Point cloud neural network-based airfoil profile inverse design method

Abstract

The invention relates to the technical field of aerospace engineering, in particular to an airfoil profile inverse design method based on a point cloud neural network. Airfoil surface pressure distribution in CFD calculation is directly used as input, and airfoil surface grid coordinates are used as output. Compared with a convolutional neural network, the point cloud neural network does not need to carry out image processing on the data, artificial errors are prevented from being introduced, and the steps of data preprocessing are simplified. Meanwhile, grid-based CFD calculation data naturally has point cloud characteristics and can be directly used for a point cloud network.

Description

Technical Field

[0001] The present invention belongs to the field of aerospace engineering technology, and specifically is an airfoil inverse design method based on point cloud neural network. Background Art

[0002] Airfoil design plays a vital role in aerospace engineering, and its performance directly affects the aerodynamic characteristics of the aircraft. The goal of airfoil design is to optimize its geometry to achieve the best lift, drag, and stability under specific flight conditions. Among the many design parameters, pressure distribution is a key performance indicator, which reflects the distribution of airflow on the airfoil surface, thereby affecting lift and drag. Therefore, the inverse design method, that is, taking pressure distribution as input and reversely deriving the geometry of the airfoil, has important engineering application value.

[0003] Traditional airfoil design methods mainly rely on empirical formulas and test data, which have certain limitations when dealing with complex flow conditions. Specifically, the limitations of traditional methods are mainly reflected in the following aspects: 1. Limitations of empirical formulas: Empirical formulas are summarized based on a large amount of test data and are suitable for airfoil design under specific conditions. However, these formulas are often not accurate enough when dealing with complex flow conditions and are difficult to meet the requirements of modern aircraft for high performance and multi-function. For example, although the NACA series airfoils perform well under low-speed flight conditions, they may not provide sufficient lift at high-speed flight. 2. Limitations of test data: Test data is an important basis for airfoil design, but obtaining test data usually requires a lot of time and cost. In addition, test data are often limited by test conditions and equipment, and it is difficult to fully reflect the performance of airfoils under different flight conditions. For example, although wind tunnel tests can provide detailed flow field information, their results may be limited by wind tunnel size and boundary conditions. 3. Limited processing capabilities for complex flow conditions: Traditional airfoil design methods have certain limitations when dealing with complex flow conditions. For example, during transonic and supersonic flight, shock waves and separation flows will be generated on the airfoil surface, and these phenomena are difficult to accurately predict and control through empirical formulas and experimental data.

[0004] With the development of computational fluid dynamics (CFD) and machine learning technology, data-driven airfoil inverse design methods have gradually become a research hotspot. These methods can achieve efficient and accurate airfoil design by combining CFD simulation and machine learning technology. Specifically, data-driven airfoil inverse design methods have the following advantages: 1. Efficiency: Data-driven methods can quickly generate airfoil geometries that meet specific aerodynamic performance requirements by learning a large amount of CFD simulation data. For example, by training a deep learning model, designers can generate multiple candidate airfoils in a short time, thereby improving design efficiency. 2. Accuracy: Data-driven methods can learn complex nonlinear relationships to generate more accurate airfoil geometries. For example, by learning the relationship between pressure distribution and geometry, the model can generate airfoils with good aerodynamic performance. Common machine learning methods currently include feedforward neural networks and convolutional neural networks. The former is inefficient when the data dimension is too large, and the data needs to be reduced in dimension, which introduces artificial errors in the process of dimensionality reduction. The latter usually requires the data to be visualized, and a very high image resolution is required to accurately represent the airfoil. At the same time, most of the image is blank, and only a small number of pixels are airfoils, wasting the performance of the neural network. Summary of the invention

[0005] Purpose of the present invention: In order to solve the problems of the above traditional machine learning methods in airfoil inverse design, in view of the deficiencies in the prior art, this application proposes an airfoil inverse design method based on a point cloud neural network, which directly uses the airfoil surface pressure distribution in the CFD calculation as input and uses the airfoil surface grid coordinates as output. Compared with convolutional neural networks, point cloud neural networks do not need to image the data, avoid the introduction of artificial errors, and simplify the data preprocessing steps. At the same time, the grid-based CFD calculation data naturally has point cloud characteristics and can be directly used in point cloud networks.

[0006] Technical solution of the present invention: In order to achieve the above-mentioned invention object, the present invention adopts the following technical solution: an airfoil inverse design method based on point cloud neural network, the specific steps include:

[0007] Step 1: Select an initial airfoil and establish a two-dimensional coordinate system Oxy based on the initial airfoil, where the origin O of the coordinate system is the leading edge of the airfoil, the coordinate system Ox axis passes through the trailing edge of the airfoil and points to the trailing edge of the airfoil, and the coordinate system Oy axis is perpendicular to the Ox axis and upward. Then, N random airfoils are generated based on the initial airfoil to form an airfoil library through airfoil parameterization method and random sampling method;

[0008] Step 2: Set the airfoil inverse design conditions and flow field conditions, and obtain the surface pressure coefficient distribution of each airfoil in the airfoil library based on CFD simulation calculation;

[0009] Step 3: Establish the neural network input data set Input and output data set Ouput; the input data set Input is Ns N×2 matrices; where Ns is the number of airfoils in the airfoil library, and N is the number of coordinate points on each airfoil surface; each N×2 matrix is ​​the horizontal coordinate x of the airfoil surface coordinate point and the corresponding surface pressure coefficient Cp; the output data set Output is Ns N×2 matrices; each N×2 matrix is ​​the horizontal coordinate x of the airfoil surface coordinate point and the normalized vertical coordinate

[0010] Step 4: Establish a point cloud neural network, and then randomly select 80% of the input data set and output data set in step 3 as the training data set, and the remaining 20% ​​of the data set as the test data set. Use the training data set to train the point cloud network, and the test data set is used to test and verify the prediction performance of the point cloud network;

[0011] Step 5: Select an airfoil from the test data set, use the horizontal coordinates of the airfoil surface coordinate points and the corresponding pressure coefficient distribution as input, input the trained point cloud network, use the network output value as the horizontal and vertical coordinates of the inverse-designed airfoil surface coordinate points, and complete the inverse design of the airfoil.

[0012] Preferably, in step 1, the airfoil parameterization method is a common parameter method, such as a free deformation method, a quasi-function shape function method, etc.

[0013] Preferably, in step 1, the method for generating a random airfoil library based on the initial airfoil by combining the airfoil parameterization method and the random sampling method is to parameterize the initial airfoil by applying the airfoil parameterization method, and then select Ns groups of geometric parameters by random sampling in the parameter space, deform the initial airfoil according to the geometric parameters, and use the set of Ns new airfoils after deformation as the airfoil library.

[0014] Preferably, in step 2, the airfoil inverse design condition is a fixed lift coefficient, and the flow field condition is the flow field conditions such as the Mach number, angle of attack, pressure and temperature under the lift coefficient used for CFD simulation calculation.

[0015] Preferably, based on CFD simulation calculation, the specific process of obtaining the surface pressure coefficient distribution of each airfoil in the airfoil library includes:

[0016] The airfoil flow field is meshed, and the airfoil lift coefficient, flow field Mach number, angle of attack, temperature, pressure and other parameters are set. The flow field is solved using computational fluid dynamics software, and then the airfoil surface pressure coefficient distribution is read.

[0017] Preferably, in step 3, the ordinate y of the airfoil surface coordinate point needs to be normalized, and the normalization method is:

[0018]

[0019] Among them, max is the maximum value of the vertical coordinates of all airfoils, and min is the minimum value of the vertical coordinates of all airfoils. is the normalized vertical coordinate.

[0020] Preferably, in step 4, the structure of the point cloud network is: N×2 input layer, 64×64 multilayer perceptron 1, 64×64 multilayer perceptron 2, 64×128 multilayer perceptron 3, maximum pooling layer 1, merging layer of multilayer perceptron 2 and maximum pooling layer 1, 128×128 multilayer perceptron 4, 128×2 multilayer perceptron 5, N×2 output layer; where N is the number of coordinate points on the airfoil surface. The mathematical expression of the multilayer perceptron is:

[0021]

[0022] Among them, x i is the input value of the i-th neuron; w ji is the weight connecting the i-th input neuron and the j-th output neuron; b j is the bias of the jth output neuron.

[0023] Preferably, in step 4, the model uses mean square error MSE as the loss function, and MSE is defined as:

[0024]

[0025] Where n is the number of data samples; y i is the true value of the i-th data sample; is the predicted value of the ith data sample.

[0026] Preferably, in step 4, the optimization training method is the Adam algorithm.

[0027] The advantages of the present invention are: in the method of the present application, the trained point cloud network can quickly and accurately complete the inverse design of the airfoil, and quickly predict the corresponding airfoil shape according to the airfoil pressure distribution curve, and the prediction time is less than 1 second. Compared with the traditional convolutional neural network, this method does not involve pixelation or data dimension reduction of the data, avoiding the introduction of artificial errors. The method of the present application provides a new idea for the rapid inverse design of airfoils. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 Schematic diagram of random airfoil library generated based on initial airfoil;

[0029] Figure 2 is the surface pressure coefficient distribution of the random airfoil library at 0.78 Ma and lift coefficient of 0.8;

[0030] Figure 3 Schematic diagram of the point cloud neural network structure of the present application, where MLP refers to multi-layer perceptron, and the (64,128) digital label after MLP indicates that the size of the first fully connected layer in the multi-layer perceptron is 64, and the size of the second fully connected layer is 128. The meanings of other digital labels are similar;

[0031] Figure 4 and Figure 5 The method described in this application is a comparison between a predicted airfoil and an actual airfoil for inverse design of an airfoil based on the distribution of pressure coefficient on the airfoil surface; wherein the actual airfoil is represented by a solid line and the predicted airfoil is represented by scattered points. DETAILED DESCRIPTION

[0032] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0033] The features and illustrative embodiments of various aspects of the present invention will be described in detail below. In the detailed description below, many specific details are proposed to provide a comprehensive understanding of the present invention. However, it is obvious to those skilled in the art that the present invention can be implemented without the need for some of these specific details. The following description of the embodiments is only to provide a better understanding of the present invention by illustrating examples of the present invention. The present invention is by no means limited to any specific settings and methods proposed below, but covers any improvements, replacements and modifications of structures, methods, devices without departing from the spirit of the present invention. In the accompanying drawings and the following description, known structures and technologies are not shown to avoid unnecessary ambiguity in the present invention.

[0034] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments can be combined with each other, and the embodiments can refer to and quote each other. The present invention will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0035] Example 1

[0036] According to the present invention, an airfoil inverse design method based on point cloud neural network is provided. The steps include:

[0037] Step 1: Select the RAE2822 airfoil as the initial airfoil, and establish a two-dimensional coordinate system Oxy based on the RAE2822 airfoil, where the origin O of the coordinate system is the leading edge of the airfoil, the coordinate system Ox axis passes through the trailing edge of the airfoil and points to the trailing edge of the airfoil, and the coordinate system Oy axis is perpendicular to the Ox axis and upward. Parameterize the initial airfoil based on the free deformation method. Specifically, the Latin hypercube sampling method is used to generate 500 sets of control point displacement values ​​of the free deformation method, and the initial airfoil is deformed. The set of 500 sets of new airfoils after deformation is used as the random airfoil library. The schematic diagram of the random airfoil library generated by the initial RAE2822 is shown in the figure. Figure 1 shown.

[0038] Step 2: Set the incoming flow Mach number of the flow field to 0.78, the pressure to 101325Pa, the temperature to 300K, and the airfoil lift coefficient to 0.8. Based on CFD simulation calculations, obtain the surface pressure coefficient distribution of each airfoil in the random airfoil library. The specific process includes: gridding the airfoil flow field, setting the airfoil lift coefficient, flow field Mach number, angle of attack, temperature, pressure and other parameters, using computational fluid dynamics software to solve the flow field, and then reading the airfoil surface pressure coefficient distribution. The number of coordinate points on each airfoil surface in the CFD calculation is 192. The surface pressure coefficient distribution of each airfoil is obtained as follows: Figure 2 shown.

[0039] Step 3: Create a neural network input data set Input and an output data set Ouput. The input data set Input is 500 192×2 matrices. 500 is the number of airfoils in the airfoil library, and 192 is the number of coordinate points on each airfoil surface. Each 192×2 matrix is ​​the horizontal coordinate x of the airfoil surface coordinate point and the corresponding surface pressure coefficient Cp. The output data set Output is 500 192×2 matrices. Each 192×2 matrix is ​​the horizontal coordinate x of the airfoil surface coordinate point and the normalized vertical coordinate The method for normalizing the ordinate y of the airfoil surface coordinate point is:

[0040]

[0041] Among them, max is the maximum value of the vertical coordinates of all airfoils, and min is the minimum value of the vertical coordinates of all airfoils. is the normalized vertical coordinate.

[0042] Step 4: Establish a point cloud neural network. The structure of the point cloud network is as follows Figure 3As shown. The structure of the point cloud network is: 192×2 input layer, 64×64 multilayer perceptron 1, 64×64 multilayer perceptron 2, 64×128 multilayer perceptron 3, maximum pooling layer 1, merging layer of multilayer perceptron 2 and maximum pooling layer 1, 128×128 multilayer perceptron 4, 128×2 multilayer perceptron 5, 192×2 output layer. The mathematical expression of the multilayer perceptron is:

[0043]

[0044] Among them, x i is the input value of the i-th neuron; w ji is the weight connecting the i-th input neuron and the j-th output neuron; b j is the bias of the jth output neuron.

[0045] Then randomly select 80% of the input data set and output data set of step 3 as the training data set, and the remaining 20% ​​of the data set as the test data set. The training data set is used to train the point cloud neural network, and the test data set is used to test and verify the prediction performance of the point cloud neural network. The optimization training method is the adam algorithm, and the mean square error MSE is used as the loss function. MSE is defined as:

[0046]

[0047] Where n is the number of data samples; y i is the true value of the i-th data sample; is the predicted value of the ith data sample.

[0048] Step 5: Select an airfoil from the test data set, use the horizontal coordinates of the airfoil surface coordinate points and the corresponding pressure coefficient distribution as input, input the trained point cloud neural network, and use the network output value as the horizontal and vertical coordinates of the inversely designed airfoil surface coordinate points. Complete the inverse design of the airfoil. Compare the inversely designed airfoil with the selected airfoil, such as Figure 4 It can be seen that the predicted airfoil of the inverse design is basically consistent with the actual airfoil, indicating that the method described in this application has extremely high prediction accuracy.

[0049] Example 2

[0050] According to the present invention, an airfoil inverse design method based on point cloud neural network is provided. The steps include:

[0051] Step 1: Select the RAE2822 airfoil as the initial airfoil, and establish a two-dimensional coordinate system Oxy based on the RAE2822 airfoil, where the origin O of the coordinate system is the leading edge of the airfoil, the coordinate system Ox axis passes through the trailing edge of the airfoil and points to the trailing edge of the airfoil, and the coordinate system Oy axis is perpendicular to the Ox axis and upward. Parameterize the initial airfoil based on the free deformation method. Specifically, the Latin hypercube sampling method is used to generate 500 sets of control point displacement values ​​of the free deformation method, and the initial airfoil is deformed. The set of 500 sets of new airfoils after deformation is used as the random airfoil library. The schematic diagram of the random airfoil library generated by the initial RAE2822 is shown in the figure. Figure 1 shown.

[0052] Step 2: Set the incoming flow Mach number of the flow field to 0.75, the pressure to 101325Pa, the temperature to 300K, and the airfoil lift coefficient to 0.78. Based on CFD simulation calculation, obtain the surface pressure coefficient distribution of each airfoil in the random airfoil library. The specific process includes: meshing the airfoil flow field, setting the airfoil lift coefficient, flow field Mach number, angle of attack, temperature, pressure and other parameters, using computational fluid dynamics software to solve the flow field, and then read the airfoil surface pressure coefficient distribution. The number of coordinate points on each airfoil surface in the CFD calculation is 192.

[0053] Step 3: Create a neural network input data set Input and an output data set Ouput. The input data set Input is 500 192×2 matrices. 500 is the number of airfoils in the airfoil library, and 192 is the number of coordinate points on each airfoil surface. Each 192×2 matrix is ​​the horizontal coordinate x of the airfoil surface coordinate point and the corresponding surface pressure coefficient Cp. The output data set Output is 500 192×2 matrices. Each 192×2 matrix is ​​the horizontal coordinate x of the airfoil surface coordinate point and the normalized vertical coordinate The method for normalizing the ordinate y of the airfoil surface coordinate point is:

[0054]

[0055] Among them, max is the maximum value of the vertical coordinates of all airfoils, and min is the minimum value of the vertical coordinates of all airfoils. is the normalized vertical coordinate.

[0056] Step 4: Establish a point cloud neural network. The structure of the point cloud network is as follows Figure 3 As shown. The structure of the point cloud network is: 192×2 input layer, 64×64 multilayer perceptron 1, 64×64 multilayer perceptron 2, 64×128 multilayer perceptron 3, maximum pooling layer 1, merging layer of multilayer perceptron 2 and maximum pooling layer 1, 128×128 multilayer perceptron 4, 128×2 multilayer perceptron 5, 192×2 output layer. The mathematical expression of the multilayer perceptron is:

[0057]

[0058] Among them, x i is the input value of the i-th neuron; w ji is the weight connecting the i-th input neuron and the j-th output neuron; b j is the bias of the jth output neuron.

[0059] Then randomly select 80% of the input data set and output data set of step 3 as the training data set, and the remaining 20% ​​of the data set as the test data set. The training data set is used to train the point cloud neural network, and the test data set is used to test and verify the prediction performance of the point cloud neural network. The optimization training method is the adam algorithm, and the mean square error MSE is used as the loss function. MSE is defined as:

[0060]

[0061] Where n is the number of data samples; y i is the true value of the i-th data sample; is the predicted value of the ith data sample.

[0062] Step 5: Select an airfoil from the test data set, use the horizontal coordinates of the airfoil surface coordinate points and the corresponding pressure coefficient distribution as input, input the trained point cloud neural network, and use the network output value as the horizontal and vertical coordinates of the inversely designed airfoil surface coordinate points. Complete the inverse design of the airfoil. Compare the inversely designed airfoil with the selected airfoil, such as Figure 5 It can be seen that the predicted airfoil of the inverse design is basically consistent with the actual airfoil, indicating that the method described in this application has extremely high prediction accuracy.

[0063] The above describes the specific embodiments of the present invention. It should be noted that, for those skilled in the art, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also fall within the scope of protection of the present invention.

Claims

1. A method for inverse design of airfoil based on point cloud neural network, characterized in that: The specific steps include: Step 1: Select an initial airfoil, and generate N random airfoils based on the initial airfoil to form an airfoil library through airfoil parameterization method and random sampling method; Step 2: Set the airfoil inverse design conditions and flow field conditions, and obtain the surface pressure coefficient distribution of each airfoil in the airfoil library based on CFD simulation calculation; Step 3: Establish the neural network input data set Input and output data set Ouput; the input data set Input is Ns N×2 matrices; where Ns is the number of airfoils in the airfoil library, and N is the number of coordinate points on each airfoil surface; each N×2 matrix is ​​the horizontal coordinate x of the airfoil surface coordinate point and the corresponding surface pressure coefficient Cp; the output data set Output is Ns N×2 matrices; each N×2 matrix is ​​the horizontal coordinate x of the airfoil surface coordinate point and the normalized vertical coordinate Step 4: Establish a point cloud neural network, and then randomly select 80% of the input data set and output data set in step 3 as the training data set, and the remaining 20% ​​of the data set as the test data set. Use the training data set to train the point cloud network, and the test data set is used to test and verify the prediction performance of the point cloud network; Step 5: Select an airfoil from the test data set, use the horizontal coordinates of the airfoil surface coordinate points and the corresponding pressure coefficient distribution as input, input the trained point cloud network, use the network output value as the horizontal and vertical coordinates of the inverse-designed airfoil surface coordinate points, and complete the inverse design of the airfoil.

2. The airfoil inverse design method based on point cloud neural network according to claim 1 is characterized in that: In step 1, airfoil parameterization methods include free deformation method and quasi-function shape function method.

3. The airfoil inverse design method based on point cloud neural network according to claim 1 is characterized in that: 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 to parameterize the initial airfoil by applying the airfoil parameterization method, and then select Ns groups of geometric parameters by random sampling in the parameter space, deform the initial airfoil according to the geometric parameters, and the set of Ns new airfoils after deformation is used as the airfoil library.

4. The airfoil inverse design method based on point cloud neural network according to claim 1 is characterized in that: In step 2, the airfoil inverse design condition is a fixed lift coefficient, and the flow field conditions are the flow field conditions such as Mach number, angle of attack, pressure and temperature under the lift coefficient used for CFD simulation calculation.

5. The airfoil inverse design method based on point cloud neural network according to claim 1, characterized in that: In step 3, the ordinate y of the airfoil surface coordinate point needs to be normalized. The normalization method is: Among them, max is the maximum value of the ordinate of all airfoils, min is the minimum value of the ordinate of all airfoils, and y is the normalized ordinate.

6. The airfoil inverse design method based on point cloud neural network according to claim 1, characterized in that: In step 4, the structure of the point cloud network is: N×2 input layer, 64×64 multilayer perceptron 1, 64×64 multilayer perceptron 2, 64×128 multilayer perceptron 3, maximum pooling layer 1, merging layer of multilayer perceptron 2 and maximum pooling layer 1, 128×128 multilayer perceptron 4, 128×2 multilayer perceptron 5, N×2 output layer. N is the number of coordinate points on the airfoil surface. The mathematical expression of the multilayer perceptron is: Among them, x i is the input value of the i-th neuron; W ji is the weight connecting the i-th input neuron and the j-th output neuron; b j is the bias of the jth output neuron.

7. The airfoil inverse design method based on point cloud neural network according to claim 1, characterized in that: In step 4, the model uses mean square error (MSE) as the loss function, and MSE is defined as: Where n is the number of data samples; y i is the true value of the i-th data sample; is the predicted value of the ith data sample.

8. The airfoil inverse design method based on point cloud neural network according to claim 1, characterized in that: In step 4, the optimization training method is the Adam algorithm.

Images