CFD parameter adaptive calibration method and system based on measured data and double-agent model

Abstract

The invention belongs to the technical field of CFD (computational fluid dynamics) parameter calibration, and discloses a CFD parameter adaptive calibration method and system based on measured data and a double-agent model, and the method comprises the steps: obtaining a CFD input parameter sample, inputting the CFD input parameter sample into a CFD solver, and obtaining an initial simulation result; determining an error evaluation index according to the initial simulation result based on a target actual measurement data result; constructing a double-agent model based on a Kriging model and a radial basis function neural network by taking a CFD input parameter sample as an independent variable and an error evaluation index as a dependent variable; the double-agent model is trained, the trained double-agent model takes the error evaluation index as fitness, and CFD input parameter values are obtained based on a genetic algorithm; the CFD input parameter values are input into the CFD solver for a simulation experiment, a calibrated simulation result is output, the reliability and generalization ability of prediction are improved through a double-agent model, a high-fidelity simulation result is output through the CFD solver, and the number of times of calling the CFD solver is reduced while the calibration precision is guaranteed.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of CFD parameter calibration, and relates to a CFD parameter adaptive calibration method and system based on measured data and a double-agent model. BACKGROUND

[0002] In the field of computational fluid dynamics (CFD) engineering simulation, in order to accurately simulate complex flow phenomena, key parameters such as turbulent flow model coefficients, grid resolution, and boundary condition settings need to be reasonably configured. However, the manual debugging method relying on expert experience lacks repeatability and objective standards, and the parameter combinations set by different personnel differ greatly, resulting in poor consistency of simulation results. The method of using a single agent model has problems of insufficient global exploration ability or low local fitting precision when dealing with strong nonlinearity and multimodal CFD response surfaces, affecting the convergence speed and accuracy of the final parameter solution. SUMMARY

[0003] In view of the deficiencies in the prior art, the purpose of the present application is to provide a CFD parameter adaptive calibration method and system based on measured data and a double-agent model.

[0004] In order to achieve the above-mentioned purpose, the technical scheme adopted by the present application is as follows: The present application provides a CFD parameter adaptive calibration method based on measured data and a double-agent model, comprising the following steps: obtaining a CFD input parameter sample, inputting the CFD input parameter sample into a CFD solver, and obtaining an initial simulation result; determining an error evaluation index based on the target measured data result for the initial simulation result; constructing a double-agent model based on a Kriging model and a radial basis function neural network with the CFD input parameter sample as the independent variable and the error evaluation index as the dependent variable; training the double-agent model, and obtaining the CFD input parameter value based on the genetic algorithm with the error evaluation index as the fitness of the trained double-agent model; inputting the CFD input parameter value into the CFD solver for simulation experiment, and outputting the calibrated simulation result.

[0005] Further, the double-agent model is :

[0006] wherein, is an approximate estimate of the prediction variance of the radial basis function neural network at U point, is the prediction value of the Kriging model at U point, is the prediction variance of the Kriging model at U point, is the prediction value of the radial basis function neural network at U point.

[0007] Further, the prediction value of the Kriging model at U point is:

[0008] wherein, is a regression term of global trend, is a random function with mean zero.

[0009] Further, the predicted value of the radial basis function neural network at the U point is is:

[0010] wherein, is the number of hidden layer neurons, is the weight connecting the hidden layer neuron and the output layer, is a radial basis function, is an input parameter vector, is the center point of the hidden layer neuron, is the distance between the input parameter vector and the center point .

[0011] Further, the error evaluation index is :

[0012] wherein, is the number of measuring points, is the pressure calculation value of the measuring point, is the physical measured pressure value of the measuring point.

[0013] Further, the training of the double-agent model comprises: adaptive sampling in the CFD input parameter space based on an expected improvement criterion, obtaining new CFD input parameter samples, inputting the new CFD input parameter samples into a CFD solver to obtain simulation results, calculating the root mean square error of the simulation results and the real wind tunnel experiment results, and training the double-agent model based on the new CFD input parameter samples and the root mean square error.

[0014] Further, the new CFD input parameter samples are obtained by Latin hypercube sampling; the sample of the Latin hypercube sampling is :

[0015]

[0016] in, , is the index of the sample points. Dimension index for CFD input parameters Dimension index for CFD input parameters For the first The lower bound of each CFD input parameter For set Any random arrangement of the above, A random number that follows a uniform distribution within the interval [0,1]. For the first The upper bound of the CFD input parameters.

[0017] Furthermore, the sampling points for the desired improvement criterion are determined based on the particle swarm optimization algorithm; The velocity vector of the particle swarm optimization algorithm is:

[0018] in, For particles In the The velocity vector at the next iteration For inertial weights, For particles In the The velocity vector at the next iteration As a self-awareness factor of particles, and Two random numbers generated independently within the interval [0,1] For particles In the The position vector obtained by the desired improvement criterion from the positions traversed before the next iteration. For particles In the The position vector at the next iteration. For the collective consciousness of particles, For all particles in the first The position vector obtained by the desired improvement criterion from the positions traversed before the next iteration.

[0019] Furthermore, the position vector of the particle swarm optimization algorithm is:

[0020] in, For particles In the The position vector at the next iteration. For particles In the The position vector at the next iteration. for particles at the first velocity vector at the second iteration.

[0021] The application is a CFD parameter adaptive calibration system based on measured data and a double-agent model, comprising an acquisition module: acquiring CFD input parameter samples, inputting the CFD input parameter samples into a CFD solver, and acquiring initial simulation results; a determination module: used for determining an error evaluation index based on target measured data results for the initial simulation results; a construction module: used for constructing a double-agent model based on a Kriging model and a radial basis function neural network with the CFD input parameter samples as independent variables and the error evaluation index as dependent variables; a training module: training the double-agent model, and acquiring CFD input parameter values based on a genetic algorithm with the error evaluation index as fitness; and an output module: used for inputting the CFD input parameter values into the CFD solver for simulation experiments and outputting calibrated simulation results.

[0022] Compared with the prior art, the application has the following beneficial technical effects: The application is a CFD parameter adaptive calibration method based on measured data and a double-agent model, a double-agent model is constructed based on a Kriging model and a radial basis function neural network, and genetic algorithm is combined to automatically calibrate CFD input parameters, thereby improving simulation precision and efficiency; first, initial simulation results are obtained by driving a CFD solver through sample parameters, an error evaluation index is defined based on target measured data, a mapping relationship between CFD input parameters and errors is established, the reliability and generalization ability of prediction are improved by using a double-agent model, the optimal parameter combination is intelligently searched through genetic algorithm with the error index as fitness, parameter automatic calibration is realized, high-fidelity simulation results are output through the CFD solver, the number of times of directly calling the CFD solver is reduced while ensuring calibration precision. BRIEF DESCRIPTION OF DRAWINGS

[0023] Figure 1 It is a flowchart of the CFD parameter adaptive calibration method based on measured data and a double-agent model of the application; Figure 2 It is a root mean square error (RMSE) graph under different fixed values in the embodiment of the application. DETAILED DESCRIPTION

[0024] In the following, the technical solutions in the embodiments of the present application will be described clearly and completely in conjunction with the accompanying drawings of the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort should belong to the scope of the present application.

[0025] Embodiment 1 The CFD parameter self-adaptive calibration method based on measured data and a double-agent model according to the present application comprises the following steps: obtaining a CFD input parameter sample, inputting the CFD input parameter sample into a CFD solver to obtain an initial simulation result; calculating an error evaluation index based on a target measured data result of the initial simulation result; constructing a double-agent model based on a Kriging model and a radial basis function neural network with the CFD input parameter sample as the independent variable and the error evaluation index as the dependent variable; training the double-agent model, taking the error evaluation index as the fitness of the trained double-agent model, and obtaining the value of the CFD input parameter based on a genetic algorithm; inputting the value of the CFD input parameter into the CFD solver for simulation experiment, and outputting the calibrated simulation result, as shown in FIG. Figure 1 .

[0026] The training of the double-agent model comprises: adaptive sampling in the CFD input parameter space based on an expected improvement criterion to obtain new CFD input parameter samples; inputting the new CFD input parameter samples into the CFD solver to obtain simulation results; calculating the root mean square error between the simulation results and the real wind tunnel experiment results according to the error evaluation index formula; adding the new CFD input parameter samples and their corresponding root mean square errors to the training data set, and incrementally training the double-agent model.

[0027] Specifically, in an embodiment, the static pressure distribution of 45 measuring points on the airfoil surface at t=10s is obtained as the comparison data, and the root mean square error between the CFD simulation result and the wind tunnel measured result at the 45 measuring points is calculated based on the error evaluation index; the root mean square error is taken as the output data of the double-agent model, which is used for training the error prediction model, and at the same time, the root mean square error is taken as the objective function of the genetic algorithm for searching the optimal Smagorinsky constant Cs.

[0028] The error evaluation index :

[0029] wherein, is the number of measuring points, is the Calculated pressure values ​​at each measuring point For the first The physical measured pressure values ​​at each measuring point.

[0030] Using CFD input parameter samples as independent variables and error evaluation indicators as dependent variables, a dual-surrogate model is constructed based on the Kriging model and radial basis function neural network.

[0031] The dual-agent model is :

[0032] in, This is an approximate estimate of the prediction variance of the radial basis function neural network at point U. This represents the predicted value of the Kriging model at point U. The predicted variance of the Kriging model at point U is... This represents the predicted value of the radial basis function neural network at point U.

[0033] Kriging models for arbitrary input parameters Error prediction value at It consists of two parts: a regression term representing a deterministic global trend. And a random function with a mean of zero

[0034]

[0035] in, The regression term represents the overall trend. It is a random function with a mean of zero.

[0036] random function Spatial correlation is defined by the covariance function, which describes the spatial correlation between any two input points. and The degree of correlation between them.

[0037]

[0038] in, The process variance represents the overall magnitude of variation in a stochastic process. This is a spatial correlation function with a range of [0, 1], which describes the correlation between two points as a function of distance.

[0039]

[0040] in, The dimension of the input parameter. For the first The distance scale of the influence of each input parameter on the function. , For the first The input parameter vector of each sample , For them in the The values ​​that can be taken in each dimension of the input parameters.

[0041] hyperparameters of the Kriging model and process variance By maximizing the following restricted log-likelihood function Determined:

[0042] in: Y is a The column vector contains The observation error value of each training sample point .

[0043] F is a The design matrix represents the global trend model. For the ordinary Kriging model, F is a set of all 1s. Column vector.

[0044] R is a The correlation matrix, its elements The hyperparameters are calculated from the aforementioned Gaussian correlation function. The function.

[0045] It is a generalized least squares estimate of the trend term coefficient, calculated using the following formula:

[0046] n is the total number of training sample points.

[0047] p is the trend term. The number of parameters in the equation (for ordinary kriging, p=1).

[0048] In practice, process variance can be... Substituting the analytical solution into the above equation simplifies the expression. The REML estimator is:

[0049] Will Substitution After ignoring the constant term, the objective function to be numerically optimized (i.e., minimizing the negative log-likelihood function) simplifies to:

[0050] where, is the vector of hyperparameters of the correlation function to be optimized; is the total number of samples, i.e., the number of observation data points involved in modeling; is the number of regression terms, i.e., the number of columns of the design matrix ; is the process variance estimate based on restricted maximum likelihood (REML) estimation; is the correlation matrix of , which is calculated based on the correlation function ; is the design matrix of , which is used to describe the overall trend term (regression model); is the matrix of , which is used to modify the impact of regression parameters in REML.

[0051] By using numerical optimization algorithms such as gradient descent, particle swarm optimization, etc., the above minimization problem is solved, thus automatically finding the optimal hyperparameter combination. It ensures that the Kriging model can learn the internal correlation structure between the input parameters and the error response to the greatest extent from the existing data.

[0052] Radial basis function neural network is a kind of feedforward neural network, which has a special non-linear fitting capability, especially suitable for capturing local details of gradient changes in functions, mainly responsible for fitting local non-linear features of error functions with high precision.

[0053] The predicted value of the radial basis function neural network at point U is:

[0054] where, is the number of hidden layer neurons, is the weight connecting the th hidden layer neuron and the output layer, is the radial basis function, is the input parameter vector, is the center point of the th hidden layer neuron, is the distance between the input parameter vector and the center point .

[0055]

[0056] where, is the width parameter, is the input parameter vector​​ With the center point The distance between them.

[0057] The dual-surrogate model of this invention can remain robust globally while achieving high-precision fitting in key local regions.

[0058] The Latin hypercube sampling (LHS) method is used to obtain samples of newly added CFD input parameters.

[0059] No. sample points The generation process can be described by the following formula:

[0060] in: , is the index of the sample points. Dimension index for CFD input parameters Dimension index for CFD input parameters For the first The lower bound of each CFD input parameter For set Any random arrangement of the above, A random number that follows a uniform distribution within the interval [0,1]. For the first The upper bound of the CFD input parameters.

[0061] It applies to the set of integers. A random permutation of the above. For each dimension Each of these is independently generated as a random permutation. This guarantees that in the first... On the dimensional, the first Each sub-interval is visited without repetition.

[0062] Is Random numbers that follow a uniform distribution over an interval are used to randomly select values ​​within a subinterval.

[0063] Generated via LHS initial sample points Then, the CFD solver is called to calculate their corresponding error evaluation indices. This forms the initial training dataset, which is used for the first training of the surrogate model.

[0064] After the initial proxy model is constructed, directly optimizing on this model may not be accurate enough to find the true global optimum. To further improve the model accuracy without performing a large number of blind CFD simulations, this invention introduces an adaptive sampling strategy based on expected improvement (EI).

[0065] The EI criterion calculates the value at a certain point. The error value obtained after performing a real simulation Relative to the currently known minimum value Amount of improvement that can be obtained The mathematical expectation.

[0066] because It follows a normal distribution as given by the Kriging model. , The parsing expression is:

[0067] in:

[0068] It is the smallest observed error value among all sample points that have undergone CFD simulation. Is the dual-proxy model at candidate points The prediction error value, It is the Kriging model at candidate points The standard deviation of the prediction It is the cumulative distribution function (CDF) of the standard normal distribution. It is the probability density function PDF of the standard normal distribution.

[0069] After determining the Expected Improvement (EI) function as the criterion for adaptive sampling, it is necessary to... Efficiently find in 3D parameter space that The point with the largest value Since the EI function is usually multimodal and non-convex, traditional gradient optimization methods are prone to getting trapped in local optima.

[0070] The Particle Swarm Optimization (PSO) algorithm is used for the search. The PSO algorithm simulates the foraging behavior of bird flocks and finds the optimal solution through group cooperation. It has the advantages of fast convergence speed, few parameters, and ease of implementation.

[0071] The PSO algorithm treats each candidate solution U in the parameter space as a "particle" flying in space. Each particle has its own position and velocity. Through iteration, the velocity and position of each particle are continuously updated, guiding the entire particle swarm towards the optimal solution region.

[0072] In the In the nth iteration, for the th particle swarm... Each particle, its velocity and location The update is determined by the following formula: The velocity vector is:

[0073] The position vector is:

[0074] in: For particles In the The velocity vector at the next iteration For inertial weights, For particles In the The velocity vector at the next iteration As a self-awareness factor of particles, and Two random numbers generated independently within the interval [0,1] For particles In the The position vector obtained by the desired improvement criterion from the positions traversed before the next iteration. For particles In the The position vector at the next iteration. For the collective consciousness of particles, For all particles in the first The position vector obtained by the desired improvement criterion from the positions traversed before the next iteration. For particles In the The position vector at the next iteration. For particles In the The position vector at the next iteration. For particles In the The velocity vector at the next iteration.

[0075] Specifically, Latin hypercube sampling (LHS) is used to generate 10 initial sample points within the search interval [0.05, 0.3] of Cs. For each sample point... Run a complete 10-second CFD simulation and calculate the corresponding error evaluation index. An initial training set was formed, and these 10 data points were used to train the dual-surrogate model, initiating 5 rounds of adaptive sampling. In each round, the expected improvement (EI) function was maximized using the particle swarm optimization algorithm to find the next most worthwhile Cs point to simulate. After running CFD simulations, the new data points were added to the training set and the model was retrained. After 15 CFD simulations, a high-precision dual-surrogate model was constructed, capable of accurately predicting the error evaluation index corresponding to any Cs value.

[0076] In the adaptive sampling phase, the particle swarm optimization algorithm in this invention is only used to determine the newly added CFD input parameter sample points under the desired improvement criterion, so as to improve the fitting accuracy of the dual surrogate model in the parameter space.

[0077] Once the dual-proxy model reaches the preset accuracy, it enters the final global parameter optimization stage, based on a genetic algorithm. Firstly... A random parameter space is generated from a set of parameters. An initial population consisting of individuals Each individual Represents a set of CFD input parameters for each individual in the population. Call the pre-trained dual-agent model and calculate its corresponding prediction error value. Define the fitness function for an individual. for Where C is a small constant. Higher fitness individuals represent better parameter combinations. A tournament selection strategy is used to select parent individuals for breeding the next generation. Compared to roulette wheel selection, tournament selection has lower computational complexity and effectively prevents super-individuals from prematurely dominating the population. The simulated binary crossover (SBX) operator is employed. SBX simulates the behavior of single-point binary crossover in the real number domain, generating evenly distributed offspring near the parents and exhibiting good local search capabilities. Specifically, two parent individuals are randomly selected from the mating pool. and For each of their parameters Two child generations are generated through the following steps. and Corresponding parameters and : Generate a A random number rand between the given values. Based on the random number rand and the distribution exponent... (A hyperparameter that controls the distance between offspring and parents) The larger the value, the closer the offspring is to the parent (calculate the expansion factor). :

[0078] ,

[0079] Calculate the parameter values ​​for the two offspring based on the expansion factor β:

[0080]

[0081] For each pair of parent individuals in the mating pool, a pre-set crossover probability is applied. The above cross operation is performed; if the current cross is not triggered (with a probability of ), the parent is directly copied to the next generation population.

[0082] The new offspring individual generated after the cross is subjected to a polynomial mutation operator, simulating the behavior of bit flip mutation in binary coding in the real number field. It can generate new individuals in the vicinity of the parent individual according to a polynomial probability distribution, with good local search performance.

[0083] For each dimension parameter in the offspring individual , a small mutation probability is executed to generate the final mutated parameter : First, a random number rand between 0 and 1 is generated, and a disturbance factor is calculated according to the random number rand and the distribution index (an hyperparameter that controls the disturbance size, the larger , the closer the offspring to the parent).

[0084]

[0085] The mutated parameter value is calculated according to the disturbance factor :

[0086] where are the upper and lower limits of the th dimension parameter, respectively, to ensure that the mutated parameter value is still within the feasible region. If exceeds the boundary, it is set to the boundary value.

[0087] The newly generated offspring replaces the parent population to form a new generation population , until the maximum evolution generation and other termination conditions are met.

[0088] In summary, on the trained dual-agent model, the genetic algorithm (GA) is started, the tournament selection and simulated binary cross operator are configured, and the RMSE predicted by the agent model is minimized as the goal, for efficient global optimization. It can converge within a few minutes to obtain the optimal Smagorinsky constant Cs* = 0.1.

[0089] The optimal parameter combination found by the genetic algorithm ​, input to a high-precision real CFD solver, a complete and independent CFD simulation calculation is performed, and after the calculation is completed, the key characteristic values of this verification simulation are also extracted , compared with the initial physical measured characteristic values . By calculating the final real error , the calibration effect is evaluated. As Figure 2 shown, under the same working conditions, Smagorinsky constant Cs = 0.10, 0.12, 0.15, 0.17 and 0.20 are taken for numerical calculation, and the root mean square error (RMSE) of the full time domain pressure coefficient is taken as the evaluation index. When Cs is 0.10, the global RMSE is the smallest, which is 0.8249, which is obviously smaller than 0.8292 when the traditional empirical value Cs = 0.17 and other candidate parameters, and the corresponding overall error is reduced. The optimal fixed parameter Cs* = 0.10 obtained by the double-agent model and the optimization algorithm of the present application can significantly reduce the deviation between the simulation results and the wind tunnel measured data compared with the traditional empirical parameter combination.

[0090] Embodiment 2 The present application is a CFD parameter self-adaptive calibration system based on measured data and double-agent model, comprising an acquisition module, a determination module, a construction module, a training module and an output module.

[0091] The acquisition module acquires CFD input parameter samples, inputs the CFD input parameter samples to a CFD solver, and acquires initial simulation results; the determination module is used to determine error evaluation indexes based on target measured data results for the initial simulation results; the construction module is used to construct a double-agent model based on a Kriging model and a radial basis function neural network with CFD input parameter samples as independent variables and error evaluation indexes as dependent variables; the training module trains the double-agent model, and the trained double-agent model takes error evaluation indexes as fitness, and acquires CFD input parameter values based on a genetic algorithm; the output module inputs the CFD input parameter values to the CFD solver for simulation experiment, and outputs calibrated simulation results.

[0092] The present application is a CFD parameter self-adaptive calibration system based on measured data and double-agent model, which can realize the method steps consistent with the above method implementation, and therefore will not be repeated.

[0093] It has to be noted that the terms "first", "second", etc. as used in the description and the claims and above-mentioned figures of the application are used to distinguish between similar objects, not necessarily describing a particular sequential or chronological order. It is to be understood that the use of data "first", "second", etc., to distinguish between objects in the description and the claims is not anything more than notational and is merely intended to distinguish between two similar objects. It is further understood that data used in such a way can be interchangeable under appropriate circumstances, and embodiments of the present application described herein are capable of operation in other sequences than those explicitly described or illustrated herein. Furthermore, the terms "comprise" and "include" and variations thereof as used in the description and the claims and above-mentioned figures of the application are intended to cover a non-exclusive inclusion such that a process, method, system, product, or apparatus that comprises a list of steps or units are not necessarily limited to those steps or units but can include other not expressly listed or inherent steps or units.

Claims

1. A CFD parameter adaptive calibration method based on measured data and a dual-surrogate model, characterized in that, Includes the following steps: Obtain CFD input parameter samples, input the CFD input parameter samples into the CFD solver, and obtain initial simulation results; The initial simulation results are based on the target measured data results to determine the error evaluation index; Using CFD input parameter samples as independent variables and error evaluation indicators as dependent variables, a dual-surrogate model is constructed based on the Kriging model and radial basis function neural network. The dual-agent model is trained, and the trained dual-agent model uses the error evaluation index as the fitness and obtains the CFD input parameter values ​​based on the genetic algorithm. The CFD input parameters are input into the CFD solver for simulation experiments, and the calibrated simulation results are output.

2. The CFD parameter adaptive calibration method based on measured data and a dual-surrogate model according to claim 1, characterized in that: The dual-proxy model is as follows: : in, This is an approximate estimate of the prediction variance of the radial basis function neural network at point U. This represents the predicted value of the Kriging model at point U. The predicted variance of the Kriging model at point U is... This represents the predicted value of the radial basis function neural network at point U.

3. The CFD parameter adaptive calibration method based on measured data and a dual-surrogate model according to claim 2, characterized in that: The Kriging model's prediction at point U for: in, The regression term represents the overall trend. It is a random function with a mean of zero.

4. The CFD parameter adaptive calibration method based on measured data and a dual-surrogate model according to claim 2, characterized in that: The radial basis function neural network predicts the value at point U. for: in, The number of neurons in the hidden layer. To connect the first The weights of each hidden layer neuron and the output layer For radial basis functions, For the input parameter vector, For the first The center point of each hidden layer neuron Input parameter vector With the center point The distance between them.

5. The adaptive calibration method for CFD parameters based on measured data and a dual-surrogate model according to claim 1, characterized in that: The error evaluation index is: : in, For the number of measurement points, For the first Calculated pressure values ​​at each measuring point For the first The physical measured pressure values ​​at each measuring point.

6. The CFD parameter adaptive calibration method based on measured data and a dual-surrogate model according to claim 1, characterized in that: Training the dual-agent model includes: Based on the expected improvement criterion, adaptive sampling is performed in the CFD input parameter space to obtain newly added CFD input parameter samples. The newly added CFD input parameter samples are input into the CFD solver to obtain simulation results. The root mean square error between the simulation results and the actual wind tunnel experimental results is calculated. The dual surrogate model is trained based on the newly added CFD input parameter samples and the root mean square error.

7. The CFD parameter adaptive calibration method based on measured data and a dual-surrogate model according to claim 6, characterized in that: The newly added CFD input parameter samples were obtained through Latin hypercube sampling; The first Latin hypercube sampling The sample is : in, , is the index of the sample points. Dimension index for CFD input parameters Dimension index for CFD input parameters For the first The lower bound of each CFD input parameter For set Any random arrangement of the above, A random number that follows a uniform distribution within the interval [0,1]. For the first The upper bound of the CFD input parameters.

8. The CFD parameter adaptive calibration method based on measured data and a dual-surrogate model according to claim 7, characterized in that: The sampling points for the desired improvement criteria are determined based on the particle swarm optimization algorithm; The velocity vector of the particle swarm optimization algorithm is: in, For particles In the The velocity vector at the next iteration For inertial weights, For particles In the The velocity vector at the next iteration As a self-awareness factor of particles, and Two random numbers generated independently within the interval [0,1] For particles In the The position vector obtained by the desired improvement criterion from the positions traversed before the next iteration. For particles In the The position vector at the next iteration. For the collective consciousness of particles, For all particles in the first The position vector obtained by the desired improvement criterion from the positions traversed before the next iteration.

9. The CFD parameter adaptive calibration method based on measured data and a dual-surrogate model according to claim 8, characterized in that: The position vector of the particle swarm optimization algorithm is: in, For particles In the The position vector at the next iteration. For particles In the The position vector at the next iteration. For particles In the The velocity vector at the next iteration.

10. A CFD parameter adaptive calibration system based on measured data and a dual-surrogate model, characterized in that: Acquisition module: Acquires CFD input parameter samples, inputs the CFD input parameter samples into the CFD solver, and obtains initial simulation results; Determining module: used to determine error evaluation indicators based on the target measured data results of the initial simulation results; Building module: Used to construct a dual-surrogate model based on Kriging model and radial basis function neural network, with CFD input parameter samples as independent variables and error evaluation index as dependent variable; Training module: Train the dual-agent model. The trained dual-agent model uses the error evaluation index as the fitness and obtains the CFD input parameter values ​​based on the genetic algorithm. Output module: Used to input the CFD input parameter values ​​into the CFD solver for simulation experiments, and output the calibrated simulation results.

Images