Method for rapidly predicting aerodynamic heat of point cloud neural network by considering variable geometric shape

Abstract

The invention discloses a point cloud neural network aerodynamic heat rapid prediction method considering a variable geometric shape, and belongs to the field of aerodynamics, and the method comprises the steps: determining a research object, and constructing a data set; then, a three-dimensional point cloud neural network is adopted to extract global shape features, local shape features and incoming flow condition features of different variable geometric shapes under the same basic configuration to serve as point cloud neural network input parameters, and the same model is adopted to predict wall surface pressure and heat flux density; wherein the training data of the three-dimensional point cloud neural network adopts the wall surface heat flux density and pressure under the non-uniform wall temperature boundary condition. According to the method, efficient prediction of the wall surface pressure distribution and the heat flux density under the same basic configuration and under the non-uniform wall temperature boundary condition of different variable geometric shapes can be realized, and the prediction accuracy is improved.

Description

Technical Field

[0001] The present invention relates to the field of aerodynamics, and more specifically, to a point cloud neural network aerodynamic and thermal rapid prediction method considering variable geometric shapes. Background Art

[0002] CFD numerical calculations, reduced-order surrogate models, and engineering methods are commonly used in the field of static thermoaeroelasticity research for rapid aerodynamic and thermal prediction. CFD numerical calculations offer the advantages of high computational accuracy, while reduced-order surrogate models and engineering methods offer the advantages of high computational efficiency and accuracy that meets engineering application requirements.

[0003] However, in the multi-physics coupled computational problems involved in thermo-aeroelasticity, full numerical methods suffer from drawbacks such as excessive time consumption and high computational cost. Engineering methods, due to their use of model simplifications, reduce computational accuracy to a certain extent. While commonly used reduced-order surrogate models can achieve highly accurate aerodynamic / thermal predictions, they use separate models for aerodynamic and aero-thermal predictions, resulting in excessive models and cumbersome coupled computations. Furthermore, currently, reduced-order surrogate models primarily focus on fixed shapes or combined shapes with unchanged geometry, or on variable geometry shapes that do not originate from the same basic configuration. This current state of research cannot yet be applied to the practical multi-physics coupled computational problems involved in thermo-aeroelasticity. Furthermore, in current numerical computations of thermo-aeroelasticity, the boundary conditions used in the flow field solution are primarily uniform wall temperature boundaries. However, the use of uniform temperature boundaries is an approximate method that increases errors in the thermo-aeroelastic coupled computations compared to numerical calculations using true non-uniform wall temperature boundaries, leading to deviations from the actual physical process.

[0004] Based on the above existing technical solutions, a technical solution that can achieve efficient simultaneous prediction of aerodynamic / thermal forces using a unified model under non-uniform wall temperature boundary conditions with variable geometric shapes is very important for the study of static thermal aeroelasticity problems. Summary of the Invention

[0005] The purpose of the present invention is to overcome the shortcomings of the existing technology and provide a point cloud neural network aerodynamic / thermal rapid prediction method considering variable geometric shapes, which can achieve efficient prediction of wall pressure distribution and heat flux density under non-uniform wall temperature boundary conditions of different variable geometric shapes under the same basic configuration, thereby improving the prediction accuracy.

[0006] The object of the present invention is achieved through the following solutions: A point cloud neural network aerodynamic and thermal rapid prediction method considering variable geometric shapes includes the following steps: Determine the research subjects and construct the data set; A three-dimensional point cloud neural network is then used to extract the global shape features, local shape features, and incoming flow condition features of different variable geometric shapes under the same basic configuration in the data set, which are used as input parameters of the point cloud neural network. The same model is then used to predict the wall pressure and heat flux density. Among them, the training data of the three-dimensional point cloud neural network uses the wall heat flux density and pressure under non-uniform wall temperature boundary conditions.

[0007] Furthermore, the research object includes a wing.

[0008] Furthermore, the determination of the research object and the construction of the data set specifically include the following sub-steps: Step (1), using an aerodynamic thermal environment solver to solve the initial shape aerodynamic thermal environment, and obtaining the wall heat flux density and pressure distribution in the undeformed state of the wing; Step (2) interpolates the wall heat flux density and pressure distribution to the structural heat transfer and structural stress and strain field calculation grid, and uses the structural thermal coupling calculation solver to carry out the wing surface structural heat transfer and structural stress and strain calculation, and calculates the wing geometric shape change at different times under the continuous action of the wall heat flux density and pressure; Step (3) is to interpolate the wing shape changes at different times through the grid deformation program to generate the deformed flow field calculation grid, and then return it to the aerodynamic thermal environment solver to carry out the aerodynamic thermal environment solution of the variable geometry wing at different times, and obtain the wall heat flux density and pressure of the different variable geometry wings; Step (4) extracts aerodynamic and thermal data at multiple deformation moments, pre-processes the data, and then divides them into training set, test set, and validation set according to the ratio to form a point cloud neural network aerodynamic and thermal prediction training data set considering variable geometric shape.

[0009] Furthermore, in step (4), the training data set specifically includes: the heat flux density numerically calculated under the non-uniform wall temperature boundary condition obtained by calculating the structural heat transfer using the structural thermal coupling calculation solver; or the heat flux density numerically calculated under the uniform wall temperature boundary condition corrected by the thermal wall correction formula.

[0010] Furthermore, in step (4), extracting aerodynamic thermal data at multiple deformation moments specifically includes extracting aerodynamic thermal data at 100 different deformation moments.

[0011] Furthermore, the three-dimensional point cloud neural network is used to extract global shape features, local shape features, and incoming flow condition features of different variable geometric shapes under the same basic configuration as point cloud neural network input parameters, and the same model is used to predict wall pressure and heat flux density. The training data of the three-dimensional point cloud neural network uses wall heat flux density and pressure under non-uniform wall temperature boundary conditions, and specifically includes the following sub-steps: S1, shape feature extraction: The three-dimensional coordinates of the wing surface grid points are used as input. Assuming that the wing surface contains a total of N coordinate points, the shape features of the wing are input into the network in the format of N*3, and "*" represents multiplication; then the input parameters are dimensionally upgraded through the multi-layer perceptron, and the number of output channels is 64, 128, 256, and 512 respectively. The N*512 features are used as the local feature vector of the wing, and N*512 is input into the maximum pooling layer. The maximum value of the feature vector is extracted and saved in the feature vector to obtain a 1*512 global feature vector. The global feature vector saves all the features of the wing shape, and the tensor operation is repeated N times to expand the dimension to align the feature vector dimension; S2, incoming flow feature extraction module: The incoming flow condition altitude, angle of attack, and Mach number are used as neural network inputs. The input parameters are then upgraded through a multi-layer perceptron network. The multi-layer perceptron network extracts the nonlinear relationship between altitude, angle of attack, and Mach number, ultimately forming a feature vector of the incoming flow condition. S3, three-dimensional heat flux and pressure prediction: The point cloud neural network predicts the wall heat flux density and pressure through the decoder, splices the local features, global features and incoming flow condition feature vectors of the wing shape to obtain a complete feature vector; then the feature vector is input into the decoder, and the decoder predicts the heat flux density and pressure of the wing wall by parsing the feature vector. The final output vector is the heat flux density or pressure of each wing wall point.

[0012] Furthermore, after obtaining the final output vector, the process also includes a prediction data reconstruction step: aligning the final output vector with the original data file and matching the topological relationship, and writing it into a data format that can be directly read by the visualization software, thereby completing the prediction data reconstruction.

[0013] Furthermore, in step S2, after the input parameters are dimensionally upgraded through the multi-layer perceptron network, the number of output channels is 64, 128, and 256 respectively.

[0014] Furthermore, in step S3, the size of the complete feature vector is N*1280.

[0015] Furthermore, in step S3, the decoder analyzes the feature vector to predict the wing wall heat flux density and pressure, and the output channels are 512, 512, 128, and 1 respectively.

[0016] The beneficial effects of the present invention include: (1) In response to the need for efficient multi-physics coupling calculations in static thermo-aeroelastic problems, the present invention proposes a method for rapid aerodynamic / thermal prediction of variable geometry shapes with non-uniform wall temperature boundary conditions based on a three-dimensional point cloud neural network. This method can achieve efficient prediction of wall pressure and heat flux density under non-uniform wall temperature boundary conditions for variable geometry shapes of the same basic configuration using the same reduced-order proxy model. The method of the present invention is suitable for efficient prediction of wall pressure distribution and heat flux density under non-uniform wall temperature boundary conditions for different variable geometry shapes under the same basic configuration.

[0017] (2) Compared with the traditional reduced-order model which can only predict the aerodynamic force and aerothermal function of a constant geometric shape and uses different proxy models for efficient prediction, the present invention uses a three-dimensional point cloud neural network to achieve efficient prediction of aerodynamic force and aerothermal function of a variable geometric shape with a constant basic configuration. Moreover, the efficient prediction of aerodynamic force / heat uses the same set of reduced-order proxy models, thus achieving simultaneous and efficient prediction of aerodynamic force and aerothermal function.

[0018] (3) The method of the present invention is not limited to one wing shape and one prediction target parameter, and is applicable to the efficient prediction of aerodynamic / thermal forces of other variable geometric shapes with unchanged basic configurations using the same model.

[0019] (4) The method of the present invention takes the wall heat flux and pressure under the non-uniform wall temperature boundary condition as the input parameters of the neural network intelligent prediction to carry out model training on the basis of considering the variable geometry shape, and takes the non-uniform wall temperature distribution as the training parameter into consideration during the training. Therefore, the aerodynamic / thermal efficient prediction of the variable geometry wing under the non-uniform wall temperature boundary condition can be achieved.

[0020] (5) The method of the present invention is not limited to the wall heat flux density and pressure distribution obtained by numerical calculation considering the non-uniform wall temperature boundary. The heat wall correction formula can also be used to perform heat wall correction on the heat flux density under the uniform wall temperature boundary condition, so that the corrected wall heat flux density and the original pressure distribution can be used as training data.

[0021] (6) Different from the traditional full numerical CFD calculation method, which takes too long to calculate, and the traditional reduced-order proxy model only predicts aerodynamics / thermals for fixed shapes and combined shapes of fixed shapes, and uses different models for prediction of aerodynamics and aerodynamic heat respectively. In the scheme of the present invention, a three-dimensional point cloud neural network can be used to extract the global shape features, local shape features and incoming flow condition features of different variable geometric shapes under the same basic configuration as neural network input parameters, so as to quickly predict the wall pressure and heat flux density. The prediction of wall pressure and heat flux density is predicted using the same model, realizing efficient prediction of aerodynamics / thermal unified framework with high measurement accuracy. At the same time, the non-uniform wall temperature boundary is taken into account in the training data, and the wall heat flux density and pressure under the non-uniform wall temperature boundary condition are used as training data. The method of the present invention is applicable to the rapid prediction of aerodynamics / thermal environment of other variable geometric configurations with unchanged basic configuration, and the efficient prediction of aerodynamics / thermal environment is predicted using the same model. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0023] Figure 1 Prepare a flowchart for a point cloud neural network-based dataset; Figure 2 This is the data pre-processing flow chart; Figure 3 This is the structure diagram of the point cloud neural network; Figure 4a A diagram of a wing model with a complex structure; Figure 4b Grid diagram for aerodynamic / thermal calculation of wing model; Figure 4c is the finite element mesh diagram of the wing model; Figure 5a Schematic diagram of deformation in the x direction; Figure 5b Schematic diagram of deformation in the y direction; Figure 5c Schematic diagram of deformation in the z direction; Figure 5d Schematic diagram of the overall deformation of the wing; Figure 6a This is the first schematic diagram of the relative error of heat flux density prediction on the leeward side; Figure 6b The second schematic diagram of the relative error of heat flux density prediction on the leeward side; Figure 7 The predicted results of heat flux density along the wing span and the relative error distribution curve; Figure 8a This is the first schematic diagram of the relative error distribution of leeward pressure prediction; Figure 8b This is the second schematic diagram of the relative error distribution of leeward pressure prediction; Figure 9 The graph shows the pressure prediction results and relative error distribution along the wing span. DETAILED DESCRIPTION

[0024] All features disclosed in all embodiments in this specification, or steps in all methods or processes implicitly disclosed, except for mutually exclusive features and / or steps, can be combined and / or expanded or replaced in any manner.

[0025] In a preferred embodiment, the present invention specifically provides a point cloud neural network aerodynamic / thermal rapid prediction method considering variable geometry, comprising the following steps: Step 1: Dataset preparation. Specifically, a complex structure wing is selected as the research object of the point cloud neural network aerodynamic / thermal rapid prediction method. The steps for constructing the dataset are as follows: (1) If Figure 1 As shown, the aerodynamic / thermal environment solver is selected to solve the initial shape aerodynamic / thermal environment, and the wall heat flux density and pressure distribution in the undeformed state of the wing are obtained; (2) Secondly, the wall heat flux density and pressure distribution are interpolated to the structural heat transfer and structural stress and strain field calculation grid, and the structural heat transfer and structural stress / strain calculation of the wing surface are carried out using the structural thermal coupling calculation solver. Through calculation, the change of the wing geometry at different times under the continuous action of the wall heat flux density and pressure can be obtained; (3) The wing shape changes at different times are interpolated through the grid deformation program to generate the deformed flow field calculation grid, which is then returned to the aerodynamic / thermal environment solver to carry out the aerodynamic / thermal environment solution of the variable geometry wing at different times, and the wall heat flux density and pressure of the different variable geometry wings are obtained; (4) Extract 100 aerodynamic / thermal data at different deformation moments and sort the data according to Figure 2 Data pre-processing is performed using a process described in the previous section. Finally, the data is divided into training, test, and validation sets to form a training dataset for point cloud neural network rapid aerodynamic / thermal prediction that considers variable geometry. This process includes processing two types of wall temperature data for training. The non-uniform wall temperature file contains the non-uniform wall temperature distribution file obtained by calculating structural heat transfer using the structural thermal coupling solver, or the corrected heat flux at the corresponding temperature obtained using the thermal wall correction formula using this non-uniform wall temperature distribution file.

[0026] Step 2: Construct a point cloud neural network structure for aerodynamic / thermal rapid prediction. Specifically, Figure 3 As shown in the figure, the point cloud neural network is used as the aerodynamic / thermal prediction model, which mainly includes the shape feature extraction process, the incoming flow condition feature extraction process and the three-dimensional thermal flow pressure intelligent prediction process.

[0027] Step S1, during the shape feature extraction process, the three-dimensional coordinates of the wing surface grid points are used as input, where the wing surface contains a total of N coordinate points, and the shape features of the wing are input into the network in the format of N*3. Then, the input parameters are dimensionally upgraded through a multi-layer perceptron, and the number of output channels is 64, 128, 256, and 512 respectively. The N*512 features are used as the local feature vector of the wing, and N*512 is input into the maximum pooling layer. The maximum value of the feature vector is extracted and saved in the feature vector to obtain a 1*512 global feature vector, which saves all the features of the wing shape, and the tensor operation is repeated N times to expand the dimension to align the feature vector dimension.

[0028] In step S2, during the oncoming flow feature extraction process, the oncoming flow condition altitude, angle of attack, and Mach number are used as neural network inputs. The input parameters are then upgraded through a multi-layer perceptron (MLP) network. The number of output channels is 64, 128, and 256, respectively. The nonlinear relationship between altitude, angle of attack, and Mach number is extracted through the MLP network, ultimately forming a feature vector of the oncoming flow condition.

[0029] In step S3, during the intelligent prediction of three-dimensional heat flux and pressure, the point cloud neural network intelligently predicts the wall heat flux density and pressure through the decoder, splicing the local features, global features, and incoming flow condition feature vectors of the wing shape to obtain a complete feature vector, where the size of the complete feature vector is N*1280. The feature vector is then input into the decoder, which parses the feature vector to predict the heat flux density and pressure of the wing wall. Its output channels are 512, 512, 128, and 1, respectively, that is, the final output vector is the heat flux density or pressure of each wing wall point. Finally, after data alignment and matching the topological relationship with the original data file, it is written into a data format that can be directly read by the visualization software, and the predicted data reconstruction is completed.

[0030] In one embodiment, Figure 4a 、 Figure 4b and Figure 4cFigure 2 shows the computational model used for rapid aerodynamic / thermal prediction. The material used is GH1015 alloy. The total number of computational grids used for the aerodynamic / thermal environment solution is 3,656,632, while the total number of grids used for structural heat transfer and stress / strain calculations is 102,776, with 24,385 grids on the walls. The calculation methods for each physical field are as follows: the structural temperature field is obtained by numerically solving the three-dimensional transient heat conduction equation using the finite volume method; the structural stress / strain is obtained by numerically solving the structural statics equation using the finite element method; and the aerodynamic / thermal environment calculation is obtained by numerically solving the three-dimensional compressible Navier-Stokes equations using the finite volume method. The physical properties of the GH1015 material are shown in Tables 1 and 2.

[0031] Table 1 Physical properties of high-temperature alloy GH1015 (I)

[0032] Table 2 Physical properties of high-temperature alloy GH1015 (II)

[0033] Figure 5a 、 Figure 5b 、 Figure 5c and Figure 5d The initial wall heat flux and pressure calculated using the aerodynamic / thermal environment solver and the coupled structural and thermal solver, respectively, are shown. A schematic diagram of the wing's variable geometry (measured in meters) is shown under continuous action using a one-way coupled calculation strategy. Based on the wing's variable geometry, the characteristic coordinates corresponding to the variable geometry at different times are extracted and used as one of the input parameters.

[0034] The wing with 20s time-varying geometry was selected as the aerodynamic / thermal prediction model, and this model did not appear in the training set and test set to ensure the validity of the model. The prediction results of heat flux density and pressure are as follows: Figure 6a and Figure 6b shown.

[0035] Figure 6a and Figure 6b The upwind and leeward distribution diagrams show the relative error between the value predicted by the point cloud neural network and the value numerically calculated by the aerodynamic / thermal environment solver. Figure 7 The predicted heat flux density and the calculated heat flux density distribution at 22%, 44%, 66% and 99% along the wing span are shown in Figure 2. Figure 7In the example, ori represents the actual numerical value, and pred represents the predicted value. As shown in the above figure, this aerodynamic / thermal prediction method can achieve relatively accurate predictions of the heat flux density on the wing wall. The relative error between the predicted and calculated heat flux values ​​is less than 5% for most of the predicted heat flux values. This demonstrates the high effectiveness of this method for predicting heat flux density on variable-geometry wings.

[0036] Figure 8a and Figure 8b Upwind and leeward distribution of the relative error between the wing wall pressure predicted by the point cloud neural network and the value numerically calculated by the aerodynamic / thermal environment solver. Figure 9 The predicted heat flux and calculated pressure distributions are plotted at 22%, 44%, 66%, and 99% along the wing's span. These figures demonstrate that this aerodynamic / thermal prediction method can accurately predict the wing wall pressure distribution, with the relative error between the predicted and calculated pressure values ​​being less than 5% for most cases. This demonstrates the effectiveness of this method for predicting pressure on variable-geometry wings.

[0037] In summary, the prediction method of the present invention can effectively realize the rapid prediction of aerodynamic / thermal forces of variable geometry wings, and the relative errors of the predictions of most wall heat flux densities and pressures are within 5%.

[0038] The units involved in the embodiments of the present invention may be implemented in software or hardware, and the units described may also be provided in a processor. In some cases, the names of these units do not limit the units themselves.

[0039] According to one aspect of an embodiment of the present invention, a computer program product or computer program is provided, comprising computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the methods provided in the various optional implementations described above.

[0040] As another aspect, embodiments of the present invention further provide a computer-readable medium, which may be included in the electronic device described in the above embodiments, or may exist independently and not incorporated into the electronic device. The computer-readable medium carries one or more programs, and when executed by the electronic device, the electronic device implements the methods described in the above embodiments.

Claims

1. A point cloud neural network aerodynamic and thermal rapid prediction method considering variable geometry, characterized by: The following steps are involved: Determine the research subjects and construct the data set; A three-dimensional point cloud neural network is used to extract the global shape features, local shape features and incoming flow condition features of different variable geometric shapes under the same basic configuration in the data set, which are used as the input parameters of the point cloud neural network. The same model is used to predict the wall pressure and heat flux density; among them, the training data of the three-dimensional point cloud neural network uses the wall heat flux density and pressure under non-uniform wall temperature boundary conditions.

2. The point cloud neural network aerodynamic and thermal rapid prediction method considering variable geometry according to claim 1 is characterized in that: The research object includes an aircraft wing.

3. The point cloud neural network aerodynamic and thermal rapid prediction method considering variable geometry according to claim 1 is characterized in that: Determining the research object and constructing the data set specifically includes the following sub-steps: Step (1), using an aerodynamic thermal environment solver to solve the initial shape aerodynamic thermal environment, and obtaining the wall heat flux density and pressure distribution in the undeformed state of the wing; Step (2) interpolates the wall heat flux density and pressure distribution to the structural heat transfer and structural stress and strain field calculation grid, and uses the structural thermal coupling calculation solver to carry out the wing surface structural heat transfer and structural stress and strain calculation, and calculates the wing geometric shape change at different times under the continuous action of the wall heat flux density and pressure; Step (3) is to interpolate the wing shape changes at different times through the grid deformation program to generate the deformed flow field calculation grid, and then return it to the aerodynamic thermal environment solver to carry out the aerodynamic thermal environment solution of the variable geometry wing at different times, and obtain the wall heat flux density and pressure of the different variable geometry wings; Step (4) extracts aerodynamic and thermal data at multiple deformation moments, pre-processes the data, and then divides them into training set, test set, and validation set according to the ratio to form a point cloud neural network aerodynamic and thermal prediction training data set considering variable geometric shape.

4. The point cloud neural network aerodynamic and thermal rapid prediction method considering variable geometry according to claim 3 is characterized in that: In step (4), the training data set specifically includes: the heat flux density numerically calculated under the non-uniform wall temperature boundary condition obtained by calculating the structural heat transfer using the structural thermal coupling calculation solver; or the heat flux density numerically calculated under the uniform wall temperature boundary condition corrected by the thermal wall correction formula.

5. The point cloud neural network aerodynamic and thermal rapid prediction method considering variable geometry according to claim 3 is characterized in that: In step (4), extracting aerodynamic thermal data at multiple deformation moments specifically includes extracting aerodynamic thermal data at 100 different deformation moments.

6. The point cloud neural network aerodynamic and thermal rapid prediction method considering variable geometry according to claim 1 is characterized in that: The method uses a three-dimensional point cloud neural network to extract global shape features, local shape features, and incoming flow condition features of different variable geometric shapes under the same basic configuration as point cloud neural network input parameters, and uses the same model to predict wall pressure and heat flux density. The training data of the three-dimensional point cloud neural network uses wall heat flux density and pressure under non-uniform wall temperature boundary conditions, and specifically includes the following sub-steps: S1, shape feature extraction: The three-dimensional coordinates of the wing surface grid points are used as input. Assuming that the wing surface contains a total of N coordinate points, the shape features of the wing are input into the network in the format of N*3, where "*" represents multiplication. The input parameters are then dimensionally upgraded through a multilayer perceptron, with the number of output channels being 64, 128, 256, and 512, respectively. The N*512 features are used as the local feature vector of the wing, and N*512 is input into the maximum pooling layer. The maximum value of the feature vector is extracted and saved in the feature vector to obtain a 1*512 global feature vector. The global feature vector stores all the features of the wing's shape, and the tensor operation is repeated N times to expand the dimension to align the feature vector dimension. S2, incoming flow feature extraction module: The incoming flow condition altitude, angle of attack, and Mach number are used as neural network inputs. The input parameters are then upgraded through a multi-layer perceptron network. The multi-layer perceptron network extracts the nonlinear relationship between altitude, angle of attack, and Mach number, ultimately forming a feature vector of the incoming flow condition. S3, three-dimensional heat flux and pressure prediction: The point cloud neural network predicts the wall heat flux density and pressure through the decoder, splices the local features, global features and incoming flow condition feature vectors of the wing shape to obtain a complete feature vector; then the feature vector is input into the decoder, and the decoder predicts the heat flux density and pressure of the wing wall by parsing the feature vector. The final output vector is the heat flux density or pressure of each wing wall point.

7. The point cloud neural network aerodynamic and thermal rapid prediction method considering variable geometry according to claim 6 is characterized in that: After obtaining the final output vector, the prediction data reconstruction step is also included: the final output vector is aligned with the original data file and the topological relationship is matched, and it is written into a data format that can be directly read by the visualization software, thereby completing the prediction data reconstruction.

8. The point cloud neural network aerodynamic and thermal rapid prediction method considering variable geometry according to claim 6 is characterized in that: In step S2, after the input parameters are dimensionally upgraded through the multi-layer perceptron network, the number of output channels is 64, 128, and 256 respectively.

9. The point cloud neural network aerodynamic and thermal rapid prediction method considering variable geometry according to claim 6 is characterized in that: In step S3, the size of the complete feature vector is N*1280.

10. The point cloud neural network aerodynamic and thermal rapid prediction method considering variable geometry according to claim 6, characterized in that: In step S3, the decoder analyzes the feature vector to predict the wing wall heat flux density and pressure, and the output channels are 512, 512, 128, and 1 respectively.

Images