Airfoil profile generative design method based on implicit diffusion model
Patent Information
- Application Number
- CN202510169290.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-02-17
- Publication Date
- 2025-07-01
Smart Images
Figure CN120234890A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of airfoil design, and specifically to a generative design method for airfoils based on implicit diffusion models. Background Art
[0002] Discriminative models and generative models are two main types of models in deep learning. Among them, discriminative models were the first to be developed. They attempt to learn from given samples and output the labels corresponding to the samples. Generative models, on the other hand, can directly generate samples based on real samples or generate predicted samples from labels. Generative models learn more data features than discriminative models. In fact, they learn the features of the input data distribution. At the same time, their naturally scalable generation process provides a basis for more powerful and complex applications. Diffusion models, especially denoising diffusion probabilistic models (DDPMs), have achieved excellent results as generative models and have outperformed state-of-the-art generative adversarial network (GAN)-type generative models on multiple different metrics and datasets. They are also easy to scale and train. Peebles et al. replaced the backbone architecture of the diffusion model from the UNet architecture with the Transformer architecture and proposed the Diffusion Transformer (DiT), achieving advanced sample quality on the class-conditional 256×256 ImageNet generation benchmark. In recent years, generative models, especially diffusion models, have shown excellent capabilities in generating high-quality samples. The exploration of tasks such as text-to-image and text-to-video in the field of artificial intelligence has promoted the research on conditional generation.
[0003] Airfoil inverse design is a typical method for aircraft airfoil design. Its task is to generate an aircraft shape that meets given physical or geometric characteristics. Therefore, inverse design can be regarded as a type of conditional generation. Currently, there are studies and proposed deep learning-based airfoil inverse design methods, but they only perform inverse design based on basic aerodynamic characteristic requirements. Moreover, since an airfoil is characterized by a large amount of coordinate point data, when performing inverse design, the airfoil needs to be parameterized, and the parameterized feature data is used for design. However, traditional airfoil parameterization methods have reconstruction errors and inductive biases in principle and are not suitable for increasingly complex multi-disciplinary and multi-objective airfoil designs. Summary of the Invention
[0004] Aiming at the problems existing in the prior art, the present invention proposes a generative airfoil design method based on the latent diffusion model. For the multidisciplinary and multi-objective airfoil inverse design task that combines aerodynamic performance and low detectability performance, a diffusion model based entirely on the Transformer architecture and applied to the latent space is constructed, and an airfoil parameterization method based on the autoencoder model is used to realize the parametric characterization of the airfoil, improve the airfoil parameterization accuracy, and achieve high-quality and diverse conditional airfoil generation.
[0005] The technical solution of the present invention is as follows:
[0006] The generative airfoil design method based on the latent diffusion model includes the following steps:
[0007] Step 1: Perform data structure unification processing on the existing airfoil point cloud dataset and expand it to obtain a multi-source airfoil point cloud dataset;
[0008] Step 2: Calculate the performance of each airfoil in the multi-source airfoil point cloud dataset obtained in Step 1, including aerodynamic performance and low detectability performance;
[0009] Step 3: Parametrize each airfoil in the multi-source airfoil point cloud dataset obtained in Step 1 using the trained airfoil autoencoder model, thereby reducing the dimensionality of the airfoil data and obtaining the latent representation of each airfoil;
[0010] Step 4: Based on the performance calculation results of each airfoil in the multi-source airfoil point cloud dataset obtained in Step 2 and the latent representation of each airfoil obtained in Step 3, construct airfoil-condition sample pairs;
[0011] Step 5: Establish a latent diffusion model, use the airfoil-condition sample pairs obtained in Step 4 as samples to train the latent diffusion model. During training, the input is the latent representation of the airfoil and the corresponding conditional vector label;
[0012] Step 6: Given a conditional vector, input it into the trained latent diffusion model to obtain the latent representation of the airfoil that satisfies the conditional vector, and then use the airfoil autoencoder to obtain the airfoil point cloud data.
[0013] Further, in Step 1, the data structure unification processing refers to: adjusting each initial airfoil point cloud data in the existing airfoil point cloud dataset so that the total number of coordinate points used to represent the airfoil in the initial airfoil point cloud data is the same.
[0014] Further, the specific process of achieving data structure consistency is as follows: Set the number of coordinate points representing the airfoil after shaping to N. For the existing airfoil point cloud dataset, if the number of coordinate points of a certain initial airfoil point cloud data is less than N, then expand the number of coordinate points of this airfoil point cloud data to N through interpolation. If the number of coordinate points of a certain initial airfoil point cloud data is greater than N, then reduce the number of coordinate points of this airfoil point cloud data to N through sampling.
[0015] Further, in step 1, the process of expanding the existing airfoil point cloud dataset is as follows: For a certain initial airfoil point cloud data, parameterize it using traditional parameterization methods, and then perturb the airfoil design parameters after parameterization within the set perturbation range.
[0016] Further, the traditional parameterization methods include CST parameterization method and B - Spline parameterization method.
[0017] Further, in step 2, for aerodynamic performance, the aerodynamic forces and moment coefficients of the airfoil under several flow conditions are calculated. The flow conditions are characterized by Reynolds number Re, Mach number Ma, and angle of attack α. The aerodynamic forces and moment coefficients include lift coefficient C l 、drag coefficient C d and moment coefficient C m , and the drag coefficient is further divided into viscous drag coefficient C df and form drag coefficient C dp ; For low - detectability performance, the forward radar cross - section area at the set radar frequency is calculated.
[0018] Further, the airfoil auto - encoder model includes an encoder and a decoder;
[0019] The encoder performs dimensionality reduction on the airfoil point cloud data to obtain the latent representation of the airfoil point cloud data. The encoder includes a serialization processing module, a stacked ViT module, and a convolutional module. The serialization processing module serializes the input airfoil point cloud data to obtain serialized data. Using the serialized data as the input of the stacked ViT module, extract the features of the airfoil point cloud data through the stacked ViT module, and input the features of the airfoil point cloud data into the convolutional module for dimensionality reduction to obtain the latent representation of the airfoil point cloud data;
[0020] The decoder reconstructs the latent representation to obtain airfoil point cloud data; the decoder includes a transposed convolution module, a stacked ViT module, a deserialization module, and a Softmax module; the transposed convolution module upsamples the input latent representation, then extracts serialized feature data through the stacked ViT module, and uses the deserialization module to map the serialized feature data back to the original airfoil point cloud data structure; finally, the Softmax module reconstructs the mapped data into airfoil point cloud data.
[0021] Further, the Softmax module in the decoder performs the following operations: first, separate the channel data representing the x coordinate and the channel data representing the y coordinate from the data output by the deserialization module; then perform a one-dimensional convolution operation on the channel data of the x coordinate; then divide the data after the one-dimensional convolution operation into two parts with equal numbers of data points along the x coordinate dimension, and perform a Softmax operation on each part to obtain the x coordinate spacing between adjacent data points; according to the x coordinate spacing between adjacent data points, calculate the x coordinate of the reconstructed data points, and combine it with the channel data representing the y coordinate in the data output by the deserialization module to obtain the reconstructed airfoil point cloud data.
[0022] Further, in the airfoil-condition sample pairs constructed in step 4, the airfoil is described by a latent representation, and the conditional sample is described by a conditional vector label, and the conditional vector label c is
[0023] c = [Ma, α, Re, L / D, C l , C d , C df , C dp , C m , A, T max , T TE , σ 120°
[0024] where each parameter is: lift coefficient C l , drag coefficient C d , moment coefficient C m , viscous drag coefficient C df , form drag coefficient C dp , average radar cross section σ within the first 120° forward range 120° , flow condition Reynolds number Re, Mach number Ma, angle of attack, lift-to-drag ratio L / D, airfoil cross-sectional area A, maximum thickness T max and trailing edge thickness T TE .
[0025] Beneficial effects
[0026] The present invention proposes a generative airfoil design method based on a latent diffusion model. For the multidisciplinary and multi-objective airfoil inverse design task, with flow field parameters, airfoil aerodynamic characteristics, low detectability, and geometric constraints as conditional labels, a diffusion model based entirely on the Transformer architecture and applied to the latent space is constructed, and an airfoil parameterization method based on an autoencoder model is used to achieve airfoil parameterization representation, improve the airfoil parameterization accuracy, and realize high-quality and diverse conditional airfoil generation.
[0027] Additional aspects and advantages of the present invention will be given in part in the following description, become apparent in part from the following description, or be learned through the practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS
[0028] The above and / or additional aspects and advantages of the present invention will become apparent and be readily understood from the following description of the embodiments in conjunction with the accompanying drawings, in which:
[0029] Figure 1 is a flowchart of the aerodynamic stealth generative design method according to an embodiment of the present invention
[0030] Figure 2 is a range graph of the calculated average RCS according to an embodiment of the present invention (represented by the blue arc in the figure)
[0031] Figure 3 is a randomly generated airfoil graph according to an embodiment of the present invention
[0032] Figure 4 is a result graph after t-SNE dimensionality reduction according to an embodiment of the present invention
[0033] Figure 5 is an airfoil graph obtained by interpolation sampling according to an embodiment of the present invention
[0034] Figure 6 is the input label c and the true label c according to an embodiment of the present invention * MAE distribution graph between
[0035] Figure 7 is an overlay display graph of 1024 airfoils generated under fixed conditions according to an embodiment of the present invention DETAILED DESCRIPTION OF THE EMBODIMENTS
[0036] Embodiments of the present invention will be described in detail below. The embodiments are exemplary and are intended to explain the present invention, and should not be construed as limiting the present invention.
[0037] In this embodiment, a generative airfoil design method based on a latent diffusion model proposed by the present invention is used for the multidisciplinary and multi-objective airfoil inverse design for aerodynamic performance and low detectability performance, as Figure 1 shown, including the following steps:
[0038] Step 1: Uniformize the data structure of the existing airfoil point cloud dataset and expand it to obtain a multi-source airfoil point cloud dataset.
[0039] The data structure uniformization process refers to: adjusting each initial airfoil point cloud data in the existing airfoil point cloud dataset to make the total number of coordinate points used to represent the airfoil in the initial airfoil point cloud data the same. The specific processing procedure is as follows: Set the number of coordinate points representing the airfoil after shaping to N. For the existing airfoil point cloud dataset, if the number of coordinate points of a certain initial airfoil point cloud data is less than N, then expand the number of coordinate points of this airfoil point cloud data to N through interpolation; if the number of coordinate points of a certain initial airfoil point cloud data is greater than N, then reduce the number of coordinate points of this airfoil point cloud data to N through sampling. In this embodiment, set the number of coordinate points representing the airfoil after shaping to 256, and each coordinate point has two coordinate values of x and y, so each initial airfoil point cloud data has 512 coordinate values.
[0040] The process of expanding the existing airfoil point cloud dataset is as follows: For a certain initial airfoil point cloud data, parameterize it using traditional parameterization methods, and then perturb the airfoil design parameters after parameterization within the set perturbation range. In this embodiment, for a certain initial airfoil point cloud data, parameterize it using the CST parameterization method and the B-Spline parameterization method respectively, and then perturb the airfoil design parameters corresponding to each parameterization method to obtain the perturbed airfoil point cloud data respectively. To ensure that the perturbed airfoil meets the requirements, the set perturbation range is 0.005 ≤ Δw ≤ 0.005, where Δw is the perturbation amount; in this embodiment, there are 906 airfoils in the existing airfoil point cloud dataset. The existing airfoil point cloud dataset plus the airfoil point cloud data obtained after perturbation, the finally obtained multi-source airfoil point cloud dataset includes 5000 airfoils, and each airfoil in the multi-source airfoil point cloud dataset is represented by airfoil point cloud data (i.e., x, y coordinates).
[0041] Step 2: Calculate the performance of each airfoil in the multi-source airfoil point cloud dataset obtained in Step 1, including aerodynamic performance and low detectability performance.
[0042] In this embodiment, for the aerodynamic performance, calculate the aerodynamic force and moment coefficients of the airfoil under 20 flow conditions, where the flow conditions are characterized by the Reynolds number Re, Mach number Ma, and angle of attack α. The aerodynamic force and moment coefficients include the lift coefficient C l , drag coefficient C d and moment coefficient C m . Further divide the drag coefficient into viscous drag coefficient C df and form drag coefficient C dp . For the low detectability performance, calculate the forward radar cross-section area at a radar frequency of 9 GHz.
[0043] Step 3: Parametrize each airfoil in the multi-source airfoil point cloud dataset obtained in Step 1 using the trained airfoil autoencoder model, thereby reducing the dimensionality of the airfoil data and obtaining the latent representation of each airfoil. The low-dimensional space obtained by dimensionality reduction is called the latent space, and subsequently, a diffusion model is applied in the latent space to perform the training and generation processes.
[0044] The airfoil autoencoder model includes an encoder and a decoder.
[0045] The encoder performs dimensionality reduction on the airfoil point cloud data to obtain the latent representation of the airfoil point cloud data. The encoder includes a serialization processing module, a stacked ViT (Visual Transformer) module, and a convolutional module. The serialization processing module serializes the input airfoil point cloud data to obtain serialized data; using the serialized data as the input to the stacked ViT module, the airfoil point cloud data features are extracted by the stacked ViT module, and the airfoil point cloud data features are input into the convolutional module for dimensionality reduction to obtain the latent representation of the airfoil point cloud data.
[0046] The serialization process treats the airfoil point cloud data as image data. When serializing the image data, for a multi-channel image, its storage structure in the computer is a three-dimensional array, and the three dimensions are: channel (size denoted as C), length (size denoted as H), and width (size denoted as W). The serialization process for the image data is to equally divide the image at equal intervals in the length and width dimensions. Assuming the distances in the length and width dimensions are both P, then the original image can be divided into a total of L = (H / P)·(W / P) = HW / P 2 sub-images, and each sub-image has a dimension Subsequently, all the pixels in each sub-image are then unfolded into a one-dimensional array, and its size is E = C·P 2 , in this way, the image has completed the following mapping:
[0047]
[0048] The finally obtained serialized data is X = [x ＜1＞T , x ＜2＞T ,..., x ＜L＞T , its sequence length is L, and each element x ＜i＞ is a one-dimensional array formed by unfolding the sub-image,
[0049] In the present invention, the airfoil point cloud data is regarded as a single-channel three-dimensional array, and the three dimensions are: channel, length, and width; where the size C of the channel dimension is 1, the size H of the length dimension is 256, which is the number of coordinate points after shaping in step 1, and the size W of the width dimension is 2, indicating that each coordinate point has two coordinate values, x and y. The specific dimension is 1×256×2. Since the size of the width dimension is very small at this time, in this embodiment, only equidistant segmentation is performed in the length dimension, which is equivalent to dividing every P coordinate points. Therefore, the sequence length corresponding to the airfoil data is L = 256 / P, and the embedding dimension is E = 1·2P = 2P. Since this is single-channel data and the size of the width dimension is only 2, the value of the embedding dimension E is actually small. In order to use the multi-head attention mechanism in practice, in this embodiment, the embedding dimension is mapped to a higher hidden dimension D after serialization. Therefore, the airfoil point cloud data has completed the following mapping:
[0050]
[0051] The stacked ViT modules are composed of several ViT Blocks connected in sequence, and each ViT Block is composed of a multi-head attention module connected to an FFN module.
[0052] The decoder reconstructs the latent representation to obtain the airfoil point cloud data. The decoder includes a transposed convolution module, a stacked ViT module, an Unpatchify module, and a Softmax module. The transposed convolution module upsamples the input latent representation, then extracts serialized feature data through the stacked ViT module, and uses the Unpatchify module to map the serialized feature data back to the original airfoil point cloud data structure; finally, the Softmax module reconstructs the mapped data into the airfoil point cloud data.
[0053] The Softmax module performs the following operations: First, separate the channel data representing the x coordinate and the channel data representing the y coordinate in the data output by the deserialization module; then perform a one-dimensional convolution operation on the channel data of the x coordinate; then divide the data after the one-dimensional convolution operation into two parts with equal numbers of data points along the x coordinate dimension, and perform a Softmax operation on each part to obtain the x coordinate spacing between adjacent data points; since the x coordinate value of the data point ranges from 0 to 1, according to the x coordinate spacing between adjacent data points, calculate the x coordinate of the reconstructed data point, and combine it with the channel data representing the y coordinate in the data output by the deserialization module to obtain the reconstructed airfoil point cloud data.
[0054] The input of the airfoil autoencoder model is airfoil point cloud data, and the output of the autoencoder is the reconstructed airfoil point cloud data. The airfoil autoencoder is trained using a multi-source airfoil point cloud dataset to obtain a trained airfoil autoencoder.
[0055] The loss function during the training of the airfoil autoencoder is specifically:
[0056]
[0057] where \(z\) is the input airfoil point cloud data, is the reconstructed airfoil point cloud data, and the first term of the loss function represents the mean square error between the input airfoil point cloud data and the reconstructed airfoil point cloud data; \(\lambda\) is a set weight; the second term of the loss function represents the smoothness of the reconstructed airfoil data, is a characteristic quantity representing the smoothness of the reconstructed airfoil point cloud data. In this embodiment, the characteristic quantity is taken as the average value of the absolute value of the third-order discrete difference along the \(y\)-coordinate dimension .
[0058]
[0059] During optimization, the AdamW optimizer is used, and the learning rate is fixed at \(1e - 5\) during the training process of the model. The airfoil autoencoder is trained for approximately 1.2 million steps, and the global batch size is 1024.
[0060] Each airfoil in the multi-source airfoil point cloud dataset is input into the trained airfoil autoencoder model for airfoil parameterization to obtain the latent representation of each airfoil.
[0061] Step 4: Based on the performance calculation results of each airfoil in the multi-source airfoil point cloud dataset obtained in Step 2 and the latent representation of each airfoil obtained in Step 3, construct airfoil-condition sample pairs.
[0062] In this embodiment, there are 5000 airfoils in the multi-source airfoil point cloud dataset, and each airfoil corresponds to 20 flow conditions during performance calculation. Therefore, a total of 100,000 airfoil-condition sample pairs are finally obtained. The conditional vector labels among them include the lift coefficient \(C_{l}\), l drag coefficient \(C_{d}\), d moment coefficient \(C_{m}\), m viscous drag coefficient \(C_{vf}\), df form drag coefficient \(C_{ff}\), dp average radar cross-sectional area \(\sigma\) within the first \(120^{\circ}\) forward, 120° flow conditions Reynolds number \(Re\), Mach number \(Ma\), and angle of attack \(\alpha\), as well as lift-to-drag ratio \(L / D\), airfoil cross-sectional area \(A\), maximum thickness \(T\), max and trailing edge thickness \(T_{e}\). TE, a total of 13 parameters:
[0063] c = [Ma, α, Re, L / D, C l , C d , C df , C dp , C m , A, T max , T TE , σ 120°
[0064] Step 5: Establish a latent diffusion model. Using the airfoil-condition sample pairs obtained in Step 4 as samples, train the latent diffusion model. During training, the input is the latent representation of the airfoil and the corresponding conditional vector label. The latent diffusion model includes a forward process and a reverse process, and the DiT architecture is selected for model implementation.
[0065] In this embodiment, the forward process and the reverse process included in the diffusion model in Step 4 are specifically as follows: In the forward process, Gaussian noise is gradually added to the original data x0, and finally it is transformed into Gaussian noise x T . The forward process follows the probability density function:
[0066] q(x t |x0) = N(x t-1 ; μ θ (x t , t), ∑ θ (x t , t))
[0067] In the diffusion model, the neural network is trained to model μ θ and Σ θ in the above formula. The goal of the reverse process is to maximize the log-likelihood value of x0. However, directly optimizing the log-likelihood value is quite difficult. Therefore, instead, optimize the variational lower bound of the log-likelihood, and the loss function is:
[0068]
[0069] If Σ θ is regarded as a hyperparameter, and this model is only used to learn μ θ , then the reparameterization trick can be used to simplify the loss function:
[0070]
[0071] Here, the constructed model is used to predict the true Gaussian noise ∈ t , where ∈ θ (x t ) represents the predicted Gaussian noise. To make the model learn μ θ and Σ θ , training ∈ using a simplified loss function θ , training Σ using a loss function based on the variational lower bound of the original log-likelihood θ .
[0072] Once p θ is trained, the reverse process can use the reparameterization trick starting from Gaussian noise to iteratively sample x t-1 ~p θ (x t-1 |x t ), and finally generate the sample x0.
[0073] Step 6: Given the conditional vector, input it into the trained latent diffusion model to obtain the airfoil latent representation that satisfies the conditional vector, and then use the airfoil autoencoder to obtain the airfoil point cloud data.
[0074] Performance test:
[0075] 1. In the DiT architecture, a set of DiT blocks are used to process the input data. This architecture uses classifier-free guidance for conditional generation, which involves randomly removing the conditional information of the input samples and replacing the original condition with a special condition . This enables the model to distinguish between unconditional and conditional information. In this embodiment, the special condition is a fixed vector with all 13 elements set to zero. To verify the effectiveness of the fixed vector and whether the model can distinguish between unconditional and conditional information, 4000 samples are randomly selected from the multi-source airfoil point cloud dataset in this embodiment, and the trained latent diffusion model is used with the fixed vector as the input to generate 4000 airfoils. Observing the 4000 samples and the generated 4000 airfoils, they have a similar distribution structure, as Figure 4 shown, indicating the effectiveness of the fixed vector as a special label.
[0076] 2. Perform linear interpolation between two input conditions, and then use the interpolated condition to guide the trained latent diffusion model for generation during the sampling process. If the model effectively captures a robust representation of the data distribution, then theoretically it can be expected that it can generate high-quality outputs between the two original input conditions to conform to the interpolated condition. Given the input conditions c1 and c2, the interpolated condition is expressed as:
[0077] c * =(1 - w)·c1 + w·c2
[0078] where w is the interpolation coefficient, indicating the degree to which the interpolation condition favors any input condition. In this embodiment, five interpolation coefficients are selected: w = 0, 0.25, 0.5, 0.75, 1.0. When w = 0, the interpolation condition degenerates to c1, and when w = 1, it degenerates to c2. The interpolation sampling results are as Figure 5 shown, where each color represents a different set of interpolation results. It can be observed that within each set, the airfoil shape transitions smoothly from left to right, indicating that the model has successfully learned a well-structured latent space with good continuity.
[0079] 3. In this embodiment, the performance evaluation of the latent diffusion model is based on the degree of compliance of the generated samples with the given conditions. By calculating the true physical characteristics of the airfoils generated under condition c on the test set to evaluate this, and then calculating the average absolute error (MAE) between and c to measure the difference. Except for the three flow condition parameters, the MAE distribution of each component of the condition label vector is as Figure 6 shown, and the overall relative error on the test set is shown in the brackets below each subplot in Figure 6 This error is calculated using the following formula to avoid division by zero:
[0080]
[0081] where i is the index of the condition in c and j is the sample index.
[0082] As Figure 6 shown, under condition c, for each characteristic c i , the generated samples of the model match most of the test samples. Except for the trailing edge thickness T TE , the overall relative error of other characteristics remains within 7%. However, when the thickness of the trailing edge changes slightly, it has no significant impact on the physical characteristics of the airfoil. Generally speaking, as long as the trailing edge thickness meets the minimum engineering requirements, its importance is relatively low. The above results show that the model of the present invention performs well in airfoil generation under multi-disciplinary and multi-index conditions.
[0083] In this embodiment, the model is conditioned on 13 characteristics, which imposes relatively strong restrictions on sample generation. This embodiment explores the characteristics of the generated results of the model under these conditions. To study the characteristics of the model's generated results, it is necessary to ensure that the model's generated results follow the given conditions as much as possible. Otherwise, the diversity of the generated results may come from insufficient compliance with the conditions rather than the inherent diversity of the model. Therefore, four condition labels are selected from the test set, which can generate samples that best fit the given conditions. For each condition, 1024 random generations are performed, and the results are as Figure 7As shown, all the generated results corresponding to each condition are superimposed and plotted in the same subfigure. It can be observed that the airfoil shapes in each subfigure almost completely overlap, showing very similar geometric features. This indicates that the conditions defined by 13 features effectively determine a unique solution.
[0084] In summary, the airfoil generative design method based on the implicit diffusion model proposed by the present invention has high design accuracy when considering multiple index requirements such as aerodynamics, low detectability, and geometric constraints for airfoil design.
[0085] Although the embodiments of the present invention have been shown and described above, it can be understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those of ordinary skill in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of the present invention without departing from the principles and purposes of the present invention.
Claims
1. A generative airfoil design method based on an implicit diffusion model, characterized in that: The following steps are involved: Step 1: Perform data structure consistency processing on the existing airfoil point cloud dataset and expand it to obtain a multi-source airfoil point cloud dataset; Step 2: Calculate the performance of each airfoil in the multi-source airfoil point cloud data set obtained in step 1, including aerodynamic performance and low detectable performance; Step 3: Parameterize each airfoil in the multi-source airfoil point cloud data set obtained in step 1 using the trained airfoil autoencoder model, thereby reducing the dimension of the airfoil data and obtaining the potential representation of each airfoil; Step 4: Based on the performance calculation results of each airfoil in the multi-source airfoil point cloud dataset obtained in step 2 and the potential representation of each airfoil obtained in step 3, an airfoil-condition sample pair is constructed; Step 5: Establish a latent diffusion model, and use the airfoil-condition sample pairs obtained in step 4 as samples to train the latent diffusion model. During the training, the input is the potential representation of the airfoil and the corresponding conditional vector label; Step 6: Given a conditional vector, input it into the trained latent diffusion model to obtain the airfoil potential representation that satisfies the conditional vector, and then use the airfoil autoencoder to obtain the airfoil point cloud data.
2. The airfoil generative design method based on an implicit diffusion model according to claim 1, characterized in that: In step 1, the data structure consistency processing refers to: adjusting each initial airfoil point cloud data in the existing airfoil point cloud data set so that the total number of coordinate points used to characterize the airfoil in the initial airfoil point cloud data is the same.
3. The airfoil generative design method based on implicit diffusion model according to claim 2, characterized in that: The specific process of achieving data structure consistency is as follows: set the number of coordinate points representing the airfoil after shaping to N. For the existing airfoil point cloud data set, if the number of coordinate points of a certain initial airfoil point cloud data is less than N, then the number of coordinate points of the airfoil point cloud data is expanded to N through interpolation; if the number of coordinate points of a certain initial airfoil point cloud data is greater than N, then the number of coordinate points of the airfoil point cloud data is reduced to N through sampling.
4. The airfoil generative design method based on an implicit diffusion model according to claim 2, characterized in that: In step 1, the process of expanding the existing airfoil point cloud data set is as follows: for a certain initial airfoil point cloud data, a traditional parameterization method is used for parameterization, and then the parameterized airfoil design parameters are disturbed within a set disturbance range.
5. The airfoil generative design method based on implicit diffusion model according to claim 4 is characterized by: The traditional parameterization method includes a CST parameterization method and a B-Spline parameterization method.
6. The airfoil generative design method based on implicit diffusion model according to claim 1, characterized in that: In step 2, for aerodynamic performance, the aerodynamic force and moment coefficients of the airfoil under several flow conditions are calculated, where the flow conditions are characterized by the Reynolds number Re, the Mach number Ma and the angle of attack α, and the aerodynamic force and moment coefficients include the lift coefficient C l , drag coefficient C d and moment coefficient C m , and the drag coefficient is further divided into the viscous drag coefficient C df and shape drag coefficient C dp ; For low detectability performance, the forward radar cross section is calculated at a given radar frequency.
7. The airfoil generative design method based on implicit diffusion model according to claim 1, characterized in that: The airfoil autoencoder model includes an encoder and a decoder; The encoder performs dimensionality reduction processing on the airfoil point cloud data to obtain a potential representation of the airfoil point cloud data; the encoder includes a serialization processing module, a stacked ViT module and a convolution module; the serialization processing module performs serialization processing on the input airfoil point cloud data to obtain serialized data; the serialized data is used as an input of the stacked ViT module, the features of the airfoil point cloud data are extracted through the stacked ViT module, and the features of the airfoil point cloud data are input into the convolution module for dimensionality reduction to obtain a potential representation of the airfoil point cloud data; The decoder reconstructs the potential representation to obtain airfoil point cloud data; the decoder includes a transposed convolution module, a stacked ViT module, a deserialization module and a Softmax module; the transposed convolution module upsamples the input potential representation, then extracts serialized feature data through the stacked ViT module, and uses the deserialization module to map the serialized feature data back to the original airfoil point cloud data structure; finally, the mapped data is reconstructed into airfoil point cloud data through the Softmax module.
8. The airfoil generative design method based on implicit diffusion model according to claim 7, characterized in that: The Softmax module in the decoder performs the following operations: first, the channel data representing the x-coordinate and the channel data representing the y-coordinate in the data output by the deserialization module are separated; then a one-dimensional convolution operation is performed on the channel data of the x-coordinate; the data after the one-dimensional convolution operation is then divided into two parts with an equal number of data points along the x-coordinate dimension, and a Softmax operation is performed on each part to obtain the x-coordinate spacing between adjacent data points; based on the x-coordinate spacing between adjacent data points, the x-coordinate of the reconstructed data point is calculated and combined with the channel data representing the y-coordinate in the data output by the deserialization module to obtain the reconstructed airfoil point cloud data.
9. The airfoil generative design method based on implicit diffusion model according to claim 1, characterized in that: In the airfoil-condition sample pair constructed in step 4, the airfoil is described by a potential representation, and the condition sample is described by a conditional vector label, where the conditional vector label c is c=[Ma,α,Re,L / D,C l ,C d ,C df ,C dp ,C m ,A,T max ,T TE ,σ 120° ] The parameters are: lift coefficient C l , drag coefficient C d , torque coefficient C m , Viscous drag coefficient C df , shape drag coefficient C dp , the average radar cross-section area σ within the forward 120° range 120° , flow conditions Reynolds number Re, Mach number Ma, angle of attack, lift-to-drag ratio L / D, airfoil cross-sectional area A, maximum thickness T max and the trailing edge thickness T TE .