A reverse design method for biomimetic fiber composite materials based on VAE-SVR

Abstract

This invention discloses a reverse design method for biomimetic fiber composite materials based on VAE-SVR, belonging to the interdisciplinary field of composite material design and deep learning. This invention constructs a sample database relating five-dimensional structural parameters to stress-strain curve characteristic properties; then, it trains a multi-output SVR forward surrogate model from structural parameters to mechanical properties, and constructs an enhanced training set by combining Latin hypercube sampling and geometric constraint screening; next, it constructs a VAE reverse generation network with the target mechanical properties as input conditions, and performs joint training using construction loss, KL divergence loss, and SVR-based performance consistency loss; finally, it generates candidate structural parameters based on the target stress-strain curve, obtains the final design scheme through error evaluation and physical constraint screening, and verifies it through co-curing sample preparation and three-point bending tests. This invention has high design efficiency, generates structural parameters with strong physical rationality, and can accurately match the target performance.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of composite material design and deep learning, and relates to a reverse design method for biomimetic fiber composite materials based on variational autoencoder (VAE) and support vector regression (SVR). Background Technology

[0002] Biomimetic fiber composites mimic the fibrous layered structure of organisms in nature (such as the cross-layered structure of shells and the fiber-oriented structure of plant stems), possessing the synergistic advantages of high strength, high toughness, and lightweight, and are widely used in engineering fields. Their performance hinges on the precise control of microstructural parameters such as fiber inclination angle and lamellar size. However, traditional design methods have significant drawbacks: they rely on trial-and-error experiments and empirical judgment, requiring repeated adjustments to structural parameters and mechanical testing, resulting in long design cycles and high costs. Furthermore, obtaining experimental data is difficult, with sample sizes typically limited to only a few hundred data points, making direct training of deep learning models prone to overfitting and weak generalization ability. Traditional forward prediction models can only achieve a one-way mapping from "parameters to performance," unable to generate parameter combinations from the target performance, making it difficult to meet customized needs. Existing reverse design models often ignore the physical constraints of structural parameters, potentially leading to geometrically unreasonable parameters and resulting in failure or substandard performance.

[0003] Deep learning technology provides a new approach for the reverse design of composite materials. Variational autoencoders (VAEs), as probabilistic generative models, can effectively learn data distributions and generate physically reasonable structural parameters. Support vector regression (SVR) has excellent generalization ability in small-sample, nonlinear regression tasks. However, existing research mainly focuses on the positive prediction of the mechanical properties of composite materials, and there is still little research on the reverse generation of structural parameters based on the target mechanical properties. The main technical difficulties are: (1) there is a complex nonlinear coupling relationship between the structural parameters and mechanical properties of biomimetic fiber composite materials; (2) the experimental sample size is limited, and deep generative models are prone to overfitting under small sample conditions; (3) non-physical solutions that do not meet geometric and manufacturing constraints are easily generated during the reverse generation process; (4) traditional generative models are difficult to simultaneously take into account the diversity of generation and the consistency of target performance.

[0004] Therefore, developing a reverse design method that integrates VAE and SVR to solve the problem of generating high-precision parameters with small sample sizes has significant engineering value. Summary of the Invention

[0005] To address the problems of long design cycles, weak generalization ability with small sample data, and low inverse mapping accuracy in traditional biomimetic fiber composite materials, this invention aims to provide a reverse design method for biomimetic fiber composite materials based on variational autoencoder (VAE) and support vector regression (SVR). By introducing target performance condition inputs, SVR forward surrogate model constraints, and a geometric constraint screening mechanism, a closed-loop reverse design process is established from the target stress-strain curve to the combination of structural parameters. This invention is applicable to the customized design of lightweight, high-performance biomimetic fiber composite materials in aerospace, automotive engineering, biomedicine, and other fields.

[0006] This invention provides a reverse design method for biomimetic fiber composite materials based on variational autoencoder (VAE) and support vector regression (SVR). It constructs a sample database containing five-dimensional structural parameters (fiber inclination angle, fiber sheet height, fiber sheet length, fiber sheet width, and top and bottom layer thickness) and stress-strain curve characteristics. Next, a multi-output SVR forward surrogate model is trained from structural parameters to mechanical properties, and an enhanced training set is constructed by combining Latin hypercube sampling and geometric constraint screening. Then, a VAE inverse generation network is constructed with the target mechanical properties as input conditions, and jointly trained using construction loss, KL divergence loss, and SVR-based performance consistency loss. Finally, candidate structural parameters are generated based on the target stress-strain curve, and the final design scheme is obtained through error evaluation and physical constraint screening. The design is then validated through co-curing sample preparation and three-point bending tests.

[0007] The aforementioned reverse engineering method for biomimetic fiber composite materials based on VAE-SVR specifically includes the following steps: S1. Constructing a structural-performance sample database: The fiber tilt angle θ, fiber sheet height H, fiber sheet length L, fiber sheet width W, and top and bottom layer thickness T are determined as five core design variables. Stress-strain curves corresponding to different parameter combinations are obtained through finite element simulation and mechanical experiments. The curves are discretized into performance vectors of a preset dimension. Orthogonal experimental design and Latin hypercube sampling are used to generate structural parameter combinations to construct a sample database between structural parameters and mechanical properties.

[0008] S2. Construct an SVR forward surrogate model and perform data augmentation: Train a multi-output SVR forward surrogate model from structural parameters to performance vectors based on the original sample database; generate candidate structural parameters using Latin hypercube sampling within the design variable constraints, and then use the SVR forward surrogate model to generate pseudo-label performance vectors for the candidate structural parameters; after removing samples that do not meet the geometric and manufacturing constraints, merge the effective pseudo-samples with the original samples to obtain an augmented training set.

[0009] S3. Construct a performance-constrained VAE inverse generative network: The encoder takes a combination of structural parameter vector x and performance vector y as input and outputs the mean vector and variance parameter of the latent variable distribution; the latent variable z is obtained through reparameterization techniques; the decoder takes the target performance vector and latent variable z as joint input and outputs the reconstructed structural parameter x. hat ; will x hat Input the pre-trained SVR positive surrogate model to obtain the prediction performance vector y hat This serves as a performance consistency constraint.

[0010] S4. Training the inverse generator network using a joint loss function: The joint loss function includes structural parameter reconstruction loss, latent variable KL divergence loss, and performance consistency loss, so that the generated structural parameters both conform to the distribution of training samples and can approximate the target performance requirements.

[0011] S5. Perform reverse design: Discretize or encode the target stress-strain curve into a target performance vector y*, sample multiple latent variables from a standard Gaussian distribution, and input them together with y* into the decoder to generate multiple sets of candidate structural parameters; after inverse normalization of the candidate results, use the SVR forward surrogate model to verify the error, and combine geometric constraints, manufacturing constraints and comprehensive error indicators to select the final design scheme.

[0012] S6. Experimental verification: According to the structural parameters obtained by screening, biomimetic fiber reinforced composite material samples were prepared by co-curing process, and mechanical tests were carried out by three-point bending test or drop hammer impact test. The measured stress-strain curves were compared with the target curves to verify the effectiveness of the proposed reverse design method.

[0013] Further, in step S1, the stress-strain curve is discretized into 21 feature points according to the uniform strain interval. When the curve has inconsistent sampling point spacing or many local peaks and valleys, spline interpolation is first used to reconstruct the original curve, and then a performance vector with fixed dimensions is output. The value ranges of the five structural parameters are set as follows: fiber inclination angle θ is 0°~180°, fiber sheet height H is 5~20 mm, fiber sheet length L is 10~50 mm, fiber sheet width W is 5~30 mm, and top and bottom layer thickness T is 1~5 mm. At the same time, H≤L; W≤L; T≤0.3H; θ∈[0°,180°] are applied. There is no geometric interference between adjacent fiber layers. The minimum structural size meets the forming accuracy requirements of the processing equipment to avoid generating unmanufacturable or geometrically unstable structures.

[0014] Further, step S2 specifically involves: the SVR forward surrogate model in step S2 uses radial basis functions as kernel functions, and searches for kernel parameters γ and penalty factors C through cross-validation; for the discretized multidimensional performance vector, a multi-output surrogate model can be constructed by adopting the method of "one-dimensional performance component corresponds to one SVR sub-model" to ensure the predictive ability of the entire stress-strain curve.

[0015] Furthermore, the VAE network in step S3 adopts a conditional input form, that is, the target performance vector directly participates in the decoding process, so that the reverse generation process no longer relies on simple posterior selection, but explicitly constrains the target performance in both the training and generation stages.

[0016] Furthermore, the SVR performance consistency constraint module in step S4 adopts a forward proxy form with frozen parameters, and its output directly participates in the loss calculation to construct a closed-loop optimization path of "generating structure - forward prediction - error feedback".

[0017] The joint loss function constructed in step S4 is a weighted coupling of the traditional VAE reconstruction loss, KL divergence loss, and the SVR performance consistency loss newly introduced in this invention. Its expression is as follows: , , , ; Among them, L total For the total loss, L rec For structural parameter reconstruction loss, L KL L represents the KL divergence loss of the latent variable distribution relative to the standard Gaussian distribution. perf f(·) represents the performance consistency loss calculated by the SVR forward surrogate model; α, β, and γ are the weight coefficients of the three types of losses, where α controls the structure fitting accuracy, β prevents posterior collapse, and γ determines the performance matching strength. The values ​​are obtained by Bayesian optimization or grid search on the validation set.

[0018] Furthermore, the candidate structural parameter screening in step S5 can use either the curve mean square error (MSE) or the relative area error (RAE), or combine MSE and RAE according to preset weights to form a comprehensive evaluation index; preferably, at least 100 sets of candidate structural parameters are generated, and the 10 sets with the smallest error are retained as candidate design schemes.

[0019] Furthermore, step S6 experimental verification is the process of preparing a specimen and conducting mechanical tests according to the structural parameters obtained by reverse design in step S5, specifically including the following steps: S61. First, sub-laminated components arranged at a certain angle (0~180°) are sequentially stacked and cut into strips of a set width and 0.5mm thickness using a laminate cutter. The sub-laminated components are then stacked on an alignment plate with pins that match the alignment holes to ensure precise positioning of the cuts. Before adding new sub-laminated components to the stack, the cuts on each sub-laminated component are manually extended to extend the entire length of the laminate. When the set stacking height is reached, the stack is removed from the alignment plate, and the strips are separated one by one along the cutting line with a thin scalpel blade. Each separated strip is rotated 90° to its side. After rotation, the rotated strips are aligned and joined using an alignment tool until the set nominal layer size is reached. The three separately prepared sub-laminated parts are compacted under vacuum for 24 hours. The top surface of the layer is protected by plates placed on feet at the corners of the laminate to prevent changes in fiber and lamination orientation due to the applied vacuum. A three-layer microstructure is assembled from these layers, with a 50μm PES film (i.e., polyethersulfone film, used as an interlayer bonding medium) between each layer. Lateral supports of RTV-101 silicone rubber are cast around the prototype laminate to accommodate the non-uniformity of the laminate edges and to provide structural stability during curing. Their height is adjusted to correspond to the thickness of the uncured laminate by adding metal sheets on top of them (three sheets on top of each foot, each 0.5mm thick). According to the standard curing process specifications provided by the prepreg supplier, the laminate material is cured in an autoclave, except that the heating and cooling rates are set to a low rate of 0.5~1.0 ℃ / min.

[0020] S62. Perform a static three-point bending test on the sample obtained in step S61 to obtain the stress-strain curve, and then compare it with the target stress-strain curve input in step S5 to check the effectiveness of the model.

[0021] The beneficial effects of this invention are: (1) The reverse design method for carbon fiber biomimetic composite materials disclosed in this invention demonstrates significant advantages and provides an effective solution for the design of biomimetic materials. By establishing a dataset mapping the relationship between geometric parameters and stress-strain curves, and introducing support vector regression to enhance the authenticity and credibility of the data, the scale of training data is effectively expanded, while the characteristics of the original parameters are strengthened. The established SVR-VAE neural network learning model can deeply learn the complex nonlinear relationship between carbon fiber biomimetic composite materials and stress-strain curves, and reverse design the structural parameters of carbon fiber biomimetic composite materials given a target stress-strain curve. This process significantly reduces the links that rely on manual trial and error and experience judgment in traditional design, and greatly improves design efficiency and parameter accuracy. By limiting five structural parameters, the reliability, consistency and success rate of subsequent machine learning reverse design can be guaranteed. The application of the two-stage training strategy further optimizes the performance balance between the encoder and the predictor, ensuring that the generated structural parameters not only meet the mathematical constraints, but also closely match the actual physical constraints, thereby improving the reliability and rationality of the design results, and providing an efficient, intelligent and accurate approach for the design of carbon fiber biomimetic composite materials.

[0022] (2) The reverse design method for carbon fiber biomimetic composite materials disclosed in this invention establishes a complete verification system in the design stage, ensuring the effectiveness and accuracy of the design results. Co-curing is used to conduct experiments on the predicted structure, obtaining the actual stress-strain curves of the physical object, which are then verified against the given assumptions, realizing the transition from virtual design to physical testing. This demonstrates the feasibility and accuracy of the proposed reverse design method.

[0023] (3) The innovativeness of the reverse design method for carbon fiber biomimetic composite materials disclosed in this invention lies in the deep integration of advanced machine learning technology with material structure design, opening up a new path for reverse design; practicality: it provides a complete toolchain from design to verification, which can be directly applied to the research and development and structural optimization process of new materials; operability: it provides detailed and clear step guidance and the combination of mature software and hardware technologies, making the complex design and verification process clear and easy to implement. Moreover, this invention solves the problem of training deep learning under small sample conditions by constructing an experimental database and establishing pseudo-labels; it solves the problem of traditional machine learning being difficult to reverse predict by optimizing the conditional generative adversarial network architecture and training strategy; and it does not rely on trial and error experiments, solving the problems of long design cycles and lack of universality of traditional methods, providing a brand-new idea for the design and application of biomimetic composite materials. Attached Figure Description

[0024] Figure 1 A schematic diagram of the overall process for inventing a reverse design method.

[0025] Figure 2This is a schematic diagram of the VAE-SVR collaborative constraint network in this invention.

[0026] Figure 3 A schematic diagram of finite element modeling for biomimetic fiber-reinforced composite materials.

[0027] Figure 4 This is a comparison chart of predicted mechanical curves and experimental mechanical curves. Detailed Implementation

[0028] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand the advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific examples, and the details in this specification can also be modified and changed based on different viewpoints and applications without departing from the spirit of the present invention. The illustrations provided in the following examples are only for illustrating the basic concept of the present invention. Unless otherwise specified, the following examples and features can be combined with each other. Example 1

[0029] This embodiment provides a reverse design method for biomimetic fiber composite materials based on variational autoencoder (VAE) and support vector regression (SVR), the process of which is as follows: Figure 1 As shown; this design method specifically includes the following steps: S1: Construction of the Structure-Performance Sample Database: See Figure 3 This study selected long, strip-shaped biomimetic fiber-reinforced composite materials as the research object, and used five design variables: fiber inclination angle θ, fiber sheet height H, fiber sheet length L, fiber sheet width W, and top and bottom layer thicknesses T. The preferred value ranges were set to 0°~90°, 5~20 mm, 10~50 mm, 5~30 mm, and 1~5 mm, respectively. Original parameter combinations were generated through orthogonal design and Latin hypercube design, and the corresponding stress-strain curves were obtained using finite element simulation.

[0030] In this embodiment, the finite element simulation uses tetrahedral elements to divide the mesh, with an element size of 0.45 mm. The bottom is fixed and the top is loaded at a speed of 0.5 mm / s.

[0031] When the material system is a carbon fiber reinforced epoxy resin composite material, the longitudinal elastic modulus is 140 GPa, the transverse elastic modulus is 10 GPa, and the Poisson's ratio is 0.32.

[0032] Each stress-strain curve is discretized into 21 feature points according to the uniform strain interval, forming a performance vector of length 21. For curves with multiple peaks, valleys, or local fluctuations, spline interpolation can be used for smooth reconstruction before discretization to ensure the fidelity of the curve information in the fixed-dimensional performance vector.

[0033] S2, SVR Forward Proxy Model and Data Augmentation: First, a forward surrogate model for SVR, which transforms structure parameters into performance vectors, is trained based on the original sample database. For the 21-dimensional performance vector, an SVR sub-model can be trained separately for each discrete point, or a unified training approach using multi-output regression can be employed.

[0034] The SVR model uses radial basis functions as kernel functions and determines the optimal parameters through grid search combined with cross-validation. In this embodiment, the search range of kernel parameter γ is 0.001~10, and the search range of penalty factor C is 0.1~100.

[0035] After obtaining the SVR forward surrogate model, a large number of candidate parameter combinations are generated using Latin hypercube sampling within the allowable range of five-dimensional structural parameters, and the corresponding pseudo-label performance vectors are generated for them by the SVR forward surrogate model. Subsequently, the following conditions are applied: H≤L, W≤L; T≤0.3H; θ∈[0°,180°]; there is no geometric interference between adjacent fiber layers; the minimum structural size meets the forming accuracy requirements of the processing equipment, so as to avoid generating unmanufacturable or geometrically unstable structures.

[0036] This process can significantly expand the scale of training data while preserving the statistical characteristics of the original data.

[0037] S3. Construct a VAE-SVR collaborative constraint inverse generation network.

[0038] See Figure 1 and Figure 2 A conditional VAE inverse generative network is constructed. The encoder input is the joint vector of the structural parameter vector x and the performance vector y. After processing through two fully connected layers, the output is the mean vector μ and the variance parameter logσ² of the latent variable distribution. The network is then constructed using reparameterization techniques. The latent variables are obtained, where ε follows a standard Gaussian distribution. The decoder input is a concatenated vector of the latent variable z and the target performance vector y, which, after passing through two fully connected layers, outputs the normalized structure parameters x. hat Preferably, the encoder has 256 and 128 hidden layer neurons, the decoder has 128 and 256 hidden layer neurons, the latent variable dimension is 12, the hidden layer activation function is ReLU, and the output layer activation function is Tanh.

[0039] To ensure consistency between the generated results and the target performance, a pre-trained and parameter-frozen SVR forward surrogate model is embedded in the training loop: the structural parameters x output by the decoder are... hat Inputting the SVR forward proxy model yields the prediction performance vector y. hat And through y hat The deviation between the actual performance vector y and the performance consistency loss is constructed. Therefore, both "structural reconstruction accuracy" and "performance matching accuracy" are constrained during the training phase.

[0040] S4, Joint Training: The conditional VAE network is trained using the Adam optimizer on the augmented training set. Preferably, the learning rate is set to 0.0001, the batch size to 64, and the number of training epochs to 200. Convergence is determined by validation set loss or an early stopping strategy. When the total loss fluctuates below a preset threshold (1×10⁻⁶ in this embodiment) over several consecutive training epochs, convergence is achieved. -4 Stop training when ( ).

[0041] The joint loss function constructed in this invention is a weighted coupling of the reconstruction loss and KL divergence loss of traditional VAE and the newly introduced SVR performance consistency loss. Its expression is as follows: (α, β, and γ are the weight coefficients of the three types of losses (α controls the structure fitting accuracy, β prevents posterior collapse, and γ determines the performance matching strength), and their values ​​are obtained by Bayesian optimization or grid search on the validation set): , , , ;

[0042] Among them, L total For the total loss, L rec For structural parameter reconstruction loss, L KL L represents the KL divergence loss of the latent variable distribution relative to the standard Gaussian distribution. perf f(·) represents the performance consistency loss calculated using the SVR forward proxy model. In this embodiment, α=1.0, β=0.1, and γ=0.8.

[0043] S5, Reverse Engineering: In the actual reverse engineering phase, the target stress-strain curve is first discretized into a target performance vector y*. Then, at least 100 latent variables are sampled from a standard Gaussian distribution and input together with y* into the decoder to generate candidate structural parameters. After inverse normalization of the candidate results, the corresponding stress-strain curves are predicted using the SVR forward surrogate model. The results are then reviewed and screened using MSE, RAE, geometric constraints, and manufacturing constraints. Finally, the 10 schemes with the smallest errors are retained as candidate design results.

[0044] S6. Experimental Verification: Based on the selected structural parameters, biomimetic fiber-reinforced composite material samples were prepared using a co-curing process, and three-point bending tests were conducted. The measured stress-strain curves were compared with the target curves to verify the effectiveness of the proposed reverse design method. The specific steps are as follows: S61. First, the sublaminated components with the predicted angular structure are sequentially stacked and cut into strips of a predetermined width and 0.5 mm thickness using a laminar flow cutter. The sublaminated components are then stacked on an alignment plate with pins that match the alignment holes to ensure precise positioning of the partial cuts. Before adding new sublaminated components to the stack, the cuts on each sublaminated component are manually extended to extend the entire length of the laminate. When the predetermined stack height is reached, the stack is removed from the alignment plate, and the strips are separated one by one along the cut lines using a thin scalpel blade. Each separated strip is rotated 90° to its side. After rotation, the rotated strips are aligned and joined using an alignment tool until the predetermined nominal layer size is achieved. The three separately prepared sublaminated components are compacted under vacuum for 24 hours each. The top surface of the layer is protected by plates placed on supports at the corners of the laminate to prevent changes in fiber and laminar flow orientation due to the applied vacuum. A three-layer microstructure was assembled from these layers, with a 50 μm PES film (polyethersulfone film, used as an interlayer bonding medium) between each layer. Lateral supports of RTV-101 silicone rubber were cast around the prototype laminate to accommodate the unevenness at the laminate edges and to provide structural stability during curing. Their height was adjusted to correspond to the thickness of the uncured laminate by adding metal sheets on top (three sheets, 0.5 mm thick, at the top of each support). The laminate was cured in an autoclave according to the standard curing process specifications provided by the prepreg supplier, except that the heating and cooling rates were set at a low rate of 0.5–1.0 °C / min.

[0045] S62. Perform a static three-point bending test on the sample obtained in step S61 to obtain the stress-strain curve, and then compare it with the prediction in S5 to check the effectiveness of the model.

[0046] The predicted curve is obtained in S5, and the experimental curve is obtained in S6. The experimental results are as follows: Figure 4As shown, the results indicate that the design scheme obtained by the method of the present invention can maintain a high degree of consistency between the measured stress-strain curve and the target curve, verifying the effectiveness and engineering feasibility of the reverse design method under small sample conditions.

Claims

1. A reverse design method for biomimetic fiber composite materials based on VAE-SVR, characterized in that, First, a sample database was constructed that includes five-dimensional structural parameters (fiber tilt angle, fiber sheet height, fiber sheet length, fiber sheet width, and top and bottom layer thickness) and stress-strain curve characteristics. Second, a multi-output SVR positive surrogate model was trained to translate structural parameters into mechanical properties, and an enhanced training set was constructed by combining Latin hypercube sampling and geometric constraint screening. Third, a VAE inverse generation network with target mechanical properties as input is constructed and jointly trained by construction loss, KL divergence loss and SVR-based performance consistency loss. Finally, candidate structural parameters were generated based on the target stress-strain curve. The final design scheme was obtained through error evaluation and physical constraint screening, and verified through co-curing sample preparation and three-point bending test.

2. The reverse design method for biomimetic fiber composite materials based on VAE-SVR according to claim 1, characterized in that, Specifically, the following steps are included: S1. Constructing a structural-performance sample database: The fiber tilt angle θ, fiber sheet height H, fiber sheet length L, fiber sheet width W, and top and bottom layer thickness T are determined as five core design variables; stress-strain curves corresponding to different parameter combinations are obtained through finite element simulation and mechanical experiments, and the curves are discretized into performance vectors of a preset dimension. Orthogonal experimental design and Latin hypercube sampling are used to generate structural parameter combinations to construct a sample database between structural parameters and mechanical properties. S2. Construct an SVR forward surrogate model and perform data augmentation: Train a multi-output SVR forward surrogate model from structural parameters to performance vectors based on the original sample database; generate candidate structural parameters using Latin hypercube sampling within the design variable constraints, and then use the SVR forward surrogate model to generate pseudo-label performance vectors for the candidate structural parameters; after removing samples that do not meet the geometric and manufacturing constraints, merge the effective pseudo-samples with the original samples to obtain the augmented training set; S3. Construct a performance-constrained VAE inverse generative network: The encoder takes a combination of structural parameter vector x and performance vector y as input and outputs the mean vector and variance parameter of the latent variable distribution; the latent variable z is obtained through reparameterization techniques; the decoder takes the target performance vector and latent variable z as joint input and outputs the reconstructed structural parameter x. hat ; will x hat Input the pre-trained SVR positive surrogate model to obtain the prediction performance vector y hat , as a performance consistency constraint; S4. Training the inverse generator network using a joint loss function: The joint loss function includes structural parameter reconstruction loss, latent variable KL divergence loss, and performance consistency loss, so that the generated structural parameters both conform to the training sample distribution and can approximate the target performance requirements. S5. Perform reverse design: Discretize or encode the target stress-strain curve into a target performance vector y*, sample multiple latent variables from a standard Gaussian distribution, and input them together with y* into the decoder to generate multiple sets of candidate structural parameters; after inverse normalization of the candidate results, use the SVR forward surrogate model to verify the error, and combine geometric constraints, manufacturing constraints and comprehensive error indicators to select the final design scheme; S6. Experimental verification: According to the structural parameters obtained by screening, biomimetic fiber reinforced composite material samples were prepared by co-curing process, and mechanical tests were carried out by three-point bending test or drop hammer impact test. The measured stress-strain curves were compared with the target curves to verify the effectiveness of the proposed reverse design method.

3. The reverse design method for biomimetic fiber composite materials based on VAE-SVR according to claim 2, characterized in that, In step S1, the stress-strain curve is discretized into 21 feature points according to the uniform strain interval. When the curve has inconsistent sampling point spacing or many local peaks and valleys, spline interpolation is first used to reconstruct the original curve, and then a performance vector with fixed dimensions is output. The value ranges of the five structural parameters are set as follows: fiber inclination angle θ is 0°~180°, fiber sheet height H is 5~20 mm, fiber sheet length L is 10~50 mm, fiber sheet width W is 5~30 mm, and top and bottom layer thickness T is 1~5 mm. At the same time, H≤L; W≤L; T≤0.3H; θ∈[0°,180°] are applied. There is no geometric interference between adjacent fiber layers. The minimum structural size meets the forming accuracy requirements of the processing equipment to avoid generating unmanufacturable or geometrically unstable structures.

4. The reverse design method for biomimetic fiber composite materials based on VAE-SVR according to claim 2, characterized in that, Step S2 specifically involves: the SVR forward surrogate model in step S2 uses radial basis functions as kernel functions, and searches for kernel parameters γ and penalty factors C through cross-validation; for the discretized multidimensional performance vector, a multi-output surrogate model is constructed by adopting the method of "one-dimensional performance component corresponds to one SVR sub-model" to ensure the predictive ability of the entire stress-strain curve.

5. The reverse design method for biomimetic fiber composite materials based on VAE-SVR according to claim 2, characterized in that, The VAE network in step S3 adopts a conditional input form, that is, the target performance vector directly participates in the decoding process, so that the reverse generation process no longer relies on simple posterior selection, but explicitly constrains the target performance in both the training and generation stages.

6. The reverse design method for biomimetic fiber composite materials based on VAE-SVR according to claim 2, characterized in that, The SVR performance consistency constraint module in step S4 adopts a forward proxy form with frozen parameters, and its output directly participates in the loss calculation to construct a closed-loop optimization path of "generating structure - forward prediction - error feedback". The joint loss function constructed in step S4 is a weighted coupling of the reconstruction loss and KL divergence loss of the traditional VAE and the introduced SVR performance consistency loss, and its expression is as follows: ； ； ； ； Among them, L total For the total loss, L rec For structural parameter reconstruction loss, L KL L represents the KL divergence loss of the latent variable distribution relative to the standard Gaussian distribution. perf f(·) represents the performance consistency loss calculated by the SVR forward surrogate model; α, β, and γ are the weight coefficients of the three types of losses, where α controls the structure fitting accuracy, β prevents posterior collapse, and γ determines the performance matching strength. The values ​​are obtained by Bayesian optimization or grid search on the validation set.

7. The reverse design method for biomimetic fiber composite materials based on VAE-SVR according to claim 2, characterized in that, In step S5, the candidate structural parameters are selected using the curve mean square error (MSE), or the relative area error (RAE), or by combining MSE and RAE with preset weights to form a comprehensive evaluation index; at least 100 sets of candidate structural parameters are generated, and the 10 sets with the smallest error are retained as candidate design schemes.

8. The reverse design method for biomimetic fiber composite materials based on VAE-SVR according to claim 2, characterized in that, Step S6, experimental verification, is the process of preparing a specimen and conducting mechanical tests according to the structural parameters obtained by reverse design in step S5. Specifically, it includes the following steps: S61. First, sub-laminated components with a 0~180° arrangement are sequentially stacked and cut into strips of a set width and 0.5mm thickness using a laminar flow cutter. Then, the sub-laminated components are stacked on an alignment plate with pins that match the alignment holes to ensure precise positioning of the partial cuts. When the set stacking height is reached, the stacked components are removed from the alignment plate, and the strips are separated one by one along the cutting line with a thin scalpel blade. Each separated strip is rotated 90° to its side. After rotation, the rotated strips are aligned and joined using an alignment tool until the set nominal layer size is reached. The three separately prepared sub-laminated components are compacted under vacuum for 24 hours. The top surface of the layer is protected by a plate placed on a support at the corner of the laminate to prevent the fiber and laminar flow orientation from changing due to the applied vacuum. These layers are assembled into a three-layer microstructure, with a 50 μm PES film between each layer; Lateral supports of RTV-101 silicone rubber were cast around the prototype laminate to accommodate the unevenness of the laminate edges and to provide structural stability during the curing process. Their height was adjusted to correspond to the thickness of the uncured laminate by adding metal sheets on top of them; the laminate was cured in an autoclave with heating and cooling rates set at a low rate of 0.5~1.0 ℃ / min; S62. Perform a static three-point bending test on the sample obtained in step S61 to obtain the stress-strain curve, and then compare it with the target stress-strain curve input in step S5 to check the effectiveness of the model.

Images