function V(G,D) for model optimization is: Where G represents the generator, D represents the recognizer, F represents the expected value of the distribution function, u represents the real sample, v represents the noise sample, and p data (u) represents the distribution of the real sample, p noise G(v) is a noise distribution defined in a low dimension; G(v) represents the output sample generated by the generator G after receiving a random noise sample v; D(u) represents the probability that the recognizer identifies u as a real sample; D(G(v)) represents the probability that the recognizer identifies the generated sample G(v) as a real sample; and 1-D(G(v)) represents the probability that the recognizer identifies the generated sample G(v) as a fake sample.

7. The method according to claim 6, characterized in that, The yarn type in step one includes one of carbon fiber, glass fiber, aramid fiber, or ultra-high molecular weight polyethylene fiber; the yarn fineness, fabric warp and weft density, and the upper and lower boundaries of the warp path are determined according to the actual situation and different process combinations.

8. The method of claim 7, wherein, The labels of the dataset in step four are obtained by the methods described in steps two to three or by experimental preparation and testing.

9. The method of claim 8, wherein, The criteria for the optimal generative adversarial neural network model in step seven are that the warp and weft yarns in the generated yarn path are interwoven and the loss function of the model optimization is less than 0.

1. If the requirements are not met, step six is ​​repeated and the generative adversarial neural network is retrained by combining Bayesian optimization of hyperparameters until the requirements are met. The hyperparameters include learning rate, minimum batch size, number of training epochs, number of neural network units, number of neural network layers, and optimizer parameters.

10. An apparatus for producing a three-dimensional woven composite material, characterized by comprising: The processor of the device is implemented based on any one of the methods described in claims 1-9.

Images