A Deep Learning-Based Method for Predicting Airfoil Aerodynamic Forces

Abstract

The present invention proposes an airfoil aerodynamic force prediction method based on deep learning. Design parameters, surface pressure coefficients, and aerodynamic force coefficients of the airfoil are extracted for the training and testing of the neural network. The pressure characteristics of the airfoil are further extracted through the surface pressure coefficients and used as the input of the neural network, thereby predicting the aerodynamic force and airfoil design parameters. Compared with the prior art, a multi-layer perceptron neural network is used to construct an airfoil aerodynamic force prediction model, which can efficiently and accurately obtain the aerodynamic force. Moreover, the method of predicting the aerodynamic force coefficient and the airfoil geometric shape by using the airfoil pressure characteristics as the input through the multi-layer perceptron model avoids the numerical solution of the flow variables at a large number of discrete points, effectively improving the efficiency and accuracy of the airfoil aerodynamic force prediction. In addition, the convolutional autoencoder and multi-layer perceptron neural network model built in the present invention can describe more complex non-linear relationships, which helps to accurately identify the airfoil pressure characteristics and accurately predict the aerodynamic force.

Description

Technical Field

[0001] The present invention relates to the fields of aerodynamics and artificial intelligence, and specifically to an airfoil aerodynamic force prediction method based on deep learning. Background Art

[0002] Airfoil aerodynamic force evaluation is an important part of the airfoil optimization design process. In the airfoil optimization design mainly based on computational fluid dynamics (CFD), the widely used aerodynamic force evaluation method is the Reynolds-averaged Navier-Stokes equation (RANS). Since CFD needs to be called multiple times for flow field solution during the optimization design process, a large number of flow field calculation problems are generated, which requires a large amount of computing time and resources. And as a system, the airfoil flow field has its own characteristics, and repeated CFD calculations ignore this point, reducing the efficiency. Deep learning has a powerful learning ability for high-order complex functions and has unique advantages in feature extraction, and can perform fast and accurate predictions. Summary of the Invention

[0003] Technical Problems to be Solved

[0004] Currently, the relevant research on predicting aerodynamic forces through neural networks mainly uses airfoil geometric shape parameters as inputs. These existing studies only use geometric information and fail to fully utilize a large amount of physical solution information contained in the flow field, such as the airfoil surface pressure coefficient, etc., resulting in limited actual prediction accuracy.

[0005] To solve the problem of low efficiency in airfoil optimization design caused by a large number of flow field calculations in the traditional airfoil optimization design method based on CFD, the present invention proposes an airfoil aerodynamic force prediction method based on deep learning, extracts airfoil design parameters, surface pressure coefficients, and aerodynamic force coefficients for neural network training and testing, further extracts the pressure characteristics of the airfoil through the surface pressure coefficient as the input of the neural network, so as to predict aerodynamic forces and airfoil design parameters.

[0006] The technical solution of the present invention is as follows:

[0007] The airfoil aerodynamic force prediction method based on deep learning includes the following steps:

[0008] Step 1: Generate a sample data set required for building a neural network; the parameters in the sample data set include airfoil design parameters, airfoil surface pressure coefficients, and aerodynamic force coefficients of each airfoil sample;

[0009] Step 2: Build and train a deep learning neural network model based on the sample data set;

[0010] Step 3: Use the built deep neural network for rapid prediction of airfoil aerodynamic forces.

[0011] Further, the generation of the sample data set required for building the neural network in step 1 includes the following steps:

[0012] Step 1.1: Parametrize the baseline airfoil, and superimpose perturbations on the baseline airfoil to derive new airfoils, obtaining a series of airfoil samples;

[0013] Step 1.2: Generate an airfoil computational grid; through coordinate transformation, map the grid from the physical space to the computational space; perform RANS numerical simulation on the airfoil samples obtained in step 1.1 to obtain the flow parameters of the airfoil samples; extract the airfoil design parameters, airfoil surface pressure coefficients, and aerodynamic coefficients of each airfoil sample as the sample data set for the training and testing of the neural network model.

[0014] Further, in step 1.1, the class shape function transformation method is used to parameterize the baseline airfoil, and the CST perturbation method is used to superimpose perturbations on the CST equation design parameters of the baseline airfoil to derive new airfoils, obtaining a series of airfoil samples.

[0015] Further, in step 1.2, the C-H type topology is used to generate the airfoil computational grid.

[0016] Further, in step 2, building and training the deep learning neural network model based on the sample data set includes the following steps:

[0017] Step 2.1: Use a convolutional neural network and a fully connected neural network to build a one-dimensional convolutional autoencoder, using the airfoil surface pressure coefficients in the sample data set obtained in step 1 as the input and output to extract pressure features;

[0018] Step 2.2: Train the one-dimensional convolutional autoencoder: use the root mean square error of the airfoil surface pressure coefficients as the loss function, and use the Adam optimization algorithm to iteratively optimize the neural network, with the optimization goal of minimizing the loss function until the loss function of the training sample data set no longer decreases, completing the training;

[0019] Step 2.3: Use the one-dimensional convolutional autoencoder to extract the pressure features of the airfoil samples and add them to the sample data set obtained in step 1;

[0020] Step 2.4: Use a fully connected neural network to build the first multi-layer perceptron model, using the airfoil pressure features in the sample data set obtained in step 2.3 as the input and the airfoil aerodynamic coefficients in the sample data set obtained in step 1 as the output;

[0021] Step 2.5: Train the first multi-layer perceptron model: Use the root mean square error of the airfoil aerodynamic coefficients as the loss function, and use the Adam optimization algorithm to iteratively optimize the neural network. The optimization goal is to minimize the loss function until the loss function of the training sample dataset no longer decreases, and the training is completed;

[0022] Step 2.6: Build a second multi-layer perceptron model using a fully connected neural network. Use the airfoil pressure characteristics in the sample dataset obtained in Step 2.3 as the input, and use the airfoil design parameters in the sample dataset obtained in Step 1 as the output;

[0023] Step 2.7: Train the second multi-layer perceptron model: Use the root mean square error of the airfoil design parameters as the loss function, and use the Adam optimization algorithm to iteratively optimize the neural network. The optimization goal is to minimize the loss function until the loss function of the training sample dataset no longer decreases, and the training is completed.

[0024] Furthermore, in Step 2.1, the one-dimensional convolutional autoencoder consists of an encoder and a decoder. The encoder contains two convolutional layers; the input channel number of the first convolutional layer is 1, the output channel number is 2, the convolutional kernel size is 10, the sliding step is 10, and the zero-padding width is 1. Subsequently, a normalization layer and a ReLU non-linear activation layer are set; the input channel number of the second convolutional layer is 25, the output channel number is 50, the convolutional kernel size is 3, the sliding step is 3, and the zero-padding width is 0. Subsequently, a normalization layer and a ReLU non-linear activation layer are set; the end of the encoder is a fully connected layer with 10 neurons; the beginning of the decoder is a fully connected layer with 850 neurons, and then two transposed convolutional layers are connected. The input channel number of the first transposed convolutional layer is 50, the output channel number is 25, the convolutional kernel size is 4, the sliding step is 3, and the zero-padding width is 1. Subsequently, a normalization layer and a ReLU non-linear activation layer are set. The input channel number of the second transposed convolutional layer is 25, the output channel number is 1, the convolutional kernel size is 10, the sliding step is 10, and the zero-padding width is 1. Subsequently, a Sigmoid non-linear activation layer is set.

[0025] Furthermore, in Step 1.1, the upper and lower edges of the airfoil surface are respectively fitted with 6th-order shape functions. 14 design parameters are used to describe the airfoil. The perturbation range of each design parameter is ±0.02. The Latin hypercube sampling method is used to extract 3000 airfoils in the design space as airfoil samples.

[0026] Furthermore, in Step 2.1, both the input layer and the output layer contain 508 parameters, which are the airfoil surface pressure coefficients (P 1 ,...,P 508 ) obtained in Step 1; in Step 2.4, the input layer contains 10 parameters, which are the airfoil pressure characteristics (f 1 ,...,f10 ), where f i represents the i-th pressure feature; the output layer contains 3 neurons, and the output is the airfoil aerodynamic coefficient (C L , C d , C m ), where C L represents the lift coefficient, C d represents the drag coefficient, C m represents the moment coefficient; in step 2.6, the input layer contains 10 parameters, which are the airfoil pressure features (f 1 ,..., f 10 ) extracted in step 2.3, where f i represents the i-th pressure feature; the output layer contains 14 neurons, and the output is the airfoil design parameters (x 1 ,..., x 14 ).

[0027] Furthermore, in step 2.4, the hidden layer of the first multi-layer perceptron model contains 2 layers, and the number of neurons is 800 and 800 respectively; in step 2.6, the hidden layer of the second multi-layer perceptron model contains 2 layers, and the number of neurons is 800 and 800 respectively.

[0028] Furthermore, in step 3, using the pressure feature as the input, the first multi-layer perceptron model trained in step 2.5 is used to predict the aerodynamic coefficient, and the second multi-layer perceptron model trained in step 2.7 is used to predict the airfoil design parameters, and then through the CST function, the airfoil geometric shape is obtained.

[0029] Beneficial effects

[0030] Compared with the prior art, the present invention has the following advantages:

[0031] 1. The present invention uses a multi-layer perceptron neural network to construct an airfoil aerodynamic prediction model, which can efficiently and accurately obtain aerodynamic forces compared with the prior art.

[0032] 2. The method of the present invention using the airfoil pressure feature as the input to predict the aerodynamic coefficient and the airfoil geometric shape through the multi-layer perceptron model avoids the numerical solution of the flow variables at a large number of discrete points, and effectively improves the prediction efficiency and accuracy of the airfoil aerodynamic force compared with the method of predicting the aerodynamic force through the airfoil shape.

[0033] 3. The convolutional autoencoder and multi-layer perceptron neural network models built in the present invention can depict more complex non-linear relationships compared with the prior art, which helps to accurately identify the airfoil pressure features and accurately predict the aerodynamic forces.

[0034] Additional aspects and advantages of the present invention will be given in part in the following description, become apparent in part from the following description, or be learned through the practice of the present invention. Brief Description of the Drawings

[0035] The above and / or additional aspects and advantages of the present invention will become apparent and be readily understood from the following description of embodiments in conjunction with the accompanying drawings, in which:

[0036] Figure 1 is a flowchart of the method of the present invention.

[0037] Figure 2 is the outline of the airfoil sample set.

[0038] Figure 3 is the airfoil computational grid.

[0039] Figure 4 is a one-dimensional convolutional autoencoder.

[0040] Figure 5 is the first multi-layer perceptron model.

[0041] Figure 6 is the second multi-layer perceptron model.

[0042] Figure 7 is the relative error distribution of the prediction of aerodynamic coefficients. Count represents the number of samples in the corresponding error interval. (a) is the relative error distribution of the prediction of lift coefficient on the training set, (b) is the relative error distribution of the prediction of drag coefficient on the training set, (c) is the relative error distribution of the prediction of moment coefficient on the training set, (d) is the relative error distribution of the prediction of lift coefficient on the test set, (e) is the relative error distribution of the prediction of drag coefficient on the test set, (f) is the relative error distribution of the prediction of moment coefficient on the test set.

[0043] Figure 8 is the scatter plot of the prediction of aerodynamic coefficients. (a) is the scatter plot of the prediction of lift coefficient on the training set, (b) is the scatter plot of the prediction of drag coefficient on the training set, (c) is the scatter plot of the prediction of moment coefficient on the training set, (d) is the scatter plot of the prediction of lift coefficient on the test set, (e) is the scatter plot of the prediction of drag coefficient on the test set, (f) is the scatter plot of the prediction of moment coefficient on the test set.

[0044] Figure 9 is the pressure feature extraction effect of airfoil No. 322 in the training case and airfoil No. 27 in the test case. (a) is the normalized predicted / true pressure coefficient of airfoil No. 322 in the training case, (b) is the normalized predicted / true pressure coefficient of airfoil No. 27 in the test case.

[0045] Figure 10Prediction results for airfoil No. 322 in the training case and No. 27 in the test case. (a) shows the predicted / true shape of airfoil No. 322 in the training case, and (b) shows the predicted / true shape of airfoil No. 27 in the test case. Detailed implementation manners

[0046] The embodiments of the present invention will be described in detail below. The embodiments are exemplary and are intended to explain the present invention, and should not be construed as limiting the present invention.

[0047] The airfoil aerodynamic force prediction method based on deep learning described in the embodiment includes: generating a sample data set; building a deep learning neural network model based on this data set; and using the built deep neural network for rapid prediction of airfoil aerodynamic forces. The specific steps are as follows:

[0048] Step 1: Generate a sample data set required for building the neural network:

[0049] 1) In this embodiment, the Rae2822 airfoil is used as the reference airfoil. The class shape function transformation (CST) method is used to parameterize the reference airfoil to reduce the number of variables. The 6th-order CST method is adopted, that is, 14 design parameters are used to describe the airfoil. The perturbation range of each design parameter is ±0.2. The Latin hypercube sampling method is used to extract 3000 airfoils in the design space as the sample set, as Figure 2 shown. The airfoils in the sample set are numbered, among which those numbered 1 - 2400 are used as the training set, and those numbered 2401 - 3000 are used as the validation set.

[0050] 2) The C-H type topology is adopted to generate the airfoil computational grid, as Figure 3 shown.

[0051] 3) The open-source solver NASA CFL3D is used to perform numerical simulations on the airfoil samples to obtain a series of flow parameters of the sample set. The calculation conditions in this embodiment are: Re = 6.5×10 6 , Ma = 0.73, T∞ = 460°R, C L = 0.824; the airfoil geometric shape parameters, airfoil surface pressure coefficients, and airfoil aerodynamic force coefficients (including lift coefficient, drag coefficient, and moment coefficient) are extracted for training and testing the neural network model. Among them, the airfoil geometric shape parameters are 14-dimensional, the airfoil surface pressure coefficients are 508-dimensional, and the airfoil aerodynamic force coefficients are 3-dimensional.

[0052] Step 2: Build a deep learning neural network model:

[0053] 1) The neural network model of the present invention includes three parts: a one-dimensional convolutional autoencoder for extracting airfoil pressure features; two multi-layer perceptron models for predicting airfoil aerodynamic forces and solving airfoil shapes respectively.

[0054] 2) The convolutional autoencoder consists of an encoder and a decoder, specifically including a one-dimensional convolutional layer and a normalization layer. Its input and output are both the airfoil surface pressure coefficients (P 1 ,..., P 508 ) obtained in step 1, and it contains 508 parameters.

[0055] The encoder contains two convolutional layers. The input channel number of convolutional layer 1 is 1, the output channel number is 2, the convolutional kernel size is 10, the sliding step is 10, and the zero-padding width is 1. Subsequently, a normalization layer and a ReLU non-linear activation layer are set. The input channel number of convolutional layer 2 is 25, the output channel number is 50, the convolutional kernel size is 3, the sliding step is 3, and the zero-padding width is 0. Subsequently, a normalization layer and a ReLU non-linear activation layer are set. The end of the encoder is a fully connected layer with 10 neurons; the beginning of the decoder is a fully connected layer with 850 neurons. Subsequently, two transposed convolutional layers are connected. The input channel number of transposed convolutional layer 1 is 50, the output channel number is 25, the convolutional kernel size is 4, the sliding step is 3, and the zero-padding width is 1. Subsequently, a normalization layer and a ReLU non-linear activation layer are set. The input channel number of transposed convolutional layer 2 is 25, the output channel number is 1, the convolutional kernel size is 10, the sliding step is 10, and the zero-padding width is 1. Subsequently, a Sigmoid non-linear activation layer is set. The model schematic diagram is as Figure 4 shown.

[0056] 3) Train the convolutional autoencoder, using the root mean square error of the airfoil surface pressure vector as the loss function, and use the Adam optimization algorithm to iteratively optimize the neural network. The optimization goal is to minimize the loss function, and the initial learning rate is set to 1×10 -4 , until the loss function of the training set no longer decreases, and the training is completed.

[0057] 4) Use the one-dimensional convolutional autoencoder to extract the pressure features of the airfoil samples and add them to the sample dataset obtained in step 1.

[0058] The first multi-layer perceptron model includes an input layer, a hidden layer, and an output layer. The input layer contains 10 parameters, which are the airfoil pressure features (f 1 ,..., f 10 ) extracted by the convolutional autoencoder (CAE) in step 2.3. The output layer is the aerodynamic coefficients (C L , C d , C m ), where C L represents the lift coefficient, C d represents the drag coefficient, and C m represents the moment coefficient. The two hidden layers each contain 800 neurons. The model schematic diagram is as Figure 5 shown.

[0059] The second multi-layer perceptron model, including an input layer, a hidden layer, and an output layer. The input layer contains 10 parameters, which are the airfoil pressure features (f 1 ,..., f 10 ) extracted by the convolutional autoencoder (CAE) in step 2.3. The output layer contains 14 airfoil design parameters (x 1 ,..., x 14 ). Each of the two hidden layers contains 800 neurons. The schematic diagram of the model is as shown in Figure 6 .

[0060] Train the first multi-layer perceptron model and the second multi-layer perceptron model. Use the root mean square error of the predicted aerodynamic coefficients and airfoil design parameters as the loss function, and use the Adam optimization algorithm to iteratively optimize the neural network. The optimization goal is to minimize the loss function, and the initial learning rate is set to 1×10 -4 , until the loss function of the training set no longer decreases, and the training is completed.

[0061] Step 3: Fast prediction of airfoil aerodynamic force:

[0062] Use the pressure features as the input. Use the first multi-layer perceptron model trained in step 2.5 to predict the aerodynamic coefficients, and use the second multi-layer perceptron model trained in step 2.7 to predict the airfoil design parameters. Then, through the CST function, obtain the airfoil geometric shape.

[0063] Use the data of the training set and the validation set to test the trained neural network model. If the test is successful, it can be used for fast prediction of airfoil aerodynamic force. Among them, the average relative errors of the lift coefficient prediction on the training set and the test set are 0.0015% and 0.0022% respectively, the average relative errors of the drag coefficient prediction are 1.16% and 1.41% respectively, and the average relative errors of the moment coefficient prediction are 0.26% and 0.35% respectively. The relative error distribution of the aerodynamic coefficient prediction is as shown in Figure 7 , the scatter plot of the aerodynamic coefficient prediction is as shown in Figure 8 , the extraction effect of the airfoil pressure features of training case No. 322 and test case No. 27 is as shown in Figure 9 , and the prediction results of the airfoil are as shown in Figure 10As shown. In addition, according to the same hidden layer settings, a model for predicting aerodynamic coefficients through airfoil profiles was established, and its prediction accuracy was tested. The average relative errors of lift coefficient prediction on the training set and the test set were 0.011% and 0.014% respectively, the average relative errors of drag coefficient prediction were 1.44% and 1.86% respectively, and the average relative errors of moment coefficient prediction were 0.44% and 0.48% respectively. The comparison of the prediction errors with the method of the present invention is shown in Table 1, where Method A represents the method of the present invention and Method B represents the method of predicting aerodynamic forces through airfoil profiles. The above results show that the model for predicting aerodynamic coefficients through pressure characteristics established by the present invention has a higher prediction accuracy than the model for predicting aerodynamic coefficients through airfoil profiles.

[0064] Table 1 Comparison of prediction accuracies between the method of the present invention and the method of predicting aerodynamic forces through profiles

[0065]

[0066] Although the embodiments of the present invention have been shown and described above, it can be understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those of ordinary skill in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of the present invention without departing from the principles and spirit of the present invention.

Claims

1. A method for predicting airfoil aerodynamic force based on deep learning, characterized in that: It includes the following steps: Step 1: Generate a sample data set required for building a neural network; the parameters in the sample data set include the airfoil design parameters, airfoil surface pressure coefficients, and aerodynamic force coefficients of each airfoil sample; Step 2: Build and train a deep learning neural network model based on the sample data set; It includes the following processes: Step 2.1: Build a one-dimensional convolutional autoencoder using a convolutional neural network and a fully connected neural network, using the airfoil surface pressure coefficient in the sample data set obtained in Step 1 as the input and output to extract pressure features; Step 2.2: Train the one-dimensional convolutional autoencoder: use the root mean square error of the airfoil surface pressure coefficient as the loss function, and use the Adam optimization algorithm to iteratively optimize the neural network. The optimization goal is to minimize the loss function until the loss function of the training sample data set no longer decreases, and the training is completed; Step 2.3: Use the one-dimensional convolutional autoencoder to extract the pressure features of the airfoil sample and add them to the sample data set obtained in Step 1; Step 2.4: Build a first multi-layer perceptron model using a fully connected neural network, using the airfoil pressure features in the sample data set obtained in Step 2.3 as the input and the airfoil aerodynamic force coefficients in the sample data set obtained in Step 1 as the output; Step 2.5: Train the first multi-layer perceptron model: use the root mean square error of the airfoil aerodynamic force coefficient as the loss function, and use the Adam optimization algorithm to iteratively optimize the neural network. The optimization goal is to minimize the loss function until the loss function of the training sample data set no longer decreases, and the training is completed; Step 2.6: Build a second multi-layer perceptron model using a fully connected neural network, using the airfoil pressure features in the sample data set obtained in Step 2.3 as the input and the airfoil design parameters in the sample data set obtained in Step 1 as the output; Step 2.7: Train the second multi-layer perceptron model: use the root mean square error of the airfoil design parameters as the loss function, and use the Adam optimization algorithm to iteratively optimize the neural network. The optimization goal is to minimize the loss function until the loss function of the training sample data set no longer decreases, and the training is completed; Step 3: Use the built deep neural network for rapid prediction of airfoil aerodynamic force. Specifically: use the pressure features as the input, use the first multi-layer perceptron model trained in Step 2.5 to predict the aerodynamic force coefficient, use the second multi-layer perceptron model trained in Step 2.7 to predict the airfoil design parameters, and then obtain the airfoil geometric shape through the CST function.

2. The method for predicting airfoil aerodynamic force based on deep learning according to claim 1, characterized in that: The steps for generating the sample data set required for building the neural network in Step 1 include the following steps: Step 1.1: Parametrize the baseline airfoil and superimpose perturbations on the baseline airfoil to derive new airfoils, obtaining a series of airfoil samples; Step 1.2: Generate an airfoil computational grid; through coordinate transformation, map the grid from the physical space to the computational space; perform RANS numerical simulation on the airfoil samples obtained in Step 1.1 to obtain the flow parameters of the airfoil samples; extract the airfoil design parameters, airfoil surface pressure coefficients, and aerodynamic coefficients of each airfoil sample as a sample dataset for the training and testing of the neural network model.

3. The method for predicting aerodynamic forces of an airfoil based on deep learning according to claim 2, characterized in that: In Step 1.1, the class shape function transformation method is used to parameterize the baseline airfoil, and the CST perturbation method is used to superimpose perturbations on the CST equation design parameters of the baseline airfoil to derive new airfoils, obtaining a series of airfoil samples.

4. The method for predicting aerodynamic forces of an airfoil based on deep learning according to claim 2, characterized in that: In Step 1.2, a C-H type topology is used to generate the airfoil computational grid.

5. The method for predicting aerodynamic forces of an airfoil based on deep learning according to claim 1, characterized in that: In Step 2.1, the one-dimensional convolutional autoencoder consists of an encoder and a decoder. The encoder contains two convolutional layers; the input channel number of the first convolutional layer is 1, the output channel number is 2, the convolutional kernel size is 10, the sliding step is 10, and the zero-padding width is 1. Subsequently, a normalization layer and a ReLU non-linear activation layer are set; the input channel number of the second convolutional layer is 25, the output channel number is 50, the convolutional kernel size is 3, the sliding step is 3, and the zero-padding width is 0. Subsequently, a normalization layer and a ReLU non-linear activation layer are set; the end of the encoder is a fully connected layer with 10 neurons; The beginning of the decoder is a fully connected layer with 850 neurons, and then two transposed convolutional layers are connected. The input channel number of the first transposed convolutional layer is 50, the output channel number is 25, the convolutional kernel size is 4, the sliding step is 3, and the zero-padding width is 1. Subsequently, a normalization layer and a ReLU non-linear activation layer are set. The input channel number of the second transposed convolutional layer is 25, the output channel number is 1, the convolutional kernel size is 10, the sliding step is 10, and the zero-padding width is 1. Subsequently, a Sigmoid non-linear activation layer is set.

6. The method for predicting aerodynamic forces of an airfoil based on deep learning according to claim 1, characterized in that: In Step 1.1, the upper and lower edges of the airfoil surface are respectively fitted with 6th-order shape functions, 14 design parameters are used to describe the airfoil, the perturbation range of each design parameter is ±0.02, and the Latin hypercube sampling method is used to extract 3000 airfoils in the design space as airfoil samples.

7. The method for predicting aerodynamic forces of an airfoil based on deep learning according to claim 1, characterized in that: In step 2.1, both the input layer and the output layer contain 508 parameters, which are the airfoil surface pressure coefficients (P 1 ,...,P 508 ) obtained in step 1; in step 2.4, the input layer contains 10 parameters, which are the airfoil pressure characteristics (f 1 ,...,f 10 ) obtained in step 2.3, where f i represents the i-th pressure characteristic; the output layer contains 3 neurons, and the output is the airfoil aerodynamic coefficients (C L ,C d ,C m ), where C L represents the lift coefficient, C d represents the drag coefficient, and C m represents the moment coefficient; in step 2.6, the input layer contains 10 parameters, which are the airfoil pressure characteristics (f 1 ,...,f 10 ) extracted in step 2.3, where f i represents the i-th pressure characteristic; the output layer contains 14 neurons, and the output is the airfoil design parameters (x 1 ,...,x 14 ).

8. The method for predicting aerodynamic forces of an airfoil based on deep learning according to claim 1, characterized in that: In Step 2.4, the hidden layer of the first multi-layer perceptron model contains 2 layers, and the number of neurons is 800 and 800 respectively; in Step 2.6, the hidden layer of the second multi-layer perceptron model contains 2 layers, and the number of neurons is 800 and 800 respectively.

Images