An airfoil flow field prediction method, device and equipment based on airfoil feature extraction and physical information neural network and a storage medium

Abstract

The application discloses an airfoil flow field prediction method and device based on airfoil feature extraction and physical information neural network, equipment and storage medium, relates to the flow field prediction field, and includes: obtaining airfoil coordinate data and airfoil picture data of a target airfoil shape, and extracting airfoil features of the airfoil picture data based on a preset airfoil feature extraction model; inputting the airfoil coordinate data and the airfoil features into a preset graph neural network to obtain predicted flow field data; determining real flow field data according to the airfoil coordinate data, and determining a current loss value of the preset graph neural network based on the real flow field data and the predicted flow field data; optimizing the preset graph neural network based on the current loss value until the training is stopped when a training end condition is met to obtain a target graph neural network, and the target flow field data is predicted. The application can more effectively process flow field data through the graph neural network, and the calculation accuracy in flow field prediction is enhanced by combining the airfoil feature extraction technology.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of flow field prediction, in particular to a wing profile flow field prediction method and device based on wing profile feature extraction and physical information neural network, equipment and storage medium. BACKGROUND

[0002] In aircraft design, obtaining the flow field data around the wing is an important part of optimizing the wing profile design. The traditional method mainly relies on wind tunnel test and computational fluid dynamics (CFD, Computational Fluid Dynamics) technology. However, although the wind tunnel test can provide comprehensive and accurate aerodynamic performance evaluation, it is time-consuming and costly, and the CFD process involves complex grid division, solution and visualization steps, and requires a large number of iterations in wing profile design optimization, resulting in high computational cost.

[0003] In recent years, machine learning and deep learning technologies have shown great potential in flow field prediction, especially deep neural networks have shown significant advantages in improving prediction speed. Research mainly focuses on using convolutional neural networks, generative adversarial networks, long short-term memory networks and other models to learn a large amount of flow field data, and quickly predict aerodynamic coefficients based on wing profile features and environmental parameters. However, the above image-based processing method has limitations in capturing the complexity of the flow field and generalization ability, especially when dealing with unstructured grids and capturing subtle geometric changes. Therefore, how to more accurately obtain the flow field data around the wing is a problem to be solved in the field. SUMMARY

[0004] Therefore, the purpose of the present application is to provide a wing profile flow field prediction method and device based on wing profile feature extraction and physical information neural network, which can more effectively process flow field data through graph neural networks, and at the same time combines wing profile feature extraction technology to enhance the calculation accuracy in flow field prediction. The specific scheme is as follows:

[0005] In the first aspect, the present application provides a wing profile flow field prediction method based on wing profile feature extraction and physical information neural network, comprising:

[0006] Obtaining wing profile coordinate data and corresponding wing profile picture data of a target wing profile shape, and extracting wing profile features corresponding to the wing profile picture data based on a preset wing profile feature extraction model; the preset wing profile feature extraction model is a model based on a lightweight visual transformer;

[0007] Inputting the wing profile coordinate data and the corresponding wing profile features into a preset graph neural network to obtain predicted flow field data corresponding to the target wing profile shape; the preset graph neural network is a network based on the physical information neural network;

[0008] determine real flow field data corresponding to the target airfoil shape according to the airfoil coordinate data, and determine a current loss value corresponding to the preset graph neural network based on the real flow field data and the predicted flow field data;

[0009] optimize the preset graph neural network based on the current loss value, stop training when a first preset training end condition is met, obtain a final trained target graph neural network, and use the target graph neural network to predict target flow field data of a current airfoil shape.

[0010] Optionally, the airfoil coordinate data and the corresponding airfoil picture data of the target airfoil shape are obtained by:

[0011] obtaining a plurality of target airfoil shapes from a preset database, and generating corresponding airfoil coordinate data based on airfoil coordinates of the target airfoil shapes;

[0012] drawing corresponding airfoil pictures based on the airfoil coordinate data to obtain the airfoil picture data corresponding to the target airfoil shape.

[0013] Optionally, the real flow field data corresponding to the target airfoil shape is determined according to the airfoil coordinate data, comprising:

[0014] completing the dimension of the airfoil coordinate data to obtain target coordinate data; the data format of the target coordinate data is a preset target format;

[0015] generating a corresponding grid based on the target coordinate data, and simulating based on the grid under a preset environmental condition to obtain the real flow field data corresponding to the target airfoil shape under the preset environmental condition.

[0016] Optionally, the corresponding grid is generated based on the target coordinate data, comprising:

[0017] determining a wing region and a far field region of the target airfoil shape;

[0018] drawing a non-structural grid corresponding to the wing region based on the target coordinate data, and drawing a structural grid of the far field region based on the target coordinate data;

[0019] generating the grid corresponding to the target airfoil shape according to the non-structural grid and the structural grid.

[0020] Optionally, the airfoil coordinate data and the corresponding airfoil feature are input into a preset graph neural network to obtain predicted flow field data corresponding to the target airfoil shape, comprising:

[0021] construct a node feature matrix corresponding to the airfoil feature under the preset environmental condition, and generate an adjacency relationship matrix corresponding to the target airfoil shape corresponding to the airfoil coordinate data;

[0022] input the node feature matrix and the adjacency relationship matrix into the preset graph neural network to obtain the predicted flow field data corresponding to the target airfoil shape.

[0023] Optionally, the airfoil feature corresponding to the airfoil picture data is extracted based on a preset airfoil feature extraction model, including:

[0024] The airfoil picture data is taken as initial picture data, and the initial picture data is encoded based on the preset airfoil feature extraction model to obtain initial features of the airfoil picture data.

[0025] The initial features are processed by using a corresponding decoder of the preset airfoil feature extraction model to reconstruct the airfoil picture data, and updated picture data is obtained.

[0026] A first error of the initial picture data and the updated picture data is determined, and whether the preset airfoil feature extraction model meets a second preset training end condition is judged based on the first error.

[0027] If yes, the initial features are taken as the airfoil features.

[0028] If no, the preset airfoil feature extraction model is optimized based on the first error, the optimized preset airfoil feature extraction model is taken as a new preset airfoil feature extraction model, and the updated picture data is taken as new initial picture data, and the step of encoding the initial picture data based on the preset airfoil feature extraction model is jumped to until the current preset airfoil feature extraction model meets the second preset training end condition, and the current initial features are taken as the airfoil features.

[0029] Optionally, the current loss value corresponding to the preset graph neural network is determined based on the real flow field data and the predicted flow field data, including:

[0030] A mean square error between the real flow field data and the predicted flow field data is determined, and a prediction loss of the preset graph neural network is determined based on the mean square error.

[0031] A physical information loss of the preset graph neural network is determined based on the prediction loss.

[0032] The current loss value of the preset graph neural network is determined according to the prediction loss, the physical information loss and a preset weight.

[0033] In a second aspect, the present application provides a wing profile flow field prediction device based on wing profile feature extraction and physical information neural network, comprising:

[0034] a feature extraction module, configured to obtain wing profile coordinate data and corresponding wing profile picture data of a target wing profile shape, and extract wing profile features corresponding to the wing profile picture data based on a preset wing profile feature extraction model; the preset wing profile feature extraction model is a model based on a lightweight visual transformer;

[0035] a flow field prediction module, configured to input the wing profile coordinate data and the corresponding wing profile features into a preset graph neural network to obtain predicted flow field data corresponding to the target wing profile shape; the preset graph neural network is a network based on the physical information neural network;

[0036] a loss determination module, configured to determine real flow field data corresponding to the target wing profile shape according to the wing profile coordinate data, and determine a current loss value corresponding to the preset graph neural network based on the real flow field data and the predicted flow field data;

[0037] a network optimization module, configured to optimize the preset graph neural network based on the current loss value, stop training when a first preset training end condition is met, obtain a final trained target graph neural network, and predict target flow field data of a current wing profile shape by using the target graph neural network.

[0038] In a third aspect, the present application provides an electronic device, comprising a processor and a memory; wherein the memory is used to store a computer program, the computer program is loaded and executed by the processor to realize the wing profile flow field prediction method based on wing profile feature extraction and physical information neural network as described above.

[0039] In a fourth aspect, the present application provides a computer readable storage medium for saving a computer program, the computer program is executed by a processor to realize the wing profile flow field prediction method based on wing profile feature extraction and physical information neural network as described above.

[0040] The application first acquires airfoil coordinate data and corresponding airfoil picture data of a target airfoil shape, extracts airfoil features corresponding to the airfoil picture data by using a preset airfoil feature extraction model based on a lightweight visual transformer, then inputs the airfoil coordinate data and the corresponding airfoil features into a preset graph neural network based on a physical information neural network, obtains predicted flow field data corresponding to the target airfoil shape, determines real flow field data corresponding to the target airfoil shape according to the airfoil coordinate data, and determines a current loss value corresponding to the preset graph neural network based on the real flow field data and the predicted flow field data, then optimizes the preset graph neural network based on the current loss value, stops training when a first preset training end condition is met, obtains a final trained target graph neural network, and predicts target flow field data of a current airfoil shape by using the target graph neural network. In this way, the application can more effectively process flow field data by using a graph neural network, and combine a physical information neural network to integrate the basic laws of fluid dynamics into the model. In this way, the graph neural network is used as the basic network model, which helps to overcome the limitations of Image form flow field prediction, improve the accuracy and interpretability of flow field prediction, and further enhance the generalization ability and calculation accuracy of the model in multi-airfoil flow field prediction by combining the airfoil feature extraction technology. BRIEF DESCRIPTION OF DRAWINGS

[0041] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or the prior art description. Obviously, the drawings in the following description are only embodiments of the present application, and for those skilled in the art, other drawings can be obtained without creative labor on the basis of the provided drawings.

[0042] Figure 1 A wing airfoil flow field prediction method based on airfoil feature extraction and physical information neural network provided by the present application is shown in the flow chart.

[0043] Figure 2 A two-dimensional wing airfoil flow field prediction method based on airfoil feature extraction and physical information neural network provided by the present application is shown in the block diagram.

[0044] Figure 3 A wing airfoil flow field prediction method based on airfoil feature extraction and physical information neural network provided by the present application is shown in the flow chart.

[0045] Figure 4 A two-dimensional wing airfoil flow field prediction method based on airfoil feature extraction and physical information neural network provided by the present application is shown in the flow chart.

[0046] Figure 5 A wing airfoil flow field prediction device structure schematic diagram provided by the present application is shown in the block diagram.

[0047] Figure 6 An electronic device structure diagram is provided. DETAILED DESCRIPTION

[0048] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0049] In recent years, machine learning and deep learning technologies have shown great potential in flow field prediction, especially deep neural networks have significant advantages in improving prediction speed. However, the current processing method has limitations in capturing the complexity of the flow field and the generalization ability, especially when dealing with unstructured grids and capturing subtle geometric changes.

[0050] To overcome the above-mentioned defects, the present application utilizes the advantages of graph neural network (GNN, Graph Neural Network) in unstructured data feature extraction, which can more effectively process flow field data based on unstructured grids, and combines physics-informed neural networks (PINNs, Physics-Informed Neural Networks) to integrate the basic laws of fluid dynamics into the model, improving the accuracy and interpretability of the prediction. In addition, advanced airfoil feature extraction techniques such as Transformer and lightweight models (such as MobileVIT (Mobile Vision Transformer, a lightweight visual transformer for mobile devices)) can be combined to further enhance the generalization ability and computational accuracy of the graph neural network model in multi-airfoil flow field prediction.

[0051] Referring to Figure 1 The embodiments of the present application disclose an airfoil flow field prediction method based on airfoil feature extraction and physics-informed neural networks, comprising:

[0052] Step S11, obtaining airfoil coordinate data of a target airfoil shape and corresponding airfoil picture data, and extracting airfoil features corresponding to the airfoil picture data based on a preset airfoil feature extraction model; the preset airfoil feature extraction model is a model based on a lightweight visual transformer.

[0053] In this embodiment, as Figure 2As shown, first, airfoil coordinate data of a target airfoil shape and corresponding airfoil picture data need to be acquired, and airfoil features corresponding to the airfoil picture data are extracted based on a preset airfoil feature extraction model; it should be noted that the preset airfoil feature extraction model is a model based on a lightweight visual transformer MobileViT. Specifically, when the airfoil coordinate data of the target airfoil shape and the corresponding airfoil picture data are acquired, a plurality of target airfoil shapes can be acquired from a preset database, and corresponding airfoil coordinate data can be generated based on airfoil coordinates of the target airfoil shapes, and then corresponding airfoil pictures can be drawn based on the airfoil coordinate data to obtain airfoil picture data corresponding to the target airfoil shape.

[0054] In a specific embodiment, 1525 different airfoil shapes from the UIUC database (University of Illinois Urbana-Champaign Airfoil Coordinates Database) can be used, and for these airfoil coordinates, 1525 airfoil pictures of size 224x224 can be drawn by the Plot function of Python, and corresponding airfoil coordinate files and airfoil picture files can be produced. It can be understood that the drawing method and picture size of the airfoil picture are not limited to the scheme disclosed in the present embodiment.

[0055] Step S12, input the airfoil coordinate data and the corresponding airfoil features into a preset graph neural network to obtain predicted flow field data corresponding to the target airfoil shape; the preset graph neural network is a network based on the physical information neural network.

[0056] In the present embodiment, as shown in Figure 2 After obtaining the airfoil features corresponding to each airfoil shape, the airfoil coordinate data and the corresponding airfoil features can be input into a preset graph neural network to obtain predicted flow field data corresponding to the target airfoil shape. The preset graph neural network is a network based on the physical information neural network PINNs.

[0057] Step S13, determining real flow field data corresponding to the target airfoil shape according to the airfoil coordinate data, and determining a current loss value corresponding to the preset graph neural network based on the real flow field data and the predicted flow field data.

[0058] In the present embodiment, as shown in Figure 2As shown, the real flow field data corresponding to the target airfoil shape can be determined according to the airfoil coordinate data, and then the current loss value of the preset graph neural network can be determined based on the real flow field data and the predicted flow field data. Specifically, the mean square error between the real flow field data and the predicted flow field data can be determined first, and the prediction loss of the preset graph neural network can be determined based on the mean square error (MSE, Mean Squared Error), and then the physical information loss of the preset graph neural network can be determined based on the prediction loss, so as to determine the current loss value of the preset graph neural network according to the prediction loss, the physical information loss and the preset weight.

[0059] That is, the difference between the predicted data and the real data can be calculated by using the mean square error loss function, and the data loss of the predicted neural network model is obtained by adding the loss values corresponding to the velocity and pressure

[0060]

[0061] ;

[0062] wherein, is the loss function of the velocity component, representing the mean square error (MSE) of the predicted velocity and the actual velocity; M1 is the total number of data points (such as the number of grid points or measurement points in the flow field); is the predicted velocity component (x and y directions) of the kth data point; is the real velocity component of the kth data point (also from high-precision simulation or experimental data); is the loss function of the pressure component, representing the mean square error of the predicted pressure and the actual pressure; M1 is consistent with the definition in the velocity loss function, which is the total number of data points; is the predicted pressure value of the kth data point; is the real pressure value of the kth data point.

[0063] The predicted results are then substituted into the continuity equation and the momentum equation respectively to calculate the loss caused by the above physical equations to obtain the physical information loss and , wherein the momentum conservation equation is divided into x direction and y direction .

[0064] ;

[0065] wherein, is the loss function of the continuity equation, used to measure whether the predicted flow field satisfies the mass conservation; , represent the spatial derivatives (representing the velocity gradient) of the predicted velocity component in x and y directions.​ Loss function representing x-direction momentum equation, measuring whether the predicted flow field satisfies the momentum conservation law; p represents fluid density; μ represents fluid dynamic viscosity coefficient (related to viscous force); represents the predicted pressure value of the kth data point; represents the convection term (inertial force) in the x-direction; represents the pressure gradient term in the x-direction; represents the viscous diffusion term (viscous force) in the x-direction; Loss function representing y-direction momentum equation; represents the convection term (inertial force) in the y-direction; represents the pressure gradient term in the y-direction; represents the viscous diffusion term (viscous force) in the y-direction.

[0066] Then, as shown in Figure 2 , the predicted output of the neural network model is passed to the physical loss module together with the corresponding real flow field data, and the data loss and the physical information loss are linearly combined according to the predetermined proportion weight to obtain the total loss used for iterative optimization.

[0067] .

[0068] Step S14, based on the current loss value, the preset graph neural network is optimized until the first preset training end condition is met, the training is stopped, the final trained target graph neural network is obtained, and the target flow field data of the current airfoil shape is predicted using the target graph neural network.

[0069] In this embodiment, the total loss obtained in the previous step, i.e. the current loss value, can be used to optimize the preset graph neural network until the first preset training end condition is met, the training is stopped, the final trained target graph neural network is obtained, and the target flow field data of the current airfoil shape is predicted using the target graph neural network. In a specific embodiment, the proportion of the physical information loss to the total loss can be set to 1:0.4, and the proportion of the data loss to the total loss can be set to 1:0.6. : The best training result is 0.001639 when the proportion is 1:0.4; at the same time, as the proportion of the physical information loss increases, the optimization of the pure data-driven model by the physical loss is constantly decreasing, for example, : When the ratio exceeds 1:1.4, the test loss fluctuates between 0.001675-0.001700, and the optimization of the model is continuously reduced compared to the previous ratio. At the same time, it can be understood that there is a certain range for the ratio setting of physical loss and data loss. If the ratio of physical loss is set too large, it will affect the optimization of the model by data loss, thereby causing negative optimization. The above first preset training end condition includes, but is not limited to, reaching a specified number of iterations, or the loss meeting a preset accuracy requirement, etc.

[0070] The embodiment can obtain airfoil coordinate data of a target airfoil shape and corresponding airfoil picture data, extract airfoil features corresponding to the airfoil picture data by using a preset airfoil feature extraction model based on a lightweight visual transformer, input the airfoil coordinate data and the corresponding airfoil features into a preset graph neural network based on a physical information neural network, obtain predicted flow field data corresponding to the target airfoil shape, determine real flow field data corresponding to the target airfoil shape according to the airfoil coordinate data, and determine a current loss value corresponding to the preset graph neural network based on the real flow field data and the predicted flow field data. Then, the preset graph neural network is optimized based on the current loss value, and the training is stopped when the first preset training end condition is met, to obtain a final trained target graph neural network, and the target graph neural network is used to predict target flow field data of a current airfoil shape. Through the above technical solution, the graph neural network can more effectively process flow field data, and the physical information neural network is combined to integrate the basic laws of fluid dynamics into the model. In this way, the graph neural network is used as the basic network model, which helps to overcome the limitations of Image form flow field prediction, improve the accuracy and interpretability of flow field prediction, and combine the airfoil feature extraction module and the physical information loss module. The neural network model based on MobileViT is used to extract airfoil features to improve the generalization of multi-airfoil flow field prediction. The fluid dynamics principle is integrated into the neural network model in the form of PINNs, and the corresponding flow conditions and airfoil data are input into the trained model to generate predicted flow field results, which improves the interpretability of the neural network model and further enhances the generalization ability and calculation accuracy of the model in multi-airfoil flow field prediction.

[0071] Based on the previous embodiment, the graph neural network can more effectively process flow field data, and the airfoil feature extraction technology is combined to enhance the calculation accuracy in flow field prediction. Next, the processing process of the flow field data will be described in detail in this embodiment. Referring to Figure 3 As shown in the figure, the embodiment of the application discloses a specific airfoil flow field prediction method based on airfoil feature extraction and physical information neural network, which includes:

[0072] Step S21: Obtain the airfoil coordinate data and corresponding airfoil image data of the target airfoil shape, and extract the airfoil features corresponding to the airfoil image data based on a preset airfoil feature extraction model; the preset airfoil feature extraction model is a model based on a lightweight visual transformer.

[0073] In this embodiment, when extracting airfoil features corresponding to airfoil image data based on a preset airfoil feature extraction model, the airfoil image data can be used as initial image data. The initial image data is encoded based on the preset airfoil feature extraction model to obtain initial features. Then, the initial features are processed using the corresponding decoder of the preset airfoil feature extraction model to reconstruct the airfoil image data, resulting in updated image data. Next, a first error between the initial image data and the updated image data is determined, and based on this first error, it is determined whether the preset airfoil feature extraction model meets a second preset training termination condition. If the preset airfoil feature extraction model meets the second preset training termination condition, the initial features are used as the airfoil features. If the preset airfoil feature extraction model does not meet the second preset training termination condition, the preset airfoil feature extraction model is optimized based on the first error. The optimized preset airfoil feature extraction model is used as the new preset airfoil feature extraction model, and the updated image data is used as the new initial image data. The process jumps back to the step of encoding the initial image data based on the preset airfoil feature extraction model, until the current preset airfoil feature extraction model meets the second preset training termination condition, at which point the current initial features are used as the airfoil features. The aforementioned second preset training termination conditions include, but are not limited to: reaching a specified number of iterations, or the loss meeting preset accuracy requirements.

[0074] In this embodiment, as Figure 2 , Figure 4 As shown, in the process of generating airfoil features, an airfoil image reconstruction module is needed to reconstruct the airfoil image. Specifically, the input airfoil image is encoded layer by layer, passing through the MV2 (MobileNetV2) layer and the MobileVITBlock layer respectively. After layer-by-layer processing, the 80-dimensional parameter vector that can represent the airfoil in the MobileVITBlock is obtained, which is the airfoil feature. Then, DecoderBlock is used to upsample and further process the features or adjust the network output to gradually restore the spatial resolution of the image and achieve the purpose of airfoil reconstruction.

[0075] Next, repeat the airfoil reconstruction steps above until the airfoil feature model converges. At this point, export the feature parameters of the corresponding airfoil and store them in the corresponding file to enable the subsequent training of the prediction model.

[0076] In a specific embodiment, the mean square error (MSE) is used as an evaluation index of the image reconstruction accuracy. In this embodiment, the MSE measures the average of the sum of squares of the pixel intensity differences between the reconstructed image and the true image. The specific calculation formula is as follows:

[0077] ;

[0078] wherein, represents the pixel value of the predicted image, represents the pixel value of the true image, and N represents the total number of pixels in the image.

[0079] For each airfoil image in the current batch, the MSE thereof is calculated, and the average of the MSE of all images in the batch is taken as the evaluation index. In a specific embodiment, the image size of each batch is (batch_size, 3, 224, 224), that is, each batch contains multiple 3-channel 224x224 pixel images. It can be understood that the smaller the MSE value, the smaller the difference between the predicted image and the true image, and the higher the prediction quality.

[0080] In step S22, the airfoil coordinate data and the corresponding airfoil features are input into a preset graph neural network to obtain predicted flow field data corresponding to the target airfoil shape; the preset graph neural network is a network based on the physical information neural network.

[0081] In this embodiment, when the predicted flow field data corresponding to the target airfoil shape is obtained by the preset graph neural network, a corresponding node feature matrix can be constructed based on the airfoil features under the preset environmental conditions, and a corresponding adjacency relationship matrix of the target airfoil shape corresponding to the airfoil coordinate data is generated, and then the node feature matrix and the adjacency relationship matrix are input into the preset graph neural network to obtain the predicted flow field data corresponding to the target airfoil shape. That is, the corresponding airfoil features and environmental conditions, i.e., flow variables (Reynolds number, Mach number, angle of attack, airfoil surface marker), can be combined to form a node feature matrix, and different grid configuration files of different CFD simulation software (SU2) can be imported for different airfoils, so as to introduce the x and y coordinates of the nodes in the airfoil grid and the node adjacency relationship, so as to further improve the node feature matrix and construct the adjacency relationship matrix. Then, the obtained node matrix and adjacency matrix are input into the graph convolutional neural network (i.e., the airfoil flow field prediction model (preset graph neural network) based on GCN (graph convolutional network) + PINNs) together, and multi-layer graph convolution is performed to obtain the predicted flow field result. In an optional specific embodiment, the number of convolution layers can be (128, 64, 32, 16, 8, 3).

[0082] Step S23, determining real flow field data corresponding to the target airfoil shape according to the airfoil coordinate data, and determining a current loss value corresponding to the preset graph neural network based on the real flow field data and the predicted flow field data.

[0083] In the embodiment, the real flow field data corresponding to the target airfoil shape is determined, and specifically, the airfoil coordinate data can be dimensionally completed to obtain target coordinate data; the data format of the target coordinate data is a preset target format; then a corresponding grid is generated based on the target coordinate data, and simulation is performed based on the preset environmental conditions to obtain the real flow field data corresponding to the target airfoil shape under the preset environmental conditions. When the corresponding grid is generated based on the target coordinate data, the wing region and the far-field region of the target airfoil shape can be determined, and a non-structured grid corresponding to the wing region is drawn based on the target coordinate data, and a structured grid of the far-field region is drawn based on the target coordinate data, and then the grid corresponding to the target airfoil shape is generated according to the non-structured grid and the structured grid.

[0084] Specifically, in the embodiment, the airfoil coordinates obtained by UIUC can be dimensionally completed and adjusted to a format supported by PointWise for processing, and then Pointwise is used to draw a grid around the airfoil based on the processed airfoil coordinates. In a specific embodiment, a non-structured grid can be used to draw a grid around the wing to capture the details of the fluid, and a structured grid can be used to draw a grid in the remaining region (far field) to improve the calculation speed. Then the drawn grid file is imported into the CFD simulation software (SU2), and the corresponding flow field data is generated according to the different environmental conditions (angle of attack, Mach number, Reynolds number) of the airfoil, and then the generated data is used as the training data and test data of the prediction model. In the embodiment, the flow variables are as follows:

[0085] MACH_LIST = [0.2, 0.25, 0.3, 0.35, 0.4, 0.45, 0.5, 0.55, 0.6, 0.7, 0.8];

[0086] AOA_LIST = [-10.0, -9.0, -8.0, -7.0, -6.0, -5.0, -4.0, -3.0, -2.0, -1.0, 0.0, 1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0, 9.0, 10.0];

[0087] REYNOLDS_LIST = [0.5, 1.0, 2.0, 3.0] x 1000000.

[0088] Then, the real flow field image and the corresponding external flow condition obtained in the above step and the corresponding airfoil can be taken as a set of training data, and by repeating the above process of obtaining the real flow field data and the predicted data under different environmental conditions, a training data set containing m sets of training data can be constructed to serve as the training data and test data for the model optimization.

[0089] That is, in the embodiment, the airfoil coordinates obtained by UIUC are first dimensionally completed, adjusted to the format supported by PointWise, and then PointWise is used to draw the grid around the airfoil according to the processed airfoil coordinates, the drawn grid file is imported into the CFD simulation software, and then the corresponding flow field data of the airfoil under different environmental conditions (angle of attack, Mach number, Reynolds number) is generated, so as to take the generated data as the training data and test data of the prediction model. When performing dimension completion, the coordinate point density on a single cross section can be increased (such as interpolation), or missing data can be repaired (such as leading edge and trailing edge closure).

[0090] Step S24: optimizing the preset graph neural network based on the current loss value, stopping training when a first preset training end condition is met, obtaining a final trained target graph neural network, and predicting target flow field data of the current airfoil shape by using the target graph neural network.

[0091] In the above step S24, the more specific processing process can refer to the corresponding content disclosed in the foregoing embodiments, which will not be described here again.

[0092] Based on the above embodiments, the flow field prediction original data for training and the airfoil picture data for training of the airfoil feature extraction module can be obtained, and then the encoded airfoil information is taken as the input, the corresponding airfoil picture is taken as the target, the MobileVIT is used to extract the airfoil features and reconstruct the airfoil from the open source data, the airfoil features are repeatedly trained, then the airfoil features representing the airfoil and the graph structure data corresponding to the airfoil are spliced, input into the graph neural network, the predicted flow field image is obtained, and the predicted flow field image and the real flow field image are input into the physical loss module to obtain the physical loss and the data loss required for updating the network, and then the final loss is obtained by combining in a certain proportion, and after a specified number of iterations or when the loss meets the accuracy requirement, the final neural network model for flow field prediction is obtained. The two-dimensional airfoil flow field prediction method based on airfoil feature extraction and physical information neural network in the embodiment uses the graph neural network as the basic network model, which can overcome the limitations of the Image-based processing method in capturing the complexity of the flow field and the generalization ability, especially in dealing with unstructured grids and capturing subtle geometric changes. Compared with the traditional image-based reduced-order model, the graph neural model integrates the principles of fluid dynamics into the neural network model, improves the interpretability of the neural network model, and combines the airfoil feature extraction module and the physical information loss module to extract the airfoil features using the neural network model based on MobileViT, thereby improving the generalization of the multi-airfoil flow field prediction. In combination with the technical solutions in the previous embodiment and the present embodiment, a reduced-order model can be established to efficiently predict the flutter boundary. Once the model is established, repeated calculation is not required, and the aircraft design and development cycle can be greatly shortened.

[0093] Referring to Figure 5 As shown in the figure, the application embodiment also discloses an airfoil flow field prediction device based on airfoil feature extraction and physical information neural network, comprising:

[0094] The feature extraction module 11 is configured to obtain airfoil coordinate data of a target airfoil shape and corresponding airfoil picture data, and extract airfoil features corresponding to the airfoil picture data based on a preset airfoil feature extraction model; the preset airfoil feature extraction model is a model based on a lightweight visual transformer;

[0095] The flow field prediction module 12 is configured to input the airfoil coordinate data and the corresponding airfoil features into a preset graph neural network to obtain predicted flow field data corresponding to the target airfoil shape; the preset graph neural network is a network based on the physical information neural network;

[0096] The loss determination module 13 is configured to determine real flow field data corresponding to the target airfoil shape according to the airfoil coordinate data, and determine a current loss value corresponding to the preset graph neural network based on the real flow field data and the predicted flow field data.

[0097] a network optimization module 14, configured to optimize the preset graph neural network based on the current loss value until a first preset training end condition is met to stop training, to obtain a final trained target graph neural network, and to use the target graph neural network to predict target flow field data of a current airfoil shape.

[0098] The embodiment can obtain airfoil coordinate data and corresponding airfoil picture data of a target airfoil shape, extract airfoil features corresponding to the airfoil picture data by using a preset airfoil feature extraction model based on a lightweight visual transformer, input the airfoil coordinate data and the corresponding airfoil features into a preset graph neural network based on a physical information neural network, obtain predicted flow field data corresponding to the target airfoil shape, determine real flow field data corresponding to the target airfoil shape according to the airfoil coordinate data, determine a current loss value corresponding to the preset graph neural network based on the real flow field data and the predicted flow field data, optimize the preset graph neural network based on the current loss value until a first preset training end condition is met to stop training, obtain a final trained target graph neural network, and use the target graph neural network to predict target flow field data of a current airfoil shape. Through the above technical solutions, the graph neural network can more effectively process flow field data, and the physical information neural network is combined to integrate the basic laws of fluid dynamics into the model. In this way, the graph neural network is used as the basic network model, which helps to overcome the limitations of Image form flow field prediction, improve the accuracy and interpretability of flow field prediction, and further enhance the generalization ability and calculation accuracy of the model in multi-airfoil flow field prediction.

[0099] In some specific embodiments, the feature extraction module 11 specifically includes:

[0100] a data acquisition unit, configured to acquire a plurality of target airfoil shapes from a preset database, and generate corresponding airfoil coordinate data based on airfoil coordinates of the target airfoil shapes;

[0101] a picture drawing unit, configured to draw corresponding airfoil pictures based on the airfoil coordinate data to obtain the airfoil picture data corresponding to the target airfoil shapes.

[0102] In some specific embodiments, the loss determination module 13 specifically includes:

[0103] a data completion sub-module, configured to complete the dimensions of the airfoil coordinate data to obtain target coordinate data; the data format of the target coordinate data is a preset target format;

[0104] a grid simulation submodule configured to generate a corresponding grid based on the target coordinate data, and simulate based on a preset environmental condition according to the grid to obtain the real flow field data corresponding to the target airfoil shape under the preset environmental condition.

[0105] In some embodiments, the grid simulation submodule specifically comprises:

[0106] a region determination unit configured to determine a wing region and a far-field region of the target airfoil shape;

[0107] a grid drawing unit configured to draw a non-structured grid corresponding to the wing region based on the target coordinate data, and draw a structured grid of the far-field region based on the target coordinate data;

[0108] a grid generation unit configured to generate the grid corresponding to the target airfoil shape according to the non-structured grid and the structured grid.

[0109] In some embodiments, the flow field prediction module 12 specifically comprises:

[0110] a matrix generation unit configured to construct a corresponding node feature matrix based on the airfoil features under the preset environmental condition, and generate a corresponding adjacency relationship matrix of the target airfoil shape corresponding to the airfoil coordinate data;

[0111] a data generation unit configured to input the node feature matrix and the adjacency relationship matrix into the preset graph neural network to obtain the predicted flow field data corresponding to the target airfoil shape.

[0112] In some embodiments, the feature extraction module 11 specifically comprises:

[0113] a data encoding unit configured to take the airfoil picture data as initial picture data, and encode the initial picture data based on the preset airfoil feature extraction model to obtain initial features of the airfoil picture data;

[0114] a picture reconstruction unit configured to process the initial features by using a corresponding decoder of the preset airfoil feature extraction model to reconstruct the airfoil picture data and obtain updated picture data;

[0115] an error determination unit configured to determine a first error of the initial picture data and the updated picture data, and determine whether the preset airfoil feature extraction model meets a second preset training end condition based on the first error; and when the second preset training end condition is met, take the initial feature as the airfoil feature; and when the second preset training end condition is not met, optimize the preset airfoil feature extraction model based on the first error, take the optimized preset airfoil feature extraction model as a new preset airfoil feature extraction model, and take the updated picture data as new initial picture data, and jump to the step of encoding the initial picture data based on the preset airfoil feature extraction model, until the preset airfoil feature extraction model meets the second preset training end condition, and take the initial feature as the airfoil feature.

[0116] In some embodiments, the loss determination module 13 specifically comprises:

[0117] a first loss determination unit configured to determine a mean square error between the real flow field data and the predicted flow field data, and determine a prediction loss of the preset graph neural network based on the mean square error;

[0118] a second loss determination unit configured to determine a physical information loss of the preset graph neural network based on the prediction loss;

[0119] a third loss determination unit configured to determine the current loss value of the preset graph neural network according to the prediction loss, the physical information loss and a preset weight.

[0120] Further, the embodiment of the present application further discloses an electronic device, Figure 6 is the structure diagram of the electronic device 20 according to an exemplary embodiment, and the contents in the figure cannot be considered as any limitation on the use range of the present application.

[0121] Figure 6 A structural schematic diagram of an electronic device 20 provided by the embodiment of the present application. The electronic device 20 specifically can include at least one processor 21, at least one memory 22, a power supply 23, a communication interface 24, an input / output interface 25 and a communication bus 26. Wherein, the memory 22 is used to store computer programs, the computer programs are loaded and executed by the processor 21, to realize the related steps in the airfoil flow field prediction method based on airfoil feature extraction and physical information neural network disclosed by any of the preceding embodiments. In addition, the electronic device 20 in the embodiment specifically can be an electronic computer.

[0122] In this embodiment, the power supply 23 is configured to provide operating voltage for each hardware device on the electronic device 20; the communication interface 24 is configured to create a data transmission channel between the electronic device 20 and external devices, and the communication protocol followed by the communication interface 24 can be any communication protocol applicable to the technical solution of the present application, which will not be specifically limited herein; the input and output interface 25 is configured to obtain external input data or output data to the outside, and the specific interface type can be selected according to the specific application needs, which will not be specifically limited herein.

[0123] In addition, the memory 22 as a carrier for storing resources can be a read-only memory, a random access memory, a magnetic disk or an optical disk, etc., and the resources stored thereon can include an operating system 221, a computer program 222, etc., and the storage mode can be temporary storage or permanent storage.

[0124] The operating system 221 is configured to manage and control each hardware device and the computer program 222 on the electronic device 20, and can be Windows Server, Netware, Unix, Linux, etc. In addition to the computer program capable of completing the airfoil flow field prediction method based on airfoil feature extraction and physical information neural network executed by the electronic device 20 disclosed in any of the foregoing embodiments, the computer program 222 can further include a computer program capable of completing other specific work.

[0125] Further, the present application also discloses a computer readable storage medium for storing a computer program; wherein the computer program is executed by a processor to implement the airfoil flow field prediction method based on airfoil feature extraction and physical information neural network disclosed above. For the specific steps of the method, please refer to the corresponding contents disclosed in the foregoing embodiments, which will not be repeated here.

[0126] Each embodiment in the specification is described in a progressive manner, and each embodiment focuses on the difference from other embodiments. For the same or similar parts between each embodiment, please refer to each other. For the device disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple, and the relevant parts are described in the method part.

[0127] Those skilled in the art will further appreciate that the units and algorithm steps of the various examples described in connection with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or combinations of both. To clearly illustrate this interchangeability of hardware and software, various examples have been described herein in terms of their functionality, which has been described generally and symbolically in flow charts. Having thus described the functionality of the examples, a person of ordinary skill in the art will be able to implement such functionality in hardware and / or software, and will recognize that the bounds of the examples are not limited by one approach or the other. The various examples can be realized in a centralized fashion in one computer system or network, or in a distributed fashion where different elements are spread across several computer systems or sub-networks. Any kind of computer system or other apparatus adapted for carrying out the methods described herein is suited. A typical combination of hardware and software could be a general purpose computer system with a computer program that, when being loaded and executed, carries out the methods described herein.

[0128] The steps of a method or algorithm described in connection with the embodiments disclosed herein can be embodied directly in hardware, in a software module executed by a processor, or in a combination of the two. A software module can reside in random access memory (RAM), flash memory, read-only memory (ROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), registers, hard disk, a removable disk, a CD-ROM, or any other form of storage medium known in the art. An exemplary storage medium is coupled to the processor such that the processor can read information from, and write information to, the storage medium. In the alternative, hard disk can be used as a storage medium.

[0129] Finally, it should be noted that the terms "first", "second", and the like, herein do not denote any order, quantity, combination, or importance, but rather are used to distinguish one element from another, and do not imply singular or plural. Moreover, the terms "include", "have", or any other variant thereof are intended to encompass non-exclusive inclusions, such that processes, methods, articles, or apparatuses that comprise a set of elements not expressly listed are also within the scope of the present application. In addition, the articles "a" and "an" are used herein to refer to one or to more than one (i.e., to one or at least one) of the grammatical object of the article. By way of example, "an element" means one element or one or more elements.

[0130] The above detailed description of the technical solutions provided by the present application has been described in detail, and the principles and implementation modes of the present application have been described in the above examples. The above example is only used to help understand the method and core idea of the present application; at the same time, for those skilled in the art, according to the idea of the present application, the specific implementation mode and application range will be changed; in view of the above, the content of the specification should not be understood as limiting the present application.

Claims

1. A method for predicting airfoil flow fields based on airfoil feature extraction and physical information neural networks, characterized in that, include: The airfoil coordinate data and corresponding airfoil image data of the target airfoil shape are obtained, and the airfoil features corresponding to the airfoil image data are extracted based on a preset airfoil feature extraction model; the preset airfoil feature extraction model is a model based on a lightweight visual transformer. The airfoil coordinate data and the corresponding airfoil features are input into a preset graph neural network to obtain the predicted flow field data corresponding to the target airfoil shape; The preset graph neural network is a network based on the physical information neural network; The actual flow field data corresponding to the target airfoil shape is determined based on the airfoil coordinate data, and the current loss value corresponding to the preset graph neural network is determined based on the actual flow field data and the predicted flow field data. The preset graph neural network is optimized based on the current loss value until the first preset training end condition is met, at which point training stops, resulting in the final trained target graph neural network. The target graph neural network is then used to predict the target flow field data for the current airfoil shape. The step of extracting the airfoil features corresponding to the airfoil image data based on the preset airfoil feature extraction model includes: The airfoil image data is used as the initial image data, and the initial image data is encoded based on the preset airfoil feature extraction model to obtain the initial features of the airfoil image data. The initial features are processed using the decoder corresponding to the preset airfoil feature extraction model to reconstruct the airfoil image data and obtain updated image data. Determine the first error between the initial image data and the updated image data, and determine whether the preset airfoil feature extraction model meets the second preset training termination condition based on the first error; If satisfied, the initial feature is taken as the airfoil feature; If the condition is not met, the preset airfoil feature extraction model is optimized based on the first error. The optimized preset airfoil feature extraction model is used as the new preset airfoil feature extraction model, and the updated image data is used as the new initial image data. The process then jumps to the step of encoding the initial image data based on the preset airfoil feature extraction model, until the current preset airfoil feature extraction model meets the second preset training termination condition. Finally, the current initial feature is used as the airfoil feature.

2. The airfoil flow field prediction method based on airfoil feature extraction and physical information neural network according to claim 1, characterized in that, The process of acquiring the airfoil coordinate data and corresponding airfoil image data of the target airfoil shape includes: Several target airfoil shapes are obtained from a preset database, and corresponding airfoil coordinate data are generated based on the airfoil coordinates of the target airfoil shapes; Based on the airfoil coordinate data, a corresponding airfoil image is drawn to obtain the airfoil image data corresponding to the target airfoil shape.

3. The airfoil flow field prediction method based on airfoil feature extraction and physical information neural network according to claim 1, characterized in that, The step of determining the actual flow field data corresponding to the target airfoil shape based on the airfoil coordinate data includes: The airfoil coordinate data is augmented with dimensions to obtain target coordinate data; the target coordinate data is in a preset target format. A corresponding mesh is generated based on the target coordinate data, and simulation is performed based on the mesh under preset environmental conditions to obtain the real flow field data corresponding to the target airfoil shape under the preset environmental conditions.

4. The airfoil flow field prediction method based on airfoil feature extraction and physical information neural network according to claim 3, characterized in that, The process of generating a corresponding mesh based on the target coordinate data includes: Determine the wing region and far-field region of the target airfoil shape; An unstructured mesh corresponding to the wing region is drawn based on the target coordinate data, and a structured mesh of the far-field region is drawn based on the target coordinate data. The mesh corresponding to the target airfoil shape is generated based on the unstructured mesh and the structured mesh.

5. The airfoil flow field prediction method based on airfoil feature extraction and physical information neural network according to claim 4, characterized in that, The step of inputting the airfoil coordinate data and the corresponding airfoil features into a preset graph neural network to obtain the predicted flow field data corresponding to the target airfoil shape includes: Based on the airfoil features under the preset environmental conditions, a corresponding node feature matrix is ​​constructed, and an adjacency relation matrix corresponding to the target airfoil shape corresponding to the airfoil coordinate data is generated. The node feature matrix and the adjacency relation matrix are input into the preset graph neural network to obtain the predicted flow field data corresponding to the target airfoil shape.

6. The airfoil flow field prediction method based on airfoil feature extraction and physical information neural network according to claim 1, characterized in that, Determining the current loss value corresponding to the preset graph neural network based on the real flow field data and the predicted flow field data includes: Determine the mean square error between the actual flow field data and the predicted flow field data, and determine the prediction loss of the preset graph neural network based on the mean square error; The physical information loss of the preset graph neural network is determined based on the prediction loss; The current loss value of the preset graph neural network is determined based on the prediction loss, the physical information loss, and the preset weights.

7. An airfoil flow field prediction device based on airfoil feature extraction and physical information neural network, characterized in that, include: The feature extraction module is used to acquire the airfoil coordinate data and corresponding airfoil image data of the target airfoil shape, and extract the airfoil features corresponding to the airfoil image data based on a preset airfoil feature extraction model; the preset airfoil feature extraction model is a model based on a lightweight visual transformer. The flow field prediction module is used to input the airfoil coordinate data and the corresponding airfoil features into a preset graph neural network to obtain the predicted flow field data corresponding to the target airfoil shape; the preset graph neural network is a network based on the physical information neural network. The loss determination module is used to determine the real flow field data corresponding to the target airfoil shape based on the airfoil coordinate data, and to determine the current loss value corresponding to the preset graph neural network based on the real flow field data and the predicted flow field data. The network optimization module is used to optimize the preset graph neural network based on the current loss value until the first preset training end condition is met, and then stop training to obtain the final trained target graph neural network. The target graph neural network is then used to predict the target flow field data of the current airfoil shape. The feature extraction module includes: A data encoding unit is used to take the airfoil image data as initial image data and encode the initial image data based on the preset airfoil feature extraction model to obtain the initial features of the airfoil image data. The image reconstruction unit is used to process the initial features using the decoder corresponding to the preset airfoil feature extraction model in order to reconstruct the airfoil image data and obtain updated image data. An error determination unit is used to determine a first error between the initial image data and the updated image data, and to determine whether the preset airfoil feature extraction model meets a second preset training termination condition based on the first error; if it meets the condition, the initial feature is used as the airfoil feature; and if it does not meet the condition, the preset airfoil feature extraction model is optimized based on the first error, the optimized preset airfoil feature extraction model is used as a new preset airfoil feature extraction model, and the updated image data is used as new initial image data. The process then jumps to the step of encoding the initial image data based on the preset airfoil feature extraction model, until the current preset airfoil feature extraction model meets the second preset training termination condition, and the current initial feature is used as the airfoil feature.

8. An electronic device, characterized in that, The electronic device includes a processor and a memory; wherein the memory is used to store a computer program, which is loaded and executed by the processor to implement the airfoil flow field prediction method based on airfoil feature extraction and physical information neural network as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, Used to store a computer program, which, when executed by a processor, implements the airfoil flow field prediction method based on airfoil feature extraction and physical information neural network as described in any one of claims 1 to 6.

Images