Intelligent assimilation method for high-speed shock tunnel experimental data and CFD simulation

Abstract

This invention relates to the interdisciplinary fields of high-speed aerodynamics, experimental fluid mechanics, and artificial intelligence, and particularly to an intelligent assimilation method for high-speed shock tunnel experimental data and CFD simulation. The method includes: acquiring raw shock tunnel experimental signals; extracting steady-state features using a temporal adversarial network to obtain the experimental ground truth at the sensor location; constructing an interactive CFD simulation environment; constructing a deep reinforcement learning agent and defining a reward function; interactively training the agent with the CFD environment, where the agent outputs actions based on its current state, the CFD environment executes the actions and returns a new state and reward, storing experience and updating the network until convergence; solidifying the trained agent, outputting optimal CFD parameters, and performing error prediction and uncertainty quantification for unmeasured regions. This invention enables the simulation results to reproduce experimental measurements to the greatest extent possible, while simultaneously achieving flow field prediction and uncertainty quantification for unmeasured regions.

Description

Technical Field

[0001] This invention relates to the interdisciplinary fields of high-speed aerodynamics, experimental fluid mechanics and artificial intelligence, and in particular to an intelligent assimilation method for high-speed shock tunnel experimental data and CFD simulation. Background Technology

[0002] Current research on the aerodynamic characteristics of high-speed aircraft (Mach number > 5) heavily relies on ground-based wind tunnel experiments and numerical simulations. Shock tunnels, as typical ground-based high-speed testing equipment, inherently present the following challenges: 1. Extremely short experimental time: The effective experimental time is only in the millisecond range, the sensor data acquisition window is narrow, and it is difficult to obtain full flow field information.

[0003] 2. Sparse and noisy measurement data: Sensors are prone to failure under strong impact environments, and the measured data often only have a limited number of measurement points (such as wall pressure and heat flow), and include transient interference such as diaphragm rupture and vibration.

[0004] 3. Uncertainties exist in CFD models: The parameters of physical models such as turbulence models, wall catalysis, and real gas effects are difficult to determine accurately, which leads to deviations between simulation results and experiments.

[0005] 4. Traditional data assimilation methods are inefficient: Existing methods such as Bayesian calibration and Kalman filtering require multiple CFD iterations, which are computationally intensive and difficult to handle high-dimensional parameter spaces and dynamic adjustments.

[0006] Therefore, there is an urgent need for an intelligent assimilation method that can automatically and efficiently use sparse experimental data to correct CFD model parameters and quantify uncertainties. Summary of the Invention

[0007] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method for intelligent assimilation of high-speed shock tunnel experimental data and CFD simulation. By constructing an "experiment-simulation" closed-loop interactive framework, a deep reinforcement learning agent is trained to automatically adjust the parameters of the CFD model, so that the simulation results can reproduce the experimental measurements to the greatest extent, while realizing flow field prediction and uncertainty quantification in unmeasured areas.

[0008] To achieve the above objectives, this invention provides a method for intelligent assimilation of high-speed shock wave wind tunnel experimental data and CFD simulation, comprising the following steps: S1. Obtain the original signals from the shock tunnel experiment, extract steady-state features using a time-series adversarial network, and obtain the true experimental values ​​at the sensor locations. S2. Construct an interactive CFD simulation environment, define the state space as the predicted values ​​of CFD at the sensor location and the key features of the flow field, and define the action space as the CFD model correction parameters. S3. Construct a deep reinforcement learning agent, including an Actor network and a Critic network, and define the reward function as a negative value based on sensor error and a parameter regularization term; S4. Train the agent by interacting with the CFD environment. The agent outputs an action based on the current state. The CFD environment executes the action and returns a new state and reward. The experience is stored and the network is updated until convergence. S5. Solidify the trained agent, output the optimal CFD parameters, and perform error prediction and uncertainty quantification for the unmeasured area.

[0009] In some embodiments of this application, in S1, acquiring the original signal from the shock tunnel experiment, extracting steady-state features using a temporal adversarial network, and obtaining the experimental true value at the sensor location includes: Acquire the raw pressure / heat flux sensor signals from the shock tunnel experiment. The signal length is 10ms and the sampling frequency is ≥1MHz. A temporal adversarial network is used to denoise and extract steady-state features from the original signal; Construct generator G: Input random noise and wind tunnel operating conditions, output denoised steady-state signal segment; Construct a discriminator D to distinguish between real experimental steady-state segments and generator outputs; through adversarial training, enable the generator to extract time-series curves reflecting the steady-state characteristics of the real flow field from the unsteady raw signal; The steady-state characteristics are used as the experimental truth values ​​for subsequent assimilation, denoted as... Corresponding sensor position .

[0010] In some embodiments of this application, in step S2, constructing an interactive CFD simulation environment includes: The CFD solver is encapsulated as an interactive simulation environment that receives action parameters, runs automatically, and returns flow field results.

[0011] In some embodiments of this application, in S2, the state space includes the predicted values ​​from the CFD simulation at the sensor location. The simplified representation of the key features of the flow field constitutes the state vector.

[0012] In some embodiments of this application, in S3, the reward function further includes a physical consistency penalty term, including shock wave position continuity constraints or entropy increase conditions.

[0013] In some embodiments of this application, in step S4, the agent is trained by interacting with the CFD environment. The agent outputs an action based on the current state, the CFD environment executes the action and returns a new state and reward, and the experience is stored and the network is updated. Initialize the CFD environment and set initial parameters. Run CFD once to obtain the initial state. In each training round, the agent adjusts its actions based on the current state. Output Action The motion parameters are passed to the CFD solver, and the CFD is run again to obtain a new flow field; Among them, the agent adopts the Deep Deterministic Policy Gradient (DDPG) or Proximal Policy Optimization (PPO) algorithm; Extract sensor position predictions from the new flow field and calculate the reward. and update status , experience Store the data in the experience replay pool, and randomly sample from the replay pool to update the Actor-Critic network. Repeat until convergence, i.e., the reward value is stable and the error between the CFD prediction and the experiment is below the threshold.

[0014] In some embodiments of this application, in step S5, the trained agent is solidified, the optimal CFD parameters are output, and error prediction and uncertainty quantification are performed on the unmeasured region, including: After training, the optimal policy network is obtained, and the optimal CFD parameters can be quickly output by inputting any initial state; During training, the parameter distribution of multiple convergence rounds is recorded, and the confidence interval of the corrected CFD prediction is calculated using Monte Carlo Dropout or ensemble methods. Using a trained agent, flow field prediction is performed on unmeasured regions, and aerothermal data with uncertainty estimates are output.

[0015] In some embodiments of this application, it also includes: S6. Engineering Application and Verification: Integrate the intelligent assimilation system into the post-processing workflow of wind tunnel tests to achieve automatic calibration of CFD models within minutes after the experiment; The assimilated CFD results are compared with those of independent measurement points to evaluate their effectiveness.

[0016] The advantages and beneficial effects of this invention compared to the prior art are: 1. Efficient use of sparse experimental data: CFD model parameters are automatically adjusted through deep reinforcement learning, requiring only a small amount of sensor data to complete model correction, avoiding tedious manual parameter tuning.

[0017] 2. Temporal noise suppression: T-GAN is used to extract steady-state features from millisecond-level transient signals, which solves the problems of high noise and strong non-steady-state interference in the original shock tunnel data.

[0018] 3. Closed-loop intelligent optimization: The DRL agent interacts with the CFD environment, enabling it to explore high-dimensional parameter space and find the globally optimal parameter combination, overcoming the defect of traditional gradient optimization that is prone to getting trapped in local optima.

[0019] 4. Uncertainty Quantification: Provides the uncertainty range of the prediction results, providing a reliable boundary for the thermal protection design of aircraft.

[0020] 5. High versatility: It can be applied to different wind tunnel equipment and different aircraft shapes, and can be transferred with only retraining or fine-tuning.

[0021] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0022] Figure 1 This is a flowchart of the overall algorithm of the present invention, which is used to show the complete process of experimental data input, T-GAN processing, CFD environment construction, reinforcement learning training and result output. Figure 2 The diagram shows the structure of a temporal adversarial network, which includes a generator and a discriminator, used to extract steady-state features from wind tunnel transient signals. Figure 3 A diagram illustrating the interaction framework between a deep reinforcement learning agent and a CFD environment; Figure 4 A schematic diagram of the reward function is designed, including the error term (MAE) and the regularization term, as well as the physical constraint penalty term; Figure 5 A comparison diagram of aerodynamic assimilation of the HB2 standard model, including a comparison of axial, normal, and pitching moment coefficients; Figure 6 This is a surface heat flux distribution diagram of the aerodynamic heat transfer of the HB2 standard model. Figure 7 This is a diagram showing the distribution of hotspots on the assimilated surface of the HB2 standard model's aerodynamic heat. Figure 8 A comparison of the aerothermal assimilation results of the HB2 standard model; Figure 9 The graph shows the change in the reward function during the training process, demonstrating the algorithm's rapid convergence. Detailed Implementation

[0023] In the description of this invention, it should be noted that the terms "upper," "lower," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, or the orientation or positional relationship commonly used when the product is in use. They are used only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. In the description of this invention, it should also be noted that, unless otherwise explicitly specified and limited, the terms "set," "install," and "connect" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal communication between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0024] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0025] like Figure 1 As shown, this invention provides a method for intelligent assimilation of high-speed shock wave wind tunnel experimental data and CFD simulation, comprising the following steps: S1. Obtain the original signals from the shock tunnel experiment, extract steady-state features using a time-series adversarial network, and obtain the true experimental values ​​at the sensor locations. S2. Construct an interactive CFD simulation environment, define the state space as the predicted values ​​of CFD at the sensor location and the key features of the flow field, and define the action space as the CFD model correction parameters. S3. Construct a deep reinforcement learning agent, including an Actor network and a Critic network, and define the reward function as a negative value based on sensor error and a parameter regularization term; S4. Train the agent by interacting with the CFD environment. The agent outputs an action based on the current state. The CFD environment executes the action and returns a new state and reward. The experience is stored and the network is updated until convergence. S5. Solidify the trained agent, output the optimal CFD parameters, and perform error prediction and uncertainty quantification for the unmeasured area.

[0026] This invention automatically adjusts CFD model parameters through deep reinforcement learning, requiring only a small amount of sensor data to complete model correction, avoiding tedious manual parameter tuning. It employs T-GAN to extract steady-state features from millisecond-level transient signals, solving the problems of high noise and strong non-steady-state interference in raw shock tunnel data. Furthermore, it uses a DRL agent to interact with the CFD environment, enabling the exploration of high-dimensional parameter spaces and the finding of globally optimal parameter combinations, overcoming the tendency of traditional gradient optimization to get trapped in local optima, and providing a reliable boundary for aircraft thermal protection design.

[0027] In some embodiments of this application, in S1, acquiring the original signal from the shock tunnel experiment, extracting steady-state features using a temporal adversarial network, and obtaining the experimental true value at the sensor location includes: Acquire the raw pressure / heat flux sensor signals from the shock tunnel experiment. The signal length is 10ms and the sampling frequency is ≥1MHz. A temporal adversarial network is used to denoise and extract steady-state features from the original signal. The structure of the temporal adversarial network is as follows: Figure 2 As shown; Construct generator G: Input random noise and wind tunnel operating conditions, output denoised steady-state signal segment; Construct a discriminator D to distinguish between real experimental steady-state segments and generator outputs; through adversarial training, enable the generator to extract time-series curves reflecting the steady-state characteristics of the real flow field from the unsteady raw signal; The steady-state characteristics are used as the experimental truth values ​​for subsequent assimilation, denoted as... Corresponding sensor position .

[0028] In some embodiments of this application, in step S2, constructing an interactive CFD simulation environment includes: The CFD solver is encapsulated as an interactive simulation environment that receives action parameters, runs automatically, and returns flow field results.

[0029] In some embodiments of this application, in S2, the state space includes the predicted values ​​from the CFD simulation at the sensor location. The simplified representation of the key features of the flow field constitutes the state vector.

[0030] In some embodiments of this application, such as Figure 4 As shown in S3, the reward function also includes a physical consistency penalty term, including shock wave position continuity constraints or entropy increase conditions.

[0031] In some embodiments of this application, in step S4, the agent is trained by interacting with the CFD environment. The agent outputs an action based on the current state, the CFD environment executes the action and returns a new state and reward, and the experience is stored and the network is updated. Initialize the CFD environment and set initial parameters. Run CFD once to obtain the initial state. In each training round, the agent adjusts its actions based on the current state. Output Action The motion parameters are passed to the CFD solver, and the CFD is run again to obtain a new flow field; Among them, the agent adopts the Deep Deterministic Policy Gradient (DDPG) or Proximal Policy Optimization (PPO) algorithm; Extract sensor position predictions from the new flow field and calculate the reward. and update status , experience Store the data in the experience replay pool, and randomly sample from the replay pool to update the Actor-Critic network. Repeat until convergence, i.e., the reward value is stable and the error between the CFD prediction and the experiment is below the threshold.

[0032] In some embodiments of this application, in step S5, the trained agent is solidified, the optimal CFD parameters are output, and error prediction and uncertainty quantification are performed on the unmeasured region, including: After training, the optimal policy network is obtained, and the optimal CFD parameters can be quickly output by inputting any initial state; During training, the parameter distribution of multiple convergence rounds is recorded, and the confidence interval of the corrected CFD prediction is calculated using Monte Carlo Dropout or ensemble methods. Using a trained agent, flow field prediction is performed on unmeasured regions, and aerothermal data with uncertainty estimates are output.

[0033] In some embodiments of this application, it also includes: S6. Engineering Application and Verification: Integrate the intelligent assimilation system into the post-processing workflow of wind tunnel tests to achieve automatic calibration of CFD models within minutes after the experiment; The assimilated CFD results are compared with those of independent measurement points to evaluate their effectiveness.

[0034] The following verification is performed using specific embodiments: Example 1: Intelligent assimilation verification of the HB-2 standard model.

[0035] This embodiment is based on the intelligent assimilation method of high-speed shock wave wind tunnel experimental data and CFD simulation based on deep reinforcement learning proposed in this invention. The JAXA standard model HB-2 is selected as the research object. Combined with actual wind tunnel experimental data and program implementation, the effectiveness of the method in high-speed aerodynamic and aerothermal prediction is verified.

[0036] Step 1: Establish an experimental data preprocessing module.

[0037] In this embodiment, experimental data from the HB-2 shock tunnel were acquired and stored in TecplotASCII format, including aerodynamic data (CAF, CN, Cm) and aerothermal data (wall heat flux q). The experimental conditions were: Mach number Ma = 9.46, and total pressure approximately 1.0 × 10⁻⁶. 6 The total temperature is approximately 950 K, the static pressure is approximately 32.7 Pa, and the Reynolds number is approximately 4.35 × 10⁻⁶. 4The aerodynamic data covers an angle of attack range of -8.07° to 15.97°, with a total of 24 angle-of-attack sample points; the aerothermal data includes 27 wall measurement points with spatial coordinates ranging from x∈[0,0.44m] to z∈[-0.07,0.07m].

[0038] Since the acquired data is already a steady-state statistical result, this embodiment employs an equivalent T-GAN processing method. Noise is denoised using a sliding window averaging technique (window length of 5 sampling points) and a 3σ criterion outlier removal method, and the average value of the stable segment is extracted as the experimental true value. The final experimental data vector is obtained. and the corresponding measuring point location Establish a mapping relationship.

[0039] Step 2: Build a CFD simulation environment.

[0040] In this embodiment, the CFD solution process is encapsulated as an interactive environment. Considering that high-fidelity CFD computation time is relatively long (approximately several hours per run), this embodiment adopts the HB-2 surrogate CFD model built based on empirical models and physical constraints. Its computation time is controlled within the range of 0.01-0.05 seconds, which can be used for rapid iteration in reinforcement learning.

[0041] Define the action space as the CFD model correction parameter vector: ; Among them, the turbulence model correction coefficient The value range is [0.5, 1.5], representing the wall catalytic efficiency. The value range is [0,1], and the Mach number correction is... The value range is [-0.3, 0.3], and the wall temperature Tw ranges from [300K, 800K]. Initial parameters are set as follows: .

[0042] The state space is defined as: ; in, This represents the average heat flux at 27 measuring points. This represents the shock wave location characteristics estimated based on the wall pressure gradient.

[0043] Step 3: Construct a deep reinforcement learning agent, such as Figure 3 As shown.

[0044] In this embodiment, the DDPG algorithm is used to construct the Actor-Critic structure. The Actor network has an input dimension of 5, corresponding to the state vector dimension, and an output dimension of 4, corresponding to the action space. The network structure is a three-layer fully connected network with 128, 256, and 128 neurons in the hidden layers, respectively, and ReLU is used as the activation function. The Critic network takes a concatenated vector of state and action (9 dimensions in total) as input, has a hidden layer structure of [256, 256, 128], and outputs a single Q-value.

[0045] The optimizer uses the Adam algorithm, with learning rates for both the Actor and Critic set to 1×10⁻³, a discount factor γ=0.99, an experience replay pool size of 1×10⁵, and a batch size of 32. The reward function is defined as follows: ; Where N is the total number of measurement points (51 data points in total for aerodynamics and heat flux), and a0 is the initial parameter vector. Φ is the physical constraint indicator function, which is set to 1 when the predicted result shows a negative heat flux or a shock wave position jump exceeding 5%, otherwise Φ is set to 0; λ is the penalty coefficient, which is set to 0.1 in this embodiment.

[0046] Step 4: Training the agent to interact with the CFD environment.

[0047] In the initial training phase, parameter a0 is set, and the initial state s0 is calculated. In each training epoch, the Actor network outputs action parameters based on the current state, inputs these parameters into the CFD environment to obtain prediction results, and then calculates the error between the predicted value and the experimental data, using this error to calculate the reward value and update the state vector.

[0048] During training, the interaction data at each step is expressed as (s t ,a t ,R t ,s t+1 The data is stored in the experience replay pool in the form of a Proxy model and randomly sampled for network updates. The total number of training epochs is set to 300, with each epoch containing one CFD call. The computation time for a single proxy model is approximately 0.02 seconds, and the total training time is approximately 1–2 hours. In the later stages of training, the error gradually converges from approximately -0.15 to -0.04, indicating a significant reduction in error.

[0049] Step 5: Model solidification and uncertainty quantification.

[0050] After training, the Actor network parameters are fixed to obtain the optimal policy model. This model can quickly output the optimal CFD correction parameters after inputting any initial state, achieving real-time prediction. The prediction computation time is approximately 0.01 seconds.

[0051] To quantify uncertainty, the Monte Carlo Dropout method was employed. During the prediction phase, the Dropout probability was kept at 0.1, and 10 forward propagation samples were performed to obtain the mean and standard deviation of the prediction results, from which a 95% confidence interval was constructed. The results show that the uncertainty range for heat flux prediction is approximately ±8%, and the uncertainty for aerodynamic coefficients is approximately ±3%. A comparison of aerodynamic assimilation using the HB2 standard model is shown below. Figure 5 As shown.

[0052] Step 6: Engineering Application and Verification.

[0053] In this embodiment, the trained intelligent assimilation model was applied to validate the HB-2 standard model data. The results show that after assimilation, the error between the CFD prediction results and the experimental data was significantly reduced. Specifically, the mean absolute errors of CAF, CN, Cm, and heat flux q were reduced by approximately 65.7%, 58.2%, 81.6%, and 30.4%, respectively, verifying the effectiveness of the proposed method in the joint prediction of aerodynamics and aerothermal processes.

[0054] Visualization of the verification results, such as Figures 5 to 9 As shown. Among them, Figure 5 The results show the comparison of aerodynamic coefficients (including axial force, normal force, and pitching moment coefficients) of the HB2 standard model before and after assimilation, demonstrating that the CFD prediction curves are in high agreement with the experimental measurement data after deep reinforcement learning assimilation. Figure 6 The original heat flux distribution at the aerothermal surface measuring points before assimilation is shown. Figure 7 The surface heat flux distribution cloud map after assimilation and correction by this method is shown, which can accurately capture the location of the heat peak and the trend of heat flux distribution; Figure 8 Furthermore, a quantitative comparison of the aerothermal predictions before and after assimilation is provided, showing a significant improvement in prediction accuracy; Figure 9 The curve showing the change of the reward function during training intuitively reflects the process of the algorithm rapidly converging and stabilizing after about 200 iterations.

[0055] Furthermore, this method offers significant advantages in computational efficiency. After model training, for any given working condition, only state information needs to be input to obtain prediction results within milliseconds to seconds. Compared to traditional CFD methods, which require hours to tens of hours to complete a single-case calculation, this method has significant engineering application value in rapid design evaluation and multi-parameter analysis. Simultaneously, the prediction results can be automatically output as Tecplot format data files and visualizations, facilitating direct use in engineering analysis and design optimization processes.

[0056] In this application, unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. In case of any inconsistency, the meaning set forth in this specification or derived from the content described herein shall prevail. Furthermore, the terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit the scope of this application.

[0057] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for intelligent assimilation of high-speed shock wave wind tunnel experimental data and CFD simulation, characterized in that, Includes the following steps: S1. Obtain the original signals from the shock tunnel experiment, extract steady-state features using a time-series adversarial network, and obtain the true experimental values ​​at the sensor locations. S2. Construct an interactive CFD simulation environment, define the state space as the predicted values ​​of CFD at the sensor location and the key features of the flow field, and define the action space as the CFD model correction parameters. S3. Construct a deep reinforcement learning agent, including an Actor network and a Critic network, and define the reward function as a negative value based on sensor error and a parameter regularization term; S4. Train the agent by interacting with the CFD environment. The agent outputs an action based on the current state. The CFD environment executes the action and returns a new state and reward. The experience is stored and the network is updated until convergence. S5. Solidify the trained agent, output the optimal CFD parameters, and perform error prediction and uncertainty quantification for the unmeasured area.

2. The intelligent assimilation method for high-speed shock wave wind tunnel experimental data and CFD simulation according to claim 1, characterized in that, In step S1, the raw signals from the shock tunnel experiment are acquired, and steady-state features are extracted using a temporal adversarial network to obtain the experimental ground truth at the sensor location, including: Acquire the raw pressure / heat flux sensor signals from the shock tunnel experiment. The signal length is 10ms and the sampling frequency is ≥1MHz. A temporal adversarial network is used to denoise and extract steady-state features from the original signal; Construct generator G: Input random noise and wind tunnel operating conditions, output denoised steady-state signal segment; Construct a discriminator D to distinguish between real experimental steady-state segments and generator outputs; through adversarial training, enable the generator to extract time-series curves reflecting the steady-state characteristics of the real flow field from the unsteady raw signal; The steady-state characteristics are used as the experimental truth values ​​for subsequent assimilation, denoted as... Corresponding sensor position .

3. The intelligent assimilation method for high-speed shock wave wind tunnel experimental data and CFD simulation according to claim 1, characterized in that, In step S2, constructing an interactive CFD simulation environment includes: The CFD solver is encapsulated as an interactive simulation environment that receives action parameters, runs automatically, and returns flow field results.

4. The intelligent assimilation method for high-speed shock wave wind tunnel experimental data and CFD simulation according to claim 1, characterized in that, In S2, the state space includes the predicted values ​​from the CFD simulation at the sensor location. The simplified representation of the key features of the flow field constitutes the state vector.

5. The intelligent assimilation method for high-speed shock wave wind tunnel experimental data and CFD simulation according to claim 1, characterized in that, In S3, the reward function also includes a physical consistency penalty term, including shock wave position continuity constraints or entropy increase conditions.

6. The intelligent assimilation method for high-speed shock wave wind tunnel experimental data and CFD simulation according to claim 1, characterized in that, In step S4, the agent interacts with the CFD environment for training. The agent outputs an action based on its current state, the CFD environment executes the action and returns a new state and reward, and the experience is stored and the network is updated. This includes: Initialize the CFD environment and set initial parameters. Run CFD once to obtain the initial state. In each training round, the agent adjusts its actions based on the current state. Output Action The motion parameters are passed to the CFD solver, and the CFD is run again to obtain a new flow field. Among them, the agent adopts the Deep Deterministic Policy Gradient (DDPG) or Proximal Policy Optimization (PPO) algorithm; Extract sensor position predictions from the new flow field and calculate the reward. and update status , experience Store the data in the experience replay pool, and randomly sample from the replay pool to update the Actor-Critic network. Repeat until convergence, i.e., the reward value is stable and the error between the CFD prediction and the experiment is below the threshold.

7. The intelligent assimilation method for high-speed shock wave wind tunnel experimental data and CFD simulation according to claim 1, characterized in that, In step S5, the trained agent is solidified, outputting optimal CFD parameters, and error prediction and uncertainty quantification are performed on the unmeasured region, including: After training, the optimal policy network is obtained, and the optimal CFD parameters can be quickly output by inputting any initial state; During training, the parameter distribution of multiple convergence rounds is recorded, and the confidence interval of the corrected CFD prediction is calculated using Monte Carlo Dropout or ensemble methods. Using a trained agent, flow field prediction is performed on unmeasured regions, and aerothermal data with uncertainty estimates are output.

8. The intelligent assimilation method for high-speed shock wave wind tunnel experimental data and CFD simulation according to claim 1, characterized in that, Also includes: S6. Engineering Application and Verification: Integrate the intelligent assimilation system into the post-processing workflow of wind tunnel tests to achieve automatic calibration of CFD models within minutes after the experiment; The assimilated CFD results are compared with those of independent measurement points to evaluate their effectiveness.

Images