A neural network-based tail separation scene aerodynamic modeling method
By using a neural network-based aerodynamic modeling method for tail-retraction separation scenarios, the aerodynamic forces of the parent missile and the submunition can be quickly predicted. This solves the problem of high computational resources and time overhead in aerodynamic force prediction in tail-retraction separation scenarios, provides a fast aerodynamic force prediction method, and supports the design of separation trajectories and safety boundaries.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- CHINA ACAD OF AEROSPACE AERODYNAMICS
- Filing Date
- 2023-03-01
- Publication Date
- 2026-06-23
AI Technical Summary
Existing technologies in tail-retreat separation scenarios incur high computational costs for aerodynamic prediction, making it difficult to quickly and accurately predict the aerodynamic forces of the parent missile and the submunition, which affects the design of separation trajectories and safety boundaries.
By employing a neural network-based approach, an aerodynamic dataset is established, samples and validation sets are formed, a preliminary machine learning model is optimized, and an aerodynamic model for the tail-retraction separation scenario is built to quickly predict the aerodynamic forces of the parent missile and the submunition.
It achieves rapid (on the order of one second) aerodynamic prediction, avoiding a large amount of CFD numerical simulation work, providing a fast aerodynamic prediction method for tail retraction separation scenarios, and supporting the design of separation trajectory and safety boundary.
Smart Images
Figure CN116484713B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to an aerodynamic modeling method for tail-reverse separation scenarios based on neural networks, belonging to the field of aircraft aerodynamic design technology. Background Technology
[0002] Currently, the widely used nonlinear unsteady aerodynamic modeling methods in engineering mainly fall into two categories: one is to establish traditional mathematical aerodynamic models (such as algebraic models, step response models, etc.) related to aerodynamic forces and flight physics; the other is intelligent learning-based aerodynamic models (such as fuzzy logic methods, support vector machines (SVM), etc.). Traditional mathematical methods involve piecewise linear aerodynamic modeling using large amounts of aerodynamic data, resulting in low model accuracy and difficulty in parameter identification, which is gradually failing to meet current engineering needs. Intelligent learning-based methods, on the other hand, can establish high-precision multi-input multi-output nonlinear aerodynamic models, making them highly suitable for nonlinear aerodynamic modeling.
[0003] In the tail-retardation separation scenario, the severe aerodynamic interference between the parent missile and the submunition causes their aerodynamic forces to exhibit very obvious nonlinear aerodynamic characteristics during separation. Using traditional computational fluid dynamics (CFD) methods to calculate the separation process under different conditions would incur significant computational resource and time overhead (several hours of computation per state), posing considerable challenges to the design and evaluation of the separation trajectory in the tail-retardation separation scenario. Therefore, it is necessary to leverage the high modeling accuracy of intelligent learning methods to establish an aerodynamic model for the tail-retardation separation scenario. This model can quickly predict the aerodynamic forces of the parent missile and the submunition (computation time on the order of one second), thus providing a rapid aerodynamic force prediction method for designing the separation trajectory and safety boundaries in the tail-retardation separation scenario. Summary of the Invention
[0004] The technical problem solved by this invention is to overcome the shortcomings of the prior art and provide a neural network-based aerodynamic modeling method for tail-retraction separation scenarios, establish an aerodynamic model of the tail-retraction separation scenario, and provide a fast aerodynamic prediction means for the design of separation trajectory and safety boundary in tail-retraction separation scenarios.
[0005] The technical solution of this invention is:
[0006] This invention discloses an aerodynamic modeling method for tail retraction separation scenarios based on neural networks, comprising the following steps:
[0007] S1. Establish aerodynamic data set;
[0008] S2. Take a portion of the aerodynamic data set as the sample set and the remaining data as the verification data to form a verification set;
[0009] S3. Establish a preliminary machine learning model;
[0010] S4. Optimize the preliminary machine learning model and establish a preliminary aerodynamic model for the tail-retraction separation scenario;
[0011] S5. Verify the preliminary tail-retraction separation scenario aerodynamic model. If the model accuracy meets the requirements, the final tail-retraction separation scenario aerodynamic model is obtained; otherwise, repeat steps S3 to S4.
[0012] Furthermore, in the above modeling method, step S1, establishing the aerodynamic dataset, specifically involves:
[0013] S11. Generate the aerodynamic calculation grid required for CFD numerical simulation based on the characteristics of the tail-retarded separation vehicle.
[0014] S12. Based on the characteristics of the tail-retardation separation aircraft, select aerodynamic settings to perform preliminary aerodynamic calculations and obtain aerodynamic forces and flow fields.
[0015] S13. Based on the aforementioned aerodynamic forces and flow field, refine the mesh until the aerodynamic forces remain almost unchanged;
[0016] S14. Simulate the flow field of the tail-reversing separation vehicle under relative attitude and obtain aerodynamic data;
[0017] S15. Sort the aerodynamic data according to relative attitude to form an aerodynamic dataset.
[0018] Furthermore, in the above modeling method, in step S11, generating the aerodynamic calculation grid required for CFD numerical simulation based on the characteristics of the tail-reversing aircraft is specifically as follows:
[0019] S31: A mesh is generated using mesh generation software and imported into the tail-recoil separation vehicle digital model, including the parent missile and the submunition;
[0020] S32: Generates mesh lines at the edge of the aircraft digital model connector;
[0021] S33: Distribute grid points on each grid line connector to densify the flow in areas with more separate flows;
[0022] S34: Generate a mesh domain through the mesh line connector and project the mesh domain onto the digital model surface to make the mesh fit the body;
[0023] S35: Generate a spatial mesh by generating a mesh block from the mesh domain;
[0024] S36: Keep the parent missile stationary, adjust the position of the submunition according to the relative attitude, repeat steps S31 to S35, and finally form the aircraft grid under different relative attitudes.
[0025] Furthermore, in the above modeling method, the relative attitude in step S36 is described using four-dimensional variables; the four-dimensional variables include relative distance in the x-direction, relative distance in the y-direction, incoming flow angle of attack α, and relative attitude angle az.
[0026] Furthermore, in the above modeling method, step S3 involves establishing a preliminary machine learning model, and the specific steps are as follows:
[0027] S51: Establish the neural network framework;
[0028] S52: Based on the neural network framework, set the number of neural network layers;
[0029] S53: Set the number of neurons in each layer of the neural network;
[0030] S54: Normalize the relative distance in the x direction, the relative distance in the y direction, the angle of attack of the incoming flow α, and the relative attitude angle;
[0031] S55: Select a neural network loss function optimization method to form a preliminary machine learning model.
[0032] Furthermore, in the above modeling method, step S4 involves optimizing the preliminary machine learning model, with the following specific steps:
[0033] S61: Input the sample set into the preliminary machine learning model for learning, and evaluate the machine learning efficiency and accuracy;
[0034] S62: Adjust the number of neurons, number of layers, loss function optimization method, and step size according to the learning accuracy;
[0035] S63: Repeat steps S61 to S62 until the learning efficiency and accuracy are optimal, and establish a preliminary aerodynamic model of the tail-retraction separation scenario.
[0036] Furthermore, in the above modeling method, in step S63, the learning efficiency and accuracy are optimized, specifically: within 60,000 learning steps, the fitting residual decreases by more than 4 orders of magnitude, and the mean square error between the validation set and the model prediction reaches the order of 10⁻⁵.
[0037] Furthermore, in the above modeling method, step S5 verifies the preliminary tail-retraction separation scenario aerodynamic model, specifically as follows:
[0038] Select several points from the validation set and input them into the preliminary tail retraction separation scenario aerodynamic model;
[0039] The accuracy of the model is judged based on the mean square error between several points and the fitted points.
[0040] Furthermore, in the above modeling method, in step S4, the preliminary tail-retraction separation scenario aerodynamic model includes input quantities and output quantities. The input quantities include relative pitch angle, relative roll angle, and relative displacement. The output quantities include the mother missile normal force, mother missile axial force, mother missile pitching moment, bullet normal force, bullet axial force, and bullet pitching moment.
[0041] Furthermore, in the above modeling method, in step S12, the aerodynamic settings include spatial discretization format, time-progression format, turbulence model, preprocessing, and entropy correction.
[0042] The advantages of this invention over the prior art are as follows:
[0043] (1) This invention establishes an aerodynamic model for tail-retreat separation scenario based on neural networks. This model can quickly predict (the calculation time is on the order of one second) the aerodynamic forces of the mother bomb and the bullet, avoiding a large amount of CFD numerical simulation work (the calculation time for each working condition is several hours), and providing a fast aerodynamic force prediction method for the design of separation trajectory and safety boundary in tail-retreat separation scenario. Attached Figure Description
[0044] Figure 1 This is the tail retraction separation scenario of the present invention;
[0045] Figure 2 This is the definition of neural network input quantities in this invention;
[0046] Figure 3 This is the neural network framework of the present invention;
[0047] Figure 4 This invention analyzes the influence of neural network neurons and layer number.
[0048] Figure 5 This is the neural network iterative process of the present invention;
[0049] Figure 6 This is a pneumatic model verification test of the present invention; (a) is the pitching moment coefficient diagram of the mother missile, and (b) is the pitching moment coefficient diagram of the bullet. Detailed Implementation
[0050] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.
[0051] like Figure 1 As shown, this invention discloses an aerodynamic modeling method for tail retraction separation scenarios based on neural networks, comprising the following steps:
[0052] Step S1: Establish aerodynamic data set;
[0053] Step S2: Take a portion of the aerodynamic data set as the sample set and the remaining data as the validation data to form a validation set;
[0054] Step S3: Establish a preliminary machine learning model;
[0055] Step S4: Optimize the preliminary machine learning model and establish a preliminary aerodynamic model for the tail-retraction separation scenario;
[0056] Step S5: Verify the preliminary tail-retraction separation scenario aerodynamic model. If the model accuracy meets the requirements, the final tail-retraction separation scenario aerodynamic model is obtained; otherwise, repeat steps S3 to S4.
[0057] Preferably, in step S1, an aerodynamic data set is established, specifically as follows:
[0058] Step S11: Generate the aerodynamic calculation grid required for CFD numerical simulation based on the characteristics of the tail-reversing separation vehicle;
[0059] Step S12: Select aerodynamic settings based on the characteristics of the tail-retracting separation aircraft to perform preliminary aerodynamic calculations and obtain aerodynamic forces and flow fields;
[0060] Step S13: Based on the aerodynamic forces and flow field, refine the mesh until the aerodynamic forces remain almost unchanged;
[0061] Step S14: Simulate the flow field of the tail-reverse separation vehicle under relative attitude and obtain aerodynamic data;
[0062] Step S15: Sort the aerodynamic data according to relative attitude to form an aerodynamic dataset.
[0063] Preferably, in step S11, the aerodynamic calculation grid required for CFD numerical simulation is generated based on the characteristics of the tail-retracting separation vehicle, specifically as follows:
[0064] Step S31: Generate a mesh using mesh generation software and import it into the tail-recoil separation vehicle digital model, which includes the mother missile and the submunition;
[0065] Step S32: Generate mesh line connectors at the edges of the aircraft digital model;
[0066] Step S33: Distribute grid points on each grid line connector and densify the grid in areas with more separated flows;
[0067] Step S34: Generate a mesh domain through the mesh line connector and project the mesh domain onto the digital model surface to make the mesh fit the body;
[0068] Step S35: Generate a grid block using the grid domain to complete the spatial grid generation;
[0069] Step S36: Keep the parent missile stationary, adjust the position of the submunition according to the relative attitude, and repeat steps S31 to S35 to finally form the aircraft grid under different relative attitudes.
[0070] Preferably, in step S36, the relative attitude is described using four-dimensional variables; the four-dimensional variables include relative distance in the x-direction, relative distance in the y-direction, angle of attack of the incoming flow α, and relative attitude angle az.
[0071] Preferably, in step S3, a preliminary machine learning model is established, and the specific steps are as follows:
[0072] Step S51: Establish the neural network framework;
[0073] Step S52: Based on the neural network framework, set the number of neural network layers;
[0074] Step S53: Set the number of neurons in each layer of the neural network;
[0075] Step S54: Normalize the relative distance in the x direction, the relative distance in the y direction, the angle of attack of the incoming flow α, and the relative attitude angle;
[0076] Step S55: Select a neural network loss function optimization method to form a preliminary machine learning model.
[0077] Preferably, in step S4, the preliminary machine learning model is optimized, and the specific steps are as follows:
[0078] Step S61: Input the sample set into the preliminary machine learning model for learning, and evaluate the machine learning efficiency and accuracy;
[0079] Step S62: Adjust the number of neurons, number of layers, loss function optimization method, and step size according to the learning accuracy;
[0080] Step S63: Repeat steps S61 to S62 until the learning efficiency and accuracy are optimal, and establish a preliminary aerodynamic model of the tail-retraction separation scenario.
[0081] Preferably, in step S63, the learning efficiency and accuracy are optimized, specifically: within 60,000 learning steps, the fitting residual decreases by more than 4 orders of magnitude, and the mean square error between the validation set and the model prediction reaches the order of 10⁻⁵.
[0082] Preferably, in step S5, the aerodynamic model of the preliminary tail-retraction separation scenario is verified, specifically as follows:
[0083] Select several points from the validation set and input them into the preliminary tail retraction separation scenario aerodynamic model;
[0084] The accuracy of the model is judged based on the mean square error between several points and the fitted points.
[0085] Preferably, in step S4, the preliminary tail-retraction separation scenario aerodynamic model includes input quantities and output quantities. The input quantities include relative pitch angle, relative roll angle and relative displacement. The output quantities include the mother missile normal force, mother missile axial force and mother missile pitching moment, and the bullet normal force, bullet axial force and bullet pitching moment.
[0086] Preferably, in step S12, the aerodynamic settings include spatial discretization format, time-progression format, turbulence model, preprocessing, and entropy correction.
[0087] Example
[0088] This invention establishes an aerodynamic model for tail-retardation separation scenarios based on neural networks. This model can quickly predict (with a computation time on the order of one second) the aerodynamic forces of the mother missile and the submunition, avoiding a large amount of CFD numerical simulation work (with a computation time of several hours for each scenario). It provides a rapid aerodynamic prediction method for the design of separation trajectories and safety boundaries in tail-retardation separation scenarios, has strong engineering application background, and is of great significance for the development of advanced aircraft in my country in the future.
[0089] The specific implementation steps of this invention are as follows:
[0090] (1) Based on the characteristics of the tail-retracting separation vehicle, the aerodynamic calculation grid required for CFD numerical simulation is generated. The axial grid spacing of the tail of the tail-retracting vehicle needs to be set to 1 / 500 of the length of the mother missile to ensure the simulation accuracy.
[0091] (2) Based on the characteristics of the tail-retreat separation aircraft, select aerodynamic settings such as spatial discretization format, time propagation format, turbulence model, preprocessing, and entropy correction to perform preliminary aerodynamic calculations and obtain aerodynamic forces and flow fields. The flow field should include basic information such as pressure, density, temperature, energy, and velocity.
[0092] (3) Based on the preliminary flow field and aerodynamic force obtained in step (2), the grid is refined until the aerodynamic force remains almost unchanged, proving that the grid size is reasonable and reliable.
[0093] (4) Simulate the flow field of the tail-retracting separation vehicle under relative attitudes (including relative distance in the x-direction, relative distance in the y-direction, incoming flow angle of attack α, and relative attitude angle az). The maximum relative distance in the x-direction and the relative distance in the y-direction can be set to twice the length of the mother missile to meet the modeling requirements. Obtain the static aerodynamic forces and moments of the mother missile and the submunition.
[0094] (5) Based on the aerodynamic data obtained in step (4), organize the aerodynamic data into the dataset required for the neural network.
[0095] (6) Take 80% of the data in the dataset from step (5) as the sample set and the remaining 20% as the validation data to form the validation set.
[0096] (7) Develop a preliminary machine learning model and initially set key parameters such as the number of learning layers, the number of neurons in each layer, and the loss function optimization method.
[0097] (8) Substitute the sample set formed in step (6) into the preliminary machine learning model formed in step (7) to evaluate the machine learning efficiency and accuracy.
[0098] (9) Adjust the number of learning layers, the number of neurons, and the loss function optimization method until the learning efficiency and accuracy reach the optimal level (within 60,000 learning steps, the fitting residual decreases by 4 orders of magnitude) to establish a tail-retraction separation scenario aerodynamic model.
[0099] (10) Select several points from the validation set as validation inputs and input them into the aerodynamic model in step (9). The root mean square error between the validation set and the model predictions reaches 10. -5 The scale was used to obtain the final aerodynamic model of the tail-reverse separation scenario.
[0100] This embodiment employs a neural network-based aerodynamic modeling method for tail-reverse separation scenarios to model and predict the aerodynamic forces in the longitudinal plane under tail-reverse separation scenarios.
[0101] Figure 1 This is a typical description of a tail-repel separation scenario. As can be seen from the figure, the tail-repel separation vehicle contains a parent missile and a submunition. The submunition is ejected from the parent missile. Since the submunition is in the flow interference zone within the parent missile, there is a relatively complex aerodynamic interference between the submunition and the parent missile. Therefore, the prediction of the aerodynamic forces and moments of the parent missile and the submunition is particularly important.
[0102] First, a preliminary plan for the inputs and outputs of the aerodynamic model is needed. Since this scenario only considers aerodynamic force prediction in the longitudinal plane, four quantities are sufficient to clearly describe the scenario. Therefore, there are four inputs, such as... Figure 2 , respectively, dx (relative x-displacement), dy (relative y-displacement), α (angle of attack of incoming flow), and az (relative pitch angle). Because the aerodynamic forces and moments of the mother missile and the submunition need to be predicted, the model outputs six quantities: mother missile normal force, mother missile axial force, mother missile pitch moment, submunition normal force, submunition axial force, and submunition pitch moment.
[0103] Then, the aerodynamic forces corresponding to different input quantities (dx, dy, α, az) are calculated, with 5 points taken for each dimension, and the values are as follows:
[0104] dx = 1m, 2m, 3m, 4m, 5m
[0105] dy = 1m, 2m, 3m, 4m, 5m
[0106] α=-10°,-5°,0°,5°,10°
[0107] az=0°, 4°, 8°, 12°, 16°
[0108] Therefore, a total of 4 calculations are required. 5 =625 computational states. CFD numerical simulation was used to calculate the aerodynamic forces and moments under different working conditions, forming an aerodynamic data set. 80% of the data was randomly selected as the sample set and 20% as the validation set.
[0109] Next, a preliminary aerodynamic model based on neural networks was established, such as... Figure 3 The input layer contains the model's input values, the output layer contains the model's output values, and the hidden layers in between require further testing and confirmation.
[0110] To allow for appropriate settings of the number of layers and neurons in the neural network, the fitting accuracy of the neural network was tested using 4, 8, 16, 24, and 32 neurons, and 1, 2, 3, and 4 hidden layers, respectively. Figure 4 As can be seen from the figure, a good fitting accuracy can be achieved when the number of hidden layers is 2 and the number of neurons is 16. Using more hidden layers and neurons will not further improve the fitting accuracy. Therefore, the number of neurons is 16 and the number of hidden layers is 2.
[0111] Then the aerodynamic model is further iterated, such as Figure 5 The model uses a sample set to obtain fitted values and a validation set to obtain predicted values. As shown in the figure, the fitted values and predicted values gradually converge after sufficient iterations, and the neuron parameters in the model are adjusted to the optimal values.
[0112] Finally, the validation set is input into the aerodynamic model for validation, such as... Figure 6 As can be seen from the figure, the aerodynamic model makes good predictions of the aerodynamic torques of both the parent missile and the submunition in this tail-recoil scenario, verifying that the aerodynamic model is reasonable.
[0113] The above is a preliminary test of the aerodynamic modeling method for tail-retraction separation scenarios based on neural networks. From the test results, we can see that the aerodynamic model framework established by this method has a strong aerodynamic prediction capability.
[0114] Although the present invention has been described in detail through the preferred embodiments above, it should be understood that the above description should not be considered as a limitation of the present invention. Various modifications and substitutions to the present invention will be apparent to those skilled in the art after reading the above description. Therefore, the scope of protection of the present invention should be defined by the appended claims.
[0115] The contents not described in detail in this specification are common knowledge to those skilled in the art.
Claims
1. A neural network-based aerodynamic modeling method for tail-retraction separation scenarios, characterized in that, Includes the following steps: S1. Establish aerodynamic data set; S2. Take a portion of the aerodynamic data set as the sample set and the remaining data as the verification data to form a verification set; S3. Establish a preliminary machine learning model; S4. Optimize the preliminary machine learning model and establish a preliminary aerodynamic model for the tail-retraction separation scenario; S5. Verify the preliminary tail-retraction separation scenario aerodynamic model. If the model accuracy meets the requirements, the final tail-retraction separation scenario aerodynamic model is obtained; otherwise, repeat steps S3 to S4. In step S4, the preliminary aerodynamic model of the tail-retraction separation scenario includes input quantities and output quantities. The input quantities include relative pitch angle, relative roll angle and relative displacement. The output quantities include the normal force of the mother missile, the axial force of the mother missile, the pitching moment of the mother missile, the normal force of the bullet, the axial force of the bullet and the pitching moment of the bullet.
2. The aerodynamic modeling method for tail retraction separation scenarios based on neural networks according to claim 1, characterized in that: In step S1, the aerodynamic data set is established, specifically as follows: S11. Generate the aerodynamic calculation grid required for CFD numerical simulation based on the characteristics of the tail-retarded separation vehicle. S12. Based on the characteristics of the tail-retardation separation aircraft, select aerodynamic settings to perform preliminary aerodynamic calculations and obtain aerodynamic forces and flow fields. S13. Based on the aforementioned aerodynamic forces and flow field, refine the mesh until the aerodynamic forces remain almost unchanged; S14. Simulate the flow field of the tail-reversing separation vehicle under relative attitude and obtain aerodynamic data; S15. Sort the aerodynamic data according to relative attitude to form an aerodynamic dataset.
3. The aerodynamic modeling method for tail retraction separation scenarios based on neural networks according to claim 2, characterized in that: In step S11, the aerodynamic calculation grid required for CFD numerical simulation is generated based on the characteristics of the tail-retracting separation vehicle. Specifically: S31: A mesh is generated using mesh generation software and imported into the tail-recoil separation vehicle digital model, including the parent missile and the submunition; S32: Generates mesh lines at the edge of the aircraft digital model connector; S33: Distribute grid points on each grid line connector to densify the flow in areas with more separate flows; S34: Generate a mesh domain through the mesh line connector and project the mesh domain onto the digital model surface to make the mesh fit the body; S35: Generate a spatial mesh by generating a mesh block from the mesh domain; S36: Keep the parent missile stationary, adjust the position of the submunition according to the relative attitude, repeat steps S31 to S35, and finally form the aircraft grid under different relative attitudes.
4. The aerodynamic modeling method for tail retraction separation scenario based on neural networks according to claim 3, characterized in that: In step S36, the relative attitude is described using four-dimensional variables; the four-dimensional variables include relative distance in the x direction, relative distance in the y direction, angle of attack of the incoming flow α, and relative attitude angle az.
5. The aerodynamic modeling method for tail retraction separation scenarios based on neural networks according to claim 1, characterized in that, In step S3, a preliminary machine learning model is established, and the specific steps are as follows: S51: Establish the neural network framework; S52: Based on the neural network framework, set the number of neural network layers; S53: Set the number of neurons in each layer of the neural network; S54: Normalize the relative distance in the x direction, the relative distance in the y direction, the angle of attack of the incoming flow α, and the relative attitude angle; S55: Select a neural network loss function optimization method to form a preliminary machine learning model.
6. The aerodynamic modeling method for a tail-retraction separation scenario based on a neural network according to claim 1, characterized in that, In step S4, the preliminary machine learning model is optimized, and the specific steps are as follows: S61: Input the sample set into the preliminary machine learning model for learning, and evaluate the machine learning efficiency and accuracy; S62: Adjust the number of neurons, number of layers, loss function optimization method, and step size according to the learning accuracy; S63: Repeat steps S61~S62 until the learning efficiency and accuracy are optimal, and establish a preliminary aerodynamic model of the tail-retraction separation scenario.
7. The aerodynamic modeling method for a tail-retraction separation scenario based on a neural network according to claim 6, characterized in that, In step S63, the learning efficiency and accuracy are optimized, specifically: within 60,000 learning steps, the fitting residual decreases by more than 4 orders of magnitude, and the mean square error between the validation set and the model prediction reaches the order of 10⁻⁵.
8. The aerodynamic modeling method for a tail-retraction separation scenario based on a neural network according to claim 1, characterized in that, In step S5, the preliminary tail-retraction separation aerodynamic model is verified, specifically as follows: Select several points from the validation set and input them into the preliminary tail retraction separation scenario aerodynamic model; The accuracy of the model is judged based on the mean square error between several points and the fitted points.
9. The aerodynamic modeling method for tail retraction separation scenario based on neural networks according to claim 2, characterized in that: In step S12, the aerodynamic settings include spatial discretization format, time-progression format, turbulence model, preprocessing, and entropy correction.