Aerodynamic parameter prediction method based on wing three-dimensional effect physical information neural network
By injecting physical information of the wing's three-dimensional effects into the neural network, the problem of insufficient accuracy of traditional neural networks in predicting the lift coefficient of trapezoidal wings under limited samples is solved, achieving higher prediction accuracy and design efficiency.
Patent Information
- Application Number
- CN202510692105.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-05-27
- Publication Date
- 2025-09-16
AI Technical Summary
Traditional neural networks have difficulty in achieving rapid and accurate prediction of the lift coefficient of a trapezoidal wing under limited sample conditions, especially in the design of complex three-dimensional wing shapes, where the prediction error is large and it is difficult to meet engineering accuracy requirements.
By introducing physical information of the three-dimensional effects of the wing, such as the relationship between the leading edge sweep angle and the lift coefficient, the root-tip ratio and the lift coefficient, and the root torsion angle and the lift coefficient, it is injected as physical information into the neural network training process to improve the prediction accuracy.
The prediction accuracy of the neural network has been significantly improved under finite sample conditions, and the lift coefficient of the trapezoidal wing can be predicted more accurately, meeting the rapid iteration requirements of engineering design.
Smart Images
Figure CN120654596A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of aircraft aerodynamic performance prediction, and in particular to an aerodynamic parameter prediction method based on a wing three-dimensional effect physical information neural network. Background Art
[0002] With the rapid development of aviation technology, the requirements for aerodynamic performance of aircraft are constantly increasing. Especially in the preliminary design stage, there is a higher demand for the ability to quickly predict aerodynamic parameters. Among them, the trapezoidal wing is widely used in various aircraft due to its simple structure and high aerodynamic efficiency. It is currently one of the most representative wing layouts. The lift coefficient of the trapezoidal wing is a key aerodynamic parameter that has a direct impact on the aerodynamic efficiency, maneuverability, and overall design of the aircraft. Therefore, achieving rapid and accurate prediction of the lift coefficient of the trapezoidal wing is of great significance to improving the efficiency of aircraft design.
[0003] Currently, traditional aerodynamic performance prediction relies primarily on high-precision CFD calculations or wind tunnel experiments. While highly accurate, these methods are computationally expensive and time-consuming, making them difficult to meet the rapid iteration requirements of modern engineering design. To improve efficiency, commonly used interpolation surrogate models (such as Kriging and RBF) and regression models (such as LS-SVR) are widely used in aerodynamic modeling. However, these methods experience increased prediction errors when the sample size is limited or the design space is slightly complex, making it difficult to meet engineering accuracy requirements. This is particularly true for complex three-dimensional wing shape design.
[0004] Artificial intelligence (AI) technology has advanced rapidly in recent years, with neural networks, in particular, demonstrating significant potential for handling nonlinear, high-dimensional problems and becoming a key tool for aerodynamic parameter prediction. Neural networks can automatically extract features from raw data, flexibly adapt to a variety of input-output relationships, and leverage modern computing platforms for efficient training, making them suitable for large-scale design optimization tasks.
[0005] However, traditional neural networks typically require a large number of samples to ensure accurate predictions. In engineering practice, obtaining sufficient high-quality samples is often costly, especially for high-precision simulations of typical configurations such as trapezoidal wings, which consume significant computational resources. Furthermore, obtaining sufficient training data can be challenging when faced with complex design spaces. Therefore, improving the prediction accuracy of neural networks within limited sample sizes has become a key challenge in current aerodynamic modeling. Summary of the Invention
[0006] In order to solve the problems existing in the prior art, the present invention proposes an aerodynamic parameter prediction method based on a wing three-dimensional effect physical information neural network. By introducing the relationship between the leading edge sweep angle and the lift coefficient, the relationship between the root-tip ratio and the lift coefficient, and the relationship between the root torsion angle and the lift coefficient as physical information into the neural network, the physical information of the three-dimensional effects caused by the leading edge sweep angle, the root-tip ratio, and the root torsion angle are injected into the training of the neural network. Under limited sample conditions, the prediction accuracy of this method is significantly improved compared with the traditional neural network model.
[0007] The present invention is achieved through the following technical solutions:
[0008] A method for predicting aerodynamic parameters based on a wing three-dimensional effect physical information neural network comprises the following steps:
[0009] Step S1: Select the airfoil at the wing root and wing tip from the airfoil library, parameterize the wing shape through five three-dimensional shape parameters: span, leading edge sweep angle, root-tip ratio, root twist angle, and tip twist angle, and generate a set number of wing shape samples based on the parameters;
[0010] Step S2: Generate a computational grid for the wing shape sample, obtain the lift coefficient of the wing shape sample under the set working condition through numerical simulation, and obtain sample data points;
[0011] Step S3: Acquire physical information of the three-dimensional effect of the wing, establish a fully connected neural network model, and train the neural network model based on the sample data points obtained in step S2; when training the neural network model, inject the obtained physical information of the three-dimensional effect of the wing into the knowledge to obtain a neural network model that takes into account the physical knowledge injection;
[0012] Step S4: Use the neural network model trained in step S3 and injected with physical knowledge to predict aerodynamic parameters, input the new span, leading edge sweep angle, root-tip ratio, root twist angle, and tip twist angle, and output the lift coefficient corresponding to the new shape.
[0013] Furthermore, in step S2, the numerical simulation method adopts the SA-equation model as the turbulence model, the Roe spatial discretization format, and the LU-SGS implicit format as the time marching format.
[0014] Furthermore, in step S3, the physical information of the wing three-dimensional effect is the relationship between the leading edge sweep angle and the lift coefficient, the relationship between the root-tip ratio and the lift coefficient, and the relationship between the root torsion angle and the lift coefficient.
[0015] Furthermore, step S3 specifically includes the following sub-steps:
[0016] Step S31: Acquire physical information of the wing's three-dimensional effect, i.e., establish the relationship between the leading edge sweep angle and the lift coefficient, the root-tip ratio and the lift coefficient, and the root torsion angle and the lift coefficient; specifically:
[0017] The lift line slope C is obtained according to the empirical formula in the overall design manual Lα The calculation expression is as follows:
[0018]
[0019] Where: C Lα is the lift line slope, A is the aspect ratio of the wing, η is the airfoil efficiency, which is taken as 0.95; β=1-Ma 2 , Ma is the incoming flow Mach number; S exp is the exposed area of the wing, S ref is the reference area of the wing, calculated as follows:
[0020] S exp =S ref =c0(1+λ)B
[0021] Where c0 is the wing root chord length, which is uniformly set to 1000 mm, λ is the root-to-tip ratio, B is the span, Λ is the wing leading edge sweep angle, and the coefficient F is set to 1.07.
[0022] The relationship between the lift coefficient and the lift line slope is established using the following formula:
[0023]
[0024] Where C L is the lift coefficient, α ∞ is the incoming flow angle of attack, is the root twist angle, α0 is the zero lift angle of attack, where α0 is calculated as follows:
[0025]
[0026] Where, α 0,root is the zero-lift angle of attack of the wing root airfoil, α 0,tip is the zero-lift angle of attack of the wingtip airfoil, and the zero-lift angle of attack of the wingroot airfoil and the zero-lift angle of attack of the wingtip airfoil are all obtained through numerical simulation;
[0027] Substituting the lift line slope expression and the zero lift angle of attack expression into the lift coefficient expression, the relationship between the lift coefficient and the aspect ratio, leading edge sweep angle, and root twist angle is obtained as follows:
[0028]
[0029] Transform the above formula to obtain the relationship between the leading edge sweep angle and the lift coefficient:
[0030]
[0031] And the relationship between the root torsion angle lift coefficient and the lift coefficient:
[0032]
[0033] For a trapezoidal wing, the relationship between aspect ratio and root-to-tip ratio is:
[0034]
[0035] Where A is the aspect ratio. Substitute the above equation into the relationship between the lift coefficient and the aspect ratio, the leading edge sweep angle, and the root twist angle, and then transform to obtain the relationship between the root-tip ratio and the lift coefficient:
[0036]
[0037] Step S32: Establish a fully connected neural network model, with the five wing shape parameters in step S1 as input and the wing lift coefficient as output. The number of hidden layers is 3, and the number of neurons in each layer is 64.
[0038] Step S33: training the neural network model established in step S32 based on the sample data points obtained in step S2; injecting knowledge into the physical information of the three-dimensional effect of the wing obtained in step S31 during the training of the neural network model to obtain a neural network model that takes physical knowledge injection into account;
[0039] The specific process of injecting knowledge into the physical information of the wing's three-dimensional effects is as follows:
[0040] ① Based on the model established in step S32 and the sample data points obtained in step 2, the root mean square error (RMSE) of the lift coefficient, leading edge sweep angle, root twist angle, and root-to-tip ratio is calculated using the following formula:
[0041]
[0042] Where N is the number of sample data points. When calculating the RMSE of the lift coefficient, y i is the true value of the lift coefficient, that is, the lift coefficient value in the sample data point, The lift coefficient value predicted by the model established in step S32; when calculating the RMSE of the leading edge sweep angle, root twist angle, and root-tip ratio, y i is the true value of the wing shape, that is, the leading edge sweep angle, root twist angle, and root-tip ratio in the sample data points, The lift coefficient value predicted by the model established in step S32 is then calculated based on the relationship between the leading edge sweep angle and the lift coefficient, the relationship between the root-tip ratio and the lift coefficient, and the relationship between the root torsion angle and the lift coefficient in step S31 to obtain the leading edge sweep angle, root torsion angle, and root-tip ratio;
[0043] ②Inject knowledge of physical information of wing three-dimensional effects. The specific process is as follows:
[0044] First, the root mean square error (RMSE) of the lift coefficient, leading edge sweep angle, root torsion angle, and root-tip ratio are weighted to establish the loss function as follows:
[0045]
[0046] Where: loss total It is the loss function that integrates the physical information of the wing's three-dimensional effect. data is the RMSE of the lift coefficient, loss Λ is the RMSE of the leading edge sweep angle, loss λ is the RMSE of the root-to-shoot ratio, is the RMSE of the root torsion angle; λ1, λ2, λ3, and λ4 are weight coefficients;
[0047] Then, according to the loss function loss that integrates the physical information of the wing's three-dimensional effect total The gradients of the neurons in the output layer and hidden layer are calculated, and the weights of the neurons are updated to train the neural network.
[0048] Furthermore, the value range of the weight coefficients λ1, λ2, λ3, and λ4 in the loss function is 0.01 to 0.1.
[0049] Beneficial effects
[0050] The present invention proposes a method for predicting aerodynamic parameters based on a wing three-dimensional effect physical information neural network. By introducing the relationship between the leading edge sweep angle and the lift coefficient, the relationship between the root-tip ratio and the lift coefficient, and the relationship between the root torsion angle and the lift coefficient as physical information into the neural network, the physical information of the three-dimensional effect caused by the leading edge sweep angle, the root-tip ratio, and the root torsion angle is injected into the training of the neural network. Under limited sample conditions, the prediction accuracy of the neural network based on the wing three-dimensional effect physical information is significantly improved compared with the traditional neural network model. BRIEF DESCRIPTION OF THE DRAWINGS
[0051] Figure 1 This is a flow chart of a method for predicting aerodynamic parameters based on a wing three-dimensional effect physical information neural network according to an embodiment of the present invention;
[0052] Figure 2This is a schematic diagram of the wing root and wing tip airfoil selection according to an embodiment of the present invention;
[0053] Figure 3 Schematic diagram of the range of variation of the wing planform according to an embodiment of the present invention;
[0054] in, Figure 3 (a) The range of variation of the wing planform with a span of 1000 mm; Figure 3 (b) The range of variation of the planform of a wing with a span of 1400 mm; Figure 3 (c) The range of variation of the planform of a wing with a span of 2000 mm; Figure 3 (d) The range of variation of the planform of a wing with a span of 2400 mm;
[0055] Figure 4 This is an example diagram of a wing sample according to an embodiment of the present invention;
[0056] in, Figure 4 (a) is the sample wing example 1; Figure 4 (b) is sample wing example 2;
[0057] Figure 5 This is a schematic diagram of the physical information neural network structure of an embodiment of the present invention;
[0058] Figure 6 Schematic diagram of lift coefficient predicted by five different neural network models according to an embodiment of the present invention;
[0059] in, Figure 6 (a) is the lift coefficient predicted by model1; Figure 6 (b) is the lift coefficient predicted by model2; Figure 6 (c) is the lift coefficient predicted by model3; Figure 6 (d) is the lift coefficient predicted by model4; Figure 6 (e) is the lift coefficient predicted by model5;
[0060] Figure 7 Schematic diagram of prediction error values of five different neural network models according to an embodiment of the present invention. DETAILED DESCRIPTION
[0061] In order to make the technical problems, technical solutions and beneficial effects solved by the present invention more clearly understood and to enable those skilled in the art to better understand the solutions of the present invention, the present invention is further described and fully explained below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only intended to illustrate the present invention and are not intended to limit the present invention.
[0062] This embodiment uses the aerodynamic parameter prediction method based on the wing three-dimensional effect physical information neural network proposed by the present invention to predict the lift coefficient of a certain trapezoidal wing based on the wing three-dimensional effect physical information neural network, such as Figure 1 As shown, the aerodynamic parameter prediction method based on the wing three-dimensional effect physical information neural network of the embodiment of the present invention includes the following steps:
[0063] Step S1: Select the airfoil at the wing root and wing tip from the airfoil library, parameterize the wing shape through five three-dimensional shape parameters: span, leading edge sweep angle, root-tip ratio, root twist angle, and tip twist angle, and generate a set number of wing shape samples based on the parameters;
[0064] In this embodiment, the airfoil selected at the wing root and wing tip is as follows: Figure 2 As shown in the figure, considering the stealth performance requirements of the wing, the wing root and wing tip airfoils both adopt double S airfoils. The leading edge of this type of airfoil is sharp and hawk-beak-shaped, which has good stealth performance and can meet the stealth constraints.
[0065] In this embodiment, the ranges of the five shape parameters, span, leading edge sweep angle, root-to-tip ratio, root twist angle, and tip twist angle, are set as shown in Table 1. The root chord length of the wing is fixed at 1000 mm. Within the set parameter range, 100 wing sample points are obtained using the Latin hypercube sampling method. Figure 3 The range of variation of the wing planform with spans of 1000mm, 1400mm, 2000mm and 2400mm respectively. Figure 4 The schematic diagram of the shape of two sample points in the wing sample. The parameter values corresponding to the shape of these two sample points are shown in Table 2;
[0066] Table 1 Range of wing shape parameters
[0067]
[0068] Table 2 Example parameter values of wing samples
[0069]
[0070] Step S2: Generate a computational grid for the wing shape sample, obtain the lift coefficient of the wing shape sample under the set working condition through numerical simulation, and obtain sample data points;
[0071] In this embodiment, an unstructured grid is generated for the wing shape sample obtained in step S1; the numerical simulation method adopts the SA-equation turbulence model, the Roe spatial discretization format, and the LU-SGS implicit time marching format; the operating conditions are set as follows: Mach number 0.8, wing angle of attack 2°, and altitude 10 km;
[0072] Step S3: Acquire physical information of the three-dimensional effect of the wing, establish a fully connected neural network model, and train the neural network model based on the sample data points obtained in step S2; when training the neural network model, inject the obtained physical information of the three-dimensional effect of the wing into the knowledge to obtain a neural network model that takes into account the physical knowledge injection; the physical information of the three-dimensional effect of the wing includes the relationship between the leading edge sweep angle and the lift coefficient, the relationship between the root-tip ratio and the lift coefficient, and the relationship between the root torsion angle and the lift coefficient;
[0073] Step S3 specifically includes the following sub-steps:
[0074] Step S31: Acquire physical information of the wing's three-dimensional effect, i.e., establish the relationship between the leading edge sweep angle and the lift coefficient, the root-tip ratio and the lift coefficient, and the root torsion angle and the lift coefficient; specifically:
[0075] The lift line slope C is obtained according to the empirical formula in the overall design manual Lα The calculation expression is as follows:
[0076]
[0077] Where: C Lα is the lift line slope, A is the aspect ratio of the wing, η is the airfoil efficiency, which is taken as 0.95; β=1-Ma 2 , Ma is the incoming flow Mach number; S exp is the exposed area of the wing, S ref is the reference area of the wing, calculated as follows:
[0078] S exp =S ref =c0(1+λ)B
[0079] Where c0 is the wing root chord length, which is uniformly set to 1000 mm, λ is the root-to-tip ratio, B is the span, Λ is the wing leading edge sweep angle, and the coefficient F is set to 1.07.
[0080] The lift coefficient is related to the lift line slope using the following formula:
[0081]
[0082] Where C L is the lift coefficient, α ∞ is the incoming flow angle of attack, is the root twist angle, α0 is the zero lift angle of attack, where α0 is calculated as follows:
[0083]
[0084] Where, α 0,root is the zero-lift angle of attack of the wing root airfoil, α 0,tipis the zero-lift angle of attack of the wingtip airfoil, and the zero-lift angle of attack of the wingroot airfoil and the zero-lift angle of attack of the wingtip airfoil are all obtained through numerical simulation;
[0085] Substituting the lift line slope expression and the zero lift angle of attack expression into the lift coefficient expression, the relationship between the lift coefficient and the aspect ratio, leading edge sweep angle, and root twist angle is obtained as follows:
[0086]
[0087] Transform the above formula to obtain the relationship between the leading edge sweep angle and the lift coefficient:
[0088]
[0089] And the relationship between the root torsion angle lift coefficient and the lift coefficient:
[0090]
[0091] For a trapezoidal wing, the relationship between aspect ratio and root-to-tip ratio is:
[0092]
[0093] Where A is the aspect ratio. Substitute the above equation into the relationship between the lift coefficient and the aspect ratio, the leading edge sweep angle, and the root twist angle, and then transform to obtain the relationship between the root-tip ratio and the lift coefficient:
[0094]
[0095] Step S32: Establish a fully connected neural network model, with the five wing shape parameters in step S1 as input and the wing lift coefficient as output. The number of hidden layers is 3, and the number of neurons in each layer is 64.
[0096] Step S33: training the neural network model established in step S32 based on the sample data points obtained in step S2; injecting knowledge into the physical information of the three-dimensional effect of the wing obtained in step S31 during the training of the neural network model to obtain a neural network model that takes physical knowledge injection into account;
[0097] Specifically:
[0098] ① Based on the model established in step S32 and the sample data points obtained in step 2, the root mean square error (RMSE) of the lift coefficient, leading edge sweep angle, root twist angle, and root-to-tip ratio is calculated using the following formula:
[0099]
[0100] Where N is the number of sample data points. When calculating the RMSE of the lift coefficient, y iis the true value of the lift coefficient, that is, the lift coefficient value in the sample data point, The lift coefficient value predicted by the model established in step S32; when calculating the RMSE of the leading edge sweep angle, root twist angle, and root-tip ratio, y i is the true value of the wing shape, that is, the leading edge sweep angle, root twist angle, and root-tip ratio in the sample data points, The lift coefficient value predicted by the model established in step S32 is then calculated based on the relationship between the leading edge sweep angle and the lift coefficient, the relationship between the root-tip ratio and the lift coefficient, and the relationship between the root torsion angle and the lift coefficient in step S31 to obtain the leading edge sweep angle, root torsion angle, and root-tip ratio;
[0101] ②Inject knowledge of physical information of wing three-dimensional effects. The specific process is as follows:
[0102] First, the root mean square error (RMSE) of the lift coefficient, leading edge sweep angle, root torsion angle, and root-tip ratio are weighted to establish the loss function as follows:
[0103]
[0104] Where: loss total It is the loss function that integrates the physical information of the wing's three-dimensional effect. data is the RMSE of the lift coefficient, loss Λ is the RMSE of the leading edge sweep angle, loss λ is the RMSE of the root-to-shoot ratio, is the RMSE of the root torsion angle; λ1, λ2, λ3, and λ4 are weight coefficients; the value range of the weight coefficients λ1, λ2, λ3, and λ4 is 0.01 to 0.1;
[0105] In this embodiment, λ1=1, λ2=λ3=λ4=0.1;
[0106] Then, according to the loss function loss that integrates the physical information of the wing's three-dimensional effect total Calculate the gradients of neurons in the output layer and hidden layer and update the weights of neurons to train the neural network;
[0107] In this embodiment, the schematic diagram of the neural network model structure considering physical knowledge injection is as follows: Figure 5 As shown, the process of neural network training considering physical knowledge injection is shown in Table 3, where the superscript ~ represents the predicted value and the subscript k represents the sample number;
[0108] Table 3 Physical knowledge injection into neural network training
[0109]
[0110]
[0111] In this embodiment, in order to verify the effectiveness of the method of the present invention, five different neural network models were constructed with or without physical knowledge injection for comparison; the five neural network models are marked as model1 to model5, among which model1 is a neural network without knowledge injection, model2 is a neural network injected with three three-dimensional effects, model3 is a neural network injected with only the relationship between the root-tip ratio and the lift coefficient, model4 is a neural network injected with only the relationship between the torsion angle and the lift coefficient, and model5 is a neural network injected with only the relationship between the leading edge sweep angle and the lift coefficient; the 100 sample data points obtained in step S2 are divided into a training set and a test set in a ratio of 8:2, the training set is trained by different neural networks, and then each neural network is tested on the test set, and finally the value of the lift coefficient predicted by each network and the corresponding RMSE error value are obtained until the training is completed.
[0112] Step S4: Use the neural network model trained in step S3 to predict aerodynamic parameters, input the new span, leading edge sweep angle, root-tip ratio, root twist angle, and tip twist angle, and output the lift coefficient corresponding to the new shape;
[0113] In this embodiment, after the training is completed, the verification results on the test set are as follows Figure 6 and Figure 7 shown. Figure 6 The lift coefficients predicted by different neural networks for the test set are shown. As can be seen from the figure, model1 has the lowest degree of fit with the original data and the largest difference. Although the neural networks of model3, model4, and model5, which only inject one three-dimensional effect, have improved prediction accuracy compared to model1, model2, which injects three three-dimensional effects at the same time, has the highest fitting accuracy. Comparing model3, model4, and model5 again, the degree of fit of the three models is not much different. The differences in each sample are large or small, and the overall difference is not obvious.
[0114] Figure 7 The relative error R, mean absolute error MAE and root mean square error RMSE of the prediction results of different models are given. The MAE of the four models after embedding physical knowledge, model2, model3, model4 and model5, are better than the neural network model1 without knowledge injection. The RMSE are all better than the neural network model1 without knowledge injection. Among them, the model2 that embeds three three-dimensional effects at the same time has the greatest improvement in prediction accuracy compared with model1. Figure 6 The results of the fit shown correspond to the results of Figure 7An anomaly can be seen in the figure. Model 4's relative error R is larger than that of model 1, which uses no knowledge injection. However, the other two error values are smaller. This is because relative error and absolute error evaluate the accuracy and reliability of the measured data from different perspectives. Furthermore, the lift coefficient in this example is small, and this error is the average of twenty sample errors. The two models predict different lift coefficients for each sample, so this anomaly is reasonable.
[0115] It can be seen from this embodiment that by injecting relevant physical knowledge into the training of the neural network model, the accuracy of the neural network in predicting aerodynamic parameters can be improved under limited samples, which can provide strong support for more complex aerodynamic prediction and optimization.
[0116] Although the embodiments of the present invention have been shown and described above, it will be understood that the above embodiments are illustrative and are not to be construed as limitations on the present invention. A person skilled in the art may change, modify, replace and modify the above embodiments within the scope of the present invention without departing from the principles and purpose of the present invention.
Claims
1. An aerodynamic parameter prediction method based on a wing three-dimensional effect physical information neural network, characterized by: The following steps are involved: Step S1: Select the airfoil at the wing root and wing tip from the airfoil library, parameterize the wing shape through five three-dimensional shape parameters: span, leading edge sweep angle, root-tip ratio, root twist angle, and tip twist angle, and generate a set number of wing shape samples based on the parameters; Step S2: Generate a computational grid for the wing shape sample, obtain the lift coefficient of the wing shape sample under the set working condition through numerical simulation, and obtain sample data points; Step S3: Acquire physical information of the three-dimensional effect of the wing, establish a fully connected neural network model, and train the neural network model based on the sample data points obtained in step S2; when training the neural network model, inject the obtained physical information of the three-dimensional effect of the wing into the knowledge to obtain a neural network model that takes into account the physical knowledge injection; Step S4: Use the neural network model trained in step S3 and injected with physical knowledge to predict aerodynamic parameters, input the new span, leading edge sweep angle, root-tip ratio, root twist angle, and tip twist angle, and output the lift coefficient corresponding to the new shape.
2. The aerodynamic parameter prediction method based on the wing three-dimensional effect physical information neural network according to claim 1 is characterized by: In step S2, the numerical simulation method adopts the SA-equation model as the turbulence model, the Roe spatial discretization format, and the LU-SGS implicit time marching format.
3. The aerodynamic parameter prediction method based on the wing three-dimensional effect physical information neural network according to claim 1 is characterized by: In step S3, the physical information of the wing three-dimensional effect is the relationship between the leading edge sweep angle and the lift coefficient, the relationship between the root-tip ratio and the lift coefficient, and the relationship between the root torsion angle and the lift coefficient.
4. The aerodynamic parameter prediction method based on the wing three-dimensional effect physical information neural network according to claim 1 is characterized by: Step S3 includes the following sub-steps: Step S31: Acquire physical information of the wing's three-dimensional effect, i.e., establish the relationship between the leading edge sweep angle and the lift coefficient, the root-tip ratio and the lift coefficient, and the root torsion angle and the lift coefficient; specifically: The lift line slope C is obtained according to the empirical formula in the overall design manual Lα The calculation expression is as follows: Where: C Lα is the lift line slope, A is the aspect ratio of the wing, η is the airfoil efficiency, which is taken as 0.95; β=1-Ma 2 , Ma is the incoming flow Mach number; S exp is the exposed area of the wing, S ref is the reference area of the wing, calculated as follows: S exp =S ref =c0(1+λ)B Where c0 is the wing root chord length, which is uniformly set to 1000 mm, λ is the root-to-tip ratio, B is the span, Λ is the wing leading edge sweep angle, and the coefficient F is set to 1.
07. The relationship between the lift coefficient and the lift line slope is established using the following formula: Where C L is the lift coefficient, α ∞ is the incoming flow angle of attack, is the root twist angle, α0 is the zero lift angle of attack, where α0 is calculated as follows: Where, α 0,root is the zero-lift angle of attack of the wing root airfoil, α 0,tip is the zero-lift angle of attack of the wingtip airfoil, and the zero-lift angle of attack of the wingroot airfoil and the zero-lift angle of attack of the wingtip airfoil are all obtained through numerical simulation; Substituting the lift line slope expression and the zero lift angle of attack expression into the lift coefficient expression, the relationship between the lift coefficient and the aspect ratio, leading edge sweep angle, and root twist angle is obtained as follows: Transform the above formula to obtain the relationship between the leading edge sweep angle and the lift coefficient: And the relationship between the root torsion angle lift coefficient and the lift coefficient: For a trapezoidal wing, the relationship between aspect ratio and root-to-tip ratio is: Where A is the aspect ratio. Substitute the above equation into the relationship between the lift coefficient and the aspect ratio, the leading edge sweep angle, and the root twist angle, and then transform to obtain the relationship between the root-tip ratio and the lift coefficient: Step S32: Establish a fully connected neural network model, with the five wing shape parameters in step S1 as input and the wing lift coefficient as output. The number of hidden layers is 3, and the number of neurons in each layer is 64. Step S33: training the neural network model established in step S32 based on the sample data points obtained in step S2; injecting knowledge into the physical information of the three-dimensional effect of the wing obtained in step S31 when training the neural network model to obtain a neural network model that takes physical knowledge injection into consideration.
5. The aerodynamic parameter prediction method based on the wing three-dimensional effect physical information neural network according to claim 4 is characterized by: The specific process of injecting knowledge into the physical information of the wing's three-dimensional effects is as follows: ① Based on the model established in step S32 and the sample data points obtained in step 2, the root mean square error (RMSE) of the lift coefficient, leading edge sweep angle, root twist angle, and root-to-tip ratio is calculated using the following formula: Where N is the number of sample data points. When calculating the RMSE of the lift coefficient, y i is the true value of the lift coefficient, that is, the lift coefficient value in the sample data point, The lift coefficient value predicted by the model established in step S32; when calculating the RMSE of the leading edge sweep angle, root twist angle, and root-tip ratio, y i is the true value of the wing shape, that is, the leading edge sweep angle, root twist angle, and root-tip ratio in the sample data points, The lift coefficient value predicted by the model established in step S32 is then calculated based on the relationship between the leading edge sweep angle and the lift coefficient, the relationship between the root-tip ratio and the lift coefficient, and the relationship between the root torsion angle and the lift coefficient in step S31 to obtain the leading edge sweep angle, root torsion angle, and root-tip ratio; ②Inject knowledge of physical information of wing three-dimensional effects. The specific process is as follows: First, the root mean square error (RMSE) of the lift coefficient, leading edge sweep angle, root torsion angle, and root-tip ratio are weighted to establish the loss function as follows: Where: loss total It is the loss function that integrates the physical information of the wing's three-dimensional effect. data is the RMSE of the lift coefficient, loss Λ is the RMSE of the leading edge sweep angle, loss λ is the RMSE of the root-to-shoot ratio, is the RMSE of the root torsion angle; λ1, λ2, λ3, and λ4 are weight coefficients; Then, according to the loss function loss that integrates the physical information of the wing's three-dimensional effect total The gradients of the neurons in the output layer and hidden layer are calculated, and the weights of the neurons are updated to train the neural network.
6. The aerodynamic parameter prediction method based on the wing three-dimensional effect physical information neural network according to claim 5 is characterized by: The value range of weight coefficients λ1, λ2, λ3, and λ4 in the loss function is 0.01 to 0.1.