An AI-CFD-based intelligent rapid prediction method for hypersonic aerodynamic thermal environment

Abstract

The application discloses an AI-CFD-based intelligent rapid prediction method for hypersonic aerodynamic thermal environment, and comprises the following steps: importing a geometric model file and a cloud picture file of a hypersonic blunt body into a CFD numerical simulation software database, and building a python environment; processing the geometric model file to obtain binary data; processing the binary data to obtain directed distance field data; processing the binary data to obtain flow region channel data; processing the cloud picture file to obtain cloud picture grid topological structure data; taking the directed distance field, the flow region channel and the cloud picture grid topological structure data as inputs of CNN training, obtaining physical boundary condition parameters at different altitudes through investigation, adding the cloud picture grid topological structure data as inputs of RBFNN training, and utilizing the trained model to predict a flow field, a temperature field, a velocity field and a gas phase concentration distribution field. Compared with traditional numerical simulation, the method is faster in prediction speed, can be used for various prediction scenes, and has higher freedom degree.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of artificial intelligence and computational fluid dynamics, specifically relating to an intelligent and rapid prediction method for hypersonic aerothermal environments based on AI-CFD. Background Technology

[0002] Currently, hypersonic aerodynamics research mainly includes engineering estimation, experimental measurement, and numerical computation. Engineering estimation for hypersonic problems has low accuracy, while experimental measurements, such as wind tunnel testing, are more accurate but expensive. Computational Fluid Dynamics (CFD) is an important interdisciplinary field of numerical computation, involving mathematics, fluid mechanics, and computer science, balancing accuracy and cost. CFD can numerically solve the governing equations for mass, momentum, and energy conservation. It is an effective tool for studying aerodynamics and heat and mass transfer. However, traditional CFD methods are typically time-consuming, so breakthroughs in reducing computation time are of great significance. Many scholars both domestically and internationally have made numerous attempts to accelerate and optimize CFD methods.

[0003] In recent years, an increasing number of researchers have focused on artificial intelligence (AI) algorithms and applied them to the field of CFD. Chen et al. used a genetic algorithm (GA) to calculate the drag and lift coefficients of an aerodynamic model. Zhang et al. used multi-scale Gaussian progression (GP) regression to calculate the Reynolds-averaged Navier-Stokes (RANS) product parameters of a turbulence model. Roshani and Feghhi used a multilayer perceptron (MLP) to calculate the oil-water ratio of the specific flow rate. Khosravi and Malekan used a multilayer feedforward neural network (MLFNN) to calculate the thermal conductivity. Khazaei et al. used a Levenberg-Marquardt neural network to study CO2 concentration. The popularity of artificial intelligence has many reasons, such as hardware improvements, the massive amounts of data generated annually, and accuracy approaching that of CFD. Due to the advantages of AI algorithms, they have become an effective tool for solving CFD problems and have received increasing attention. However, the aforementioned domestic and international scholars have simply considered the changes in the flow field at low Reynolds numbers, which is not applicable to the prediction of changes in the flow field, temperature field, and air component concentration under hypersonic (high Reynolds number) conditions. Summary of the Invention

[0004] To address the aforementioned problems, this invention proposes an intelligent and rapid prediction method for hypersonic aerodynamic thermal environments. This method is based on the concept of artificial neural networks, offering faster prediction speeds and computational accuracy comparable to numerical calculations. Specifically, a deep learning model based on convolutional neural networks (CNNs) can quickly obtain a mapping library of different simulation results by modifying the model structure; a machine learning model based on radial basis function neural networks (RBFNNs) can also quickly obtain a mapping library of different simulation results by changing physical property parameters. This invention employs the following technical solutions:

[0005] A method for intelligent and rapid prediction of hypersonic aerothermal environment based on AI-CFD includes the following steps:

[0006] S1: Import the geometric model file and contour file of the hypersonic blunt-nosed body into the CFD numerical simulation software database and set up the Python environment;

[0007] S2: Process the geometric model file of the hypersonic blunt-nosed body in the CFD numerical simulation software database to obtain binarized data;

[0008] S3: Process the binarized data to obtain directed distance field data;

[0009] S4: Process the binarized data to obtain the flow channel data;

[0010] S5: Process the cloud map file of the hypersonic blunt-nosed body in the CFD numerical simulation software database to obtain the cloud map grid topology data corresponding to the cloud map file;

[0011] S6: Directed distance field, flow channel, and cloud map grid topology data are used as inputs for CNN training. The CNN is trained using a CNN coupled with a U-net image segmentation structure. The trained CNN model is then used to predict the flow field, temperature field, velocity field, and gas phase concentration distribution field on the surface of a blunt body under hypersonic aerodynamic thermal environment. Physical boundary condition parameters at different altitudes are obtained through surveys. These parameters, along with cloud map grid topology data, are used as inputs for RBFNN training. The trained RBFNN model is then used to predict the flow field, temperature field, velocity field, and gas phase concentration distribution field on the surface of a blunt body under hypersonic aerodynamic thermal environment.

[0012] Furthermore, step S2 includes the following sub-steps:

[0013] 2.1 In the Python program, call the Python packages required for the calculation, including: cv2, numpy, matplotlib, and pandas;

[0014] 2.2 Use the function cv2.imread() to read the input image information, and use the function cv2.cvtColor(*,cv2.COLOR_BGR2GRAY) to solve for the binarized data of each input image;

[0015] 2.3 Use the df.to_csv() function in the pandas package to output binary data.

[0016] Furthermore, step S3 includes the following sub-steps:

[0017] 3.1 Use the function cv2.distanceTransform() on the binarized data in step S2 to obtain the distance field data;

[0018] 3.2 Define a loop to iterate through the distance field data and change the data point values ​​inside the blunt-headed body in the distance field data to the corresponding negative values, thereby transforming the distance field data into directed distance field data;

[0019] 3.3 Use the function df.to_csv() in the pandas package to output the directed distance field data.

[0020] Furthermore, step S4 includes the following sub-steps:

[0021] 4.1 Define an overall loop that iterates through the binarized data in step S2, changing the data point values ​​inside the blunt body to 0, the flow field data point values ​​around the blunt body to 1, the data point values ​​of the no-slip wall to 2, the data point values ​​of the inlet face to 3, and the data point values ​​of the outlet face to 4.

[0022] 4.2 Use the function df.to_csv() in the pandas package to output the stream channel data.

[0023] Furthermore, step S5 includes the following sub-steps:

[0024] 5.1 In the Python program, call the Python packages required for the calculation, including: cv2, numpy, matplotlib, and pandas;

[0025] 5.2 Use cv2.imread() to read cloud map data, where the color of each data point in the cloud map has its fixed RGB value and physical value;

[0026] 5.3 Use .item(x,y,n) to get the nth channel value of the cloud map at the (x,y) coordinates. n=0 is the R value, n=1 is the G value, and n=2 is the B value.

[0027] 5.4 After reading the RGB channel values ​​of the cloud map, query the mapping relationship between the RGB values ​​and the specific physical values, and deduce the values ​​of each data point in the cloud map to obtain the cloud map grid topology data corresponding to the cloud map file;

[0028] 5.5 Use the pandas package to output the CSV file corresponding to the cloud map grid topology data.

[0029] Furthermore, step S6, the CNN training and prediction, includes the following sub-steps:

[0030] 6.1 In a Python program, call the Python packages required for computation, including: numpy, matplotlib, pickle, and pandas;

[0031] 6.2 Use the functions np.stack() and np.expand_dims() in the numpy package to convert the oriented distance field data, flow region channel data and cloud map grid topology data in CSV format output in steps S3 to S5 into an integrated data set, and use pkl.dump() to output a file in .pkl format;

[0032] 6.3 After obtaining the .pkl format integrated file, perform CNN training by calling the required Python packages for computation, including: torch, pickle, panda, os, and json. Use pickle.load to load the data, use split_tensors() to divide the dataset and test set, use torch.optim.AdamW to define the optimizer, and define the DeepCFD function for training.

[0033] 6.4. The trained CNN model is used to predict the flow field, temperature field, velocity field and gas phase concentration distribution field on the surface of a blunt body under hypersonic aerodynamic thermal environment.

[0034] Furthermore, step S6, the RBFNN training and prediction, includes the following sub-steps:

[0035] 6.1 In the Python program, call the Python packages required for the calculation, including: pod_rbf, numpy, matplotlib, and pandas;

[0036] 6.2 Set up the training physical environment. Use the function np.array() to create an array and input the physical parameters at each altitude.

[0037] 6.3 Load the dataset and train the RBFNN model to obtain the trained model;

[0038] 6.4. The trained RBFNN model is used to predict the flow field, temperature field, velocity field and gas phase concentration distribution field on the surface of a blunt body under hypersonic aerodynamic thermal environment.

[0039] The advantages of this invention over the prior art are:

[0040] The intelligent and rapid prediction method for hypersonic aerothermal environments of this invention is based on artificial neural networks. Compared with traditional numerical simulation, it offers faster prediction speed and computational accuracy no less than that of numerical calculations. Specifically, the deep learning model based on convolutional neural networks can quickly obtain a mapping library of different simulation results by changing the model structure; the machine learning model based on radial basis function neural networks can also quickly obtain a mapping library of different simulation results by changing the physical property parameters. Furthermore, the prediction method of this invention can be applied to various prediction scenarios, offering a high degree of flexibility. Attached Figure Description

[0041] Figure 1 This is a flowchart of the intelligent and rapid prediction method for hypersonic aerothermal environment based on AI-CFD of the present invention.

[0042] Figure 2 A flowchart for obtaining binarized data;

[0043] Figure 3 A flowchart for obtaining cloud map mesh topology data;

[0044] Figure 4 A flowchart for packaging oriented distance field data and flow region channel data;

[0045] Figure 5 A flowchart for packaging cloud map mesh topology data;

[0046] Figure 6 A flowchart for training a dataset using a CNN;

[0047] Figure 7 A flowchart for training a dataset using RBFNN;

[0048] Figure 8 The prediction results after training with CNN;

[0049] Figure 9 The prediction results are obtained after training with RBFNN. Detailed Implementation

[0050] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are merely illustrative and are not intended to limit the present invention.

[0051] This invention provides an intelligent and rapid prediction method for hypersonic aerothermal environments based on AI-CFD, such as... Figure 1 As shown, it includes the following steps:

[0052] S1: Import the geometric model file and contour file of the hypersonic blunt-nosed body into the CFD numerical simulation software database and set up the Python environment;

[0053] S2: Process the geometric model file of the hypersonic blunt-nosed body in the CFD numerical simulation software database to obtain binarized data.

[0054] S3: Process the binarized data to obtain the directed distance field data.

[0055] S4: Process the binarized data to obtain the flow channel data.

[0056] S5: Process the cloud map file of the hypersonic blunt-nosed body in the CFD numerical simulation software database to obtain the cloud map grid topology data corresponding to the cloud map file.

[0057] S6: Directed distance field, flow channel, and cloud map grid topology data are used as inputs for CNN training. The training employs a CNN coupled with a U-net image segmentation structure to predict the flow field, temperature field, velocity field, and gas phase concentration distribution field outside the test set. The U-net structure, based on the CNN structure, feeds back the parameters obtained from training convolutional pooling layers to the decoding part, reducing training time. Physical boundary condition parameters at different altitudes are obtained through surveys, and combined with cloud map grid topology data as inputs for RBFNN training, to predict the flow field, temperature field, velocity field, and gas phase concentration distribution field outside the test set.

[0058] In one possible implementation, in the AI-CFD-based intelligent rapid prediction method for hypersonic aerothermal environment provided by the present invention, step S2, obtaining binarized data, specifically includes:

[0059] 2.1 In the Python programs used, such as Figure 2 As shown, the necessary Python packages for the calculation are first called, including: cv2, numpy, matplotlib, and pandas. NumPy is used for general mathematical matrix calculations, and matplotlib is used to check the correctness of the output data.

[0060] 2.2 Use the function cv2.imread() to read the input image information, and use the function cv2.cvtColor(*,cv2.COLOR_BGR2GRAY) to solve for the binarized data of each input image.

[0061] 2.4 Use the function df.to_csv() in the pandas package to output binary data.

[0062] In one possible implementation, in the AI-CFD-based intelligent rapid prediction method for hypersonic aerothermal environment provided by the present invention, step S3, obtaining oriented distance field data, specifically includes:

[0063] 3.1 Use the function cv2.distanceTransform() to obtain the distance field data from the binarized data in S2.

[0064] 3.2 Define an overall loop to traverse the distance field data and change the data point values ​​inside the blunt-headed body in the distance field data to the corresponding negative values, thereby transforming the distance field data into directed distance field data.

[0065] 3.3 Use the function df.to_csv() in the pandas package to output the directed distance field (SDF) data.

[0066] In one possible implementation, in the AI-CFD-based intelligent rapid prediction method for hypersonic aerothermal environment provided by the present invention, step S4, obtaining flow channel data, specifically includes:

[0067] 4.1 Define an overall loop that iterates through the binarized data in step S2, changing the data point values ​​inside the blunt body to 0, the flow field data point values ​​around the blunt body to 1, the data point values ​​of the no-slip wall to 2, the data point values ​​of the inlet face to 3, and the data point values ​​of the outlet face to 4.

[0068] 4.2 Use the function df.to_csv() in the pandas package to output flow region channel data (FRC).

[0069] In one possible implementation, in the AI-CFD-based intelligent rapid prediction method for hypersonic aerodynamic thermal environment provided by the present invention, step S5, obtaining the cloud map grid topology data corresponding to the cloud map file, specifically includes:

[0070] 5.1 In the Python programs used, such as Figure 3 As shown, the Python packages required for the calculation are called first, including: cv2, numpy, matplotlib, and pandas.

[0071] 5.2 Use cv2.imread() to read contour map data. Each data point in the contour map has a fixed RGB value and physical value for its color.

[0072] 5.3 Use .item(x,y,n) to get the nth channel value of the cloud map at the (x,y) coordinates. n=0 is the R value, n=1 is the G value, and n=2 is the B value.

[0073] 5.4 After reading the RGB channel values ​​of the cloud map, query the mapping relationship between the RGB values ​​and the specific physical values, and deduce the values ​​of each data point in the cloud map to obtain the cloud map grid topology data corresponding to the cloud map file.

[0074] 5.5 Use the pandas package to output the CSV file corresponding to the cloud map grid topology data.

[0075] In one possible implementation, in the AI-CFD-based intelligent rapid prediction method for hypersonic aerodynamic thermal environments provided by the present invention, step S6, CNN training and prediction, specifically includes:

[0076] 6.1 In the Python programs used, such as Figure 4 As shown in Figure 5, the Python packages required for the calculation are called first, including: numpy, matplotlib, pickle, and pandas.

[0077] 6.2 Use the `np.stack()` and `np.expand_dims()` functions from the NumPy package to convert the SDF, FRC, and cloud map grid topology data output in steps S3-S5 into integrated data. Then, integrate the data sets together and use `pkl.dump()` to output a file in .pkl format.

[0078] 6.3 After obtaining the integrated file in .pkl format, perform CNN training, such as... Figure 6 As shown, the Python packages required for computation include: torch, pickle, panda, os, and json. The specific steps of the training procedure are as follows: loading data, splitting the dataset and test set, defining the optimizer, and building the training model. Data is loaded using pickle.load; the dataset and test set are split using split_tensors(); the optimizer is defined using torch.optim.AdamW; and the DeepCFD function is defined for training.

[0079] 6.4 The trained CNN model is used to predict the flow field, temperature field, velocity field and gas phase concentration distribution field on the surface of a blunt body under hypersonic aerodynamic thermal environment.

[0080] In one possible implementation, in the AI-CFD-based intelligent rapid prediction method for hypersonic aerodynamic thermal environments provided by the present invention, step S6, RBFNN training and prediction, specifically includes:

[0081] 6.1 In the Python programs used, such as Figure 7 As shown, the Python packages required for the calculation are called first, including: pod_rbf, numpy, matplotlib, and pandas.

[0082] 6.2 Set up the training physical environment. Use the function np.array() to create an array and input the physical parameters at each altitude.

[0083] 6.3 Load the dataset and train it to obtain the trained RBFNN model.

[0084] 6.4 The trained RBFNN model is used to predict the flow field, temperature field, velocity field and gas phase concentration distribution field on the surface of a blunt body under hypersonic aerodynamic thermal environment.

[0085] The following detailed embodiments illustrate the specific implementation of the AI-CFD-based intelligent rapid prediction method for hypersonic aerothermal environment provided by the present invention.

[0086] Example 1:

[0087] Using the CNN method in this approach, with blunt-nosed cone angles of 10°, 15°, and 20° as training samples, the results of predicting the flow field, temperature field, velocity field, and gas phase concentration distribution field at a cone angle of 25° are obtained.

[0088] First step, refer to Figure 2 First, the necessary Python packages for the calculation are called, including: cv2, numpy, matplotlib, and pandas. The function `cv2.imread()` is used to read the input image information. After reading the input image information, the function `cv2.cvtColor(*, cv2.COLOR_BGR2GRAY)` is used to solve for the binarized data of each input image. After the calculation is completed, the binarized data of the 10°, 15°, and 20° geometric models are obtained.

[0089] The second step is to change the data point values ​​inside the blunt-headed body in the binarized data to the corresponding negative values ​​to obtain SDF data.

[0090] The third step is to change the data point values ​​inside the blunt body to 0, the flow field data point values ​​around the blunt body to 1, the data point values ​​of the no-slip wall to 2, the data point values ​​of the inlet face to 3, and the data point values ​​of the outlet face to 4 in the binarized data to obtain the FRC data.

[0091] Step 4, refer to Figure 3First, the necessary Python packages for computation are called, including: cv2, numpy, matplotlib, and pandas. Then, the contour plot data is read. After reading the contour plot data, the mapping relationship between RGB color values ​​and specific physical values ​​is defined, and the values ​​of each data point in the contour plot are derived. Finally, the derived contour plot data is output as a CSV file using the called Python packages.

[0092] Step 5, refer to Figure 4 and Figure 5 First, the necessary Python packages for computation are called, including numpy, pickle, and pandas. These packages are then used to convert the SDF and FRC data output from steps two and three into corresponding CSV files. Finally, the data sets are combined and output as a .pkl file. (Reference) Figure 6 The program calls the necessary Python packages for computation, including: torch, pickle, panda, os, and json. The specific steps are as follows: loading data, splitting the dataset into a test set, defining the optimizer, and building the training model. The results obtained after 2000 iterations are as follows. Figure 8 As shown.

[0093] Example 2:

[0094] Using the RBFNN method in this paper, simulation results at altitudes of 40, 45, 50, and 60 km with a blunt cone angle of 25° were used as training samples to predict the flow field, temperature field, velocity field, and gas phase concentration distribution field at an altitude of 55 km.

[0095] First step, refer to Figure 3 First, the necessary Python packages for computation are called, including: cv2, numpy, matplotlib, and pandas. Then, the contour plot data is read. After reading the contour plot data, the mapping relationship between RGB color values ​​and specific physical values ​​is defined, and the values ​​of each data point in the contour plot are derived. Finally, the derived contour plot data is output as a CSV file using the called Python packages.

[0096] Step 2, refer to Figure 7 First, the necessary Python packages for computation are called, including: pod_rbf, numpy, matplotlib, and pandas. Then, the training physical environment is set, i.e., the physical parameters at each altitude. Afterward, the dataset is loaded and trained, outputting the trained images and errors. The temperature field distribution results obtained after 10,000 iterations are as follows. Figure 9 As shown.

[0097] The two examples above are merely some embodiments of the present invention. For those skilled in the art, various modifications and improvements can be made without departing from the inventive concept of the present invention, and these all fall within the protection scope of the present invention.

Claims

1. A method for intelligent and rapid prediction of hypersonic aerothermal environment based on AI-CFD, characterized in that, Includes the following steps: S1: Import the geometric model file and contour file of the hypersonic blunt-nosed body into the CFD numerical simulation software database and set up the Python environment; S2: Process the geometric model file of the hypersonic blunt-nosed body in the CFD numerical simulation software database to obtain binarized data; S3: Process the binarized data to obtain directed distance field data; S4: Process the binarized data to obtain the flow channel data; S5: Process the cloud map file of the hypersonic blunt-nosed body in the CFD numerical simulation software database to obtain the cloud map grid topology data corresponding to the cloud map file; S6: Directed distance field, flow channel, and cloud map grid topology data are used as inputs for CNN training. The CNN is trained using a U-net image segmentation structure coupled with a CNN. The trained CNN model is then used to predict the flow field, temperature field, velocity field, and gas phase concentration distribution field on the surface of a blunt body under hypersonic aerodynamic thermal environment. Physical boundary condition parameters at different altitudes are obtained through surveys. These parameters, along with cloud map grid topology data, are used as inputs for RBFNN training. The trained RBFNN model is then used to predict the flow field, temperature field, velocity field, and gas phase concentration distribution field on the surface of a blunt body under hypersonic aerodynamic thermal environment. Step S3 includes the following sub-steps: 3.1 Use the function cv2.distanceTransform() on the binarized data in step S2 to obtain the distance field data; 3.2 Define a loop to iterate through the distance field data and change the data point values ​​inside the blunt-headed body in the distance field data to the corresponding negative values, thereby transforming the distance field data into directed distance field data; 3.3 Use the function df.to_csv() in the pandas package to output the directed distance field data; Step S4 includes the following sub-steps: 4.1 Define an overall loop that iterates through the binarized data in step S2, changing the data point values ​​inside the blunt body to 0, the flow field data point values ​​around the blunt body to 1, the data point values ​​of the no-slip wall to 2, the data point values ​​of the inlet face to 3, and the data point values ​​of the outlet face to 4. 4.2 Use the function df.to_csv() in the pandas package to output the stream channel data; Step S5 includes the following sub-steps: 5.1 In the Python program, call the Python packages required for the calculation, including: cv2, numpy, matplotlib, and pandas; 5.2 Use cv2.imread() to read cloud map data, where the color of each data point in the cloud map has its fixed RGB value and physical value; 5.3 Use .item(x,y,n) to get the nth channel value of the cloud map at the (x,y) coordinates. n=0 is the R value, n=1 is the G value, and n=2 is the B value. 5.4 After reading the RGB channel values ​​of the cloud map, query the mapping relationship between the RGB values ​​and the specific physical values, and deduce the values ​​of each data point in the cloud map to obtain the cloud map grid topology data corresponding to the cloud map file; 5.5 Use the pandas package to output the CSV file corresponding to the cloud map grid topology data.

2. The intelligent and rapid prediction method for hypersonic aerothermal environment based on AI-CFD as described in claim 1, characterized in that, Step S2 includes the following sub-steps: 2.1 In the Python program, call the Python packages required for the calculation, including: cv2, numpy, matplotlib, and pandas; 2.2 Use the function cv2.imread() to read the input image information, and use the function cv2.cvtColor(*,cv2.COLOR_BGR2GRAY) to solve for the binarized data of each input image; 2.3 Use the df.to_csv() function in the pandas package to output binary data.

3. The intelligent and rapid prediction method for hypersonic aerothermal environment based on AI-CFD as described in claim 2, characterized in that, Step S6, the CNN training and prediction, includes the following sub-steps: 6.1 In a Python program, call the Python packages required for the computation, including: numpy, matplotlib, pickle, pandas; 6.2 Use the functions np.stack() and np.expand_dims() in the numpy package to convert the oriented distance field data, flow region channel data and cloud map grid topology data in CSV format output in steps S3 to S5 into an integrated data set, and use pkl.dump() to output a file in .pkl format; 6.3 After obtaining the integrated file in .pkl format, perform CNN training by calling the necessary Python packages for computation, including: The program uses torch, pickle, panda, os, and json to load data, uses pickle.load to load data, uses split_tensors() to split the dataset and test set, uses torch.optim.AdamW to define the optimizer, and defines the DeepCFD function for training. 6.

4. The trained CNN model is used to predict the flow field, temperature field, velocity field and gas phase concentration distribution field on the surface of a blunt body under hypersonic aerodynamic thermal environment.

4. The intelligent and rapid prediction method for hypersonic aerothermal environment based on AI-CFD as described in claim 3, characterized in that, Step S6, the RBFNN training and prediction, includes the following sub-steps: 6.1 In a Python program, call the Python packages required for the computation, including: pod_rbf,numpy,matplotlib,pandas; 6.2 Set up the training physical environment. Use the function np.array() to create an array and input the physical parameters at each altitude. 6.3 Load the dataset and train the RBFNN model to obtain the trained model; 6.

4. The trained RBFNN model is used to predict the flow field, temperature field, velocity field and gas phase concentration distribution field on the surface of a blunt body under hypersonic aerodynamic thermal environment.

Images