Airfoil profile adaptive design and optimization method combining reinforcement learning and conditional generative adversarial network

Abstract

The invention provides an airfoil profile adaptive design and optimization method based on reinforcement learning and a conditional generative adversarial network, and belongs to the field of airfoil profile optimization design. The method aims at achieving the two core functions of self-adaptive optimal airfoil generation and rapid optimization. According to the first function, through interaction of a reinforcement learning agent and a conditional generative adversarial network generator, generation strategies for generating optimal airfoil profile potential variables under different design conditions are continuously learned, so that the airfoil profile with the optimal aerodynamic performance is adaptively generated, intelligent and adaptive design of the airfoil profile is achieved, and the optimization period is remarkably shortened. Secondly, a self-adaptive agent model based on meta-learning is introduced, the model not only establishes a mapping relation among design conditions, aerodynamic performance and potential variables and directly predicts the optimal potential variables of a specific airfoil profile, but also can quickly adapt to an updated database through simple adjustment; the agent model predicts the optimal potential variable corresponding to the specific airfoil profile for the self-adaptive generation framework, so that the agent can quickly optimize the specific airfoil profile under the condition of fine adjustment.

Description

Technical Field

[0001] The present invention belongs to the technical field of airfoil aerodynamic optimization design, and specifically relates to an airfoil adaptive design and optimization method combining reinforcement learning and conditional generative adversarial network (CGAN). Background Art

[0002] Airfoil design and optimization is an important research direction in the fields of modern aerospace, fluid machinery and renewable energy. Its core goal is to optimize aerodynamic performance by accurately adjusting the geometric shape of the airfoil, such as improving lift-to-drag ratio, reducing drag, and enhancing stability. For example, patent [CN118296736A] provides an airfoil optimization design method. By parameterizing the airfoil, a control point and a proxy model of aerodynamic performance based on a support vector machine (SVM) are constructed, and a multi-population genetic algorithm (MPGA) is combined for optimization to achieve high efficiency and reliability in airfoil optimization design. However, it lacks the ability to dynamically adjust and quickly adapt to changing design conditions. Patent [CN112231836A] uses high-order Bezier curves to parameterize the airfoil, and combines genetic algorithms with numerical simulations to optimize the objective function value within a larger optimization range, thereby determining the target airfoil and its corresponding angle of attack. Its disadvantage is that it relies on a fixed optimization strategy and cannot quickly adapt to changing design conditions.

[0003] These airfoil design methods based on traditional optimization methods are highly dependent on the professional experience of designers. The optimization design process usually requires senior engineers to manually adjust the optimization parameters and iterate repeatedly to obtain a better design solution. However, this optimization process is often one-time. Once the design conditions change, designers must readjust the parameters and iterate again, resulting in a lot of repetitive work, low optimization response speed, and low efficiency. At the same time, this method is highly dependent on the experience and knowledge level of designers, and it is difficult to make full use of past optimization results, resulting in airfoil optimization design being limited by professional level, low efficiency and limited optimization effect. With the increasing demand for efficient airfoil design and optimization in modern industry, especially in response to diverse design conditions, methods for quickly generating optimal airfoils are particularly important.

[0004] In order to cope with the rapid design and adaptive optimization of airfoils, the existing airfoil optimization technology is gradually developing in the direction of intelligence. Patent [CN117150680A] is an airfoil optimization design method based on deep learning and reinforcement learning. It realizes the parametric dimension reduction of the airfoil through deep learning, describes the airfoil geometry as a small number of fitting parameters, and combines the reinforcement learning optimization design model to automatically adjust the fitting parameters to optimize the airfoil lift-to-drag ratio and achieve efficient airfoil optimization design. Patent [CN115618497A] is an airfoil optimization design method based on deep reinforcement learning. It adjusts the control points of the parameterized airfoil through deep reinforcement learning, and can quickly optimize when the design conditions are changed. Although these two airfoil optimization methods based on reinforcement learning algorithms can achieve adaptive optimization of airfoils, these two methods rely on fine-tuning the parameterized control points to optimize the airfoil. The optimization results of this method may not be able to jump out of the local optimal solution during the optimization process due to the limitation of the preset control point range, making it difficult to explore a better design solution, and the optimization effect depends on the setting of artificial control points. Patent [CN111814246B] proposes an airfoil inverse design method based on a generative adversarial network. By building a database of airfoils and aerodynamic curves, the generative adversarial network is trained to directly map the pressure coefficient curve to the airfoil shape and predict its corresponding Mach number, Reynolds number and angle of attack. This method is efficient and accurate, and can directly derive the airfoil shape and generate a smooth airfoil surface. However, GAN lacks effective optimization guidance when generating airfoils, and cannot achieve adaptive generation and optimization of the optimal airfoil under different experimental conditions.

[0005] In summary, although there have been studies on using generative adversarial networks (GAN) for airfoil inverse design and using reinforcement learning methods to achieve adaptive airfoil design, the airfoil design method that uses GAN alone can achieve airfoil generation under set conditions, but lacks guidance for the generation of the optimal airfoil, and the airfoil selection results cannot be separated from human experience. The airfoil optimization method that combines reinforcement learning with airfoil parameterized control points not only relies on the manual setting of control points, but also the airfoil generation results are limited by the setting of the control point range, which limits the generation of new airfoils under specific conditions. Summary of the invention

[0006] In order to solve the above technical problems, the present invention proposes an airfoil adaptive design and optimization method combining reinforcement learning and conditional generative adversarial network (CGAN) to construct a conditional generative adversarial neural network, wherein the input of the conditional generative adversarial neural network is the airfoil thickness and latent variables, and the output is the airfoil shape data;

[0007] The latent variables are iteratively optimized through reinforcement learning agents, and the conditional generative adversarial network is used to generate corresponding airfoil shape data according to the latent variables iteratively adjusted during the optimization process; the iteratively optimized airfoil shape data is evaluated according to the optimized design configuration data, where the optimized design configuration data includes design conditions and aerodynamic performance indicators; the evaluation result feedback guides the reinforcement learning agent to adjust the latent variables; finally, the optimal airfoil shape data is obtained, and the airfoil adaptive design is realized. Through several rounds of iterative optimization, it is ensured that the generated airfoil shape achieves the best aerodynamic performance.

[0008] Optionally, the conditional generative adversarial neural network includes a generator and a discriminator, the generator includes a fully connected layer, a deconvolution layer and a Bessel output layer connected in sequence, and the discriminator includes a convolution layer and a fully connected layer connected in sequence, wherein the input of the generator is the airfoil thickness and latent variables, and the output is the airfoil shape data, the input of the discriminator is the airfoil appearance data generated by the generator, the real airfoil appearance data and the airfoil thickness, and the output is the data distribution score, the generator is evaluated according to the data distribution score, and according to the evaluation result, the generator is trained to achieve that it can generate an airfoil that is almost the same as the actual airfoil.

[0009] Optionally, in the optimized design configuration data, the design conditions include Reynolds number, angle of attack and airfoil thickness, and the aerodynamic performance indicators include lift coefficient and drag coefficient.

[0010] Optionally, the process of obtaining the optimal airfoil shape data includes: initializing design conditions and latent variables through a reinforcement learning agent; inputting the latent variables into a conditional generative adversarial neural network for generation to obtain current airfoil shape data; simulating the current airfoil shape data to obtain different reward values; according to the different reward values, obtaining the expected cumulative reward for the current airfoil shape data through a value network; adjusting the latent variables according to the immediate reward agent through a strategy network to obtain adjusted latent variables; based on the adjusted latent variables, inputting them into the conditional generative adversarial neural network again, iterating the process of generation, simulation, immediate reward calculation and latent variable adjustment until the termination condition is reached, obtaining the optimal latent variables with the best immediate reward, and using the airfoil shape data corresponding to the optimal latent variables as the optimal shape data.

[0011] Optionally, different reward values ​​include an airfoil smoothness reward, a multi-angle-of-attack comprehensive performance reward, a maximum lift coefficient reward, and a minimum drag reward.

[0012] Optionally, the latent variables are adjusted by strategy parameters, wherein the strategy parameters are updated according to the expected cumulative reward, and the updating of the strategy parameters is achieved by maximizing an objective function, and the objective function is:

[0013]

[0014] Q represents the expected cumulative reward, si represents the action at the i-th moment, μ(s|θ μ ) represents the policy network, θ μ represents the strategy parameters, m represents the expected cumulative reward time step, and J represents the target value.

[0015] Optionally, before iteratively optimizing the latent variables through the reinforcement learning agent, the following steps are also included:

[0016] The optimal design configuration data is predicted through an adaptive agent model to obtain the optimal latent variables;

[0017] The reinforcement learning agent uses the optimal latent variables as the initial latent variables for iterative optimization, wherein the adaptive proxy model uses a neural network model to characterize the relationship between the optimized design configuration data in the historical optimal airfoil appearance data and the latent variables.

[0018] Optionally, the process of training the adaptive proxy model includes:

[0019] Constructing an optimal design library, wherein the optimal design library stores the optimal design configurations obtained after optimization under different design conditions; when a new optimization task is completed, storing the data corresponding to the optimization task in the optimal design library;

[0020] Divide the data in the optimal design library to obtain different sampling tasks;

[0021] In the inner layer training, different sampling tasks are input into the adaptive proxy model, the predicted values ​​are calculated, and the loss function L is calculated based on the predicted values:

[0022]

[0023] Where N represents the number of training samples. represents the predicted value, represents the actual optimal latent variable, i represents the training sample number;

[0024] In the outer layer training, several gradient descents are performed to update the temporary parameters θ′ of the proxy model i :

[0025]

[0026] Among them, θ represents the initial parameters of the adaptive proxy model, θ' represents the model parameters of the adaptive proxy model after optimization of the inner tasks of the meta-learning framework, α represents the learning rate, ▽ θ L new (θ) represents the loss function L on the current task new The gradient of θ;

[0027] Calculate the loss value of each sampling task according to the temporary parameters, calculate the overall loss according to the loss value of each sampling task, and update the initial parameters of the adaptive proxy model by back propagation according to the overall loss;

[0028] Repeat the task sampling, inner optimization and outer optimization steps until the prediction error of the adaptive proxy model on several tasks converges to a predetermined range, and obtain a trained adaptive proxy model.

[0029] Compared with the prior art, the present invention has the following advantages and technical effects:

[0030] The present invention combines conditional generative adversarial networks (CGAN), reinforcement learning modules and optimal design reuse mechanisms to achieve the generation of airfoil shapes with excellent aerodynamic performance under a variety of design conditions and to quickly optimize the airfoil under specific design conditions. The specific technical effects are as follows:

[0031] (1) Intelligent design: By introducing reinforcement learning algorithms, the intelligent agent can automatically adjust the design parameters according to different design conditions and the optimal airfoil design strategy, generate the optimal airfoil that meets the conditions, and realize the intelligentization of airfoil design and optimization.

[0032] (2) Adaptive and rapid response: Traditional airfoil designs often need to be re-optimized when encountering different design requirements. However, the present invention uses a reinforcement learning agent to gradually learn the optimization strategies under different conditions. It can adaptively and rapidly respond to new design requirements and generate optimized airfoil designs. No matter how the design conditions change, the model can give the best design solution within 1 minute. Compared with traditional optimization design methods, the optimization speed is increased by 60%, greatly reducing the design cycle and cost.

[0033] (3) Continuous improvement of optimization effect: The reinforcement learning algorithm agent in this optimization method can continuously learn and accumulate the experience of each optimal design. The agent will continuously accumulate the design strategy of the optimal airfoil under all design conditions. The ability of this agent to automatically learn and optimize design experience makes airfoil design and optimization free from the limitations of professional experience and can design new airfoils that have never been designed by humans. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] The drawings constituting a part of the present application are used to provide a further understanding of the present application. The illustrative embodiments and descriptions of the present application are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0035] Figure 1 A schematic diagram of a conditional generative adversarial neural network training and airfoil generation framework according to an embodiment of the present invention;

[0036] Figure 2 A schematic diagram of a flow chart of an airfoil adaptive design and optimization method according to an embodiment of the present invention;

[0037] Figure 3 A schematic diagram of the airfoil generation result of the optimal aerodynamic performance according to an embodiment of the present invention;

[0038] Figure 4 A schematic diagram of optimization results of another airfoil with optimal aerodynamic performance according to an embodiment of the present invention;

[0039] Figure 5 A structure diagram of an adversarial network is generated based on the condition of airfoil thickness according to an embodiment of the present invention;

[0040] Figure 6 A coupling relationship diagram between the reinforcement learning agent and the GCAN generator according to an embodiment of the present invention. DETAILED DESCRIPTION

[0041] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0042] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0043] In view of these problems, the advantages of reinforcement learning and conditional generative adversarial networks (CGAN) in adaptive generation and optimization of airfoils are particularly significant: reinforcement learning realizes autonomous learning by interacting with the environment, and can optimize the generation strategy under various design conditions, while CGAN can efficiently generate diversified airfoils that meet the requirements under different conditions without the limitation of the preset range of control points. Therefore, the present invention proposes a scheme for adaptively generating optimal airfoils based on reinforcement learning to guide CGAN, which uses reinforcement learning to guide CGAN to generate the optimal airfoil shape, which not only improves the optimization efficiency, but also realizes adaptive design under diversified conditions to overcome the limitations of existing methods.

[0044] In view of the shortcomings of traditional design methods, the present invention proposes an airfoil adaptive design and optimization method based on reinforcement learning and conditional generative adversarial networks. The method aims to achieve adaptive design and optimization and can achieve two functions. First, it can automatically generate the optimal airfoil according to different design conditions (such as airfoil thickness, Reynolds number), reduce manual intervention, shorten the optimization cycle, and thus achieve intelligent design of the airfoil. Second, introduce an adaptive agent model to establish a mapping relationship between design conditions, aerodynamic performance and latent variables, so that the corresponding latent variables can be predicted based on the given airfoil, and then combined with the adaptive airfoil generation function to achieve rapid optimization of specific airfoils. Thanks to the continuous learning ability and experience reuse ability of the reinforcement learning agent, it is possible to avoid starting over when facing new design conditions, effectively improving design efficiency and optimization effects.

[0045] To solve the above technical problems, the present invention proposes an airfoil adaptive design and optimization method combining reinforcement learning and conditional generative adversarial network (CGAN). The method can realize two functional modules, one of which is the function of adaptively generating the optimal airfoil, and the other is the function of rapidly optimizing a specific airfoil.

[0046] Specifically, this method uses continuous strategy learning of reinforcement learning agents to guide CGAN to generate airfoil shapes with optimal aerodynamic performance under various design conditions. At the same time, an adaptive proxy model based on meta-learning is introduced to achieve rapid optimization of airfoils under specific design conditions based on efficient reuse of historical optimization results.

[0047] 1. Adaptive optimal airfoil generation function based on reinforcement learning and CGAN

[0048] To achieve the above function 1, the present invention combines the reinforcement learning algorithm (deep deterministic policy gradient algorithm, DDPG) with the conditional generative adversarial network (CGAN) to adaptively generate the optimal airfoil shape that meets specific aerodynamic performance requirements. It mainly includes:

[0049] Construct a conditional generative adversarial neural network: The conditional generative adversarial neural network includes a generator and a discriminator. The generator that can generate a reasonable airfoil is trained. The structure is as follows: Figure 1 As shown. The input of the conditional generative adversarial neural network is the airfoil thickness t and the random latent variable z, and the output is the airfoil shape data The airfoil thickness data and coordinate data (x, y coordinates) are used to train a conditional generative adversarial neural network, and a generator G is obtained that can generate smooth and reasonable airfoils with different geometric shapes according to the airfoil thickness t. The trained generator can generate airfoil shape contour data according to a given latent variable z.

[0050] Reinforcement learning framework for adaptive airfoil generation: This reinforcement learning framework includes a DDPG agent, actions are additions and subtractions of latent variables, the state space is {Reynolds number Re, angle of attack α, airfoil thickness t, latent variable z}, and the environment includes a numerical evaluation environment and an airfoil generation environment (a trained CGAN generator).

[0051] Coupling mechanism of reinforcement learning and CGAN generator: The following steps are used to achieve deep coupling between the reinforcement learning agent and the CGAN generator to realize the generation of adaptive optimal airfoil.

[0052] ① Initial state acquisition: According to user input or design requirements, determine the initial design conditions t, Reynolds number Re, angle of attack range α, etc., and randomly initialize the potential variable z 0 , forming the initial state s 0 =[Re,α,t,z 0 ], including design conditions and current potential variables.

[0053] ②Agent output action: Change the current state s 0 =[Re,α,t,z 0 ] Input policy network μ(s|θ μ ), generating action a=Δz.

[0054] Δz=a=μ(s|θ μ )

[0055] ③ Latent variable update: adjust the latent variable through action Δz to generate a new latent variable:

[0056] z i+1 =z i +Δz

[0057] ④ Airfoil generation: Update the latent variable z i+1 The design thickness condition t is input into the generator G(t,z i+1 ), generate a new airfoil shape

[0058] ⑤ Aerodynamic performance evaluation: Input the generated airfoil shape data into the simulation environment (such as XFOIL) to calculate the aerodynamic performance index {C L ,C D}wait.

[0059] ⑥ Reward calculation: According to the aerodynamic performance index and design constraints, the immediate reward R is calculated according to the reward function i+1 .

[0060]

[0061] Among them, the first item is the airfoil smoothness reward, the second item is the average lift-to-drag ratio at multiple attack angles, and w1 and w2 are weights.

[0062] ⑦ Experience storage and network update: record experience tuples (s i ,a i ,R i+1 ,s i+1 ) and store it in the experience buffer pool. Sample small batches of experience from the buffer, update the Critic and Actor network parameters, and optimize the policy network to generate better actions.

[0063] ⑧ Iterative optimization: Repeat the above steps. The intelligent agent guides the generator to generate an airfoil with better aerodynamic performance by continuously adjusting the latent variable z until the preset optimization goal is met or the maximum number of iterations is reached. This can achieve the generation of the optimal airfoil and strategy learning under certain design conditions.

[0064] ⑨ Continuous learning of the agent: When there are continuous tasks to generate the optimal airfoil for new design conditions, steps ①-⑧ are continuously iterated. Through multiple rounds of iterative training under different design conditions, the agent continuously learns the relationship between design conditions and the aerodynamic performance of the airfoil, and optimizes and learns the strategy for generating the potential variables of the generator to generate the optimal airfoil. As the number of learning increases, the reinforcement learning agent becomes smarter and smarter, so that when encountering new design conditions, the agent can quickly give the potential variables corresponding to the airfoil with the optimal aerodynamic shape under the design conditions. The potential variables and thickness conditions are passed to the generator to generate the required airfoil shape contour data, realizing adaptive optimal airfoil generation.

[0065] 2. Rapid optimization of specific airfoils based on reinforcement learning, CGAN and optimal design reuse mechanism

[0066] Based on the above adaptive airfoil generation function, this function further designs an adaptive proxy model that can predict the corresponding potential variables according to the specific airfoil, and realizes the rapid optimization of the specific airfoil based on the adaptive airfoil generation function framework, which mainly includes:

[0067] ① The optimal design library is used to store the optimal design configuration obtained after optimization under different design conditions, including: the optimal potential variable z*, the corresponding optimized design configuration data, including design conditions (such as Reynolds number Re, angle of attack α, thickness t) and aerodynamic performance indicators (lift-drag coefficient L, drag coefficient D), in preparation for subsequent airfoil optimization, the record format is:

[0068]

[0069] Whenever a new optimization task is completed, the corresponding design conditions, aerodynamic performance indicators and optimal potential variables are stored in the optimal design library for subsequent reuse.

[0070] ②Adaptive agent model:

[0071] Agent model network architecture: The agent model uses an MLP neural network with an input layer, several hidden layers, and an output layer. The input layer receives design conditions including airfoil thickness t, Reynolds number Re, angle of attack range α, and aerodynamic performance indicators such as lift coefficient CL and drag coefficient CD. The output layer outputs the most predicted potential variable z*.

[0072] Meta-learning framework: A neural network model built based on meta-learning methods (such as Model-Agnostic Meta-Learning, MAML), with the goal of enabling the proxy model to adapt to the ever-changing historical optimal design library. That is, when the data in the database changes, the proxy model can be quickly updated and trained with a small amount of resources, thereby improving the prediction ability.

[0073] The above technical solution is described in detail:

[0074] like Figure 1-2 As shown, this embodiment provides an airfoil adaptive design and optimization method based on reinforcement learning and conditional generative adversarial network (CGAN).

[0075] The solution includes a reinforcement learning module, an airfoil generation module of a conditional generative adversarial neural network, and an adaptive proxy model prediction framework. The method can realize two functional modules, one of which is the function of adaptively generating the optimal airfoil, and the other is the function of rapidly optimizing a specific airfoil. Specifically, the method guides CGAN to generate an airfoil shape with optimal aerodynamic performance under a variety of design conditions through continuous strategy learning of a reinforcement learning agent. At the same time, an adaptive proxy model based on meta-learning is introduced to achieve rapid optimization of airfoils under specific design conditions based on efficient reuse of historical optimization results.

[0076] To achieve the above function one, the present invention combines reinforcement learning algorithms (such as deep deterministic policy gradient algorithms, DDPG) with conditional generative adversarial networks (CGAN) to achieve adaptive generation of optimal airfoil shapes that meet specific aerodynamic performance requirements. Its main process includes: in the optimization iteration process under specific design conditions, the reinforcement learning agent continuously adjusts the latent variable z. The adjusted latent variable is passed to the CGAN generator to generate corresponding airfoil shape data. The aerodynamic simulation environment evaluates the generated airfoil shape based on the optimized design configuration data (including design conditions and aerodynamic performance indicators). The evaluation results are fed back to the agent through a reward function to guide it to make the next action adjustment. This process continues to iterate until the generated airfoil shape achieves the best aerodynamic performance. Through multiple rounds of iterative training under different design conditions, the agent continuously learns the relationship between design conditions and airfoil aerodynamic performance, optimizes and learns the generation strategy of the latent variables of the generator to generate the optimal airfoil. In this way, when given new design conditions, the agent can quickly apply the learned strategy to generate the optimal airfoil that meets the requirements, and realize the adaptive optimization of airfoil design.

[0077] To achieve the above function 2, the present invention further constructs an adaptive proxy model to realize the rapid optimization of a specific airfoil using an adaptive airfoil generation framework. The main process includes: when receiving the optimization requirements under specific design conditions, calling the adaptive proxy model to predict the optimal latent variable z according to the current design conditions. Subsequently, the reinforcement learning agent performs iterative optimization based on the initial latent variable z, and the agent only needs to fine-tune on the basis of the optimal design to obtain the best optimization result.

[0078] The method provided by the present invention mainly involves the following three modules: a conditional generative adversarial neural network (CGAN) module, a reinforcement learning module and an adaptive proxy model module, each of which has the following functions:

[0079] Conditional Generative Adversarial Neural Network (CGAN): CGAN is responsible for generating an airfoil shape (airfoil coordinate data) with geometric consistency and rationality based on the airfoil thickness and latent variables provided by the reinforcement learning agent, ensuring that the generated airfoil meets the specified design conditions.

[0080] Reinforcement learning module: The reinforcement learning agent is responsible for adjusting and learning the generation strategy of the latent variables of the CGAN generator to generate the optimal airfoil under design conditions, continuously learning the generation strategy of the latent variables of the optimal airfoil under different design conditions, and guiding the CGAN generator to adaptively generate the airfoil with the best aerodynamic performance under different design conditions.

[0081] Adaptive agent model module: Receive the optimized design history data (including Reynolds number Re, angle of attack α, lift coefficient CL, drag coefficient CD, airfoil thickness t) and latent variable z explored by the reinforcement learning agent under various design conditions, use the neural network method to build an agent model that maps the Reynolds number Re, angle of attack α, lift coefficient CL, drag coefficient CD, airfoil thickness t and latent variable z, and combine the MAML meta-learning method to enable the agent model to adapt to the continuously updated optimal design library through simple parameter adjustments. When optimizing a specific airfoil, the latent variables corresponding to the specific airfoil can be quickly predicted through the provided optimized design configuration data, including airfoil aerodynamic information (such as lift and drag coefficient curve) and optimization conditions (such as thickness and Reynolds number), as the initial input of the reinforcement learning agent optimization, that is, the initial latent variables in the iterative process, to achieve adaptive and rapid optimization of the airfoil.

[0082] For the above different modules, the specific contents are as follows:

[0083] 1. Conditional Generative Adversarial Network (CGAN) Module

[0084] The conditional generative adversarial network consists of a generator and a discriminator, such as Figure 5 As shown, it is a conditional generative adversarial neural network structure for airfoil generation based on thickness conditions.

[0085] (1) Generator: The generator adopts a structure of several MLP layers, several fully connected layers, several convolutional layers connected in sequence, and a Bessel output layer to ensure that the generated airfoil shape data has a reasonable geometric shape and a smooth shape curve.

[0086] enter:

[0087] Thickness condition t: represents the thickness characteristic of the target airfoil, expressed in numerical form.

[0088] Latent variable z: The random latent variable is an important input of the generator, which directly affects the diversity and geometry of the generated airfoil. It is used to control the diversity of the generated airfoil. Different latent variables and thickness conditions in the latent space correspond to different airfoil geometry sampling.

[0089] Let G(t,z) denote the generator function, where:

[0090] t: airfoil thickness condition.

[0091] z: Random latent variable, initialized to Gaussian distribution z~N(0,1).

[0092] like Figure 5The generator network accepts the airfoil thickness condition t and the random latent variable z. The MLP first maps the thickness parameter t to obtain a feature vector, which is then concatenated with the random latent variable to obtain (t, z). The concatenated vector is passed to the convolutional layer and the fully connected layer to generate a fake airfoil x. fake .

[0093] Generated airfoil geometry coordinates x fake For: x fake =G(t,z)

[0094] Output: The generator G(t,z) generates the geometric shape of the airfoil, that is, the coordinate data of the airfoil profile.

[0095] The latent variables are used as input data of the generator and are directly related to the airfoil appearance. However, the latent variables are only used as generation variables of the airfoil appearance, mainly to generate the corresponding airfoil appearance data. The purpose of the generator is not mainly to generate the relationship between the latent variables and the airfoil appearance, but mainly to learn the rules of the airfoil appearance data itself to generate an airfoil appearance that conforms to the actual situation. The latent variables and auxiliary variables are only used as a call or to find relevant parameters for generating the airfoil appearance. Reinforcement learning and online learning adjust the latent variables and continuously select the generated airfoil appearance for evaluation and feedback to obtain the optimal airfoil appearance.

[0096] (2) Discriminator: The discriminator network D is responsible for judging whether the airfoil geometry generated by the generator conforms to the actual airfoil data distribution. The discriminator uses a combination of several MLP layers, several deconvolution layers, and several fully connected layers connected in sequence to extract the features of the airfoil geometry layer by layer and output a score. The score is used to judge whether the generated airfoil meets the design requirements.

[0097] enter:

[0098] Airfoil geometry: includes the coordinate data of the airfoil generated by the generator and the real airfoil.

[0099] Thickness Condition t: The discriminator also receives the thickness condition as input to ensure that the airfoil geometry meets the given thickness requirement.

[0100] like Figure 5 The discriminator structure is shown in Figure 1. The MLP layer in the discriminator receives thickness data and maps it into a high-dimensional feature vector, which is concatenated with the received airfoil geometry (including the generated airfoil and the real airfoil). The deconvolution layer and the fully connected layer are then used for feature extraction, and finally a probability score of the real airfoil is output.

[0101] Output: The discriminator outputs a score D(x|t) that indicates whether the input airfoil comes from the true data distribution.

[0102] (3) Training of the Conditional Generative Adversarial Network for Airfoils

[0103] The training goal of CGAN is to enable CGAN to generate various airfoil geometries based on the input thickness conditions and latent variables through adversarial training of the generator and the discriminator, thus providing a basis for subsequent airfoil design and optimization. The training goals are:

[0104]

[0105] Among them, G(z|t) means that the input is subject to z~p under the condition of thickness t z The airfoil generated by random noise is D(x|t), which indicates that the discriminator discriminates the input under the condition of thickness t and obeys x~p data The probability that the sample is a real sample, D(G(x|t)|t) represents the probability that the airfoil generated by the discriminator identification generator is a real airfoil, and E represents the mathematical expectation.

[0106] (4) Generator model training: The CGAN network is trained according to the set input and training objectives, so that the generator can generate smooth and reasonable airfoils with different geometric shapes according to the airfoil thickness t and the given latent variables. The trained generator interacts with the reinforcement learning agent through the latent variable z.

[0107] 2. Reinforcement Learning Module

[0108] The present invention adopts the Deep Deterministic Policy Gradient (DDPG) algorithm as the core optimization algorithm of the reinforcement learning agent to achieve accurate adjustment of the latent variable z, thereby guiding the trained conditional generative adversarial network (CGAN) generator to generate the airfoil shape with the best aerodynamic performance. The reinforcement learning module for latent variable adjustment is described in detail below:

[0109] (1) State space S

[0110] Including the current design conditions (such as Reynolds number Re, angle of attack range α (-5, 20), airfoil thickness t) and the potential variable z under the current iteration step, the state space S is defined as:

[0111] S=[Re,α,t,z t ]

[0112] (2) Action space A:

[0113] The action space A corresponds to the amount of adjustment the agent makes to the latent variable z. Since DDPG is applicable to continuous action spaces, the action vector a t is the incremental adjustment of the latent variable:

[0114] a=Δz

[0115] After the action is executed, the next state potential variable z is formed i+1 =z i +Δz, new latent variable z i+1 The new airfoil shape data is passed to the generator G.

[0116] (3) Generation and simulation feedback environment:

[0117] Airfoil generation: When the agent gives the current latent variable (z), the generator will output a corresponding airfoil coordinate based on (z) and the given airfoil thickness and other conditions. One latent variable corresponds to one airfoil geometry.

[0118] Simulation feedback: The generated airfoil shape is input into aerodynamic analysis software (such as XFOIL) to calculate performance indicators such as lift coefficient, drag coefficient, maximum lift, minimum drag, etc. Then, an immediate reward R is obtained based on a pre-defined reward function (taking into account lift-to-drag ratio, smoothness, etc., maximum lift, and minimum drag).

[0119] (4) Reward function:

[0120] Airfoil Smoothness Bonus: Use the airfoil smoothness bonus to ensure that the generated airfoil shape is smooth and continuous. Define limits on the curvature variation to achieve this. For designs that exceed the curvature limit, apply a negative bonus.

[0121] R s =Q·curve_smooth

[0122] R s Q is the smoothness reward, Q is the smoothness penalty factor, and curve_smooth represents the curvature penalty intensity.

[0123] Calculation of curve_smooth:

[0124] For the discrete airfoil coordinate points: (x 1 ,y 1 ),...,(x n ,y n ), by calculating the standard deviation σ of the curvature of adjacent points k A measure of airfoil smoothness.

[0125]

[0126] k is the curvature, and the calculation formula is as follows:

[0127]

[0128] where x i ', xi ",y i ',y i ",s i 'respectively the first and second derivatives of x, the first and second derivatives of y and the radian, where x represents the x-axis coordinate of the airfoil coordinate point, y represents the y-axis coordinate, and k represents the curvature, where represents the mean, k i represents the curvature of the i-th coordinate point, N represents the total number of coordinate points, i represents the coordinate point number, the superscript ' represents the first-order derivative of the corresponding parameter, the superscript " represents the second-order derivative of the corresponding parameter, and s i 'Indicates the radian of the second coordinate point.

[0129] Comprehensive performance reward for multiple angles of attack: In order to make the generated airfoil have excellent performance within the set angle of attack range, the average lift-to-drag ratio at multiple angles of attack can be used as the performance reward function. (Different performance reward functions can be set for different design and optimization goals, such as the average lift coefficient at multiple angles of attack, or the maximum stall angle of the airfoil)

[0130]

[0131] Among them, R m is the average lift-to-drag ratio of multiple angles of attack, n is the number of angles of attack, L i and D i are the lift coefficient and drag coefficient corresponding to the ith angle of attack respectively.

[0132] Maximum lift coefficient bonus: If the target requires sufficient lift, pay attention to the maximum lift coefficient

[0133] If the maximum lift coefficient of the generated airfoil increases, a positive reward is given, and if it decreases, a negative reward is given. The reward is the maximum change in lift-to-drag ratio.

[0134] Minimum resistance reward: If the goal is to reduce resistance, then focus on the minimum resistance coefficient

[0135] Total reward function R: Comprehensive consideration of aerodynamic performance and geometric smoothness

[0136]

[0137] (5) Network structure of DDPG agent

[0138] like Figure 6 The structure diagram of the intelligent agent in . Including the policy network and the value network:

[0139] Policy network μ(s|θ μ) is responsible for generating a deterministic action a, which is the adjustment to the latent variable z.

[0140] a t =μ(s|θ μ )

[0141] The value network evaluates the Q-value (expected cumulative reward) for a given state and action combination.

[0142] Q(s t ,a t |θ μ )

[0143] (6) Strategy update method

[0144] Through the policy gradient method, according to the Q value evaluated by the Critic network, adjust the policy network parameter θ μ , so that the generated action a can maximize the expected reward. The policy network parameters are updated by maximizing the following objective function:

[0145]

[0146] m represents the expected time step.

[0147] 3. Meta-learning based adaptive agent model

[0148] In order to make up for the deficiency that the adaptive optimal airfoil generation framework based solely on reinforcement learning and CGAN cannot optimize specific airfoils, the present invention proposes an adaptive agent model based on meta-learning to achieve specific airfoil optimization and realize the second function of the present invention, that is, predicting the optimal latent variables corresponding to the specific airfoil as the initial action of the intelligent agent optimization, that is, the initial latent variables in the iterative process. The intelligent agent only needs to fine-tune on the basis of the optimal design to obtain the best optimization result.

[0149] The role of the adaptive proxy model is to quickly predict latent variables based on a given airfoil. When there is new data in the database, the proxy model uses a small amount of data through meta-learning methods to quickly adapt to the new task and improve the accuracy of latent variable prediction. It mainly includes the following components:

[0150] (1) Optimal Design Library

[0151] The optimal design library is used to store the optimal design configuration obtained after optimization under different design conditions, including: the optimal potential variable z*, the corresponding optimized design configuration data, including design conditions (such as Reynolds number Re, angle of attack α, thickness t) and aerodynamic performance indicators (lift-drag coefficient L, drag coefficient D), in preparation for subsequent airfoil optimization, the record format is:

[0152]

[0153] Whenever a new optimization task is completed, the corresponding design conditions, aerodynamic performance indicators and optimal potential variables are stored in the optimal design library for subsequent reuse.

[0154] (2) Meta-learning framework

[0155] MAML is a learning method that aims to improve the model's ability to quickly adapt to new tasks. Its main purpose is to learn a good initialization parameter in multi-task training so that when faced with a new task, only a small amount of gradient update is needed to achieve good performance.

[0156] a. Task division: Divide the data in the optimal design library into multiple subtasks, each subtask corresponding to a different design condition combination {t, Re, α, CL, CD}.

[0157] b. Inner layer training:

[0158] For each sampled task, calculate the predicted value z* of the surrogate model using the design conditions and optimal latent variables for that task as input.

[0159] Based on the predicted value The actual optimal latent variable The difference between them is used to calculate the loss function L:

[0160]

[0161] Perform one or a few gradient descents to update the temporary parameters θ′ of the proxy model i :

[0162]

[0163] c. Outer training:

[0164] Use the temporary parameters θ′ obtained by inner optimization i Calculate the loss for each task

[0165] Calculate the overall loss of the proxy model on each subtask and update the initial parameters θ of the proxy model through back propagation:

[0166]

[0167] d. Iterative optimization: Repeat the task sampling, inner optimization, and outer optimization steps until the prediction error of the proxy model on multiple tasks converges to a predetermined range.

[0168] (3) Agent Model Framework

[0169] The proxy model is a neural network model trained based on the meta-learning framework, which is responsible for predicting the optimal latent variable z* under specific design conditions. The model has the ability to quickly adapt to new data. When faced with database updates, it can adjust and update the proxy model based on a small amount of data to provide highly accurate prediction results. The proxy model uses an MLP neural network with an input layer, several hidden layers, and an output layer. The input layer receives design conditions including airfoil thickness t, Reynolds number Re, angle of attack range α, and aerodynamic performance indicators such as lift coefficient CL and drag coefficient CD. The output layer outputs the predicted most potential variable z*.

[0170] a. Input new task data: provide new design conditions {t, Re, α, CL, CD}.

[0171] b. Predict latent variables: Input the new design conditions {t, Re, α, CL, CD} into the adaptive agent model and output the predicted optimal latent variables z*.

[0172] Input=(Re,α,t,CL,CD)

[0173] Output the optimal latent variable z* for prediction

[0174] Output = z*

[0175] c. Provide initial latent variables: Use the predicted z* as the initial latent variable of the reinforcement learning agent, as the initial latent variable corresponding to the specific airfoil.

[0176] (5) Training process of adaptive agent model

[0177] The training process of the adaptive agent model combines the training steps of the meta-learning framework and the agent model. The specific process is as follows:

[0178] a. Construction of the optimal design library

[0179] Data accumulation: Through multiple optimization tasks, the optimal potential variable z* and its corresponding airfoil aerodynamic performance data under different design conditions are accumulated.

[0180] Data management: Regularly update the optimal design library to ensure data diversity and representativeness, remove redundant data, and maintain the efficiency of the library.

[0181] b. Adaptive surrogate model training

[0182] Task sampling: Randomly sample multiple design tasks from the optimal design library, each task contains different design conditions and corresponding optimal latent variables.

[0183] Inner optimization: For each sampled task, use the data of the task to perform a gradient descent, update the model parameters, and obtain the temporary model parameters θ′ that are suitable for the task i .

[0184] Outer optimization: use temporary model parameters θ′ for all tasks i Calculate the loss for each task Through back-propagation, the initial parameters θ of the proxy model are updated to minimize its overall loss on all tasks.

[0185] Iterative optimization: Repeat the task sampling, inner optimization, and outer optimization steps until the proxy model performs well on multiple tasks and has the ability to quickly adapt to new design tasks.

[0186] Online update of proxy models

[0187] As the historical design database is updated, the proxy model needs to be continuously updated to improve its adaptability and prediction accuracy.

[0188] a. Data accumulation: Newly completed optimization task data is added to the optimal design library.

[0189] b. Regular retraining: Regularly use the latest optimal design library data to re-perform meta-training and update the parameters of the proxy model.

[0190] 4. Coupling mechanism between agent and generator - adaptive optimal airfoil generation function

[0191] The reinforcement learning agent does not directly change the internal structure of the generator, but iteratively adjusts the external input of the latent variable (z) to make the generator output different airfoil shapes, and then continuously updates and adjusts according to the aerodynamic simulation feedback, and finally guides the generator to produce the airfoil with the best comprehensive performance under this condition, and continuously learns the generation strategy in different iteration cycles. The following is the coupling process of the reinforcement learning agent and the generator in the adaptive optimal aerodynamic performance airfoil generation, as shown in Figure 1. Figure 6 As shown:

[0192] ① Initial state acquisition: According to user input or design requirements, determine the initial design conditions t, Reynolds number Re, angle of attack range α, etc., and randomly initialize the potential variable z 0 , forming the initial state s 0 =[Re,α,t,z 0 ], including design conditions and current potential variables.

[0193] ②Agent output action: Change the current state s 0 =[Re,α,t,z 0 ] Input policy network μ(s|θ μ ), generating action a=Δz.

[0194] Δz=a=μ(s|θ μ )

[0195] ③ Latent variable update: adjust the latent variable through action Δz to generate a new latent variable:

[0196] z i+1 =z i +Δz

[0197] ④ Airfoil generation: Update the latent variable z i+1 The design thickness condition t is input into the generator G(t,z i+1 ), generate a new airfoil shape

[0198] ⑤ Aerodynamic performance evaluation: Input the generated airfoil shape data into the simulation environment (such as XFOIL) to calculate the aerodynamic performance index {C L ,C D}wait.

[0199] ⑥ Reward calculation: According to the aerodynamic performance index and design constraints, the immediate reward R is calculated according to the reward function i+1 .

[0200]

[0201] Among them, the first item is the airfoil smoothness reward, the second item is the average lift-to-drag ratio at multiple attack angles, and w1 and w2 are weights.

[0202] ⑦ Experience storage and network update: record experience tuples (s i ,a i ,R i+1 ,s i+1 ) and store it in the experience buffer pool. Sample small batches of experience from the buffer, update the Critic and Actor network parameters, and optimize the policy network to generate better actions.

[0203] ⑧ Iterative optimization: Repeat the above steps. The intelligent agent guides the generator to generate an airfoil with better aerodynamic performance by continuously adjusting the latent variable z until the preset optimization goal is met or the maximum number of iterations is reached. This can achieve the generation of the optimal airfoil and strategy learning under certain design conditions.

[0204] ⑨ Continuous learning of the agent: When there are continuous tasks to generate the optimal airfoil for new design conditions, steps ①-⑧ are continuously iterated. Through multiple rounds of iterative training under different design conditions, the agent continuously learns the relationship between design conditions and the aerodynamic performance of the airfoil, and optimizes and learns the strategy for generating the potential variables of the generator to generate the optimal airfoil. As the number of learning increases, the reinforcement learning agent becomes smarter and smarter, so that when encountering new design conditions, the agent can quickly give the potential variables corresponding to the airfoil with the optimal aerodynamic shape under the design conditions. The potential variables and thickness conditions are passed to the generator to generate the required airfoil shape contour data, realizing adaptive optimal airfoil generation.

[0205] ⑩Result output and design library update

[0206] Show optimal airfoil: Show the generated airfoil shape and its aerodynamic performance indicators.

[0207] Store optimization results: Add the design conditions, aerodynamic performance indicators and optimal potential variables z* obtained from this optimization to the optimal design library for future reuse.

[0208] 5. Adaptive surrogate model and adaptive optimal airfoil generation framework coupling mechanism - specific airfoil rapid optimization

[0209] The coupling of the adaptive proxy model and the adaptive optimal airfoil generation framework enables the rapid optimization of a specific airfoil under specific conditions. The specific coupling process is as follows:

[0210] ①Receive optimization tasks for specific airfoils

[0211] The user inputs the attack angle range α of a specific airfoil and the corresponding lift and drag coefficient, airfoil thickness t, Reynolds number Re, etc.

[0212] ②Proxy model predicts latent variables

[0213] Input design conditions: Input the new design conditions {t, Re, α, CL, CD} into the adaptive surrogate model.

[0214] Fast adaptation and prediction: Receive user input and pass it into the adaptive agent model, and output the predicted optimal latent variable z*.

[0215] Provide initial latent variables: pass the predicted z* to the reinforcement learning agent.

[0216] ③ Optimization of reinforcement learning agents

[0217] Initialize latent variables: Use the surrogate model’s predicted z* as the initial latent variable z0.

[0218] Generate an airfoil corresponding to a specific airfoil: through the generator G(t,z 0 ) Generate an airfoil with the same shape as the given airfoil as the optimized initial airfoil

[0219] Pneumatic simulation: Enter the aerodynamic simulation environment and calculate aerodynamic performance indicators {CL, CD}, etc.

[0220] Reward calculation and feedback: Calculate the immediate reward R based on aerodynamic performance and design constraints 0 .

[0221] Strategy update: The agent generates a new action a0 through the policy network based on the current state s0 = {t, Re, α, z0} and reward R0, and adjusts the latent variable z1 = z0 + a0.

[0222] Iterative optimization: Repeat the steps of generation, simulation, reward calculation and strategy update to gradually optimize z and generate an airfoil shape with better aerodynamic performance

[0223] ④Result output and design library update

[0224] Display the optimal airfoil: The optimized airfoil shape And its aerodynamic performance indicators are displayed to users.

[0225] Store optimization results: Add the design conditions, aerodynamic performance indicators and optimal potential variables z* obtained from this optimization to the optimal design library for future reuse.

[0226] ⑤Proxy model update

[0227] As the optimal airfoil generation and optimization tasks increase, the results of each iteration are saved in the optimal design history data set. When the data set accumulates to a certain amount, in order to further improve the prediction accuracy of the proxy model, the proxy model is adjusted and updated based on the meta-learning framework using a small amount of updated data.

[0228] 6. Online model application and continuous learning

[0229] Functional implementation: This framework and model can achieve two functions. One is that according to the airfoil operating conditions, the intelligent agent adaptively generates airfoils with excellent comprehensive performance under multiple angles of attack under certain design conditions, such as Figure 2As shown in ④, according to the airfoil design conditions, the optimal airfoil appearance data is generated by the cooperation of the reinforcement learning agent and the conditional generative adversarial neural network. Another function is that when the user wants to optimize the airfoil shape under a certain operating environment, the adaptive proxy model agent model quickly predicts the optimal latent variables corresponding to the airfoil according to the Reynolds number, airfoil thickness, and lift-drag coefficient curve (angle of attack and corresponding lift-drag coefficient) corresponding to the airfoil, and uses the initial latent variables optimized by reinforcement learning. The agent fine-tunes the latent variables based on the predicted latent variables and uses the optimal airfoil generation strategy to quickly optimize the airfoil under specific operating conditions, such as Figure 2 In step ⑤, firstly, the airfoil optimization conditions, i.e., the optimized configuration data, are input into the proxy model to predict the optimal latent variables as the initial latent variables of the reinforcement learning agent. The reinforcement learning agent then performs iterative optimization based on the initial latent variables and cooperates with the generator to generate the optimal airfoil appearance data.

[0230] Online continuous learning: Because the reinforcement learning agent has the ability to continuously optimize and learn strategies, and the proxy model in the optimal design reuse mechanism proposed in the present invention also has the ability to quickly use new data, when the optimization model is used for optimal airfoil design and airfoil optimization, the agent will become smarter, the designed and optimized airfoils will become more and more excellent, and the prediction accuracy of the proxy model will gradually become more accurate.

[0231] in, Figure 3 Re = 10 4 ,The airfoil with the best aerodynamic performance generated when t=12%; Figure 4 Re = 10 3 Under the condition of Re=10 4 ,The optimization result of the airfoil with the best aerodynamic performance when t=12%. It can be seen that this method can generate the optimal airfoil shape that meets a certain condition, and can also optimize the original airfoil according to a certain condition.

[0232] It should be noted that the conditional generative adversarial network is mainly used to fit the inherent law between the airfoil coordinates, rather than to fit the law between the input data and the output data. It guarantees that the shape data of the airfoil conforms to the inherent characteristics of the normal airfoil, such as smooth appearance and the requirements between different coordinates at different positions. It should also be noted that the conditional generative adversarial network, speaking alone, is an airfoil generation network, not an airfoil optimization network. Its optimization content is mainly in the two parts of subsequent reinforcement learning and online learning. Compared with the generative adversarial network with multi-parameter input and encoder addition, the above-mentioned purpose of use of the present invention focuses on the data characteristics of the output data itself, rather than the characteristics between the multi-parameters and the output data. At the same time, due to the black box characteristics of the neural network, the relationship between the multi-parameters and the output data is not clear, and the correspondence between the multi-parameters and the output data cannot be guaranteed. The present invention combines the subsequent reinforcement learning method and the online learning method to make the relationship between the output airfoil appearance and the observed values ​​of the multi-parameters clear, and ensure that the association between the airfoil appearance and the observed values ​​of the multi-parameters can be explained. The airfoil design method based on reinforcement learning mainly limits the range of the airfoil appearance, and takes the relevant range under the limit as the design range of the airfoil. However, within this range, the appearance outside the range cannot be simulated. If the airfoil appearance outside the range is optimal, it cannot be searched. However, if the range limit is removed, it will not be able to effectively guide the adjustment of the airfoil appearance control points, resulting in the final airfoil data being unable to obtain the optimal one and the data search method resulting in an increase in the amount of data calculation. The present invention combines the generative adversarial network and reinforcement learning methods, and uses the generative adversarial network to limit the airfoil appearance data to ensure that it complies with the normal rules of the airfoil data. Reinforcement learning is used as the target parameter to be designed as a limit as a guiding optimization direction, and guides the generative adversarial network to generate corresponding airfoil data that meets the conditions. Through the above-mentioned restrictions, the corresponding airfoil appearance can be effectively and quickly generated without causing the above-mentioned impact of the prior art, and can effectively solve the problems existing in the prior art.

[0233] The above are only preferred specific implementations of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present application should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A method for adaptive design and optimization of airfoil based on reinforcement learning and conditional generative adversarial network, characterized in that: include: A conditional generative adversarial neural network is constructed, wherein the input of the conditional generative adversarial neural network is the airfoil thickness and latent variables, and the output is the airfoil shape data; The latent variables are iteratively optimized through reinforcement learning agents, and the corresponding airfoil shape data is generated using a conditional generative adversarial network based on the latent variables iteratively adjusted during the optimization process; The iteratively optimized airfoil shape data is evaluated according to the optimized design configuration data, wherein the optimized design configuration data includes design conditions and aerodynamic performance indicators; the evaluation result feedback guides the reinforcement learning agent to adjust the potential variables; finally, the optimal airfoil shape data is obtained to realize the airfoil adaptive design, and through several rounds of iterative optimization, it is ensured that the generated airfoil shape achieves the best aerodynamic performance.

2. The method according to claim 1, characterized in that: The conditional generative adversarial neural network includes a generator and a discriminator, wherein the generator includes a fully connected layer, a deconvolution layer and a Bessel output layer connected in sequence, and the discriminator includes a convolution layer and a fully connected layer connected in sequence, wherein the input of the generator is the airfoil thickness and latent variables, and the output is the airfoil shape data, the input of the discriminator is the airfoil appearance data generated by the generator, the real airfoil appearance data and the airfoil thickness, and the output is the data distribution score, the generator is evaluated according to the data distribution score, and according to the evaluation result, the generator is trained to achieve that it can generate an airfoil that is almost the same as the actual airfoil.

3. The method according to claim 1, characterized in that In the optimized design configuration data, the design conditions include Reynolds number, angle of attack and airfoil thickness, and the aerodynamic performance indicators include lift coefficient and drag coefficient.

4. The method according to claim 1, characterized in that: The process of obtaining the optimal airfoil shape data includes: Initialize design conditions and latent variables through reinforcement learning agents; Input the latent variables into the conditional generative adversarial neural network for generation to obtain the current airfoil shape data; Simulate the current airfoil shape data to obtain different reward values; According to different reward values, the expected cumulative reward of the current airfoil shape data is obtained through the value network; The latent variable is adjusted according to the immediate reward agent by the policy network to obtain an adjusted latent variable; Based on the adjusted latent variables, they are input into the conditional generative adversarial neural network again, and the process of generation, simulation, immediate reward calculation and latent variable adjustment is iteratively performed until the termination condition is reached, and the optimal latent variable with the best immediate reward is obtained, and the airfoil shape data corresponding to the optimal latent variable is used as the optimal shape data.

5. The method according to claim 4, characterized in that Different reward values ​​include airfoil smoothness reward, multi-angle attack comprehensive performance reward, maximum lift coefficient reward and minimum drag reward.

6. The method according to claim 4, characterized in that The latent variables are adjusted by the strategy parameters, wherein the strategy parameters are updated according to the expected cumulative rewards, and the updating of the strategy parameters is achieved by maximizing the objective function, and the objective function is: Q represents the expected cumulative reward, si represents the action at the i-th moment, μ(s|θ μ ) represents the policy network, θ μ represents the strategy parameters, m represents the expected cumulative reward time step, and J represents the target value.

7. The method according to claim 1, characterized in that The iterative optimization of latent variables by the reinforcement learning agent is preceded by: The optimal design configuration data is predicted through an adaptive agent model to obtain the optimal latent variables; The reinforcement learning agent uses the optimal latent variables as the initial latent variables for iterative optimization, wherein the adaptive proxy model uses a neural network model to characterize the relationship between the optimized design configuration data in the historical optimal airfoil appearance data and the latent variables.

8. The method according to claim 7, characterized in that The process of training the adaptive proxy model includes: Construct an optimal design library, wherein the optimal design library stores the optimal design configurations obtained after optimization under different design conditions; when a new optimization task is completed, the data corresponding to the optimization task is stored in the optimal design library; the data in the optimal design library is divided to obtain different sampling tasks; In the inner layer training, different sampling tasks are input into the adaptive proxy model, the predicted values ​​are calculated, and the loss function L is calculated based on the predicted values: Where N represents the number of training samples. represents the predicted value, represents the actual optimal latent variable, i represents the training sample number; In the outer layer training, several gradient descents are performed to update the temporary parameters θ of the proxy model. i ': Among them, θ represents the initial parameters of the adaptive proxy model, θ' represents the model parameters of the adaptive proxy model after optimization of the inner tasks of the meta-learning framework, α represents the learning rate, Represents the loss function L on the current task new The gradient of θ; Calculate the loss value of each sampling task according to the temporary parameters, calculate the overall loss according to the loss value of each sampling task, and update the initial parameters of the adaptive proxy model by back propagation according to the overall loss; The task sampling, inner optimization and outer optimization steps are repeated until the prediction error of the adaptive proxy model on several tasks converges to a predetermined range, thereby obtaining a trained adaptive proxy model.

Images