A three-dimensional multi-scale structure optimization and topology optimization method based on GAN network

Abstract

The present invention discloses a three-dimensional multi-scale structure optimization and topology optimization method based on a GAN network, firstly establishing a macro-generative design model and a micro-generative design model respectively, then using the obtained model to expand to three-dimensional structure generation, and realizing the generation of three-dimensional multi-scale structure based on the connectivity between microstructures; the topology optimization method, the steps include: using a topology optimization method to construct a macro-generative design model. Using a topology optimization method to construct a micro-generative design model. Based on the obtained model, a three-dimensional convolutional neural network is used to extract three-dimensional structural features, and the number of channels is increased at the same time to map the three-dimensional structural parameters, realize the generation of a three-dimensional structural model, and construct a three-dimensional multi-scale structural generative design model. The present invention not only has extremely high efficiency and realizes the rapid generation of three-dimensional multi-scale structures, but also can improve the stress concentration phenomenon of the structure and ensure the connectivity of the microstructure.

Description

Technical Field

[0001] The present invention belongs to the technical field of mechanical structure optimization design, and specifically relates to a three-dimensional multi-scale structure optimization and topology optimization method based on a GAN network. Background Art

[0002] The R&D and manufacturing of high-end equipment often involves the design of high-performance components. These components often operate under extreme loads, and traditional topology optimization methods have reached performance bottlenecks. There is an urgent need to push performance limits from multiple perspectives, both macroscopically and microscopically. This is where multiscale structural design comes in. However, multiscale structural design inevitably incurs a significant increase in computational costs. The precise and efficient generation of multiscale structures remains a major challenge in the field of mechanical structural optimization.

[0003] In recent years, the improvement of computer performance has enabled the realization of complex artificial intelligence. As the main means of realizing artificial intelligence, deep learning has inherent advantages in processing high-dimensional data due to its powerful data fitting and feature extraction capabilities. Topology optimization based on deep learning transforms the iterative optimization process of traditional structural optimization into an image processing problem. The trained model has extremely fast generation speed and high computational accuracy. However, in the research on topology optimization based on deep learning, most scientific researchers or technicians focus on aspects such as image generation accuracy and algorithm speed improvement, and research on three-dimensional multi-scale structural design is still blank. In addition, in the structural optimization process of high-performance components, the poor microstructure connection effect and the mutual influence of macroscopic and microscopic performance are also problems that need to be solved. Therefore, it is necessary to develop a three-dimensional multi-scale structural optimization and topology optimization method based on GAN network, in order to achieve high-precision and rapid generation of three-dimensional multi-scale structures, providing new technical support for breakthroughs in component performance and ultra-lightweighting. Summary of the Invention

[0004] In response to the technical problems mentioned in the background technology, the purpose of the present invention is to provide a three-dimensional multi-scale structure optimization and topology optimization method based on the GAN network.

[0005] In the present invention, the GAN network (Generative Adversarial Network) is one of the basic models commonly used in generative design. The network can extract image features and, through the training of the discriminator and the generator, can generate new data images, thereby improving the novelty and diversity of the results. However, when the network is used for structural optimization, although the innovation of the results can be improved, the generated structure cannot guarantee the performance requirements and requires subsequent verification, which loses the efficient advantage of generative design. The VAE network (Variational AutoEncoder) contains an encoder and a decoder. The characteristic of the network is that during the training process, the association between the content of the data and the image features is taken into account at the same time, thereby realizing the generation and representation learning of the data. The VAE network can establish a mapping relationship between physical properties and image features. In order to solve the problem that the GAN network is difficult to guarantee mechanical properties in structural optimization, the present invention proposes a new generative design model (GAN-VAE network), which cleverly uses the GAN network as a cabinet frame and the VAE network as a generator of the GAN network, adjusts the parameters of each layer in the network, ensures that the image features are associated with the mechanical properties, and at the same time, the generated results also have the diversity advantage of the GAN network.

[0006] The technical solutions adopted in the present invention are as follows.

[0007] A three-dimensional multi-scale structure optimization method based on a GAN network first establishes a macro-generative design model and a micro-generative design model respectively, then uses the resulting models (macro-generative design model and micro-generative design model) to expand to three-dimensional structure generation, and realizes the generation of three-dimensional multi-scale structure based on the connectivity between microstructures. Furthermore, a GAN-VAE network is adopted in the generative design model (macro-generative design model and micro-generative design model). The GAN-VAE network uses the GAN network as a framework and the VAE network as the generator of the GAN network. In the present invention, the microstructure refers to the smallest unit defined to constitute the macrostructure.

[0008] A three-dimensional multi-scale structure topology optimization method based on a GAN network, comprising the following steps: Step 1: Use topology optimization method to build a macro-generative design model; Step 2: Use topology optimization method to build a micro-generative design model; Step 3: Based on the obtained models (macro-generative design model and micro-generative design model), a 3D convolutional neural network is used to extract 3D structural features while increasing the number of channels (i.e., increasing the number of channels in the network) to map the 3D structural parameters and realize the generation of a 3D structural model. In the present invention, the two-dimensional structural feature is N in the N×N plane area.2 The pixel values ​​are multiplied by the convolution kernel and then summed to generate a single pixel value in the corresponding new feature map; the three-dimensional structural feature is N in the N×N×N cube area. 3 The pixel values ​​are multiplied by the convolution kernel and then summed to generate a single pixel value in the corresponding new feature map; Step 4: Construct a three-dimensional multi-scale structure generative design model.

[0009] As a preferred solution, in step 1, topology optimization methods are used to generate sample data (i.e., images of structures generated using topology optimization) for model (macro-generative design model) training and testing {for example, the input to the GAN-VAE network is an 81×41×6 tensor, that is, 6 sets of images of the input channel (input 6 images, output 1 new structure image)}. A GAN-VAE network is used, and the Adam algorithm is used to optimize the loss function (optimized network loss function). The model convergence and stability are judged based on the network loss, and the accuracy of the generated results is judged based on the image loss. Because the generator and discriminator will have certain errors when extracting features and making judgments (any neural network training will have certain errors), the network error is calculated in this model, including the sum of the errors of all internal networks; the image loss compares the pixels of the generated image with the pixels of the real image, which is a re-verification of the results after the discriminator determines the authenticity based on the features; the loss function is the condition that constrains the network loss.

[0010] As a preferred solution, in step 2, the channel is adjusted based on the obtained macro-generative model, the elastic constant of the microstructure is used as a constraint, and a microstructure dataset with a gradient function is established; the GAN-VAE network is adopted, and the Adam algorithm is used to optimize the loss function. The degree of model convergence and stability are judged based on the network loss, and the accuracy of the generated results is judged based on the image loss.

[0011] As a preferred solution, in step 3: During model training, the number of deep learning network channels (GAN-VAE network (the macro and micro networks are the same, and channels are added only in the three-dimensional model)) is increased from 6 to 10. The first three channels represent displacements in the X, Y, and Z directions, channels 4-6 represent normal stresses in the XY, YZ, and XZ directions, channels 7-9 represent shear strains in the XY, YZ, and XZ directions, and the tenth channel represents volume fraction, ranging from 0 to 1 (0-1 represents the range of volume fraction variation).

[0012] As a preferred solution, an interpolation model is used to expand the data (microstructure images of different volume fractions, shear elastic modulus, and compliance) according to a set step size for performance indicators such as volume fraction, shear elastic modulus, and compliance to establish a microstructure dataset with gradient function.

[0013] As a preferred solution, in step 4, a three-dimensional convolutional neural network is added to the GAN-VAE network to traverse the three-dimensional structure and perform feature extraction, and then convert it into a two-dimensional image for processing; To slice the 3D model and extract its features, a 3D convolutional neural network is added before the input of the 2D GAN-VAE discriminator and a 3D deconvolutional network is added after the generator, forming a new network (the 3DCONV in the figure is a 3D convolutional neural network). The input of the 3D GAN-VAE network is an N×N×N tensor, that is, N sets of images of input channels (the number of input channels varies depending on the macro and micro structures and the types of microstructures, so the number of channels is determined based on the characteristics). Among them, given the boundary conditions and constraints, the macroscopic configuration is obtained according to the principal stress; under the condition that each microstructure performance satisfies the stress constraint of the macrostructure in the microunit, the microstructure with the minimum volume fraction is searched and the macroscopic configuration is filled; combined with the cellular automaton algorithm, smooth transition and stable connection between adjacent microstructures are achieved (judged by whether the average strain energy of the unit in the cellular automaton is greater than 0, which proves that there are adjacent units that can be connected), and a three-dimensional multi-scale structural generative design model is constructed.

[0014] As a preferred solution, the encoder of the GAN-VAE network generator (because the discriminator also has an encoder) consists of three layers of blocks (three blocks are connected in series, and the result of each convolution layer is used as the input of the next layer). The first layer performs feature extraction and data dimensionality reduction through two convolution layers, and at the same time, the LeakyReLU activation function is connected after each convolution layer to improve convergence efficiency (the first block performs feature extraction and data dimensionality reduction through two convolution layers, and at the same time, the LeakyReLU activation function is connected after each convolution layer to improve convergence efficiency), and then the dimension of each feature is reduced through the pooling layer; the second and third layers are both composed of convolution and LeakyReLU activation functions; the second block and the third block are both composed of convolution and LeakyReLU activation functions; The decoder of the GAN-VAE network generator consists of a two-layer network (the decoder of the GAN-VAE network generator consists of two blocks). The feature data (feature data is the data result after the feature is extracted by the generator encoder) is subjected to two layers of deconvolution and then connected to the LeakyReLU activation function. Finally, the generated image is output through a fully connected layer. The discriminator of the GAN-VAE network consists of 4 layers of blocks and activation functions. Each layer of data in the 4 layers of blocks must be processed by convolution, regularization and the activation function Softmax activation function to achieve feature extraction.

[0015] Furthermore, in the process of constructing the micro-generative design model, the topology optimization method is used to find the most suitable structural distribution in the design domain, and the results of the topology optimization and the equivalent elastic matrix are calculated; among them, the objective function is set (set for the microstructure to be optimized, according to the required performance, such as the "maximum elastic modulus") and the variables are set and iterated (this iteration is the topology optimization method itself, so that the structure continues to approach the target of the required performance) to generate several image data, and the image data is divided according to the set ratio for deep learning model training and testing.

[0016] Beneficial effects: Compared with traditional mechanical structure optimization methods, the present invention not only has extremely high efficiency and realizes the rapid generation of three-dimensional multi-scale structures, but also shortens the design time by nearly 10 times. For a structure with 80×40 units, the design time only takes about 9 seconds, and for a structure with 60×20×8 units, the design time only takes about 33 seconds. It can also improve the stress concentration phenomenon of the structure and ensure the connectivity of the microstructure. More importantly, the GAN network is used as the cabinet frame and the VAE network is used as the generator of the GAN network, which cleverly solves the problem that the GAN network is difficult to ensure mechanical properties in structural optimization, ensures the correlation between image features and mechanical properties, and also has the diversity advantage of the GAN network. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 This is a flow chart of a three-dimensional multi-scale structure topology optimization method based on a GAN network in an embodiment; Figure 2 This is a generator architecture diagram of the macro-generative design model in the embodiment; Figure 3 This is a diagram of the discriminator architecture of the macro-generative design model in the embodiment; Figure 4 This is a design flow chart of the three-dimensional multi-scale structure generative design model in the embodiment; Figure 5 This is a generator architecture diagram of a three-dimensional multi-scale structure generative design model in an embodiment; Figure 6 This is a diagram of the discriminator architecture of the three-dimensional multi-scale structure generative design model in the embodiment; Figure 7 Schematic diagram of the loading condition of the three-dimensional support structure in Example 1; Figure 8 Generate a result graph for the three-dimensional support structure in Example 1; Figure 9 This is a physical image corresponding to the three-dimensional support structure in Example 1; Figure 10 This is the channel situation involved in the training of the model in Example 1. DETAILED DESCRIPTION

[0018] The following is a clear and complete description of the technical solutions of the present invention in conjunction with the accompanying drawings. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention. Example 1

[0019] A three-dimensional multi-scale structure topology optimization method based on GAN network, the overall process is as follows Figure 1 As shown, the steps are as follows: Step 1: Use topology optimization to build a macro-generative design model. Use topology optimization to generate sample data for model training and testing. A GAN-VAE network is used, and the Adam algorithm is used to optimize the loss function. The model's convergence and stability are determined based on network loss, and the accuracy of the generated results is determined based on image loss. This completes the establishment of the macro-generative design model. Specifically: (11) Macrostructure Data Generation The sample data for the macro-generative design model is generated by a topology optimization method. Since topology optimization requires a one-to-one correspondence between working conditions and results, the working conditions need to be iteratively generated as variables. These variables include volume fraction, load size, location, and direction, and constraints on fixed edges. For example, the topology optimization design domain is set to 80×40, with the left fixed end fully constrained to 6 degrees of freedom. The loading position is the rightmost plane of the cantilever beam (the optimization of the cantilever beam is a typical model for topology optimization, and the input to the GAN-VAE network is the result of topology optimization). The load angle is 0-180° with a step size of 45°. The volume fraction range is 0.5-0.7 with a step size of 0.05. A total of 1,500 sample data are generated for model (macro-generative design model) training and testing. (12) GAN-VAE Network The generator adopts the VAE network structure, such as Figure 2 As shown in the figure, the encoder of the VAE network consists of three blocks. The first layer uses two convolutional layers to extract features and reduce data dimensionality. At the same time, the LeakyReLU activation function is connected after each convolution layer to improve convergence efficiency. The dimension of each feature is then reduced through the pooling layer to prevent overfitting. The second and third layers are composed of convolution and LeakyReLU activation functions to improve the nonlinear mapping between image features and structural performance. In the latent space (latent space is a technical term in VAE networks), the encoder data is sampled and the image features are matched with mechanical properties. The decoder of the VAE network consists of a two-layer network. The feature data undergoes two layers of deconvolution and is then connected to the LeakyReLU activation function. Finally, the generated image is output through a fully connected layer. The discriminator consists of 4 layers of blocks and activation functions, such as Figure 3 As shown, each layer of data undergoes convolution, regularization, and activation functions to achieve efficient and accurate feature extraction. The convolutional layer performs feature extraction, data dimensionality reduction, and weight sharing. The LeakyReLU activation function eliminates the gradient saturation effect of the Sigmoid function and improves convergence efficiency. The regularization layer limits the number of channels to ensure stable output results. Batch Normalization further improves the model's generalization capabilities, and the Sigmoid function is selected as the final output to ensure the discriminator's judgment accuracy. To accelerate model loss calculations, the Adam algorithm was used to calculate the iteration step size. This method automatically adjusts parameters based on the values ​​of the second and first moment estimates of the gradient, extending gradient descent and eliminating the influence of irrelevant parameters. The Adam optimizer hyperparameters learning_rate and beta_1 (which adjusts the Adam optimizer parameters to accelerate iterations) were set to 0.5 to prevent oscillation and ensure a smooth gradient descent. (13) Model training According to the parameter iteration of topology optimization, a total of 1,500 topology results were generated. The data was divided into training set and test set in a ratio of 8:2. The input of the GAN-VAE network is a tensor of 81×41×6 (forces and constraints can only be added to the nodes, so 81×41 pixels are required), and the output is a topology optimization result of 80×40. The input tensor (a professional term that can be accumulated as a set of vectors, a vector is a set of data, and a tensor is two-dimensional data that can form a picture) can represent the performance parameters of the structure, such as displacement, stress, and volume fraction. A total of 6 channels are used, of which the 1st and 2nd channels are the displacement of the node in the X and Y directions, respectively, the 3rd to 5th channels are the stress and shear strain of the node (expressed in angle), and the 6th channel is the volume fraction, such as Figure 10 As shown, in channels 1 and 2, the displacements of adjacent units vary uniformly. However, due to the high stress at the applied point, concentrated stress points can be observed in channels 3-5. Because the stress and displacement values ​​differ significantly, a BatchNormalization layer is added to eliminate the magnitude effect. The neural network output is an 80×40 tensor, representing the probability density output after comparison with the ground truth. Because of the sigmoid function activation, the function's value ranges from 0 to 1. (14) Model loss determination In addition to the loss of the GAN-VAE network itself, in order to make the generated results as similar as possible to the topology optimization results, the mean square error (MSE) of the image should be considered in the loss function. Its numerical value is the mean of the sum of squares of the pixel errors between the generated image and the real image. Step 2: Use topology optimization to build a microscopic generative design model. Based on the resulting macroscopic generative model, adjust the channels and use the elastic constants of the microstructure as constraints to ensure microstructural performance. A microstructural dataset with gradient functionality is also established. A GAN-VAE network is used, and the Adam algorithm is used to optimize the loss function. The model's convergence and stability are determined based on the network loss, and the accuracy of the generated results is determined based on the image loss. Specifically: (21) Microstructure data generation Using topology optimization methods, we find a suitable structural distribution within the design domain and calculate the optimized results and equivalent elastic matrix. Taking the shear modulus as the objective function, we iteratively generate 1,500 data sets using volume fraction and elastic constant as variables. This data set is then divided into an 8:2 ratio for training and testing the deep learning model. The network architecture and loss determination refer to the description of the macro-generative design model. (22) Model training Based on topology optimization, 1,500 40×40 pixel microstructure images were generated. The material elastic constant of each equivalent volume element (RVE) is a direct factor in expressing material properties. Therefore, each elastic constant is visualized to form a tensor as the input data of the GAN-VAE network. The input of the neural network is a 40×40×5 tensor with a total of 5 channels. Four channels are used to express the characteristics of the material, including anisotropy, negative Poisson's ratio and maximum shear modulus. The fifth channel represents the volume fraction of the microstructure, which better controls the fixed form of the structure and facilitates subsequent interpolation processing to form a gradient function data set. Step 3: Based on the obtained model, a 3D convolutional neural network is used to extract 3D structural features, while increasing the number of channels to map 3D structural parameters to achieve 3D structural model generation; specifically: (31) Network Architecture The overall architecture of the model in this step is to add a 3DCNN network to the two-dimensional GAN-VAE generative model to extract three-dimensional structural information. Unlike two-dimensional convolution, three-dimensional convolution not only increases the number of channels but also ensures a close relationship between layers, making feature extraction more accurate. The model input is still a visual image of the structural characteristic parameters, and the output is the three-dimensional topology optimization result. The generator uses a 3D CNN to stack features after the VAE network output, and then outputs the generated results through a fully connected layer. The discriminator performs five downsampling operations to extract features from the 3D topology optimization results. Since 3D images can be viewed as slices of 2D images, each layer of features must output a probability density. Therefore, the Softmax activation function is ultimately used to enhance the discriminator's multi-feature classification capabilities. (32) Model training For generative design of three-dimensional macrostructures, the number of deep learning network channels needs to be increased from 6 to 10. The first three channels represent displacement in the X, Y, and Z directions, channels 4-6 represent normal stress in the XY, YZ, and XZ directions, channels 7-9 represent shear strain in the XY, YZ, and XZ directions, and the tenth channel represents volume fraction, ranging from 0 to 1. For generative design of 3D microstructures, simply change the number of input channels and the output topology optimization results. If only considering 3D tetragonal elements, simply modify the existing channels to account for the increased elastic constants. The first six channels represent the microstructure's properties, and the seventh channel represents the volume fraction. (33) Establishment of gradient function dataset In order to establish a microstructure dataset with gradient function and provide data support for multi-scale topology optimization and material design, an interpolation model is used to expand the data of performance indicators such as volume fraction, shear elastic modulus, and compliance at a certain step size. Step 4: Construct a 3D multi-scale structure generative design model. Add a 3D convolutional neural network to the GAN-VAE network, traverse the 3D structure and perform feature extraction, and then convert it into a 2D image. The process is as follows: Figure 4 As shown in the figure, given the boundary conditions and constraints, the macro configuration is obtained according to the principal stress to ensure the load-bearing performance of the generated result; based on the condition that each microstructure performance satisfies the stress constraint of the macro structure in the micro unit, the microstructure with the minimum volume fraction is searched and filled in the macro configuration to ensure the ultra-lightweight of the multi-scale structure; then, the cellular automaton algorithm is combined to achieve smooth transition and stable connection between adjacent microstructures, and finally an ultra-lightweight multi-scale structure is generated, thereby constructing a three-dimensional multi-scale structure generative design model. In this step: (41) Network Architecture The output of the model in this step is the macroscopic and microscopic topology optimization results, input data and the number of channels (see the description in steps 1 and 2). In 3D CNN, the most common deep learning model is the voxel-based three-dimensional representation. Since topology optimization forms a structure with a unit as the smallest unit, the three-dimensional neural network can be used to extract the three-dimensional topology optimization information. The generator part is as follows Figure 5As shown in Figure 1, a three-dimensional deconvolution layer is connected after the output of the VAE network to achieve image reconstruction. Before the feature input, the discriminator first performs dimensionality reduction processing on the 3D image. Figure 6 As shown in the figure, since there is more than one probability output, the output layer of the discriminator is replaced by Sigmod with Softmax activation function, so that the network can better classify multiple features. (42) Optimization and improvement of microstructure connectivity Cellular automata (CA) can adapt to changes in the surrounding environment based on a certain distance, so each cell is related to its neighboring cells. Since microstructure generation uses strain energy as an objective function, and average strain energy is a key parameter that directly expresses strain, comparing the predicted unit strain energy with the actual unit strain energy can ensure that the unit has at least one connection, thereby ensuring the connectivity of the microstructure.

[0020] In order to verify the accuracy and versatility of the model under other working conditions, a three-dimensional support structure is selected as a case study. The design domain of the three-dimensional support structure is as follows: Figure 7 As shown, the overall size is L×H×W, where length L=20cm, height H=15cm, width W=20cm, the pixels of the image are 20×15×20, the side length of the macrostructure unit is 1cm, and the side length of a single microstructure is also 1cm. The constraints are the four corners of the bottom of the three-dimensional support structure, the load is applied at the midpoint of each side of the upper surface and the center position, and the direction is vertically downward. Four microstructure units with different volume fractions are selected. This type of unit is a three-dimensional lattice structure generated with the maximum shear elastic modulus as the objective function. The results of the three-dimensional multi-scale structure generation are shown in Figure 2. Figure 8 As shown in the figure, it can be seen that the closer to the load, the larger the volume fraction of the microstructure, and the stronger the ability to resist shear load. The gray is a fully solid microstructure. Figure 9 As shown, the generated model was manufactured using PolyJet 3D printing technology, resulting in a finished product with structural integrity and high precision. The results demonstrate that this solution effectively integrates additive manufacturing technology, providing design references and dataset support for the research and development of new materials and structures. Furthermore, under the same operating conditions and volume fraction, the solution in this embodiment achieves smaller deformation than traditional topology optimization results, and also reduces stress concentration in the structure during the generation process.

[0021] Using the solution of Example 1, we compared the generation efficiency with different numbers of units and types of microstructures. The results of the comparison between Example 1 and traditional topology optimization methods are shown in Table 1. The results show that the solution of Example 1 has a significantly higher efficiency in generating three-dimensional structures than traditional topology optimization methods. Table 1 Comparison results between Example 1 and traditional topology optimization method

[0022] The present invention is not only extremely efficient, but also realizes the rapid generation of three-dimensional multi-scale structures, shortening the design time by nearly 10 times. For a structure with 80×40 units, the design takes only about 9 seconds, and for a structure with 60×20×8 units, it takes only about 33 seconds. It can also improve the stress concentration phenomenon of the structure and ensure the connectivity of the microstructure. More importantly, it uses the GAN network as the framework and the VAE network as the generator of the GAN network, cleverly solving the problem that the GAN network is difficult to ensure mechanical properties in structural optimization, ensuring the correlation between image features and mechanical properties, and also has the diversity advantage of the GAN network.

Claims

1. A three-dimensional multi-scale structure optimization method based on a GAN network, characterized by: First, a macro-generative design model and a micro-generative design model are established respectively, and then the obtained models are used to expand to three-dimensional structure generation, and the generation of three-dimensional multi-scale structures is realized based on the connectivity between microstructures.

2. The three-dimensional multi-scale structure optimization method according to claim 1, characterized in that: The GAN-VAE network is used in the generative design model. The GAN-VAE network uses the GAN network as a framework and the VAE network as the generator of the GAN network.

3. A three-dimensional multi-scale structure topology optimization method based on a GAN network, characterized in that the steps include: Step 1: Use topology optimization method to build a macro-generative design model; Step 2: Use topology optimization method to build a micro-generative design model; Step 3: Based on the obtained model, a three-dimensional convolutional neural network is used to extract three-dimensional structural features, and the number of channels is increased to map the three-dimensional structural parameters to realize the generation of a three-dimensional structural model; Step 4: Construct a three-dimensional multi-scale structure generative design model.

4. The three-dimensional multi-scale structure topology optimization method according to claim 3, characterized in that: In step 1, the topology optimization method is used to generate sample data for model training and testing. The GAN-VAE network is adopted, and the Adam algorithm is used to optimize the loss function. The degree of convergence and stability of the model are judged based on the network loss, and the accuracy of the generated results is judged based on the image loss.

5. The three-dimensional multi-scale structure topology optimization method according to claim 4, characterized in that: In step 2, the channel is adjusted based on the obtained macro-generative model, the elastic constant of the microstructure is used as a constraint, and a microstructure dataset with a gradient function is established; the GAN-VAE network is used, and the Adam algorithm is used to optimize the loss function. The degree of convergence and stability of the model is judged based on the network loss, and the accuracy of the generated results is judged based on the image loss.

6. The three-dimensional multi-scale structure topology optimization method according to claim 3, characterized in that: In step 3: During model training, the number of deep learning network channels is increased from 6 to 10, where the first three channels represent displacements in the X, Y, and Z directions, channels 4-6 represent normal stresses in the XY, YZ, and XZ directions, channels 7-9 represent shear strains in the XY, YZ, and XZ directions, and the tenth channel represents volume fractions, ranging from 0 to 1.

7. The three-dimensional multi-scale structure topology optimization method according to claim 3, characterized in that: The interpolation model is used to expand the data of performance indicators such as volume fraction, shear elastic modulus, and compliance according to the set step size to establish a microstructure data set with gradient function.

8. The three-dimensional multi-scale structure topology optimization method according to claim 3, characterized in that: In step 4, a three-dimensional convolutional neural network is added to the GAN-VAE network to traverse the three-dimensional structure and perform feature extraction, and then convert it into two-dimensional image processing; In order to slice the 3D model and extract the features of the 3D model, a 3D convolutional neural network is added before the input of the 2D GAN-VAE network discriminator, and a 3D deconvolutional network is added after the generator to form a new network. The input of the 3D GAN-VAE network is an N×N×N tensor, that is, N groups of images of the input channel. Among them, given the boundary conditions and constraints, the macroscopic configuration is obtained according to the principal stress; under the condition that each microstructure performance satisfies the stress constraint of the macrostructure in the microunit, the microstructure with the minimum volume fraction is searched and the macroscopic configuration is filled; combined with the cellular automaton algorithm, smooth transition and stable connection between adjacent microstructures are achieved, and a three-dimensional multi-scale structure generative design model is constructed.

9. The three-dimensional multi-scale structure topology optimization method according to claim 3, characterized in that: The encoder of the GAN-VAE network generator consists of three layers of blocks. The first layer performs feature extraction and data dimensionality reduction through two convolution layers. At the same time, the LeakyReLU activation function is connected after each convolution layer to improve the convergence efficiency, and then the dimension of each feature is reduced through the pooling layer. The second and third layers are composed of convolution and LeakyReLU activation functions; the second block and the third block are composed of convolution and LeakyReLU activation functions; The decoder of the GAN-VAE network generator consists of a two-layer network. The feature data is deconvolved through two layers and then connected to the LeakyReLU activation function. Finally, the generated image is output through the fully connected layer. The discriminator of the GAN-VAE network consists of 4 layers of blocks and activation functions. Each layer of data in the 4 layers of blocks must be processed by convolution, regularization, and the activation function Softmax activation function to achieve feature extraction.

10. The three-dimensional multi-scale structure topology optimization method according to claim 3, characterized in that: In the process of building a micro-generative design model, the topology optimization method is used to find the most suitable structural distribution in the design domain, and the results of topology optimization and the equivalent elastic matrix are calculated. After setting the objective function and variables, several image data are iteratively generated, and the image data are divided according to the set ratio for deep learning model training and testing.

Images