Airfoil profile optimization design method based on deep learning and reinforcement learning

Abstract

Provided is an airfoil profile optimization design method based on deep learning and reinforcement learning, relating to the technical field of mechanical optimization design. The method comprises: establishing an airfoil profile parameterization design platform based on deep learning, on the basis of the airfoil profile parameterization design platform, carrying out parameterization dimension reduction description on a basic airfoil profile to be optimized, to form at least four fitting parameters, adjusting the fitting parameters to implement geometric shape modification of the basic airfoil profile, and obtaining geometric coordinates of a first airfoil profile; obtaining a lift coefficient and a resistance coefficient of the airfoil profile on the basis of an airfoil profile performance prediction model; interacting with the airfoil profile performance prediction model and the airfoil profile parameterization design platform on the basis of an airfoil profile optimization design model; and setting a termination condition of reinforcement learning training for the airfoil profile optimization design model until a predetermined termination condition is met, and outputting geometric coordinates and an airfoil profile lift-to-drag ratio of the optimized first airfoil profile.

Description

A method for airfoil optimization design based on deep learning and reinforcement learning Technical Field

[0001] The present invention relates to the technical field of mechanical optimization design, and more specifically, to an airfoil optimization design method based on deep learning and reinforcement learning. Background Art

[0002] Fluid machinery and equipment, such as aircraft, underwater propulsion systems, and nuclear power plant main pumps, face highly parametric design requirements. Blades, as core components of these high-end fluid machinery and equipment, have a crucial impact on their performance. The airfoil is the foundation of blade design, and research into its optimal design is crucial for improving the performance of these fluid machinery and equipment.

[0003] Airfoil optimization design involves parametric modeling techniques and methods. Currently, parametric airfoil modeling methods often use too many design parameters to describe the primary objective. This creates a large design space for airfoil optimization, increasing both computational fluid dynamics prediction costs and the complexity of optimization algorithm searches. Regarding airfoil optimization methods, the application of evolutionary algorithms, such as genetic algorithms, to airfoil optimization has become a hot topic.

[0004] With the development of artificial intelligence, deep learning and reinforcement learning methods are increasingly being applied to the optimization and design of mechanical equipment. In particular, AlphaGo's impressive performance in Go and other gaming applications has made the application of reinforcement learning in control one of the hottest research areas in machine learning. Therefore, exploring the application of deep learning and reinforcement learning in the optimization and design of mechanical equipment is imperative, as it can alleviate the challenges of excessive airfoil parameters and the difficulty in finding optimization algorithms.

[0005] In view of this, the present invention provides an airfoil optimization design method based on deep learning and reinforcement learning.

[0006] Summary of the Invention

[0007] To overcome the challenges of the existing technology, this paper proposes an airfoil optimization design method based on deep learning and reinforcement learning. This method addresses the current issues in airfoil design, such as the large design space, high computational costs, and extensive algorithm search space caused by the excessively large design parameter dimensions. Furthermore, this method addresses the current limited application of reinforcement learning technology in the field of fluid machinery optimization design, resulting in unclear application processes and difficulty in determining reward values ​​and training termination conditions. This approach aims to further explore the potential of reinforcement learning applications.

[0008] According to one aspect of the present invention, a method for airfoil optimization design based on deep learning and reinforcement learning is provided, comprising the following steps:

[0009] A basic airfoil to be optimized is selected, and fitting parameters of the basic airfoil are obtained through an airfoil parameterized design platform based on deep learning. A large number of geometric coordinates of the first airfoil can be obtained by adjusting the fitting parameters of the basic airfoil.

[0010] Construct an airfoil performance prediction model based on the fluid dynamics software Xfoil; based on the geometric coordinates of the first airfoil and preset flow condition parameters, obtain the lift coefficient and drag coefficient corresponding to the first airfoil, and mark the ratio of the lift coefficient to the drag coefficient as the airfoil lift-to-drag ratio;

[0011] An airfoil optimization design model is established using a DQN reinforcement learning framework. The geometric coordinates and lift-to-drag ratio of the first airfoil are marked as the airfoil optimization state, and the reward value is set based on the change in the lift-to-drag ratio. The airfoil fitting parameters are continuously predicted based on the first adjustment interval. This process is repeated to obtain a large number of state-action-reward value samples. The DQN-based airfoil optimization design model learns based on the large number of obtained samples to maximize the reward value.

[0012] The network training process is divided into two rounds. A fixed number of steps is preset for each round. After the fixed number of steps is reached, the basic airfoil is updated to the optimal airfoil geometry and its fitting parameters obtained in the previous training process. This is used as the starting point to continue the next round of optimization and network training.

[0013] The termination conditions of reinforcement learning training are set for the airfoil optimization design model until the predetermined termination conditions are met, and the geometric coordinates and lift-to-drag ratio of the optimized first airfoil are output.

[0014] Preferably, the specific application logic of the airfoil parameterized design platform is:

[0015] The airfoil image dataset in the public airfoil database is used as the input value of the airfoil parametric design platform;

[0016] The convolutional neural network extracts the features of the airfoil image data, classifies and compresses the features of the airfoil image data through a multi-layer perceptron, and reduces the dimension of the airfoil image data features into N fitting parameters; the fitting parameters are marked as P n , P n =P0, P1, ..., P N-1 , and N ≥ 4;

[0017] Selecting a basic airfoil to be optimized, restoring the geometric coordinates of the basic airfoil corresponding to N fitting parameters, and obtaining the geometric coordinates of the initialized basic airfoil;

[0018] The N fitting parameters of the dimensionality reduction description are adjusted based on the first adjustment interval, the geometric coordinates of the adjusted basic airfoil are restored, constraints are set to prevent airfoil deformity, the airfoil that meets the constraints is marked as the first airfoil, and the geometric coordinates of the first airfoil are output.

[0019] Preferably, the fitting parameters of the basic airfoil are at least four fitting parameters.

[0020] Preferably, the specific logic of the constraint condition for preventing airfoil deformity is: up >y down ;

[0021] Among them, y up is the ordinate value of the coordinate point on the upper surface of the airfoil; down It is the ordinate value of the coordinate point on the lower surface of the airfoil.

[0022] Preferably, the specific application logic of the airfoil performance prediction model is:

[0023] Taking the first objective and flow condition parameters as input values ​​of the airfoil performance prediction model, wherein the flow condition parameters include the airfoil angle of attack and the Reynolds number;

[0024] The airfoil lift-to-drag ratio is taken as the output value of the airfoil performance prediction model.

[0025] Preferably, the airfoil optimization design model specifically includes the following logic:

[0026] The airfoil optimization state corresponding to each round is Z = [P, CL, CD], where the fitting parameters of the first airfoil are P, CL, and CD represent the lift coefficient and drag coefficient of the first airfoil under the corresponding fitting parameters, respectively;

[0027] The airfoil optimization design model determines the next action based on the airfoil optimization state corresponding to Z, thereby obtaining the fitting parameter set P_next of the next first airfoil;

[0028] Among them: the fitting parameter set P of the first airfoil is an N-dimensional vector, that is: P = [P0, P1, ..., P N-1 ], N is the number of airfoil parameters after dimensionality reduction; action is a vector with the same shape as the first airfoil fitting parameter set P, representing the influence of this action on each parameter, that is: action = [a0, a1, ..., a N-1 ]; a0-a N-1 You can take different values ​​as needed. In this design, the fixed step number is 0.001 for uniform sampling, that is, a0-a N-1 The first adjustment interval of each variable is ±0.001; after the action is determined, the next round of airfoil optimization state Zm+1 is: Z m+1 =[P next ,CL next ,CD next ] P next =P+action

[0029] Among them CL next ,CD next By the new parameter P next Calculated using computational fluid dynamics software.

[0030] Preferably, the reward value is set according to the optimization goal and various constraints, and the parameter α controls the weight of each reward in the entire reward. The specific formula is:

[0031] The optimization objective is to maximize the lift-to-drag ratio. The constraints are: Cl / Cd > Cl0 / Cd0. The lift-to-drag ratio of the airfoil corresponding to the initial state is Cl0 / Cd0; the lift-to-drag ratio of the airfoil corresponding to the current state is Cl(i) / Cd(i); and the lift-to-drag ratio of the airfoil corresponding to the previous state is Cl(i-1) / Cd(i-1). α1 and α2 are weight coefficients.

[0032] Preferably, the termination conditions of the reinforcement learning training process include:

[0033] If the first airfoil generated in each round is deformed, the trajectory is terminated and the vehicle returns, and the reward value is reward = A;

[0034] If the calculation results of the first airfoil generated in each round using the computational fluid dynamics software cannot converge, the trajectory is terminated and the game returns, and the reward value reward = B;

[0035] If the first airfoil generated in each round does not meet the additional design conditions, the trajectory is terminated and returned, and the reward value reward = C;

[0036] If the number of reinforcement learning training steps exceeds the preset fixed number of steps, the reinforcement learning training is terminated.

[0037] The technical effects and advantages of the airfoil optimization design method based on deep learning and reinforcement learning of the present invention are as follows:

[0038] This invention provides a new approach and solution for the optimization design of airfoils and, by extension, fluid machinery. By achieving dimensionality reduction of the primary objective design parameters using deep learning methods, it can reduce the computational workload and algorithm search space during optimization. Furthermore, through reinforcement learning technology, real-time optimization of airfoils can be achieved, enabling real-time interaction between simulation software and actions, reducing the overall sample size. Furthermore, the application of reinforcement learning technology in the optimization design of fluid machinery is a relatively new field, and its application process and the setting of reward values ​​and termination conditions during training are currently difficult issues. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] FIG1 is a flow chart of an airfoil optimization design method according to the present invention;

[0040] FIG2 is a network training process of the airfoil optimization design method of the present invention;

[0041] FIG3 is a diagram showing the network training effect of the airfoil optimization design method of the present invention;

[0042] Figure 4 is a window diagram of the airfoil parametric design software built based on deep learning in the present invention. DETAILED DESCRIPTION

[0043] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0044] Example 1

[0045] Referring to FIG. 1-2 , the airfoil optimization design method based on deep learning and reinforcement learning described in this embodiment includes the following steps:

[0046] A basic airfoil to be optimized is selected, and fitting parameters of the basic airfoil are obtained through an airfoil parameterized design platform based on deep learning. A large number of geometric coordinates of the first airfoil can be obtained by adjusting the fitting parameters of the basic airfoil.

[0047] It should be noted here that: after adjusting the fitting parameters of the basic airfoil, the design optimization of the basic airfoil geometry is achieved, as shown in Figure 4, and the optimized airfoil geometry image is marked as the first airfoil; in actual application, there is an applicable range for adjusting the fitting parameters of the basic airfoil, and this applicable range is marked as the first adjustment range.

[0048] The specific application logic of the airfoil parametric design platform is as follows:

[0049] The airfoil image dataset in the public airfoil database is used as the input value of the airfoil parametric design platform;

[0050] The convolutional neural network extracts the features of the airfoil image data, classifies and compresses the features of the airfoil image data through a multi-layer perceptron, and describes the features of the airfoil image data as N fitting parameters through parameterized dimensionality reduction; the fitting parameters are marked as P n , P n =P0, P1, ..., PN -1 , and N ≥ 4;

[0051] Selecting a basic airfoil to be optimized, restoring the geometric coordinates of the basic airfoil corresponding to N fitting parameters, and obtaining the geometric coordinates of the initialized basic airfoil;

[0052] Based on the first adjustment interval, N fitting parameters of the dimensionality reduction description are adjusted, the geometric coordinates of the adjusted basic airfoil are restored, and constraints are set to prevent airfoil deformity. The airfoil that meets the constraints is marked as a first airfoil, and the geometric coordinates of the first airfoil are output.

[0053] Furthermore, the number of the fitting parameters is preferably 6.

[0054] It should be noted here that the parametric design of airfoils includes 6 fitting parameters to describe the geometric characteristics of the airfoil. These 6 parameters do not belong to the traditional geometric characteristics of the airfoil, such as chord length, thickness, semi-span, leading edge sweep angle, thickness distribution, etc., but they can accurately express the airfoil geometry.

[0055] These six fitting parameters provide sufficient flexibility to adjust the airfoil during the design process without making the design space too complex, which helps maintain the efficiency and stability of the optimization algorithm. Of course, the number of parameters can also be adjusted according to the requirements and complexity of the specific problem.

[0056] Specifically, this example uses NACA0010 as the basic airfoil to be optimized, and generates 6 fitting parameters [0.1625, -0.0610, 0.1478, 0.0179, -0.2006, -0.1182] through dimensionality reduction of the airfoil parametric design platform based on deep learning.

[0057] Furthermore, the airfoil parameterized design platform includes a deep learning model that combines a convolutional neural network and a multi-layer perceptron;

[0058] Furthermore, the specific application logic of the airfoil parametric design platform is:

[0059] The airfoil image dataset in the public airfoil database is used as the input value of the airfoil parametric design platform;

[0060] It should be noted here that the airfoil image dataset in the public airfoil database is airfoil two-dimensional image data obtained by converting the airfoil coordinate point database published by the University of Illinois at Urbana-Champaign.

[0061] The convolution layer is used to extract the data features of the airfoil image, and then the image features are classified and compressed through the multi-layer perceptron, compressing the first target into 6 fitting parameters;

[0062] Based on the first adjustment interval, the six fitting parameters of the dimensionality reduction description are adjusted, the geometric coordinates of the adjusted basic airfoil are restored, and constraints are set to prevent airfoil deformity. The airfoil that meets the constraints is marked as a first airfoil, and the geometric coordinates of the first airfoil are output.

[0063] It should be noted here that the airfoil parametric design platform can achieve two major functions: obtaining 6 dimensionality reduction parameters of the basic airfoil and adjusting 6 fitting parameters to obtain a large number of geometric coordinates of the first airfoil.

[0064] The specific logic of the constraint condition for preventing airfoil deformity is: up >y down ;

[0065] Among them, y up is the ordinate value of the coordinate point on the upper surface of the airfoil; down It is the ordinate value of the coordinate point on the lower surface of the airfoil.

[0066] Construct an airfoil performance prediction model based on the fluid dynamics software Xfoil; based on the geometric coordinates of the first airfoil and preset flow condition parameters, obtain the lift coefficient and drag coefficient corresponding to the first airfoil, and mark the ratio of the lift coefficient to the drag coefficient as the airfoil lift-to-drag ratio;

[0067] Furthermore, the specific application logic of the airfoil performance prediction model is as follows:

[0068] The airfoil performance prediction model was fed with the geometric coordinates of the first airfoil and preset flow parameters, including the angle of attack and Reynolds number. For example, the airfoil angle of attack was chosen to be 3° and the Reynolds number to be 30,000. This optimization design was iterated over approximately 600 cycles, the results of which are shown in Figure 3. In Figure 3, the blue curve represents the airfoil with the highest lift-to-drag ratio in each cycle, while the orange curve represents the average of all maximum lift-to-drag ratios over the past 100 cycles. Cycles with lifts less than 100 are not included. Through the optimization design, the lift-to-drag ratio of the NACA0010 airfoil at a Reynolds number of 30,000 and an angle of attack of 3° was improved from 93 to 102, a 9.68% improvement.

[0069] The airfoil lift-to-drag ratio is taken as the output value of the airfoil performance prediction model.

[0070] An airfoil optimization design model is established using a DQN reinforcement learning framework. The geometric coordinates and lift-to-drag ratio of the first airfoil are marked as the airfoil optimization state, and the reward value is set based on the change in the lift-to-drag ratio. The airfoil fitting parameters are continuously predicted based on the first adjustment interval. This process is repeated to obtain a large number of state-action-reward value samples. The DQN-based airfoil optimization design model learns based on the large number of obtained samples to maximize the reward value.

[0071] Furthermore, the airfoil optimization design model specifically includes the following logic:

[0072] The airfoil optimization state corresponding to each round is Z = [P, CL, CD], where the fitting parameters of the first airfoil are P, CL, and CD represent the lift coefficient and drag coefficient of the first airfoil under the corresponding fitting parameters, respectively;

[0073] The airfoil optimization design model determines the next action based on the airfoil optimization state corresponding to Z, thereby obtaining the fitting parameter set P_next of the next first airfoil, as shown in Figure 3;

[0074] Among them: the fitting parameter set P of the first airfoil is an N-dimensional vector, that is: P = [P0, P1, ..., P N-1 ], N is the number of airfoil parameters after dimensionality reduction; action is a vector with the same shape as the first airfoil fitting parameter set P, representing the influence of this action on each parameter, that is: action = [a0, a1, ..., a N-1 ]; a0-a N-1 You can take different values ​​as needed. In this design, the fixed step number is 0.001 for uniform sampling, that is, a0-a N-1 The first adjustment interval of each variable is ±0.001; after the action is determined, the next round of airfoil optimization state Z m+1 is: Z m+1 =[P next ,CL next ,CD next ] P next =P+action

[0075] Among them CL next ,CD next By the new parameter P next Calculated using computational fluid dynamics software.

[0076] Specifically, the specific application logic of the airfoil optimization design model is as follows:

[0077] The airfoil optimization design model is a reinforcement learning airfoil optimization design framework based on DQN. DQN is a convolutional neural network that builds an intelligent agent. The agent is used to adjust the six fitting parameters and interacts with the computational fluid dynamics software xfoil.

[0078] When training the neural network, the agent selects the next action based on the reward value corresponding to the airfoil lift-drag ratio output by the airfoil performance prediction model, and interacts with the airfoil performance prediction model again to provide the next airfoil lift-drag ratio and the reward corresponding to the airfoil lift-drag ratio after the action.

[0079] In this way, a large number of airfoil fitting parameter-action-reward value samples are generated during the training process. The neural network uses these samples for training, continuously updating the network hyperparameters and improving the network's estimation accuracy of the reward. The training ends when the termination condition is met.

[0080] It should be noted here that the convolutional neural network acts as the intelligent agent, which is responsible for selecting the action of adjusting the six fitting parameters of the airfoil.

[0081] The agent's training is reward-based, typically based on the lift-to-drag ratio of the airfoil, as output by the airfoil performance prediction model. At each training step, the agent interacts with the airfoil performance prediction model based on the current airfoil parameter configuration, obtaining the lift-to-drag ratio as a reward. This reward guides the agent's next action.

[0082] The reward value is set according to the optimization goal and various constraints, and the parameter α controls the weight of each reward in the entire reward. The specific formula is:

[0083] Among them, the optimization goal is: maximizing the lift-to-drag ratio; the constraint condition is: Cl / Cd>Cl0 / Cd0; the airfoil lift-to-drag ratio corresponding to the initial state is Cl0 / Cd0; the airfoil lift-to-drag ratio corresponding to the current state is Cl(i) / Cd(i); the airfoil lift-to-drag ratio corresponding to the previous state is Cl(i-1) / Cd(i-1); α1 and α2 are weight coefficients. Specifically, α1 and α2 are preferably 0.5 respectively. This value and the specific reward value items can be changed according to the application scenario, optimization goal and constraint conditions of the airfoil to be optimized.

[0084] The airfoil optimization design model acquires a large number of examples of airfoil geometry parameters, actions, and reward values. The airfoil parameters are continuously adjusted to obtain new lift-to-drag ratios and reward values. These empirical examples are accumulated and used to train a reinforcement learning network to evaluate the difference between actual and estimated rewards. The network weights are continuously updated during training to improve the accuracy of reward estimates.

[0085] The network training process is divided into two rounds. A fixed number of steps is preset for each round. After the fixed number of steps is reached, the basic airfoil is updated to the optimal airfoil geometry and its fitting parameters obtained in the previous training process, and this is used as the starting point to continue the next round of optimization and network training.

[0086] Specifically, the termination conditions of the reinforcement learning training process include:

[0087] If the first airfoil generated in each round is deformed, the trajectory is terminated and the vehicle returns, and the reward value is -100.

[0088] If the calculation results of the first airfoil generated in each round cannot be converged using the computational fluid dynamics software, the trajectory is terminated and returned, and the reward value reward = -100;

[0089] If the first airfoil generated in each round does not meet additional design conditions, such as the maximum airfoil thickness and airfoil moment coefficient, the trajectory is terminated and returned, and the reward value reward = -1;

[0090] If the number of reinforcement learning training steps exceeds the preset fixed number of steps, the reinforcement learning training is terminated.

[0091] It's important to note that this type of termination condition avoids wasting time on inappropriate airfoil configurations and allows for faster learning of effective optimization strategies. It also handles situations where aerodynamic parameters cannot be calculated, allowing for timely exit and entry into the next round of training. These measures help improve model efficiency and stability.

[0092] The above formulas are all dimensionless and numerical calculations. The formulas are obtained by collecting a large amount of data and performing software simulation to obtain the most recent real situation. The preset parameters and thresholds in the formulas are set by technicians in this field according to actual conditions.

[0093] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in the present invention should be included in the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be based on the scope of protection of the claims.

[0094] Finally: The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for airfoil optimization design based on deep learning and reinforcement learning, characterized in that: The following steps are involved: A basic airfoil to be optimized is selected, and fitting parameters of the basic airfoil are obtained through an airfoil parameterization design platform based on deep learning: a large number of geometric coordinates of the first airfoil can be obtained by adjusting the fitting parameters of the basic airfoil; Construct an airfoil performance prediction model based on the fluid mechanics software Xfoil; obtain the lift coefficient and drag coefficient corresponding to the first airfoil based on the geometric coordinates of the first airfoil and the preset flow condition parameters, and mark the ratio of the lift coefficient to the drag coefficient as the airfoil lift-to-drag ratio; The DQN reinforcement learning framework is used to establish an airfoil optimization design model, and the geometric coordinates and lift-to-drag ratio of the first airfoil are marked as the airfoil optimization state; The reward value is set based on the change of the lift-to-drag ratio of the airfoil; the airfoil fitting parameter continuously predicts the action strategy based on the first adjustment interval, and a large number of state-action-reward value samples can be obtained by repeating this process; the airfoil optimization design model based on DQN is learned based on the large number of samples obtained to maximize the reward value; The network training process is divided into two rounds. A fixed number of steps is preset in each round. After the fixed number of steps is reached, the basic airfoil is updated to the optimal airfoil geometry and its fitting parameters obtained in the previous training process, and the next round of optimization and network training is continued from this point. The termination condition of the reinforcement learning training is set for the airfoil optimization design model until the predetermined termination condition is met, and the geometric coordinates and lift-to-drag ratio of the optimized first airfoil are output.

2. The airfoil optimization design method based on deep learning and reinforcement learning according to claim 1, characterized in that: The specific application logic of the airfoil parametric design platform is: The airfoil image dataset in the public airfoil database is used as the input value of the airfoil parametric design platform; The convolutional neural network extracts the features of the airfoil image data, classifies and compresses the features of the airfoil image data through a multi-layer perceptron, and reduces the dimension of the airfoil image data features to N fitting parameters; the fitting parameters are marked as P n , P n =P0, P1, ..., P N-1 , and N ≥ 4; Selecting a basic airfoil to be optimized, restoring the geometric coordinates of the basic airfoil corresponding to the N fitting parameters, and obtaining the geometric coordinates of the initialized basic airfoil; Based on the first adjustment interval, the N fitting parameters described by the dimensionality reduction are adjusted, the geometric coordinates of the adjusted basic airfoil are restored, and constraints are set to prevent airfoil deformity, the airfoil that meets the constraints is marked as a first airfoil, and the geometric coordinates of the first airfoil are output.

3. The airfoil optimization design method based on deep learning and reinforcement learning according to claim 2, characterized in that: The fitting parameters of the basic airfoil are at least four fitting parameters.

4. The airfoil optimization design method based on deep learning and reinforcement learning according to claim 2, characterized in that: The specific logic of the constraint condition for preventing airfoil deformity is: up >y down ; Among them, y up is the ordinate value of the coordinate point on the upper surface of the airfoil; down It is the ordinate value of the coordinate point on the lower surface of the airfoil.

5. The airfoil optimization design method based on deep learning and reinforcement learning according to claim 4, characterized in that: The specific application logic of the airfoil performance prediction model is: Taking the first target and flow condition parameters as input values ​​of the airfoil performance prediction model, wherein the flow condition parameters include the airfoil angle of attack and the Reynolds number; The airfoil lift-to-drag ratio is taken as the output value of the airfoil performance prediction model.

6. The airfoil optimization design method based on deep learning and reinforcement learning according to claim 5, characterized in that: The airfoil optimization design model specifically includes the following logic: The airfoil optimization state corresponding to each round is Z = [P, CL, CD], where the fitting parameters of the first airfoil are P,, CL, CD represent the lift coefficient and drag coefficient of the first airfoil under the corresponding fitting parameters respectively; The airfoil optimization design model determines the next action according to the airfoil optimization state corresponding to Z, thereby obtaining the fitting parameter set P_next of the next first airfoil; Where: The fitting parameter set P of the first airfoil is an N-dimensional vector, that is: P = [P0, P1, ..., P N-1 ], N is the number of airfoil parameters after dimension reduction; action action is a vector with the same shape as the fitting parameter set P of the first airfoil, representing the influence of this action on each parameter, that is: action = [a0, a1, ..., a N-1 ]; a0-a N-1 You can take different values ​​as needed. In this design, the fixed step number is 0.001 for uniform sampling, that is, a0-a N-1 The first adjustment interval of each variable is ±0.001; after the action is determined, the next round of airfoil optimization state Z m+1 for: Z m+1 =[P next ,CL next ,CD next ] P next =P+action Among them, CL next ,CD next By the new parameter P next Calculated using computational fluid dynamics software.

7. The airfoil optimization design method based on deep learning and reinforcement learning according to claim 6, characterized in that: The reward value is set according to the optimization goal and various constraints, and the parameter α controls the weight of each reward in the entire reward. The specific formula is: Among them, the optimization goal is: maximizing the lift-to-drag ratio; the constraint conditions are: Cl / Cd>Cl0 / Cd0; the airfoil lift-to-drag ratio corresponding to the initial state is Cl0 / Cd0; the airfoil lift-to-drag ratio corresponding to the current state is Cl(i) / Cd(i); the airfoil lift-to-drag ratio corresponding to the previous state is Cl(i-1) / Cd(i-1); α1 and α2 are weight coefficients.

8. The airfoil optimization design method based on deep learning and reinforcement learning according to claim 7, characterized in that: The termination conditions of the reinforcement learning training process include: If the first airfoil generated in each round is deformed, the trajectory is terminated and returned, and the reward value reward = A; If the calculation results of the first airfoil generated in each round cannot converge using computational fluid dynamics software, Terminate this trajectory and return, then the reward value reward = B; If the first airfoil generated in each round does not meet the additional design conditions, the trajectory is terminated and returned, and the reward value reward = C; If the number of steps of the reinforcement learning training exceeds the preset fixed number of steps, the reinforcement learning training is terminated.

Images