A method for predicting aerodynamic parameters of airfoil flow field based on point cloud network

Abstract

This invention belongs to the fields of computational fluid dynamics and deep learning, specifically relating to a method for predicting aerodynamic parameters of airfoil flow fields based on point cloud networks. It directly uses the airfoil flow field grid coordinates from CFD calculations as input and the flow field parameters corresponding to the grid points as output. Compared to convolutional neural networks, point cloud neural networks do not require image processing of the data, avoiding the introduction of human error and simplifying the data preprocessing steps. Furthermore, grid-based CFD calculation data naturally possesses point cloud characteristics and can be directly used in point cloud networks. In addition, to avoid excessively large input and output dimensions in the point cloud network, the airfoil flow field grid is downsampled, and then the gappy POD method is used to fill in the aerodynamic parameters of unsampled points in the flow field based on the predicted values ​​of the flow field aerodynamic parameters from the sampled grid points.

Description

Technical Field

[0001] This invention belongs to the fields of computational fluid dynamics and deep learning, and specifically relates to a method for predicting aerodynamic parameters of airfoil flow fields based on point cloud networks. Background Technology

[0002] Predicting airfoil flow fields is crucial in aerospace, wind power generation, and other fields. An airfoil flow field refers to the flow field formed by air or other fluids flowing around an airfoil. Accurate calculation of airfoil flow fields is essential for aircraft design, aerodynamic performance optimization, and flight safety. Traditional flow field solution methods primarily rely on Computational Fluid Dynamics (CFD) techniques. CFD solves fluid dynamics equations (such as the Navier-Stokes equations) through numerical simulation to simulate fluid flow around an airfoil. However, while accurate, CFD methods are computationally expensive and time-consuming, especially when dealing with complex geometries, where computational efficiency further decreases. Traditional CFD methods typically employ numerical methods such as the Finite Volume Method (FVM), Finite Element Method (FEM), or Finite Difference Method (FDM), which require discretizing the flow path into a mesh structure. To capture the flow field details at complex surfaces and walls, mesh refinement operations are required at the airfoil's leading and trailing edges and walls. These mesh generation and refinement steps themselves also require significant computational resources and time.

[0003] In recent years, with the development of machine learning technology, its application in flow field prediction has gradually attracted attention. Machine learning methods train neural networks to learn flow field characteristics from large amounts of flow field data, thereby achieving efficient flow field prediction. Compared with traditional CFD methods, machine learning methods have advantages such as fast computation speed and strong adaptability. Currently, in the publicly available literature, the main machine learning-based flow field prediction methods are image-processing-based convolutional neural networks (CNNs). In airfoil flow field prediction based on CNNs, airfoil data is usually pixelated and converted into an image before being used as input to the neural network. The process of converting data such as airfoil coordinates into image pixels introduces human error. To enable the CNN to capture minute deformations of the airfoil, the resolution of the input image needs to be increased. However, a higher resolution input image means more parameters to be trained on the neural network, resulting in longer training time. Secondly, to use the airfoil geometry as input to the CNN, the airfoil boundary and background are usually set to different pixel values, or the pixels in the airfoil's internal region are masked, such as setting all pixel values ​​in the airfoil's internal region to 0, to distinguish the airfoil boundary. This processing method results in not all pixels in the input image being used to describe the airfoil geometry, wasting the computational power of the convolutional neural network. Finally, in the CFD computational domain, regions of interest to researchers, such as near the wall and the airfoil leading edge, undergo mesh refinement to obtain more accurate calculation results. However, if pixelation is performed, every region in the computational domain will be treated "equally," divided into equally spaced pixels. This completely wastes the mesh refinement operation used in CFD computations to improve computational accuracy. Summary of the Invention

[0004] The purpose of this invention is to address the problems encountered by convolutional neural networks in airfoil flow field prediction. To overcome the shortcomings of existing technologies, this application proposes a method for predicting aerodynamic parameters of airfoil flow fields based on point cloud networks. This method directly uses the airfoil flow field grid coordinates from CFD calculations as input and the flow field parameters corresponding to the grid points as output. Compared to convolutional neural networks, point cloud neural networks do not require image processing of the data, avoiding the introduction of human error and simplifying the data preprocessing steps. Furthermore, grid-based CFD calculation data naturally possesses point cloud characteristics and can be directly used in point cloud networks. In addition, to avoid excessively large input and output dimensions in the point cloud network, the airfoil flow field grid is downsampled. Then, using the gappy POD method, the predicted aerodynamic parameters of the flow field at unsampled points are filled in based on the sampled grid point aerodynamic parameter predictions.

[0005] The technical solution of this invention: To achieve the above-mentioned objective, this invention adopts the following technical solution: A method for predicting aerodynamic parameters of an airfoil flow field based on point cloud networks, the specific steps of which include:

[0006] Step 1: Select an initial airfoil and establish a two-dimensional coordinate system Oxy, where the origin O is the leading edge of the airfoil, the Ox axis passes through the trailing edge of the airfoil and points towards the trailing edge, and the Oy axis is perpendicular to the Ox axis and points upward; combine the airfoil parameterization method and the random sampling method to generate N random airfoils based on the initial airfoil to form an airfoil library;

[0007] Step 2: Set the flow field conditions and obtain the flow field aerodynamic parameters of each airfoil in the airfoil library based on CFD simulation calculation;

[0008] Step 3: Sample Ns grid points from Nt grid points in the airfoil flow field; the sampling rule is: if the i-th (1≤i≤Nt) grid point is an airfoil surface grid point, then it is sampled; otherwise, the probability of it being sampled is related to the distance D from that point to the airfoil surface. i Satisfies a normal distribution:

[0009]

[0010] in D i Let be the distance from the i-th (1≤i≤Nt) grid point to the airfoil surface;

[0011] Step 4: For all airfoils in the airfoil library, use the coordinates of Ns sampling grid points of the airfoil flow field as input data and the aerodynamic parameters of the flow field of the sampling grid points as output data to establish a neural network dataset;

[0012] Step 5: Build a point cloud network, then randomly select 80% of the dataset from Step 4 as the training dataset, and the remaining 20% ​​as the test dataset. Use the training dataset to train the point cloud network, and use the test dataset to test and verify the prediction performance of the point cloud network.

[0013] Step 6: Select an airfoil from the test dataset, use the coordinates of the airfoil flow field sampling grid points as input, input the trained point cloud network, and use the network output value as the predicted value of the aerodynamic parameters of the airfoil flow field sampling grid points.

[0014] Preferably, the prediction method further includes the following steps:

[0015] Step 7: Create a matrix P with 0.8×N rows and Ns×2+Nt×n columns. Each row of matrix P represents the flow field data of an airfoil sample in the training set.

[0016] The first to Ns×2 columns of matrix P are the horizontal and vertical coordinates of Ns sampling grid points, and the Ns×2+1 to Ns×(2+n) columns are the aerodynamic parameters of the airfoil flow field corresponding to Ns sampling grid points;

[0017] The Ns×(2+n)+1 column to the last column represents the n aerodynamic parameters corresponding to the non-sampled grid points of the airfoil flow field.

[0018] Create a row vector Pr, where the data in Pr is the airfoil flow field data to be predicted. The first to Ns×2 columns are the horizontal and vertical coordinates of the Ns sampling grid points in the airfoil flow field to be predicted. The Ns×2+1 to Ns×(2+n) columns are the flow field parameters of the Ns sampling grid points in the airfoil flow field to be predicted, which are obtained through prediction in step 6. The Ns×(2+n)+1 to the last column are the n aerodynamic parameters corresponding to the non-sampling grid points in the airfoil flow field to be predicted, which are obtained by calculating the average value of the corresponding columns of matrix P.

[0019] Step 8: Using the eigenorthogonal decomposition method, calculate the basic mode matrix Pc, basic mode coefficients Ps, and mean vector Pm of the combination matrix [P; Pr]; the basic mode matrix Pc, basic mode coefficients Ps, and mean vector Pm satisfy [P; Pr] = Ps × Pc T +Pm;

[0020] Step 9: Update Pr using formula (2):

[0021]

[0022] Where b is the principal component coefficient vector;

[0023] Step 10: Repeat step 9 to iteratively update Pr until the maximum number of iterations E is reached. The Ns×2+1 columns of Pr up to the last column are the aerodynamic parameters corresponding to the flow field grid points of the airfoil to be predicted.

[0024] Preferably, in step 1, the airfoil parameterization method is a common parametric method, such as the free deformation method or the function-like shape function method.

[0025] Preferably, in step 1, the method of generating a random airfoil library based on the initial airfoil by combining the airfoil parameterization method and the random sampling method is as follows: the initial airfoil is parameterized by applying the airfoil parameterization method, and then N sets of geometric parameters are selected in the parameter space by random sampling. The initial airfoil is deformed according to the geometric parameters, and the set of N new airfoils after deformation is used as the airfoil library.

[0026] Preferably, in step 2, the aerodynamic parameters of the airfoil flow field can be aerodynamic pressure coefficient, Mach number, velocity, or other aerodynamic parameters of the airfoil flow field that can be obtained through CFD simulation calculation.

[0027] Preferably, in step 4, the input data is in the format of an N×Ns×2 matrix, meaning there are N input samples, each sample is an Ns-row, 2-column matrix, and each row of data represents the horizontal and vertical coordinates of a sampling grid point. The output data is in the format of an N×Ns×n matrix, meaning there are N output samples, each sample is an Ns-row, n-column matrix, and each row of data represents the n flow field aerodynamic parameters corresponding to a sampling grid point.

[0028] Preferably, in step 5, the point cloud network structure is as follows: an Ns×2 input layer, a 64×64 multilayer perceptron 1, a 64×64 multilayer perceptron 2, a 64×128 multilayer perceptron 3, a max-pooling layer 1, a merged layer of multilayer perceptron 2 and max-pooling layer 1, a 128×128 multilayer perceptron 4, a 128×n multilayer perceptron 5, and an N×n output layer. Here, Ns represents the number of sampling grid points for the airfoil flow field, and n represents the number of aerodynamic parameters for the airfoil flow field.

[0029] The advantages of this invention are as follows: In the method of this application, the trained point cloud network can quickly predict the aerodynamic parameters of the airfoil flow field. Compared with traditional CFD solution methods, the computation time is reduced from several minutes to 1 second, achieving a speed improvement of hundreds of times. Compared with traditional convolutional neural networks, this method does not involve data pixelation or data dimensionality reduction, preserving the accuracy of CFD calculation results and avoiding the introduction of human error. The method of this application provides a new approach for the rapid prediction of aerodynamic parameters of the airfoil flow field.

[0030] Compared to a single point cloud network, the method in this application does not predict the aerodynamic parameters of all grid points in the flow field. Instead, it models and predicts the aerodynamic parameters only for sampled grid points, and then combines the gappy POD method to calculate the aerodynamic parameters of non-sampled grid points based on the aerodynamic parameters of the sampled grid points. This reduces the data dimensionality of the point cloud network and improves the training speed. Attached Figure Description

[0031] Figure 1 This is a schematic diagram of the initial airfoil and the random airfoil library. The solid line represents the initial airfoil, denoted as RAE2822, and the dark area around the initial airfoil represents the coordinate range of the random airfoil library, denoted as airfoil library.

[0032] Figure 2 This is a schematic diagram of the process structure of the method in this application, where MLP refers to a multilayer perceptron, and the (64,128) numerical label after MLP indicates that the size of the first fully connected layer in the multilayer perceptron is 64, the size of the second fully connected layer is 128, and the meanings of other numerical labels are similar.

[0033] Figure 3 This is a comparison between the predicted values ​​of the airfoil surface pressure coefficient distribution obtained by the method described in this application and the CFD calculated values; Figure 3 (a) represents the CFD calculated value. Figure 3 (b) represents the predicted value. Figure 3 (c) represents the prediction error. Detailed Implementation

[0034] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0035] The features and illustrative embodiments of various aspects of the present invention will now be described in detail. Numerous specific details are set forth in the following detailed description to provide a thorough understanding of the invention. However, it will be apparent to those skilled in the art that the invention may be practiced without requiring some of these specific details. The following description of embodiments is merely intended to provide a better understanding of the invention by illustrating examples of the invention. The invention is by no means limited to any specific setups and methods set forth below, but covers any improvements, substitutions, and modifications to structures, methods, and devices without departing from the spirit of the invention. Well-known structures and techniques are not shown in the drawings and the following description to avoid unnecessarily obscuring the invention.

[0036] It should be noted that, unless otherwise specified, the embodiments of the present invention and the features thereof can be combined with each other, and the various embodiments can be referenced and cited in each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0037] Example 1

[0038] A method for predicting aerodynamic parameters of airfoil flow field based on point cloud networks, comprising the following steps:

[0039] Step 1: Select the RAE2822 airfoil as the initial airfoil and establish a two-dimensional coordinate system Oxy, where the origin O is the leading edge of the airfoil, the Ox axis passes through the trailing edge and points towards it, and the Oy axis is perpendicular to the Ox axis and points upwards. Parameterize the initial airfoil based on the free deformation method. The parameterization method is as follows: use the Latin hypercube sampling method to generate 250 sets of control point displacement values ​​for the free deformation method, deform the initial airfoil, and the set of 250 new airfoils after deformation is used as a random airfoil library. A schematic diagram of the initial RAE2822 and the random airfoil library is shown below. Figure 1 As shown.

[0040] Step 2: Set the inflow Mach number to 0.45 and the angle of attack to 2°. Based on CFD simulation, obtain the flow field pressure coefficient of each airfoil in the random airfoil library.

[0041] Step 3: Sample 2048 grid points from the 28560 grid points in the airfoil flow field. The sampling rule is: if the i-th (1≤i≤28560) grid point is a grid point on the airfoil surface, then it is sampled; otherwise, the probability of it being sampled is related to the distance D from that point to the airfoil surface. i Satisfies a normal distribution:

[0042]

[0043] in D i Let be the distance from the i-th (1≤i≤Nt) grid point to the airfoil surface.

[0044] Step 4: For all 250 airfoils in the airfoil library, a neural network dataset is established using the coordinates of 28,560 sampling grid points in the airfoil flow field as input data and the aerodynamic parameters of the sampling grid points as output data. The input data is a 250×2048×2 matrix, meaning there are 250 input samples, each a 2048-row, 2-column matrix, with each row representing the x and y coordinates of a sampling grid point. The output data is a 250×2048×1 matrix, meaning there are 250 output samples, each a 2048-row, 1-column matrix, with each row representing one aerodynamic parameter corresponding to a sampling grid point: the flow field pressure coefficient.

[0045] Step 5: Establish the point cloud network. The structure of the point cloud network is as follows: Figure 2 As shown, the point cloud network structure is as follows: a 2048×2 input layer, a 64×64 multilayer perceptron 1, a 64×64 multilayer perceptron 2, a 64×128 multilayer perceptron 3, a max-pooling layer 1, a merged layer of multilayer perceptron 2 and max-pooling layer 1, a 128×128 multilayer perceptron 4, a 128×1 multilayer perceptron 5, and a 2048×1 output layer. Then, 80% of the dataset from step 4 is randomly selected as the training dataset, and the remaining 20% ​​is used as the test dataset. The point cloud network is trained using the training dataset, and the test dataset is used to test and verify the prediction performance of the point cloud network. The point cloud network uses the mean squared error (MSE) as the loss function, and MSE is defined as:

[0046]

[0047] Where n is the number of data samples; y i It is the true value of the i-th data sample; This is the predicted value of the i-th data sample. The optimized training method is the Adam algorithm.

[0048] Step 6: Select an airfoil from the test dataset, use the coordinates of 2048 sampling grid points of the airfoil flow field as input, input the trained point cloud network, and use the network output value as the predicted pressure coefficient value of the sampling grid points of the airfoil flow field.

[0049] Step 7: Create a matrix P with 0.8×250 rows and 2048×2+28560×1 columns. Each row of matrix P represents the flow field data of an airfoil sample in the training set. Columns 1 to 2048×2 of matrix P are the x and y coordinates of 2048 sampling grid points. Columns 2048×2+1 to 2048×(2+1) are the airfoil flow field pressure coefficients corresponding to the 2048 sampling grid points. Columns 2048×(2+1)+1 to the last column are the pressure coefficients corresponding to the non-sampling grid points of the airfoil flow field. Create a row vector Pr, where the data of Pr is the pressure coefficient of the airfoil flow field to be predicted. Columns 1 to 2048×2 are the horizontal and vertical coordinates of 2048 sampling grid points in the airfoil flow field to be predicted. Columns 2048×2+1 to 2048×(2+1) are the flow field pressure coefficients of the 2048 sampling grid points in the airfoil flow field to be predicted, obtained through prediction in step 6. Columns 2048×(2+1)+1 to the last column are the pressure coefficients corresponding to the non-sampling grid points in the airfoil flow field to be predicted, obtained by calculating the average value of the corresponding columns in matrix P.

[0050] Step 8: Using the eigenorthogonal decomposition method, calculate the basic mode matrix Pc, basic mode coefficients Ps, and mean vector Pm of the combination matrix [P; Pr]. The basic mode matrix Pc, basic mode coefficients Ps, and mean vector Pm satisfy [P; Pr] = Ps × Pc T +Pm, specifically:

[0051]

[0052] Among them, P i，j It is the element in the i-th row and j-th column of matrix P; Ps i，j It is the element in the i-th row and j-th column of matrix Ps; Pc i，j It is the element in the i-th row and j-th column of matrix Pc; Pm i It is the i-th element of the mean vector Pm;

[0053] Step 9: Update Pr using formula (2):

[0054]

[0055] Where b is the principal component coefficient vector.

[0056] Step 10: Repeat Step 9 to iteratively update Pr until the maximum iteration step 50 is reached. The last column of Pr (2048×2+1) represents the pressure coefficient corresponding to the flow field grid point of the airfoil to be predicted. The prediction of the airfoil flow field pressure coefficient is completed.

[0057] Example 2

[0058] A method for predicting aerodynamic parameters of airfoil flow field based on point cloud networks. The steps include:

[0059] Step 1: Select the RAE2822 airfoil as the initial airfoil and establish a two-dimensional coordinate system Oxy, where the origin O is the leading edge of the airfoil, the Ox axis passes through the trailing edge and points towards it, and the Oy axis is perpendicular to the Ox axis and points upwards. Parameterize the initial airfoil based on the free deformation method. The parameterization method is as follows: use the Latin hypercube sampling method to generate 250 sets of control point displacement values ​​for the free deformation method, deform the initial airfoil, and the set of 250 new airfoils after deformation is used as a random airfoil library. A schematic diagram of the initial RAE2822 and the random airfoil library is shown below. Figure 1 As shown.

[0060] Step 2: Set the inflow Mach number to 0.5 and the angle of attack to 1.74°. Based on CFD simulation, obtain the flow field pressure coefficient of each airfoil in the random airfoil library.

[0061] Step 3: Sample 2048 grid points from the 28560 grid points in the airfoil flow field. The sampling rule is: if the i-th (1≤i≤28560) grid point is a grid point on the airfoil surface, then it is sampled; otherwise, the probability of it being sampled is related to the distance D from that point to the airfoil surface. i Satisfies a normal distribution:

[0062]

[0063] in D i Let be the distance from the i-th (1≤i≤Nt) grid point to the airfoil surface.

[0064] Step 4: For all 250 airfoils in the airfoil library, a neural network dataset is established using the coordinates of 28,560 sampling grid points in the airfoil flow field as input data and the aerodynamic parameters of the sampling grid points as output data. The input data is a 250×2048×2 matrix, meaning there are 250 input samples, each a 2048-row, 2-column matrix, with each row representing the x and y coordinates of a sampling grid point. The output data is a 250×2048×1 matrix, meaning there are 250 output samples, each a 2048-row, 1-column matrix, with each row representing one aerodynamic parameter corresponding to a sampling grid point: the flow field pressure coefficient.

[0065] Step 5: Establish the point cloud network. The structure of the point cloud network is as follows: Figure 2 As shown, the point cloud network structure is as follows: a 2048×2 input layer, a 64×64 multilayer perceptron 1, a 64×64 multilayer perceptron 2, a 64×128 multilayer perceptron 3, a max-pooling layer 1, a merged layer of multilayer perceptron 2 and max-pooling layer 1, a 128×128 multilayer perceptron 4, a 128×1 multilayer perceptron 5, and a 2048×1 output layer. Then, 80% of the dataset from step 4 is randomly selected as the training dataset, and the remaining 20% ​​is used as the test dataset. The point cloud network is trained using the training dataset, and the test dataset is used to test and verify the prediction performance of the point cloud network. The point cloud network uses the mean squared error (MSE) as the loss function, and MSE is defined as:

[0066]

[0067] Where n is the number of data samples; y i It is the true value of the i-th data sample; This is the predicted value of the i-th data sample. The optimized training method is the Adam algorithm.

[0068] Step 6: Select an airfoil from the test dataset, use the coordinates of 2048 sampling grid points of the airfoil flow field as input, input the trained point cloud network, and use the network output value as the predicted pressure coefficient value of the sampling grid points of the airfoil flow field.

[0069] Step 7: Create a matrix P with 0.8×250 rows and 2048×2+28560×1 columns. Each row of matrix P represents the flow field data of an airfoil sample in the training set. Columns 1 to 2048×2 of matrix P are the x and y coordinates of 2048 sampling grid points. Columns 2048×2+1 to 2048×(2+1) are the airfoil flow field pressure coefficients corresponding to the 2048 sampling grid points. Columns 2048×(2+1)+1 to the last column are the pressure coefficients corresponding to the non-sampling grid points of the airfoil flow field. Create a row vector Pr, where the data of Pr is the pressure coefficient of the airfoil flow field to be predicted. Columns 1 to 2048×2 are the horizontal and vertical coordinates of 2048 sampling grid points in the airfoil flow field to be predicted. Columns 2048×2+1 to 2048×(2+1) are the flow field pressure coefficients of the 2048 sampling grid points in the airfoil flow field to be predicted, obtained through prediction in step 6. Columns 2048×(2+1)+1 to the last column are the pressure coefficients corresponding to the non-sampling grid points in the airfoil flow field to be predicted, obtained by calculating the average value of the corresponding columns in matrix P.

[0070] Step 8: Using the eigenorthogonal decomposition method, calculate the basic mode matrix Pc, basic mode coefficients Ps, and mean vector Pm of the combination matrix [P; Pr]. The basic mode matrix Pc, basic mode coefficients Ps, and mean vector Pm satisfy [P; Pr] = Ps × Pc T +Pm, specifically:

[0071]

[0072] Among them, P i，j It is the element in the i-th row and j-th column of matrix P; Ps i，j It is the element in the i-th row and j-th column of matrix Ps; Pc i，j It is the element in the i-th row and j-th column of matrix Pc; Pm i It is the i-th element of the mean vector Pm;

[0073] Step 9: Update Pr using formula (2):

[0074]

[0075] Where b is the principal component coefficient vector.

[0076] Step 10: Repeat Step 9 to iteratively update Pr until the maximum iteration step 50 is reached. The last column of Pr (2048×2+1) represents the pressure coefficient corresponding to the flow field grid point of the airfoil to be predicted. The prediction of the airfoil flow field pressure coefficient is completed.

[0077] The specific embodiments of the present invention have been described above. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention. These improvements and modifications should also be within the scope of protection of the present invention.

Claims

1. A method for predicting aerodynamic parameters of an airfoil flow field based on point cloud networks, characterized in that, The specific steps include: Step 1: Select an initial airfoil, and generate N random airfoils to form an airfoil library based on the initial airfoil, using airfoil parameterization and random sampling methods; Step 2: Set flow field conditions, and obtain the flow field aerodynamic parameters of each airfoil in the airfoil library based on CFD simulation calculations; Step 3: Sample Ns grid points from Nt grid points in the flow field of each airfoil; Step 4: For all airfoils in the airfoil library, use the coordinates of the Ns sampled grid points of the airfoil flow field as input data, and the flow field aerodynamic parameters of the sampled grid points as output data to establish a neural network dataset; Step 5: Establish a point cloud network and train the point cloud network using the neural network dataset; Step 6: Select an airfoil from the test dataset, use the coordinates of the sampled grid points of the airfoil flow field as input, input the trained point cloud network, and use the network's output value as the predicted value of the aerodynamic parameters of the sampled grid points of the airfoil flow field; Step 7: Create 0.8×N A matrix P with Ns×2+Nt×n rows and Ns×2+Nt×n columns, where each row of matrix P represents the flow field data of an airfoil sample in the training set; columns 1 to Ns×2 of matrix P represent the x and y coordinates of Ns sampling grid points; columns Ns×2+1 to Ns×(2+n) represent the aerodynamic parameters of the airfoil flow field corresponding to the Ns sampling grid points; columns Ns×(2+n)+1 to the last column represent the n aerodynamic parameters corresponding to the non-sampling grid points of the airfoil flow field; a row vector Pr is created. The data is the airfoil flow field data to be predicted. Among them, the first to Ns×2 columns are the horizontal and vertical coordinates of the Ns sampling grid points of the airfoil flow field to be predicted. The Ns×2+1 to Ns×(2+n) columns are the flow field parameters of the Ns sampling grid points in the airfoil flow field to be predicted, which are obtained through prediction in step 6. The Ns×(2+n)+1 to the last column are the n aerodynamic parameters corresponding to the non-sampling grid points of the airfoil flow field to be predicted, which are obtained by calculating the average value of the corresponding columns of matrix P. Step 8: Using the eigenorthogonal decomposition method, calculate the basic mode matrix Pc, basic mode coefficients Ps, and mean vector Pm of the combination matrix [P; Pr]; The basic mode matrix Pc, basic mode coefficients Ps, and mean vector Pm satisfy [P; Pr] = Ps × PcT + Pm; Step 9: Update Pr using formula (2): Where b is the principal component coefficient vector; Step 10: Repeat step 9 to iteratively update Pr until the maximum number of iterations E is reached. The Ns×2+1 columns of Pr up to the last column are the aerodynamic parameters corresponding to the flow field grid points of the airfoil to be predicted.

2. The method for predicting aerodynamic parameters of an airfoil flow field based on point cloud networks according to claim 1, characterized in that, In step 1, the airfoil parameterization methods include the free deformation method and the function-like shape function method.

3. The method for predicting aerodynamic parameters of an airfoil flow field based on point cloud networks according to claim 1, characterized in that, In step 1, the method of generating a random airfoil library based on the initial airfoil, which combines the airfoil parameterization method and the random sampling method, is as follows: the initial airfoil is parameterized by applying the airfoil parameterization method, and then N sets of geometric parameters are selected in the parameter space by random sampling. The initial airfoil is deformed according to the geometric parameters, and the set of N new airfoils after deformation is used as the airfoil library.

4. The method for predicting aerodynamic parameters of an airfoil flow field based on point cloud networks according to claim 1, characterized in that, In step 2, the sampling rule is as follows: if the i-th grid point is an airfoil surface grid point, 1≤i≤Nt, then it is sampled; otherwise, the probability of it being sampled and the distance Di from that point to the airfoil surface satisfy a normal distribution. in Di is the distance from the i-th grid point to the airfoil surface, 1≤i≤Nt.

5. The method for predicting aerodynamic parameters of an airfoil flow field based on point cloud networks according to claim 1, characterized in that, In step 5, 80% of the dataset from step 4 is randomly selected as the training dataset, and the remaining 20% ​​is used as the test dataset. The point cloud network is trained using the training dataset, and the test dataset is used to test and verify the prediction performance of the point cloud network.

6. The method for predicting aerodynamic parameters of an airfoil flow field based on point cloud networks according to claim 1, characterized in that, The aerodynamic parameters of the airfoil flow field include aerodynamic pressure coefficient, Mach number, velocity, or other aerodynamic parameters of the airfoil flow field that can be obtained through CFD simulation calculation.

7. The method for predicting aerodynamic parameters of an airfoil flow field based on point cloud networks according to claim 1, characterized in that, In step 4, the input data is in the format of an N×Ns×2 matrix, that is, there are N input samples, each sample is an Ns-row, 2-column matrix, and each row of data represents the horizontal and vertical coordinates of a sampling grid point; the output data is in the format of an N×Ns×n matrix, that is, there are N output samples, each sample is an Ns-row, n-column matrix, and each row of data represents the n flow field aerodynamic parameters corresponding to a sampling grid point.

8. The method for predicting aerodynamic parameters of an airfoil flow field based on point cloud networks according to claim 1, characterized in that, In step 5, the point cloud network structure is as follows: an Ns×2 input layer, a 64×64 multilayer perceptron 1, a 64×64 multilayer perceptron 2, a 64×128 multilayer perceptron 3, a max pooling layer 1, a merged layer of multilayer perceptron 2 and max pooling layer 1, a 128×128 multilayer perceptron 4, a 128×n multilayer perceptron 5, and an N×n output layer; where Ns is the number of sampling grid points for the airfoil flow field, and n is the number of aerodynamic parameters for the airfoil flow field.

Images