A design system for three-dimensional bionic objects

Abstract

The present invention discloses a three-dimensional biomimetic design method based on a generative adversarial network. The method comprises the following steps: Step 1: Establishing a training set; Step 2: Establishing a deep generative model and training the deep generative model using the training set; the deep generative model comprises an implicit autoencoder (3DCNN) and a latent vector generative model; the implicit autoencoder comprises an encoder (3DCNN) and a decoder (IM-Decoder); the latent vector generative model employs a generative adversarial network; and Step 3: Using the deep generative model to generate a three-dimensional biomimetic object. The present invention also discloses a system for implementing the above method, comprising a data preprocessing module, a deep generative model, a sample vector repository, and a post-processing module.

Description

Technical Field

[0001] The present invention belongs to the field of machine learning and intelligent technology, and relates to a design system for three-dimensional bionic objects. Background Art

[0002] Bionics design involves creatively simulating the structure, function, form, and color of nature and biological systems. Starting from human psychological needs and guided by the designer's personal biomimetic philosophy, bionic design creates products imbued with biological interest and natural forms. These products, in turn, serve humanity, thus completing a natural cycle from person to person. Based on the different ways bionics is applied, bionic design can be broadly categorized into product morphological bionics, functional bionics, structural bionics, color bionics, and texture bionics. Morphological bionics is widely used in modern product design, including those for daily necessities, small appliances, furniture, and transportation. Today, designers are turning their attention to bionic design as a key area of ​​product design innovation.

[0003] To generate more outstanding and creative biomimetic designs, academia and industry have been exploring the principles of biomimetic morphological design through computer technology for many years. Before applying deep generative models to biomimetic product form design, academia used evolutionary design to establish modeling databases and extraction rules. These models used evolutionary algorithms to define fitness and product feature models, extracting characteristic elements from the contours of biomimetic objects to generate diverse design solutions for product contours. Recent research has attempted to use deep learning techniques to automatically synthesize the contours of biomimetic products. In particular, Simiao Yu et al. were the first to apply deep generative models, such as generative adversarial networks, to the biomimetic design of product shapes. They mathematically defined the biomimetic design problem from a deep learning perspective, thereby designing a deep generative model to generate biomimetic product contour designs.

[0004] Deep generative models combining autoencoders and generative adversarial networks (GANs) have demonstrated unique advantages and great potential in areas such as 2D image generation, restoration, and style transfer. In 2016, Wu et al. from MIT proposed a method for applying generative adversarial networks to 3D shape generation. The trained model can reconstruct a 3D model of a chair from a photo. Using a voxelized 3D data model, they first proposed the concept of a latent vector space for 3D shape representation. Panos et al. employed an autoencoder and an intermediate latent generative adversarial network (LatentGAN) to embed and represent 3D point cloud data. Using an autoencoder to directly train a latent vector space based on a 3D dataset, they then leveraged bijections between Euclidean space and 3D shapes to perform a series of 3D generation tasks. The presence of the intermediate LatentGAN ensured the diversity of the generated results. Chen et al. used an implicit method to optimize the 3D model fitting of the deep generative model LatentGAN, making the interpolated 3D models more continuous and smooth. They also reused voxel-based 3D data and used the MarchingCube algorithm to convert the voxel data into a more versatile 3D mesh. However, deep generative models have not yet been applied in the design of bionic product forms. Summary of the Invention

[0005] The present invention provides a three-dimensional bionic design system based on a generative adversarial network, the system comprising:

[0006] The data preprocessing module is used to input the 3D shape dataset represented by the bionic object and design target, expand the dataset through design target-based and performance-based data augmentation, and obtain the signed distance field of the 3D model required for model training through mesh2voxel and SDF sampler;

[0007] A deep generative model comprising an implicit autoencoder, a generative adversarial network / interpolation operator / creative generator solver; the implicit autoencoder comprising an encoder 3DCNN and a decoder IM-Decoder;

[0008] A sample vector repository, which is used to store latent vector point sets. The latent vectors in the repository are used to obtain three-dimensional biomimetic generated objects through an implicit decoder.

[0009] A post-processing module is used to post-process the three-dimensional bionic object to obtain a three-dimensional bionic design.

[0010] In the three-dimensional bionic design system based on a generative adversarial network proposed in the present invention, the implicit autoencoder is used to fit the signed distance field of the three-dimensional model to perform the first stage of training; in order to obtain a high-fidelity generated product that meets the design requirements, it is necessary to search for a suitable latent vector point set in the latent space; the latent vector point set is obtained by Gaussian sampling after training the generative adversarial network, or by interpolating the latent vector pairs obtained by the encoder of the three-dimensional model of the bionic object and the design target in the given training set, or by using a creative generated product solver; wherein, after the first stage of codec network training is completed, the model to be trained undergoes a second stage of training to obtain a callable latent vector generation model.

[0011] 3DCNN Encoder

[0012] layer Convolution kernel size stride length Activation Function Output shape Input voxels (64,64,64,1) conv3d (4,4,4) (2,2,2) BNLReLU (32,32,32,32) conv3d (4,4,4) (2,2,2) BNLReLU (16,16,16,64) conv3d (4,4,4) (2,2,2) BNLReLU (8,8,8,128) conv3d (4,4,4) (2,2,2) BNLReLU (4,4,4,256) conv3d (4,4,4) Sigmoid (1,1,1,128)

[0013] Implicit Decoder

[0014]

[0015]

[0016] Generative Adversarial Networks

[0017] Generator

[0018] layer Activation Function Output shape Input latent vector (128) Fully connected LReLU (2048) Fully connected LReLU (2048) Fully connected Sigmoid (128)

[0019] Discriminator

[0020] layer Activation Function Output shape Input latent vector (128) Fully connected LReLU (2048) Fully connected LReLU (2048) Fully connected (1)

[0021] The present invention proposes a three-dimensional bionic design method based on a deep generative model. The deep generative model can adopt one of a generative adversarial network, a creative generative solver, and an interpolation operator alone, or a combination of multiple ones can be adopted simultaneously.

[0022] The present invention proposes a three-dimensional bionic design method based on a generative adversarial network, comprising the following steps:

[0023] Step 1: Create a training set;

[0024] Step 2: Establish a deep generative model and train the deep generative model using the training set;

[0025] Step 3: Generate a three-dimensional biomimetic object using the deep generative model.

[0026] Step 4: The three-dimensional bionic object is post-processed in collaboration with the designer to obtain a three-dimensional bionic design.

[0027] Step 1 of the present invention includes: constructing an identification conversion path from a 3D file format to a voxel model format, specifically including the following steps:

[0028] Step 1.1: Use binvox to voxelize the .obj file of the 3D mesh and convert it into a .binvox file.

[0029] Step 1.2: The voxelized file is flood-filled to obtain point value pairs, which is equivalent to discrete sampling of the isosurface;

[0030] Step 1.3: Use the encoder in the implicit autoencoder to learn the mapping from 3D voxels to 128-dimensional latent vectors, and the decoder to learn the mapping from the 128-dimensional latent vector plus the 3D voxel position to the signed distance field of the voxel; the signed distance field is converted into 3D grid data through the marching cube algorithm; the signed distance field is thresholded to obtain voxel data.

[0031] Step 1 of the present invention further includes: a training set enhancement step, which includes design target-based data enhancement and performance-based data enhancement, increasing the total amount of the training set from 113+6000 to more than 20k.

[0032] The data enhancement based on design goals includes: design-oriented data enhancement focuses on transferring design knowledge to the preparation of data sets to meet design requirements. For example, in the designer's design process, the factor of direction is also taken into account. After evaluating existing bionic products and brainstorming, generative design considering multiple directions enables the present invention to explore more possibilities. Taking into account the symmetry of the target product and the trade-off between diversity and efficiency, the present invention chooses to apply rotation transformations in seven directions to the bionic object dataset to achieve diversified hybrid results.

[0033] Performance-based data augmentation includes: Performance-oriented data augmentation aims to improve the performance of deep generative models. Generators are impacted by imbalanced training data, particularly datasets with only a few hundred animals. Non-rigid deformations, such as stretching, are effective methods to alleviate this problem. This paper stretches the training dataset along three orthogonal axes, with degrees ranging from 0.25 to 2.

[0034] Step 1 of the present invention further includes: increasing the resolution of the generated voxel file from 16^3 to 32^3 to 64^3.

[0035] This includes: the model generates a signed distance field (SDF), and by controlling the sampling density of the SDF, it allows the generation of voxels of arbitrary resolution. However, the level of detail of the network fitting depends on the resolution of the voxel file used.

[0036] The training process uses a progressive approach, first training for 100 epochs at a 16^3 voxel resolution, then for 200 epochs at a 32^3 voxel resolution, and finally for 1000 epochs at a 64^3 voxel resolution. Training can be terminated during the training process to control the generation of a 3D model with multi-scale details.

[0037] In the present invention, the deep generative model includes an implicit autoencoder and a latent vector generation model; the implicit autoencoder includes an encoder 3DCNN and a decoder IM-Decoder; the latent vector generation model adopts a generative adversarial network.

[0038] Compared with traditional decoders based on convolutional neural networks, the IM-Decoder can make the latent vector space and the generated three-dimensional voxels more continuous, thereby achieving a more continuous interpolation effect.

[0039] The loss function of the IM-Decoder model depends on the sampling method, where w p is the weight of the weighted sampling method, which depends on the sparsity of the sampling points. The loss function is the weighted mean square error between the true value label and the predicted label of each point. Let S be the point set of the target shape, and the loss function L(θ) is:

[0040]

[0041] Where θ is the network weight, f θ For network mapping, is the true value of the actual signed distance field sampled at point p.

[0042] In the latent generative adversarial network (latent-GAN) framework, both the generator and discriminator consist of three fully connected layers, with the first two layers using leaky-relu activation functions and the final layer using a sigmoid function. The network is used to fit the latent vector point set obtained by the encoder from the training set through adversarial training. Ultimately, the generator can output latent vectors that follow the distribution of the training points in the latent space of the training set, taking a standard normal distribution as input.

[0043] The generative adversarial network includes a generator and a discriminator, and the generative adversarial network is used to fit the distribution of embedded points of the training set in the latent space.

[0044] The present invention establishes a generative adversarial network comprising the following steps:

[0045] Step 2.1: Train the autoencoder to learn low-dimensional representations;

[0046] Step 2.2: Train a generative adversarial network consisting of a generator and a discriminator in the latent vector space between the encoder and decoder of the autoencoder;

[0047] Step 2.3: After the training set is prepared, the autoencoder and the adversarial generative network are trained in two stages respectively.

[0048] The two-stage training refers to: the first stage uses progressive training to train the autoencoder to make the generated model have more details; the second stage uses the generator and discriminator to train the adversarial generative network in the latent vector space to make the results generated by the generator more realistic and diverse.

[0049] After the autoencoder and adversarial generation network are trained, the embedding map AE is obtained S→Z , which includes the encoder of the autoencoder and the generation mapping G Z→S , which includes a generator of a generative adversarial network and a decoder of an autoencoder.

[0050] In the present invention, the latent vector generation model can also employ a creative generative solver. The creative generative solver includes a heuristic objective function, which includes a heuristic sampling algorithm and a latent space optimization algorithm. The heuristic sampling algorithm samples the latent space. The generative model is trained using a weighted training method, which reduces inference and performs black-box optimization to obtain a latent vector generator.

[0051] The invention provides a method for establishing a creative product solver, comprising the following steps:

[0052] Step 2.1: Train the autoencoder to learn low-dimensional representations;

[0053] Step 2.2: Deploy the creative generator solver in the latent vector space between the encoder and decoder of the autoencoder;

[0054] Step 2.3: After the training set is prepared, the autoencoder and idea generation solver are trained in two stages respectively.

[0055] The two-stage training process involves: first, progressive training of the autoencoder to generate a more detailed model; and second, training of the creative generative solver. This involves sampling the latent space using a heuristic sampling method, then calculating the scores of the sampled points using a heuristic objective function. High-scoring sampled points are then used to train the kernel density estimator for non-parametric estimation. These high-scoring sampled points are then used as the evaluation set for subsequent training.

[0056] Considering the high computational overhead of the previous method, high-scoring sample points were used as part of the training set to train a generative model from the paper "Sample-Efficient Optimization in the Latent Space of Deep Generative Models via Weighted Retraining." Due to the effectiveness of weighted training, black-box optimization was ultimately achieved with minimal inference, resulting in a latent vector generator. The latent vector generator automatically inputs 32-dimensional training outputs that follow a standard multivariate Gaussian distribution and have high scores in the heuristic objective function. These sample points, after passing through the implicit decoder (IM-Decoder), generate creative artifacts that are considered high-quality.

[0057] The hybrid heuristic objective function f(x) consists of a connected component counting function and a binary classifier for distinguishing between the design target and the biomimetic object. The connected component counting function makes the objective function f(x) black box, that is, the general closed form or differential information is unknown.

[0058]

[0059] where δ is the threshold used to discretize the continuous signed distance field, N r is the voxel resolution. f θ (z,p) is the network prediction value at point p in the signed distance field obtained by decoding z, representing the set of integers.

[0060] Given the embedded representation of the bionic object and the design goal respectively and M and N are the number of training samples for the bionic object and design target. Zd and zb are just naming symbols, and k is a sequence number ranging from 1 to M or N. This is how ijk is basically used.

[0061] Train each perceptron classifier f c (z)=argmax i P(Y=i|z). The loss function is:

[0062]

[0063] in: is a latent vector, Y represents a binary random variable, ideally f c (z)∈{0,1}, the actual differentiable network value is continuous f c (z)∈[0,1]. The goal of this invention is to encourage the generation of a trade-off between the biomimetic object and the design target category. When the classifier uncertainty reaches its maximum value, percept() should reach the optimal value, so it is defined as follows:

[0064] Percept(z)=H(Z)=-f c (z)log f c (z)-(1-f c (z))log(1-f c (z))

[0065] The hybrid heuristic objective function is finally expressed as:

[0066] S(z)=Percept(z)+α1N cg (z)

[0067] Where α1 is the coefficient.

[0068] The Latent Space Optimization (LSO) algorithm maps the latent vector point set to the high value area of ​​the objective function S(z). The objective function f is evaluated as few times as possible, and finally the evaluation point sequence is obtained. The algorithm currently used is from Austin Tripp, Erik Daxberger, and José Miguel Hernández-Lobato (2020). Sample efficient optimization in the latent space of deep generative models via weighted retraining. Advances in Neural Information Processing Systems 33 (2020). This method ultimately produces a latent vector generator that, by sampling from a simple prior distribution, outputs latent vectors in the region of high values ​​of the objective function S(z).

[0069] In order to evaluate the effect of latent space optimization algorithm, based on Bayesian optimization, heuristic sampling method and kernel density estimation are used to form the true value point set.

[0070] The heuristic sampling method is a prime interval sampler. First, the range of the high-quality sample latent vector point set is narrowed down to the convex hull of the latent vector of the training dataset. Selecting several middle points from the interpolation is an acceptable creative exploration method suitable for the on-the-fly evaluation of the convex hull S. If the present invention is for each pair and Interpolate and sample z1 and z2 N times, and select the middle m points in each pair. The sample set As shown below:

[0071]

[0072] Where N represents the number of interpolation samples; λ ranges from 1 to m.

[0073] The amount of samples obtained in this way is very large, which is N interpolation =m×|D b ||D d |. In order to reduce the sampling amount, the present invention finds that sampling with a prime number interval can keep the sample set The algebraic structure of . Sample set is a cyclic group of order m, and the sampling The subset is still a cyclic group of order m, and the subset obtained by methods such as uniform sampling converges to this structure probabilistically.

[0074] The resulting sample set is then estimated using kernel density estimation with a bandwidth of 0.1. The resulting creative artifacts can then be sampled using kernel density estimation.

[0075] In the present invention, the latent vector generation model can also utilize an interpolation operator. The interpolation operator's structure is as follows: Given a biomimetic object latent vector x1 and a design target latent vector x2, the interpolation vector is expressed as x = γx1 + (1-γ)x2, where γ∈[0,1]. The corresponding interpolation latent vector is obtained by uniformly sampling γ, and the corresponding generated product is finally obtained through a decoder. The implicit field obtained by mapping the interpolated latent vector through the generator is the creative generated product, and a triangular mesh file is obtained through the marchingcube algorithm.

[0076] The present invention establishes a deep generative model comprising the following steps:

[0077] Step 2.1: Train the autoencoder to learn low-dimensional representations;

[0078] Step 2.2: Train an interpolator in the latent vector space between the encoder and decoder of the autoencoder.

[0079] Step 2.3: After the training set is prepared, the autoencoder is trained. The autoencoder is trained using progressive training to make the generated model have more details.

[0080] The interpolation operator can directly obtain a closed-form solution through the training set data without the need for training.

[0081] The present invention also provides an application of the above method in three-dimensional representation, wherein the application specifically includes the following process:

[0082] Step 1: Learn the embedding of the 3D shape through the encoder, which is converted into elements in a low-dimensional space;

[0083] Step 2: Searching in the low-dimensional space according to specific rules to find a series of elements in the low-dimensional space that integrate the semantic features of different three-dimensional shapes; the specific rules include interpolation, generating latent vectors using a generative adversarial network, and a creative generator solver;

[0084] Step 3: Use the implicit encoder-decoder to learn the mapping and remap the elements in the low-dimensional space back to the three-dimensional shape to obtain the three-dimensional shape of the object.

[0085] In the present invention, the design in the three-dimensional bionic design refers to the design obtained by optimizing the product. Both "product" and "synthetic" correspond to the English term synthetic, specifically referring to the product generated by the generator.

[0086] Generator. The generator is a neural network model used for generating tasks. In the model training phase, it is necessary to learn a latent representation space Z for the space X where the multimedia content in the given training set is located, and the Z space is a low-dimensional Euclidean space. In the model inference phase, the present invention has x=f θ (z), that is, given a latent space sampling point z, the generator can generate data of the corresponding modality.

[0087] The implicit autoencoder consists of a 3DCNN encoder and an implicit decoder, which are used to fit the signed distance field of the 3D model and reconstruct the signed distance field. After the first stage of training, the field is represented as a 3D shape dataset. The loss function is the reconstruction error of the implicit field sampling points. The detailed network structure is shown in the table.

[0088] The middle hidden layer of the autoencoder in this invention is a low-dimensional Euclidean space, and the semantics are extracted through deep learning methods. Specific method: Assume that the low-dimensional space is R n A subset of , defined as the latent vector space Z, given a sample of the design target domain Samples from the bionic object domain Assume that the distribution of the design target domain D is p(d) and the distribution of the bionic object domain B is p(b). The first step is to train the encoder AE S→Z , to learn from the sample to embed the distribution p(d) and p(b) into a subset Z in the low-dimensional latent vector space Z B With Z D In the example, we assume that their distribution is p(z b ) and p(z d ). The second step is to distribute p(z b ) and p(z d ) and obtain the potential representation space Z of the creative generation domain through searching. I , assuming its distribution is p(z i ), in p(z i) to obtain a series of samples The third step is to use the generator G of the trained generative model Z→S The sample Mapping into a three-dimensional shape domain, thus obtaining a series of three-dimensional shapes with bionic creative inspiration value.

[0089] This invention uses interpolation in latent space to synthesize novel biomimetic design forms. Related work on representational learning has demonstrated that three-dimensional shapes can be transformed by operating on latent vectors. The model trained in this invention can similarly generate three-dimensional shapes by using an encoder to generate two latent vectors z1 and z2 for two different objects. The interpolated latent vector z' = γz1 + (1-γ)z2 (0 < γ < 1) is then input into a generator. These three-dimensional shapes exhibit smooth transitions between objects as γ increases.

[0090] The generative adversarial network is used to map the normally distributed samples of N(0,1) into new sample points that fit the latent vector distribution of the training set.

[0091] Gaussian sampling is used to directly sample N(0,1) points in the latent space. The results were not as good as those of the generative adversarial network.

[0092] Select appropriate biomimetic objects and design goals from the voxel models in the input dataset. S→z Obtain latent vectors, interpolate between these latent vectors, and obtain a large number of intermediate products, which are then handed over to designers for selection. In the subsequent generation process, different angle combinations are tried to obtain a richer creative inspiration model.

[0093] Applying Gaussian distribution within the latent space of the latent-GAN model is another method for generating novel biomimetic design forms. After sampling, designers can identify biomimetic objects and design goals that demonstrate promising model learning results, preparing the model for subsequent use and creation of biomimetic products. To ensure both the generated results and the designer's acceptance, this method employs dataset expansion and directional adjustments to produce a richer and more diverse range of three-dimensional shapes and morphological details. BRIEF DESCRIPTION OF THE DRAWINGS

[0094] Figure 1 It is a schematic diagram of the process of the present invention.

[0095] Figure 2 It is a schematic diagram of the three-dimensional bionic design method based on the deep generative model of the present invention.

[0096] Figure 3 It is a flow chart of the two-stage training method of the present invention.

[0097] Figure 4It is a schematic diagram of the data representation method of four three-dimensional shapes and the conversion method between them in the present invention.

[0098] Figure 5 This is the overall framework diagram of the model call of the present invention.

[0099] Figure 6 It is a hypothetical diagram of the creative generation problem with the goal of bionic design in an embodiment of the present invention.

[0100] Figure 7 This is a diagram of the data format of the results generated by an embodiment of the present invention.

[0101] Figure 8 It is a schematic diagram of the three-dimensional form of a typical product, design inspiration points and design solution development direction of an embodiment of the present invention.

[0102] Figure 9 It is a design scheme diagram of the embodiment of the present invention completed under the inspiration of the generated object.

[0103] Figure 10 This is a comparison diagram of the morphology of some products generated by the embodiments of the present invention and the design morphology of internationally outstanding bionic children's home chairs.

[0104] Figure 11 This is a comparison chart of the design scheme inspired by some of the products in the embodiments of the present invention and the design of an internationally outstanding bionic children's home chair.

[0105] Figure 12 This is a schematic diagram of scoring the results generated by the creative product solver in some embodiments of the present invention.

[0106] Figure 13 It is a structural diagram of the creative generated object solver of the present invention. DETAILED DESCRIPTION

[0107] The present invention is further described in detail with reference to the following specific examples and accompanying drawings. The processes, conditions, experimental methods, etc. for implementing the present invention, except for those specifically mentioned below, are common knowledge and common common sense in the art and are not particularly limited by the present invention.

[0108] Example 1

[0109] The latent autoencoder can only fit the distribution of the training data, and may not necessarily produce a structure that integrates the generated objects. However, sometimes, simply reconstructing the structure involves a certain degree of design abstraction. In this embodiment, a generative adversarial network is used to obtain a three-dimensional biomimetic design. The specific process is as follows:

[0110] The generative adversarial network includes a generator and a discriminator, and the generative adversarial network is used to fit the distribution of embedded points of the training set in the latent space.

[0111] During the training phase, the generator inputs 128-dimensional Gaussian noise from a normal distribution N(0,1) and outputs a generated 128-dimensional vector. The discriminator inputs the generated vector and the latent space embeddings of the training set to classify whether the vector belongs to the training set or to the generator. The two sides undergo adversarial training. Ultimately, after training is complete, the generator can input Gaussian noise and output a latent vector that closely matches the distribution of the training set embeddings in latent space.

[0112] In the calling phase, the 128-dimensional vector sampled from the normal distribution of N(0,1) is mapped to a new sample point that fits the latent vector distribution of the training set.

[0113] Generator

[0114] layer Activation Function Output shape Input latent vector (128) Fully connected LReLU (2048) Fully connected LReLU (2048) Fully connected Sigmoid (128)

[0115] Discriminator

[0116] layer Activation Function Output shape Input latent vector (128) Fully connected LReLU (2048) Fully connected LReLU (2048) Fully connected (1)

[0117] like Figure 1 As shown in Figure 2, the latent autoencoder can only fit the distribution of the training data, and may not necessarily obtain a fusion structure between the generated products. However, sometimes, simply reconstructing the structure involves a certain degree of design abstraction.

[0118] Example 2

[0119] In this embodiment, an interpolation operator is used to obtain a three-dimensional bionic design. The specific process is as follows:

[0120] This embodiment of the invention targets children aged 3-6, uses a bionic home chair as the design target, uses the familiar dog as the bionic object, and uses the urban home interior as the application scenario. The designers involved include a Red Dot Product Design Award winner, a product designer who previously worked at an early childhood education institution, and a product designer with a 5-year-old child. The computer engineers are a team of graduate students and faculty from the School of Computer Science at East China Normal University.

[0121] Based on work, designers and computer engineers jointly built a bionic design resource library. The open source model library of chairs was used in training, and simple, undecorated armchairs were used to prepare for the training of the generative model. Dog models with obvious features and consistent postures were selected to build a bionic modeling library.

[0122] This project trained a predefined 3D deep learning model using interpolation calculations multiple times. Each generated object consisted of three views and a model file. After nine rounds of training and deployment, over 15,000 generated objects were generated for deployment, evaluation, and screening. Designers and computer engineers continuously refined and advanced the model in the following five areas, ensuring that subsequent generated objects possessed even greater biomimetic design value.

[0123] The first stage is the 3D deep learning generated product call and optimization measure stage, as shown in the following table:

[0124] Table 1. 3D deep learning generated calls and optimization measures

[0125]

[0126] Deep learning-generated artifacts include both 2D contours and 3D form files. The biomimetic morphological value of these artifacts is assessed based on two aspects: creative value and model quality. Designers quickly select artifacts with high biomimetic design value based on criteria such as whether the artifacts embody the key features of the original biomimetic object and the compatibility of the generated morphological structure with the biomimetic object and product structure. Algorithm engineers evaluate the artifacts based on the performance of deep learning, identifying anomalies such as distortion, fractures, and missing data, and their causes. They also observe and evaluate the impact of each interpolation, parameter adjustment, and dataset on the generated results. After multiple iterations of adjustments and screening during the generation process, approximately five types of artifacts with high biomimetic morphological value were identified, which can be further optimized into design solutions. Both the generated 3D forms and 2D contours provide valuable insights for designers.

[0127] The second stage involves further optimizing 3D artifacts with high biomimetic morphological value into design proposals. The designer's work at this stage is divided into two parts: 1. Selecting artifacts with high biomimetic design morphological value and optimizing their morphology to form a design proposal; 2. Inspired by one or more artifacts, the designer directly conducts secondary creative thinking to develop a new design proposal. The designer's sensitivity to biomimetic morphology, morphological manipulation techniques, and product design experience play a key role in this process. Computer engineers need to fully understand the designer's selection criteria and assist them in the screening process, as well as in converting artifact formats and models.

[0128] Designers adjusted and optimized the generated model based on product structure and functional design to form multiple design plans for children's home chairs.

[0129] In order to evaluate the value of deep generative models in the creative generation of three-dimensional bionic forms, the present invention compares the generated forms selected by designers in this work with the forms of well-known bionic product designs. Figure 10 shown.

[0130] Comparison revealed that, aside from issues such as surface roughness and incomplete local morphology, the resulting morphology closely resembles most existing biomimetic product designs, conforming to the principles of abstract, intentional, and figurative bionics. The resulting morphology also closely resembles the biomimetic object in terms of posture, dynamics, and typical animal characteristics. The resulting morphology is rich, encompassing virtually every type of creative form variation currently available in biomimetic design.

[0131] In the embodiment of the present invention, three product designers observed the characteristics of the generated product and integrated their own ideas to form 10 design schemes in a short period of time, as well as multiple design idea sketches, such as Figure 11 As shown in the figure, the distribution of the solutions reveals that multiple generated objects offer inspiration to designers, and after morphological optimization, they can form design solutions. From the perspective of the creative inspiration points of the solutions, designers can draw inspiration from a single generated object, or combine the creative points of two or more generated objects to form new creative forms. This process demonstrates that generated objects can effectively inspire designers and assist them in forming design solutions. In contrast, international designers' designs excel in morphological manipulation techniques. Designers also drew inspiration from the outlines of the generated objects, which helped them form their design solutions.

[0132] Example 3

[0133] like Figure 12 As shown in the figure, from left to right and from top to bottom, the scores of the creative generative solver's generated results range from high to low. High-scoring generated objects often integrate the morphological features of the bionic object and the design target. Meanwhile, intermediate-scoring generated objects merely fit the distribution of the training data and may not necessarily achieve a structure that integrates the generated objects. Low-scoring generated objects often exhibit significant fragmentation, affecting product functionality.

[0134] In this embodiment, a creative object solver is used to obtain a three-dimensional bionic design. The specific process is as follows:

[0135] First, the latent space is sampled using a heuristic sampling method, and then the scores of the sampling points are calculated using a heuristic objective function. The high-scoring sampling point set is used to train the kernel density estimate of the non-parametric estimation. At the same time, the high-scoring sampling point set is used as the set for subsequent evaluation.

[0136] Considering the high computational overhead of the previous method, high-scoring sample points were used as part of the training set to train a generative model from the paper "Sample-Efficient Optimization in the Latent Space of Deep Generative Models via Weighted Retraining." Due to the effectiveness of weighted training, black-box optimization was ultimately achieved with minimal inference, resulting in a latent vector generator. The latent vector generator automatically inputs 32-dimensional training outputs that follow a standard multivariate Gaussian distribution and have high scores in the heuristic objective function. These sample points, after passing through the implicit decoder (IM-Decoder), generate creative artifacts that are considered high-quality.

[0137] Example 4

[0138] The present invention proposes a method for generating three-dimensional biomimetic objects based on a deep generative model. The deep generative model can employ a single generative adversarial network, a creative object solver, or an interpolation operator, or a combination of multiple models. When multiple models are employed simultaneously, the accuracy of the three-dimensional biomimetic design is improved.

[0139] The protection content of the present invention is not limited to the above embodiments. Without departing from the spirit and scope of the present invention, changes and advantages that can be thought of by those skilled in the art are included in the present invention and are protected by the appended claims.

Claims

1. A design system for three-dimensional biomimetic objects, characterized in that: The system adopts a three-dimensional bionic design method based on a generative adversarial network, and the method includes the following steps: Step 1: Create a training set; Step 2: Establish a deep generative model and train the deep generative model using the training set; the deep generative model includes an implicit autoencoder and a latent vector generative model; the implicit autoencoder includes an encoder 3DCNN and a decoder IM-Decoder; the latent vector generative model uses a generative adversarial network; Building a deep generative model involves the following steps: Step 2.1: Train the autoencoder to learn low-dimensional representations; Step 2.2: Train a generative adversarial network consisting of a generator and a discriminator in the latent vector space between the encoder and decoder of the autoencoder; Step 2.3: After the training set is prepared, the autoencoder and the generative adversarial network are trained in two stages respectively; The two-stage training involves: in the first stage, the autoencoder is trained using progressive training to make the generated model more detailed; in the second stage, the adversarial generative network in the latent vector space is trained by the generator and the discriminator to make the results generated by the generator more realistic and diverse; Step 3: Generate a three-dimensional bionic object using the deep generative model; The system comprises: The data preprocessing module is used to input the 3D shape dataset represented by the bionic object and design target, expand the dataset through design target-based and performance-based data augmentation, and obtain the signed distance field of the 3D model required for model training through mesh2voxel and SDF sampler; A deep generative model comprising an implicit autoencoder and a generative adversarial network; the implicit autoencoder comprises an encoder 3DCNN and a decoder IM-Decoder; The sample vector repository is used to store latent vector point sets. The latent vectors in the repository are used to obtain three-dimensional biomimetic generated objects through an implicit decoder.

2. The design system for three-dimensional biomimetic objects according to claim 1, wherein: The system further comprises a post-processing module, which is used to obtain a three-dimensional bionic design by post-processing the three-dimensional bionic object.

3. The design system for three-dimensional biomimetic objects according to claim 2, wherein: The implicit autoencoder is used to fit the signed distance field of the three-dimensional model to perform the first stage of training. In order to obtain a high-fidelity generated product that meets the design requirements, it is necessary to search for a suitable latent vector point set in the latent space. The latent vector point set is obtained by Gaussian sampling after training a generative adversarial network. After the first stage of codec network training is completed, the model to be trained undergoes a second stage of training to obtain a callable latent vector generation model.

4. The design system for three-dimensional biomimetic objects according to claim 1, wherein: The method further comprises step 4: subjecting the three-dimensional bionic object to post-processing in collaboration with a designer to obtain a three-dimensional bionic design.

5. The design system for three-dimensional biomimetic objects according to claim 1, wherein: Step 1 includes: constructing an identification conversion path from a 3D file format to a voxel model format, specifically including the following steps: Step 1.1: Use binvox to voxelize the .obj file of the 3D mesh and convert it into a .binvox file. Step 1.2: The voxelized file is flood-filled to obtain point value pairs, which is equivalent to discrete sampling of the isosurface; Step 1.3: Use the 3DCNN encoder in the implicit autoencoder to learn the mapping from 3D voxels to 128-dimensional latent vectors, and the decoder learns the mapping from the 128-dimensional latent vector plus the 3D voxel position to the signed distance field of the voxel; the signed distance field is converted into 3D grid data through the marching cube algorithm; the signed distance field is thresholded to obtain voxel data.

6. The design system for three-dimensional biomimetic objects according to claim 1, wherein: Said step 1 further comprises: a training set enhancement step, which includes design target-based data enhancement and performance-based data enhancement; The data enhancement based on the design goal includes: applying rotation transformation in seven directions to the bionic object dataset to achieve diversified hybrid results; The performance-based data augmentation includes stretching the training dataset on three orthogonal axes with a degree ranging from 0.25 to 2.

7. The design system for three-dimensional biomimetic objects according to claim 1, wherein: The step 1 further includes: improving the resolution of the generated voxel file, which includes: generating a signed distance field by the model, and allowing the generation of voxels of arbitrary resolution by controlling the sampling density of the signed distance field; During training, progressive training is used. First, 100 rounds of training are performed on the 16^3 resolution voxel file, then 200 rounds of training are performed on the 32^3 resolution voxel file, and finally 1000 rounds of training are performed on the 64^3 resolution voxel file. The generation of 3D models with multi-scale details can be controlled by terminating the training process.

8. The design system for three-dimensional biomimetic objects according to claim 1, wherein: The generative adversarial network includes a generator and a discriminator, and the generative adversarial network is used to fit the distribution of embedded points of the training set in the latent space.

9. The design system for three-dimensional biomimetic objects according to claim 8, wherein: After the autoencoder and adversarial generation network are trained, the embedding map AE is obtained S→Z , which includes the encoder of the autoencoder and the generation mapping G Z→S , which includes a generator of a generative adversarial network and a decoder of an autoencoder.

10. The design system for three-dimensional biomimetic objects according to claim 1, wherein: The loss function of the IM-Decoder model depends on the sampling method; the loss function is the weighted mean square error between the true value label and the predicted label of each point. Let S be the point set of the target shape, and the loss function L(θ) is: Among them, θ is the network weight, f θ For network mapping, is the true value of the actual signed distance field sampled at point p; w p is the weight of the weighted sampling method, which depends on the sparsity of the sampling points.

11. The design system for three-dimensional biomimetic objects according to claim 1, wherein: The generative adversarial network includes a generator and a discriminator. Both the generator and the discriminator contain three fully connected layers, wherein the activation functions of the first two layers are leaky-relu and the last layer is a sigmoid function. The generative adversarial network is used to fit the latent vector point set obtained by the encoder of the training set through adversarial training. Finally, the generator outputs a latent vector that obeys the distribution of the training point set in the latent space of the training set by inputting a standard normal distribution.

Images